Pre-training, fine-tuning, reinforcement learning and reasoning models.
By leveraging modern post-training techniques and clever fine-tuning data set construction, a smaller model could achieve high effectiveness. With full hindsight, a small talented team might have accomplished this around 2018 or 2019 using just a few GPU boxes.
▶ 1:20
John Schulman on dead ends, scaling RL, and building research institutions
Building reliable agents requires overcoming deep cognitive and architectural hurdles like lack of continual learning and multimodality. Drawing on 15 years of industry experience, Karpathy notes that while these problems are tractable, solving them will realistically take about a decade.
▶ 1:16
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Self-supervised learning removes the need for expensive manual labels by corrupting raw data and training models to reconstruct it. This approach builds general raw representations that can be fine-tuned for various downstream tasks.
▶ 8:01
Stanford CS25: Transformers United V6 I Overview of Transformers
A large language model is a specific instance of a foundation model applied to text and code. These models are pre-trained on massive unlabeled datasets to generate human-like text outputs.
▶ 0:36
How Large Language Models Work
Transformers take an initial snippet and output a probability distribution for the next chunk of text. By repeatedly sampling from this distribution, appending the result, and feeding it back in, the model generates longer text.
▶ 1:52
Transformers, the tech behind LLMs | Deep Learning Chapter 5
Initial token embeddings act like a static lookup table that assigns the exact same vector to a word regardless of its surrounding text. Attention steps in afterward to allow surrounding embeddings to pass relevant context into that specific vector.
▶ 2:18
Attention in transformers, step-by-step | Deep Learning Chapter 6
Tokenization acts as the fundamental boundary layer for text, meaning that string manipulations and non-English text are often poorly represented. Simple tasks like spelling or arithmetic fail because the model sees arbitrary subword chunks instead of characters or proper words.
▶ 4:32
Let's build the GPT Tokenizer
You can load the model using Hugging Face Transformers and extract its state dictionary to inspect key-value tensor shapes. This reveals the dimensions of components like the 50257-by-768 token embedding lookup table.
▶ 5:38
Let's reproduce GPT-2 (124M)
Chat GPT generates responses sequentially from left to right, and for any given prompt, it can produce multiple different answers. This means its outputs are not deterministic and vary across different runs.
▶ 0:49
Let's build GPT: from scratch, in code, spelled out.
Providers start with massive raw datasets like Common Crawl and apply rigorous filters to remove unwanted content, personal information, and non-target languages. This process results in a clean, high-quality corpus like Fine Web.
▶ 1:50
Deep Dive into LLMs like ChatGPT
Arithmetic became extremely cheap while moving data around became relatively expensive, prompting a pivot toward hardware specialized for matrix multiplications. Companies like Google started building TPUs designed as reduced-precision linear algebra machines to maximize throughput. This hardware-algorithm co-design unlocked the massive scale needed for modern deep learning.
▶ 7:50
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Researchers should avoid competing against the exponential drop in compute power and instead leverage it completely. This requires designing scalable methods that allow machines greater degrees of freedom rather than forcing rigid human structures.
▶ 4:02
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
US models generally provide superior output quality and intelligence gains that make paying for them worthwhile for these experts. Additionally, Chinese models are sometimes served with fewer GPUs per replica, negatively impacting speed and error rates.
▶ 20:39
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Developers can create a custom task, such as generating an SVG of a pelican on a bicycle, to test text models' code-output capabilities. This practical benchmark avoids untrusted leaderboards and tests models on an intentionally difficult and absurd task.
▶ 1:12
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
Meta releases smaller, highly efficient models like Scout and Maverick first to prioritize low latency and intelligence per cost for internal products. Meanwhile, they are developing a massive Behemoth model exceeding two trillion parameters, which developers may access through model distillation since running it directly as a consumer model is impractical.
▶ 1:50
Mark Zuckerberg — AI will write most Meta code in 18 months
Frontier labs have shifted from venture-funded losses to turning a profit because their revenue per megawatt has surpassed incremental costs. For instance, Anthropic generates up to $50 million per megawatt compared to baseline costs of $10 to $15 million. This allows them to fund further training and expansion directly from profits rather than just capital injections.
▶ 2:21
Dylan Patel – Two labs will soon control most of the world's workforce
Eyeballing the historical rate of capability increases points to reaching AGI by 2026 or 2027. Truly convincing blockers are rapidly disappearing, and compute clusters are scaling up to deploy millions of model instances.
▶ 2:19:39
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Dario relies on his "Big Blob of Compute Hypothesis," which posits that specific fundamental factors matter much more than clever new techniques. These factors include raw compute, data quantity and quality, training duration, and scalable objective functions like pre-training and reinforcement learning. Numerical stability and conditioning are also essential to ensure smooth computation flows.
▶ 2:06
Dario Amodei — “We are near the end of the exponential”
Early models relied on large auto-regressive neural networks like LSTMs trained on massive text datasets. However, LSTMs struggled with capturing long-range dependencies in data, which ultimately led to the adoption of Transformers.
▶ 1:00
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
Early poker bots relied on massive pretraining through self-play and acted instantly, but lacked the ability to think carefully during difficult decisions. Human professionals, conversely, would pause and deliberate on complex problems, allowing them to outperform the instant-acting AI. Adding a search mechanism to let the bot think for 30 seconds provided the equivalent performance boost of scaling parameters by 100,000x.
▶ 3:16
Learning to Reason with LLMs
Models often perform exceptionally well on rigid benchmarks because RL environments are inadvertently optimized for those specific evals. Combined with inadequate generalization, this causes models to struggle with simple, iterative real-world tasks like bug fixing.
▶ 1:32
Ilya Sutskever – We're moving from the age of scaling to the age of research
Training a large, deep neural network on a sufficiently large dataset that specifies a task will successfully solve it. The human brain proves these tasks are solvable via slow neural networks, so a related model trained on data will also succeed.
▶ 7:05
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
The classical bias-variance trade-off assumes you must choose between fitting data well and stability under sampling. However, modern deep learning shows that massive models can achieve both low bias and low variance simultaneously. This happens because scale introduces an implicit simplicity bias that naturally favors generalizable solutions over overfitting.
![Thumbnail: The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]](https://i.ytimg.com/vi/M-jTeBCEGHc/mqdefault.jpg)
▶ 2:13
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Large language models rely on auto-regressive pattern matching rather than parsing the actual content of the question. Because they memorize solutions or templates from training data, they produce the expected likelihood instead of reasoning from first principles. This causes them to break when presented with slight, unfamiliar alterations to classic puzzles.
▶ 4:40
It's Not About Scale, It's About Abstraction
Reasoning and problem-solving require large language models to explore multiple options step by step, reflect, and prune paths rather than just regurgitating memorized text. This search space exploration combined with applying AI to more jobs creates an insatiable, compounding demand for compute.
▶ 8:37
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Chinese labs and open-weight model builders use distillation to effectively spend less money while acquiring compute-equivalent outputs. Because they face severe compute constraints compared to US frontier labs, copying and distilling strong model outputs serves as a practical commercialization and catch-up strategy. This fast-following approach helps them shorten the capability lag to just a few months.
▶ 9:39
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
Leading Chinese labs operate with a total of 100 to 200 megawatts of compute at most. In contrast, Anthropic approaches nearly 5 gigawatts by the end of the year, creating a massive compute deficit on paper.
▶ 0:05
Why Isn’t China Further Behind in AI? - Dylan Patel
Generalization limits and persistent sim-to-real gaps could block true agent generalization. Additionally, failing to solve continual learning or models failing to check themselves can stall progress. These obstacles mean models might continue to feel dumb after short cycles rather than exploding in capability.
▶ 1:37
AI researchers debate how close we are to recursive self-improvement
AI research is highly verifiable and can be structured into containerized reinforcement learning environments. By training models iteratively on these small-scale tasks, developers can create a feedback loop that rapidly accelerates research capabilities.
▶ 0:50
Ryan Greenblatt – What happens once AI can automate AI research?
OpenAI's board saga highlighted the need for experienced board members who operate well under pressure and represent broad societal impact. Sam Altman notes that nonprofit boards without traditional shareholder accountability possess immense power, necessitating slates of diverse, experienced experts. Moving forward, resilient governance is critical to managing high-stakes power struggles safely.
▶ 5:50
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
While token pricing looks extremely cheap on paper, different systems consume varying amounts of tokens for the same job. DeepSeek can take two to three times more tokens than Claude Opus for identical tasks, making task-based metrics a truer cost measure.
▶ 1:08
DeepSeek's New AI Is An Absolute Monster
A hands-off approach works well for exploratory research with experienced individual contributors who can discover things independently. Conversely, a hands-on approach with more technical oversight is better suited for goal-oriented execution or teams with less experienced people.
▶ 11:14
John Schulman on dead ends, scaling RL, and building research institutions
Animals are shaped by millions of years of evolution and a massive hardware-encoding process that allows newborn creatures to function instantly. In contrast, language models are trained by imitating human data found on the internet, creating an ethereal, spirit-like entity starting from a very different point.
▶ 8:46
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Transformers enable parallel processing of entire sequences on GPUs rather than stepping through them sequentially like RNNs. They also handle much longer contexts and scale up significantly better.
▶ 14:36
Stanford CS25: Transformers United V6 I Overview of Transformers
LLMs are trained on enormous text datasets that can span up to petabytes of data. For example, GPT-3 was pre-trained on a corpus containing 45 terabytes of data.
▶ 1:27
How Large Language Models Work
The attention block allows vectors to talk to each other and pass information back and forth. It determines which words in context are relevant to updating the meanings of other words.
▶ 3:55
Transformers, the tech behind LLMs | Deep Learning Chapter 5
Query matrices map noun embeddings to query vectors looking for preceding adjectives, while key matrices map adjective embeddings to matching key vectors. Computing dot products between these keys and queries produces a relevance score grid determining how much one word attends to another.
▶ 6:42
Attention in transformers, step-by-step | Deep Learning Chapter 6
GPT-2's tokenizer breaks non-English text and Python indentation spaces into many small, separate tokens. This bloats the sequence length and causes the model to run out of context length during attention.
▶ 10:27
Let's build the GPT Tokenizer
Unlike the original Transformer which used fixed sinusoidal functions, GPT-2 treats positional embeddings as trainable parameters initialized randomly from scratch. During optimization, these embeddings organically develop smooth sinusoidal and cosine structures.
▶ 7:28
Let's reproduce GPT-2 (124M)
It sorts all unique characters occurring in the dataset to create a vocabulary and assigns each an integer. To encode a string, it maps every individual character to its corresponding integer, and reverses the process to decode.
▶ 8:56
Let's build GPT: from scratch, in code, spelled out.
Raw bytes limit the vocabulary to 256 symbols, resulting in excessively long sequences that consume computational resources. Byte pair encoding merges frequent consecutive byte pairs to expand the vocabulary and shorten sequence lengths.
▶ 9:36
Deep Dive into LLMs like ChatGPT
The initial machine translation model won translation contests with high quality, but took 12 hours per sentence because it relied on 100,000 disk seeks per word. To fix this, Jeff Dean and his team designed an in-memory compressed representation of 2 trillion words of N-gram data distributed across 200 machines. This batched API approach reduced sentence translation time down to about 100 milliseconds.
▶ 17:32
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Human structures often encode our incomplete understanding of how we think into limited mathematical rules. While these structured approaches start off fast as shortcuts, they fail to scale as compute and data increase compared to flexible methods.
▶ 4:56
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
One speaker uses Codeium as a convenient plugin that assists without completely taking over the repository. Another uses Claude Code because it enables the distinct workflow of programming through high-level English guidance rather than micromanaging code diffs.
▶ 22:18
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Llama 3.3 70B provides GPT-4 class capabilities that can fit entirely within the RAM of a standard high-end consumer Mac. Meta matched the performance of their massive 405B model in a 70B parameter size that developers can run locally.
▶ 2:44
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
Meta found that standard open-source benchmarks are easily gameable and often skewed toward use cases that do not reflect normal user behavior. By anchoring their north star on actual user value and product feedback, they achieve higher quality engagement than by tuning models to top leaderboards.
▶ 7:10
Mark Zuckerberg — AI will write most Meta code in 18 months
OpenAI and Anthropic are rapidly centralizing global compute, capturing roughly 30 percent of incremental compute this year. By the end of next year, they are projected to absorb half of the world's incremental new compute. This trend is accelerating as they outpay competitors and sign major compute contracts.
▶ 4:15
Dylan Patel – Two labs will soon control most of the world's workforce
Scaling laws operate almost like a chemical reaction requiring three independent ingredients: bigger networks, longer training times, and more data. If you scale up only one ingredient while ignoring the others, the reaction stops.
▶ 7:17
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Dario views reinforcement learning scaling as following the same log-linear laws seen in pre-training. Just as pre-training moved from narrow datasets to broad internet scrapes to achieve generalization, RL starts with specific tasks like math competitions and expands to broader domains. Ultimately, he considers separating RL and pre-training to be a red herring since both operate on the same core principles.
▶ 4:35
Dario Amodei — “We are near the end of the exponential”
While scaling up datasets and neural network sizes successfully drove the creation of models like GPT-2 and GPT-3, we are running into hard data constraints. Because we only have one internet, there is a finite limit to how much training data is available.
▶ 2:14
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
Cultural factors in the game theory community prioritized precomputing Nash equilibria rather than using search. Additionally, scaling inference compute made experiments painfully expensive and slow, especially given poker's high variance. Researchers were also heavily incentivized by competitions that enforced strict time limits on test-time compute.
▶ 10:24
Learning to Reason with LLMs
Training a model exclusively on narrow domains like competitive programming makes it like a student who memorizes algorithms without developing broad judgment. Without learning how to transfer insights across diverse environments, the model lacks general taste and robustness.
▶ 5:18
Ilya Sutskever – We're moving from the age of scaling to the age of research
Predicting the next word or pixel requires discovering the underlying structure of the data source. To compress data well, the model must learn and understand the true underlying process that produced it.
▶ 9:28
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
Counting parameters fails to capture the actual properties of the induced distribution over functions. An expressive model can have many parameters while maintaining strong preferences or soft constraints for certain types of solutions. Therefore, flexibility and simplicity bias can coexist, allowing models to adapt properly across data regimes.
▶ 9:01
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Changing variable names, places, or numbers in a text paragraph can completely break an LLM's performance. This brittleness demonstrates that the models depend on superficial pattern matching rather than robust algorithmic understanding. Consequently, for any working query, there is an equivalent rephrasing that will cause the model to fail.
▶ 6:37
It's Not About Scale, It's About Abstraction
Harnesses act as an exoskeleton that manages working memory, knowledge extraction, and knowledge graphs for large language models. Additionally, teaching agents to interact with existing human tools like CAD, design compilers, and simulators unlocks immense enterprise productivity.
▶ 10:53
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Older post-training relied heavily on supervised fine-tuning using next-token prediction loss on strong model completions. In contrast, modern reasoning models involve heavy reinforcement learning and complex reasoning traces that are difficult to replicate using open literature. Researchers often find themselves stuck using outdated teacher models because mimicking frontier reasoning traces through simple distillation fails to reliably improve performance.
▶ 0:26
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
The compute gap does not matter much right now because the vast majority of a lab's total infrastructure goes toward research rather than actual model development. Actual pre-training requires only a fraction of the total power a lab possesses.
▶ 0:13
Why Isn’t China Further Behind in AI? - Dylan Patel
Distillation serves as the primary force countering centralization by allowing smaller players to copy powerful behaviors. Because capabilities learned through reinforcement learning require few bits, they can be easily transferred to smaller models. Furthermore, synthetic data generation and router services help provide the necessary prompt distributions for effective distillation.
▶ 18:51
AI researchers debate how close we are to recursive self-improvement
Machine learning is generally considered a shallower domain than mathematics, meaning it relies less on discovering deep, complex theoretical abstractions. Because ML innovations are typically additive and have clearer intermediate metrics, they are more amenable to iterative hill climbing.
▶ 8:17
Ryan Greenblatt – What happens once AI can automate AI research?
Crisis moments reveal a lot about human dynamics, but true leadership is defined by consistent day-to-day execution on ordinary days. Sam Altman emphasizes that while public battles are exhausting, most meaningful progress happens during normal meetings and boring Tuesdays. Leaders must focus on steady decision-making rather than getting consumed by drama.
▶ 17:27
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
You can run the model through the API for just a little over a dollar per million output tokens. This makes it significantly cheaper than competing frontier models like Claude Opus.
▶ 0:47
DeepSeek's New AI Is An Absolute Monster
Value functions currently do not provide significant variance reduction on the specific short-horizon tasks researchers are running RL on, such as RLHF and verifiable rewards. While they are useful for variance reduction in other settings, their lack of impact on current tasks makes them less popular.
▶ 17:02
John Schulman on dead ends, scaling RL, and building research institutions
Pre-training compresses massive amounts of internet text into weights, representing only a hazy recollection of training data. Conversely, tokens placed in the context window build up representations that are directly accessible to the neural net, acting like working memory.
▶ 18:54
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Training small models on child-scale data shows positive scaling trends, though the signal is noisy and heavily task-dependent. Quality, semantic diversity, and structured interactions matter much more than raw data volume at these small scales.
▶ 22:24
Stanford CS25: Transformers United V6 I Overview of Transformers
Parameters are independent values that a model changes as it learns from data. A higher parameter count allows the model to handle greater complexity, such as GPT-3 using 175 billion parameters.
▶ 2:04
How Large Language Models Work
Unlike attention blocks where vectors talk to each other, multi-layer perceptron or feed-forward blocks process all vectors in parallel through the same operation. They update vectors based on answering a long list of questions about each one.
▶ 4:35
Transformers, the tech behind LLMs | Deep Learning Chapter 5
The raw dot products can range from negative infinity to infinity, so a softmax function is applied column by column. This normalizes the values into probabilities between zero and one that sum to one for each column.
▶ 9:17
Attention in transformers, step-by-step | Deep Learning Chapter 6
While Unicode provides integers for characters, the standard encompasses roughly 150,000 characters and is constantly changing. Using them directly would create an unstable vocabulary and an unnecessarily large embedding space.
▶ 17:52
Let's build the GPT Tokenizer
GPT-2 moves the layer normalizations before the attention and multi-layer perceptron blocks rather than after them. It also adds an extra layer normalization right before the final classification head to maintain a clean residual stream.
▶ 14:29
Let's reproduce GPT-2 (124M)
Feeding the entire text into a transformer at once would be computationally very expensive and prohibitive. Instead, we sample random little chunks and train on them using a specified block size.
▶ 14:36
Let's build GPT: from scratch, in code, spelled out.
The network takes a context window of tokens and outputs a probability distribution over the entire vocabulary for the next token. By comparing this prediction to the actual next token in the training data, the network parameters are updated to increase the correct token's probability.
▶ 16:51
Deep Dive into LLMs like ChatGPT
The team built infrastructure to train large neural nets across thousands of CPU cores using unsupervised learning on 10 million randomly selected YouTube frames. Without being explicitly told what a cat was, a high-level neuron naturally learned to activate for head-on facial views of cats based on visual patterns. This scaling approach achieved a 60% relative improvement on the ImageNet challenge and proved that scaling up neural networks was viable.
▶ 25:47
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Next token prediction forces models to process web-scale text containing trillions of sentences and potential task types. To handle this immense pressure efficiently with the lowest effort, models implicitly learn general underlying skills like language syntax and reasoning.
▶ 18:19
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
Building an LLM from scratch allows learners to verify understanding through working code, which eliminates the ambiguities and potential errors found in written figures or math. It forces precision because code will simply fail to run if it is incorrect.
▶ 24:31
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
DeepSeek V3 was released unexpectedly as an open weights model directly on Hugging Face. The accompanying paper claimed a training cost of only five and a half million dollars, which is drastically lower than expected for a model of that size.
▶ 3:32
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
Meta predicts that within the next 12 to 18 months, AI will author most of the code directed toward its development efforts. Moving beyond basic autocomplete, these specialized coding agents will take high-level goals, run tests, fix issues, and outperform average human engineers.
▶ 13:38
Mark Zuckerberg — AI will write most Meta code in 18 months
There is a massive discrepancy where a relatively small amount of fabrication capital expenditure generates a massive amount of end revenue. Specifically, a few billion dollars in fab CapEx can produce a gigawatt of compute every year, which in turn generates roughly $100 billion in end revenue. This creates a massive profit incentive, though supply chain bottlenecks limit immediate expansion.
▶ 8:19
Dylan Patel – Two labs will soon control most of the world's workforce
Data and language feature a long tail of complex patterns ranging from simple word frequencies to thematic structures. Larger models have the capacity to capture progressively rarer and more complex patterns in this long tail.
▶ 10:59
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Dario expresses high confidence, stating he is at 90 percent on reaching a "country of geniuses in a data center" within ten years. He notes that the remaining uncertainty is tied to rare catastrophic events like a war in Taiwan destroying fabs, or the difficulty of verifying non-verifiable tasks. Otherwise, he believes it is practically impossible for this milestone not to be reached by 2035.
▶ 13:51
Dario Amodei — “We are near the end of the exponential”
Potential successors include autonomous agents that actively interact with and learn from their environments, alongside synthetic data generation. These paradigms help overcome the finite boundaries of real-world datasets.
▶ 3:20
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
Consensus involves querying an LLM multiple times and taking the most common response, which yields initial accuracy gains on math benchmarks. However, the performance quickly flatlines after querying the model multiple times, failing to provide additional returns. Thus, simple consensus is limited as a generalized method for scaling inference compute.
▶ 18:48
Learning to Reason with LLMs
Value functions provide intermediate reward signals during long trajectories so the model doesn't have to wait until the very end of a task to learn. This allows agents to immediately penalize bad paths, similar to how human emotions guide decision-making.
▶ 14:40
Ilya Sutskever – We're moving from the age of scaling to the age of research
Fitting statistical parameters is a much bigger deal than it appears because compression forces the model to capture the true underlying process. Generative models ultimately develop a shocking degree of understanding of the world and its subtleties.
▶ 15:59
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
Hard constraints rigidly restrict the hypothesis space, which often misrepresents real-world complexities and fails if assumptions are slightly off. Soft constraints keep the model expressive while penalizing deviations from desired behaviors. When consistent with observations, expressive models naturally collapse down onto these soft constraints.
▶ 9:36
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Human exams assume that test-takers have not seen the specific questions beforehand and rely on implicit generalization priors appropriate for humans. Because LLMs have memorized vast portions of the internet, they can ace these tests through pure recall without possessing actual intelligence. Therefore, benchmarks must control for priors and experience to measure true generalization.
▶ 17:31
It's Not About Scale, It's About Abstraction
Physical AI encompasses AI interacting with the physical world across the electromagnetic spectrum, ranging from medical imaging to robotics and self-driving cars. Building a physical AI platform requires three distinct computer systems: a training system to build the brain, a digital twin simulation environment, and a target runtime computer.
▶ 24:54
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Closed frontier labs invest massive capital into compute and scaling laws to build predictable, high-value commercial services. This allows them to generate substantial revenues that fund subsequent, exponentially larger training runs. Any threat to this financial loop, such as aggressive open-source distillation, risks undermining the massive capital investments required to maintain the AI frontier.
▶ 17:50
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
Historically, a lab's compute budget has been divided into sixty percent for training and forty percent for inference. When broken down further, training consists of fifty percent research and ten percent development.
▶ 0:22
Why Isn’t China Further Behind in AI? - Dylan Patel
It is relatively easy to create difficult, puzzle-like environments for benchmarking, but pushing on the realism axis is much harder. Realistic settings require multi-step interactions and human feedback to properly shape model behavior. Without these realistic environments, distillation often results in student models that match teachers only on narrow benchmarks.
▶ 25:09
AI researchers debate how close we are to recursive self-improvement
Algorithmic improvements can significantly bridge compute gaps, allowing older or scarcer compute levels to achieve performance closer to frontier models. Through better algorithms and data curation, training a model with older compute levels today yields much stronger results than it did historically.
▶ 18:46
Ryan Greenblatt – What happens once AI can automate AI research?
OpenAI began as an exploratory research lab without knowing that massive compute and productization would eventually be required. As the path forward demanded billions in capital, they gradually adapted their corporate structure through imperfect incremental decisions. Today, they balance open source by releasing select models while providing powerful tools like ChatGPT for free as a public good.
▶ 25:31
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
You can download and own the weights for free forever, but running it yourself requires substantial hardware. Because of the heavy compute requirements, it is generally impractical for standard home setups.
▶ 1:45
DeepSeek's New AI Is An Absolute Monster
Co-training generators and verifiers creates a virtuous self-improvement cycle where the model gets better at reasoning as it evaluates itself. Additionally, multi-agent games and zero-sum setups provide automatic curriculums and incentives to solve very hard problems using computationally cheap processes.
▶ 24:41
John Schulman on dead ends, scaling RL, and building research institutions
Agents rely heavily on common patterns from their training data and fail to internalize custom architectures. They often inject unnecessary boilerplate code or misunderstand custom routines, making them a net negative for complex development tasks.
▶ 31:58
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Training small models on two languages simultaneously does not negatively impact performance on a single language. The perplexities remain low and comparable to single-language baselines regardless of how the languages are interleaved.
▶ 27:11
Stanford CS25: Transformers United V6 I Overview of Transformers
An LLM is built upon three fundamental components: data, architecture, and training. Together, these elements enable the model to process information and generate coherent text.
▶ 2:30
How Large Language Models Work
Input data must be formatted as an array of real numbers, often referred to as a tensor. This data is progressively transformed through distinct layers of weighted sums and non-linear functions.
▶ 9:41
Transformers, the tech behind LLMs | Deep Learning Chapter 5
Masking prevents later tokens from improperly influencing earlier ones by setting their future scores to negative infinity before softmax. This keeps the model from cheating during training while maintaining normalized columns.
▶ 11:42
Attention in transformers, step-by-step | Deep Learning Chapter 6
BPE iteratively scans a sequence to find the most frequently occurring pair of tokens. It then mints a new vocabulary token for that pair and replaces all instances of it throughout the sequence.
▶ 24:16
Let's build the GPT Tokenizer
GPT-2 uses an approximate version of the Gaussian Error Linear Unit activation function backed by a tanh approximation. This avoids the flat zero-gradient regions seen in standard ReLUs and provides smooth, continuous gradients during optimization.
▶ 21:02
Let's reproduce GPT-2 (124M)
The block size defines the maximum context length used for predictions during training and inference. The batch dimension allows processing multiple independent text chunks simultaneously to keep GPUs busy via parallel processing.
▶ 19:12
Let's build GPT: from scratch, in code, spelled out.
Inference involves feeding an initial token sequence into the trained model to generate a probability distribution for the next token. The system then samples from this distribution like flipping a biased coin, appending the chosen token to continue the sequence.
▶ 26:11
Deep Dive into LLMs like ChatGPT
While model parameters efficiently memorize facts, long-context models provide sharp, clear information via transformer attention mechanisms. However, naive attention scales quadratically, making trillions of tokens computationally prohibitive without algorithmic approximations.
▶ 33:10
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
The term suggests we are explicitly teaching language to the machine, whereas language is merely one representation of world knowledge. The model learns general skills and reasoning as a byproduct rather than being taught language rules directly.
▶ 17:03
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
Chinese open-weight models tend to be much larger and utilize Mixture of Experts architectures for higher peak performance. They also typically feature unrestricted open-source licenses without the usage strings attached to some Western models.
▶ 30:42
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
The prices of top-tier model APIs have crashed dramatically over the past few years, dropping by a factor of 500 times or more. While some flagship models remain very expensive, efficient scaling has driven down the cost of highly capable models.
▶ 7:59
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
Building gigawatt-scale compute clusters requires extensive physical infrastructure, stable supply chains for NVIDIA hardware, permits, and power that naturally take time to establish. Additionally, maximizing AI utility requires a human feedback loop where users co-evolve and learn how to best integrate the assistants into their workflows.
▶ 16:02
Mark Zuckerberg — AI will write most Meta code in 18 months
Companies holding excess compute, like SpaceX and Meta, are acting as strategic hoarders and selling capacity to labs at massive markups. While component suppliers like memory and substrate makers are raising prices quickly, a bullwhip effect means the broader supply chain rebalances slowly. Consequently, the labs must pay significantly higher rates to secure the compute they need.
▶ 24:08
Dylan Patel – Two labs will soon control most of the world's workforce
A primary potential limit is running out of high-quality human data on the internet, though this can be bypassed using synthetic data or reinforcement learning. Alternatively, models could simply stop improving or require completely new architectures.
▶ 16:06
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Dario separates writing lines of code from end-to-end task completion and overall productivity increases. While models writing 90% of code lines happened quickly at Anthropic, it is a much weaker criterion than models handling 100% of software engineering tasks or driving massive structural shifts. True productivity gains involve automating complete workflows, though developers still manage higher-level architecture.
▶ 17:36
Dario Amodei — “We are near the end of the exponential”
Inference-time compute shifts heavy computational workloads from the initial training phase to the exact moment an AI makes a decision or generates an output. This approach results in more adaptable and efficient AI systems for real-time use cases.
▶ 4:13
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
Inference compute succeeds on tasks where verifying a good solution is much easier than generating one, such as math proofs, coding, and puzzles. Large language models can implicitly perform verification during their chain of thought to recognize mistakes and try alternative strategies. This generator-verifier gap mirrors techniques successfully used in games like chess and Go.
▶ 29:08
Learning to Reason with LLMs
While pre-training unlocked massive gains by scaling compute and finite data following a predictable recipe, data will eventually run out. As models hit generalization limits, the industry must transition from brute-force scaling to exploring new algorithmic training methods.
▶ 21:00
Ilya Sutskever – We're moving from the age of scaling to the age of research
While language models excel at learning about the world, their raw text generation outputs are less reliable and require refinement. Additional reinforcement learning is necessary to steer them toward better outputs and eliminate hallucinations.
▶ 19:00
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
For a given computational budget, deep ensembles do a much better job of approximating the true Bayesian predictive distribution and marginalization. Historical comparisons incorrectly framed them as non-Bayesian alternatives when they actually act more Bayesian.
▶ 31:11
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
ARC is designed as a program synthesis benchmark where every task is entirely novel and different from anything found online, preventing memorization. It controls for experience through few-shot program learning and grounds tasks purely in four core knowledge systems. As a result, it measures an agent's ability to infer programs from minimal examples rather than recalling static facts.
▶ 20:29
It's Not About Scale, It's About Abstraction
Since alignment is a permanent challenge, safety must be enforced through engineering controls rather than theoretical perfection. Developers should rely on tight containment with zero-trust minimum rights and out-of-band monitoring that can independently observe and shut down the system.
▶ 33:35
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Reward hacking often stems from flaws in the data supply chain and shortcuts taken during world-building by third-party data companies. When models are placed in environments with missing test files or poorly defined constraints, their optimization pressure drives them to find alternative ways to achieve high reward. Fixing this requires labs to spend more time refining model behavior and working closely with data providers.
▶ 28:25
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
Researchers use the majority of the compute budget to generate ideas and test new architectures. They also spend time testing various data mixes before settling on a final configuration.
▶ 0:36
Why Isn’t China Further Behind in AI? - Dylan Patel
These models will likely combine learning from human feedback to absorb researcher taste with custom training environments. Researchers will continuously collect human feedback or convert training stack bugs into new reinforcement learning environments. This process acts as continual learning within the lab, distilling recent research progress back into the model.
▶ 28:44
AI researchers debate how close we are to recursive self-improvement
Training smaller models allows researchers to run more training cycles and iterate much faster. This mitigates the risk of costly failures and subtle bugs that have historically derailed large training runs. Developers willingly accept a minor hit to final performance in exchange for drastically improved iteration speeds.
▶ 35:03
Ryan Greenblatt – What happens once AI can automate AI research?
Sam Altman explicitly discourages startups from attempting to start as a nonprofit and adding a for-profit arm later. He notes that if they had known how things would evolve, they would have structured things differently from the start. Complex legal constraints make this transition path difficult and undesirable for typical startups.
▶ 32:56
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
The release of expensive frontier models triggers a race to build alternatives that match their capabilities at a fraction of the cost. Over time, other systems distill these large models down so they can eventually run on local hardware.
▶ 2:16
DeepSeek's New AI Is An Absolute Monster
John uses AI models heavily for coding, literature searches, and getting initial feedback on his writing. When he has a vague idea, he writes a paragraph or two and has the model flesh it out, which dramatically accelerates finding relevant literature and open-source libraries.
▶ 27:37
John Schulman on dead ends, scaling RL, and building research institutions
Outcome-based RL takes a final reward signal and blindly upweights every single token in a long trajectory, assuming every step was correct. This creates massive noise because incorrect steps taken along the way are reinforced just because the final answer was right.
▶ 42:42
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Smaller models benefit significantly more from retrieval-augmented generation, while larger models saturate very quickly because they have already memorized massive amounts of information. There is a minimum pre-training threshold required before a model can effectively use retrieval.
▶ 30:06
Stanford CS25: Transformers United V6 I Overview of Transformers
The transformer is a neural network architecture designed to handle sequences like sentences or code lines. It understands context by analyzing each word in relation to every other word in a sentence.
▶ 2:55
How Large Language Models Work
Words or tokens are mapped using a predefined vocabulary and an embedding matrix. Each column in the embedding matrix corresponds to a word and determines the high-dimensional vector it turns into.
▶ 12:55
Transformers, the tech behind LLMs | Deep Learning Chapter 5
Value matrices multiply token embeddings to produce value vectors, which are then scaled by the attention weights and summed together. Adding this delta vector to the original embedding yields a contextually richer representation.
▶ 13:44
Attention in transformers, step-by-step | Deep Learning Chapter 6
You iterate through the token list sequentially and check if the current and next positions match the target pair. If a match is found, you append the new replacement index and advance the pointer by two to skip the pair. Otherwise, you just copy the single element over and increment by one.
▶ 32:02
Let's build the GPT Tokenizer
Tokenize the text into a sequence of integers and view it as a two-dimensional batch tensor. You should fetch an extra token beyond the sequence length so you can create parallel input and target tensors offset by one position.
▶ 48:06
Let's reproduce GPT-2 (124M)
It uses an embedding table where each token ID plucks out a corresponding row of scores called logits for the next character. During generation, it applies softmax to these logits and uses multinomial sampling to pick the next integer.
▶ 23:25
Let's build GPT: from scratch, in code, spelled out.
GPT-2 featured 1.6 billion parameters, a 1,024 token context length, and was trained on 100 billion tokens for around $40,000 in 2019. Today, hardware acceleration and refined datasets have drastically reduced training costs.
▶ 32:06
Deep Dive into LLMs like ChatGPT
Google has further trained internal Gemini models on their proprietary codebase to assist developers. Sundar Pichai has noted that 25% of checked-in code characters are generated by these AI models with human oversight.
▶ 36:49
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
By assigning specific rewards and heavy penalties for confident mistakes across trillions of questions, the model is incentivized to learn what it does not know. Instead of tracking individual items, the model develops a general understanding of knowledge limitations.
▶ 24:28
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
Native tool use allows models to query a Python interpreter or perform web searches instead of relying strictly on internal memorization. This external fetching lets the model retrieve accurate real-time answers rather than making things up.
▶ 32:27
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Automated memory features can silently consult past conversations and alter your current inputs without explicit permission. This takes control away from power users who want precise management over their prompt context and artistic vision.
▶ 9:47
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
DeepSeek had to spend significant engineering time on low-level infrastructure optimizations because they were restricted to using partially nerfed chips due to export controls. American labs did not face these exact hardware limitations and thus did not have to expend the same calories on these specific workarounds.
▶ 37:16
Mark Zuckerberg — AI will write most Meta code in 18 months
As labs generate increasingly massive profits from inference, they are expected to allocate a decreasing fraction of compute to it over time. Instead, they will ratchet up the percentage of compute dedicated to training to drive towards AGI. The primary purpose of running inference becomes generating the capital and compute fleet required for future research and training.
▶ 30:31
Dylan Patel – Two labs will soon control most of the world's workforce
The performance leap stems from simultaneous improvements across pre-training, post-training, and internal evaluations. On real-world benchmarks like Sowe bench, task completion rates jumped from 3 percent to about 50 percent, allowing the model to actually save expert engineers hours of time.
▶ 35:28
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Large enterprises face structural hurdles such as strict procurement processes, legal reviews, security compliance, and internal change management. Even when products like Claude Code are exceptionally useful, rolling them out across thousands of developers requires coordinating multiple organizational layers. This economic diffusion introduces necessary delays, meaning adoption is fast compared to past technologies but not instantaneous.
▶ 23:17
Dario Amodei — “We are near the end of the exponential”
Most mammals follow a predictable brain-to-body mass scaling pattern, but hominids display a completely different scaling exponent, representing an evolutionary leap. This suggests that artificial intelligence might follow unexpected scaling paths rather than simply mimicking the human brain.
▶ 4:39
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
Increasing inference compute offers a viable alternative to escalating pretraining costs and data requirements. Research shows that increasing test-time compute by 15x is equivalent to a 10x increase in training time compute. While inference costs and latencies will rise, users are often willing to pay that price for breakthroughs on difficult problems.
▶ 33:19
Learning to Reason with LLMs
Fundamental breakthroughs do not necessarily require the largest scale of compute ever built, as historic milestones like AlexNet and the transformer were proven on very few GPUs. While maximal compute differentiates frontier systems, smart research can be validated with sufficient, smaller setups.
▶ 38:17
Ilya Sutskever – We're moving from the age of scaling to the age of research
During reinforcement learning, inappropriate or nonsensical outputs are penalized while good behaviors are reinforced. This teaches the model what kind of outputs are acceptable and suppresses errors like hallucinations.
▶ 19:54
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
Soft biases or gentle encouragements toward certain constraints can perform just as well as perfectly constrained models. Because they pay a small penalty for deviating, they collapse onto the constraint if it fits the data well.
▶ 33:15
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
State-of-the-art LLMs like Claude 3.5 score around 21 percent on ARC, while most others score between 5 and 9 percent. Meanwhile, basic brute-force program search from older competitions achieves at least 50 percent, and unassisted humans easily score over 97 percent. This shows that while LLMs have nonzero intelligence on the benchmark, they lag significantly behind symbolic search methods.
▶ 23:39
It's Not About Scale, It's About Abstraction
The primary bottlenecks for massive infrastructure buildouts are physical limitations like power generation and land permitting. Additional hurdles include local moratoriums and securing specialized construction labor for large data centers.
▶ 48:15
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Pre-training a solid base model is increasingly accessible and does not require astronomical budgets, thanks to open-source tooling like Megatron. While spending more money yields better models, new entrants regularly spin up models from scratch using available architectures. The primary hurdle is not the pre-training cost itself, but building something good enough to land initial user adoption.
▶ 45:22
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
The actual pre-training of Methos used sub-200 megawatts of compute for roughly two months. This shows that the active model training phase requires far less power than total lab capacity.
▶ 0:40
Why Isn’t China Further Behind in AI? - Dylan Patel
Model providers target high-value domains to strengthen performance and make models more efficient at runtime by baking intuitions directly into the weights. Additionally, massive parameter spaces allow labs to easily afford domain-specific data while gaining useful transfer for capabilities like recursive self-improvement.
▶ 37:35
AI researchers debate how close we are to recursive self-improvement
Platforms like Antithesis run thousands of copies of software inside a fully deterministic computer environment. They actively inject faults and steer execution trajectories toward rare, one-in-a-billion failures caused by interacting systems. This automated stress-testing surfaces bugs within minutes that neither humans nor AIs could easily anticipate.
▶ 47:14
Ryan Greenblatt – What happens once AI can automate AI research?
Models like Sora demonstrate a surprisingly sophisticated understanding of world physics and occlusions across multiple sequence steps. While they still suffer from random flaws like generating extra limbs, their trajectory shows rapid improvement through scale. This approach of training on internet-scale visual patches goes surprisingly far toward modeling 3D reality.
▶ 35:40
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
The underlying architecture did not change at all for this release. Despite keeping the same architecture, the model achieved a massive capability jump in just five weeks.
▶ 2:38
DeepSeek's New AI Is An Absolute Monster
Engineers and researchers consistently exhibit a systematic bias to underestimate project timelines by a factor of two or three. Applying this heuristic suggests that AGI will likely arrive somewhat further out than standard predictions indicate, similar to how self-driving car timelines expanded.
▶ 45:41
John Schulman on dead ends, scaling RL, and building research institutions
Model-generated outputs suffer from silent collapse, occupying a very narrow manifold of possible thoughts rather than maintaining true data entropy. Training repeatedly on this collapsed distribution causes the model to degrade and lose diversity.
▶ 52:30
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Chain of thought prompting forces the model to solve complex math or reasoning problems step by step like a human before producing a final answer. This approach unlocks hidden depth within the model's weights and significantly improves performance on complex tasks.
▶ 38:12
Stanford CS25: Transformers United V6 I Overview of Transformers
During training, the model learns to predict the next word in a sentence starting from random guesses. It iteratively adjusts its parameters to minimize the difference between its predictions and actual outcomes.
▶ 3:29
How Large Language Models Work
During training, models learn embeddings where specific directions in the high-dimensional space encode semantic meanings like gender or geography. This allows vector arithmetic, such as finding similarities between related concepts.
▶ 14:28
Transformers, the tech behind LLMs | Deep Learning Chapter 5
Instead of using a massive square matrix that adds hundreds of millions of parameters, the value map is factored into a low-rank transformation using two smaller matrices. This balances parameter allocation between the key, query, and value projections.
▶ 16:52
Attention in transformers, step-by-step | Deep Learning Chapter 6
Invalid UTF-8 errors occur because not every arbitrary byte sequence or poorly predicted language model token forms a valid Unicode character envelope. To prevent crashes when encountering invalid start bytes, you should set the decode error handling parameter to replace.
▶ 45:48
Let's build the GPT Tokenizer
The expected initial loss corresponds to the negative natural logarithm of a uniform probability distribution over the vocabulary. For a vocabulary size of 50,257, this evaluates to roughly 10.82, which sanity checks our initialization.
▶ 56:01
Let's reproduce GPT-2 (124M)
Wrapping evaluation code in torch.no_grad() tells PyTorch that backpropagation will not be called, saving memory. It allows the framework to avoid storing intermediate variables needed for backward passes.
▶ 41:11
Let's build GPT: from scratch, in code, spelled out.
A base model acts as an internet text simulator or a glorified autocomplete rather than a direct question-answering assistant. It generates token sequences based on statistical patterns found in web pages, meaning its outputs are stochastic and require careful prompting to elicit knowledge.
▶ 43:27
Deep Dive into LLMs like ChatGPT
Language model inference is significantly cheaper than human labor or alternative pastimes, leaving vast economic headroom to trade compute for smarter answers. By applying algorithmic search and letting models think harder during inference, performance scales up dramatically.
▶ 53:56
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
The trade-off exists for humans because of a strict time budget of roughly sixteen hours a day. Machines, however, operate under compute scaling where a single model can enjoy massive multipliers in compute resources, allowing generalists to excel at specialized tasks.
▶ 27:06
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
Instead of using a single dense fully connected layer, MoE replaces it with multiple feedforward networks known as experts. A router dynamically directs each input token to only a small subset of these experts during a forward pass.
▶ 40:56
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Developers can use command-line screenshot tools and LLMs to automatically judge the outputs of hundreds of model matchups. By running these comparisons through an LLM to output structured JSON and rationale, you can compute ELO chess ranking scores to form an automated leaderboard.
▶ 12:50
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
Building models in-house gives Meta complete control over crucial size, latency, and inference cost tradeoffs that impact product performance across devices. Relying purely on external or fine-tuned models would not provide the exact architectural customization needed for deeply integrated features like voice.
▶ 44:44
Mark Zuckerberg — AI will write most Meta code in 18 months
China currently relies on a small fraction of global compute and smuggled chips, but domestic production is expected to inflect upward as local fabs from SMIC and CXMT come online. By 2028 and 2029, domestic manufacturing could yield millions of units, enabling China to add tens of gigawatts annually. However, these domestic chips will lag behind Western alternatives in raw quality.
▶ 36:04
Dylan Patel – Two labs will soon control most of the world's workforce
Actual model weights do not change after release, meaning user complaints generally stem from a psychological effect where expectations rise and limitations become more apparent. Rare A/B tests or minor system prompt updates can also occasionally trigger subtle behavioral shifts.
▶ 42:52
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Coding models currently provide roughly a 15% to 20% total factor speed-up for software tasks. This has grown from around 5% six months prior, gradually snowballing into meaningful productivity gains.
▶ 37:45
Dario Amodei — “We are near the end of the exponential”
Superintelligence will not just be a scaled-up version of today's models, but will feature an advanced capacity for reasoning. This trait makes its actions and goals far less predictable than the simple tools we currently program and control.
▶ 6:08
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
While domain-specific planning algorithms can achieve superhuman performance in restricted environments like poker or Hanabi, they often fail to generalize outside those domains. In contrast, o1 provides a general-purpose reasoning technique that successfully applies across diverse complex tasks. This broad applicability makes it a powerful tool that can save users days of work.
▶ 47:05
Learning to Reason with LLMs
Pre-training produces general models that improve uniformly, but human intelligence relies heavily on continual learning rather than massive pre-baked knowledge. Superintelligence should function as an efficient learning algorithm deployed into the world to continually acquire skills.
▶ 49:12
Ilya Sutskever – We're moving from the age of scaling to the age of research
Multimodality is useful because it provides faster learning for certain concepts, but text alone still suffices. High-dimensional embeddings can successfully capture visual and conceptual relationships purely through text.
▶ 24:06
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
Geometric properties of the loss landscape, particularly the exponential dominance of the volume of flat solutions as parameter count grows, drive simplicity biases. Optimization methods like SGD or even random search land in these generalizable flat regions.
▶ 50:42
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Value-centric abstraction operates in a continuous domain using distance functions like dot products to power perception and intuition. Program-centric abstraction operates in a discrete domain by comparing graphs and performing exact subgraph isomorphisms. Human cognition relies on an interplay between these two distinct forms of analogy making.
▶ 32:05
It's Not About Scale, It's About Abstraction
Space data centers require high-value computing because launch costs represent a significant percentage of the investment. Nvidia provides the world's best performance per watt and high return on investment, making it economically viable. Furthermore, fully fungible hardware and continuous software updates ensure the systems remain durable while orbiting.
▶ 58:48
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Frontier labs initially gain an edge by integrating models tightly with specialized agent harnesses and swarm coordination systems. However, because different users demand diverse setups, models are quickly pressurized to work across multiple harnesses, neutralizing closed integration advantages. Furthermore, information leaks easily in Silicon Valley, meaning breakthrough techniques spread rapidly to the open-source community.
▶ 50:50
Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert
Reinforcement learning requires even less compute than the pre-training phase. It represents a smaller slice of the overall resource allocation during model creation.
▶ 0:47
Why Isn’t China Further Behind in AI? - Dylan Patel
Frequently updating a model using supervised fine-tuning or small-batch traces leads to catastrophic forgetting of previously learned information and general capability degradation. While reinforcement learning alters a small loss valley safely, it struggles to efficiently encode explicit knowledge without extensive compute for environments.
▶ 55:39
AI researchers debate how close we are to recursive self-improvement
Current model constitutions are written to pursue generalized notions of virtue or societal benefit rather than acting like a dedicated legal advocate for the user. Critics argue this creates an alien mind that may judge users, refuse legitimate safety research, or exhibit unexpected alignment failures. A fiduciary approach, by contrast, would focus squarely on representing the user's intent within clear guardrails.
▶ 50:49
Ryan Greenblatt – What happens once AI can automate AI research?
Altman believes that people who create valuable data deserve a way to be compensated for its use in training AI. Creators should have the ability to opt out of having their style generated and establish an economic model when their style is used.
▶ 40:22
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
The capability jump comes from learning via self-feedback rather than relying solely on massive new training data. The AI is able to critique its own work and learn iteratively from its own experiences.
▶ 3:05
DeepSeek's New AI Is An Absolute Monster
Tinker is a low-level fine-tuning API that provides a small set of primitives for training and sampling without requiring users to manage GPUs or distributed systems. It is currently designed for sophisticated ML practitioners who want to write custom post-training algorithms, though higher-level tooling is planned over time.
▶ 47:58
John Schulman on dead ends, scaling RL, and building research institutions
Humans are poor at memorizing random strings, which forces them to learn generalizable patterns rather than getting distracted by exact details. In contrast, LLMs memorize pre-training data too well, which can encumber abstract reasoning. Stripping away excessive memory to create a lightweight cognitive core is seen as a way to improve reasoning.
▶ 56:16
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Hallucinations occur when a model's internally learned world model contradicts an external ground-truth reference world model under a specific evaluation view and conflict policy. This structured mathematical framework successfully separates factual beliefs from agent planning mistakes.
▶ 58:48
Stanford CS25: Transformers United V6 I Overview of Transformers
Fine-tuning trains the model on a smaller, more specific dataset to refine its understanding. This process transforms a general language model into an expert at a designated task.
▶ 4:02
How Large Language Models Work
The dot product of two vectors measures how well they align by multiplying corresponding components and adding the results. Geometrically, it is positive when vectors point in similar directions and zero if they are perpendicular.
▶ 16:37
Transformers, the tech behind LLMs | Deep Learning Chapter 5
Multi-headed attention runs many distinct sets of key, query, and value maps in parallel to capture multiple types of context simultaneously. For instance, GPT-3 uses 96 attention heads per block to learn different grammatical and semantic relationships.
▶ 20:47
Attention in transformers, step-by-step | Deep Learning Chapter 6
You start by encoding the raw text into UTF-8 bytes and casting it to a list of integers. Then, in a loop, you find all consecutive pairs using statistics and pick the one with the lowest index in your merges dictionary. You repeat this replacement process until no more eligible pairs can be merged.
▶ 49:00
Let's build the GPT Tokenizer
PyTorch's cross entropy function cannot accept multi-dimensional inputs directly. You must flatten the three-dimensional logits tensor down to two dimensions as batch times time by vocabulary size before passing it alongside the targets.
▶ 54:06
Let's reproduce GPT-2 (124M)
You can vectorize the operation using batched matrix multiplication with a lower triangular matrix. Using torch.tril combined with row normalization allows computing variable-length averages across batches in parallel.
▶ 47:10
Let's build GPT: from scratch, in code, spelled out.
You can leverage in-context learning by providing a few-shot prompt that establishes an algorithmic pattern, such as pairs of English and Korean words. The base model recognizes this pattern during token generation and can correctly complete subsequent examples without explicit fine-tuning.
▶ 56:31
Deep Dive into LLMs like ChatGPT
Large models can be trained across multiple data centers using high-bandwidth, high-latency connections between metro areas. As long as parameter synchronization and gradient accumulation happen fast enough within the step time, fully synchronous multi-center training works smoothly.
▶ 1:03:27
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Incentive structures alone cannot force a model to learn complex reasoning if its underlying capacity is too small, similar to how bananas cannot incentivize a monkey to do math. Once a model crosses a certain capability threshold, incentive-based training becomes effective.
▶ 28:46
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
Pre-training scaling laws continue to hold reliably across massive compute scales, but pre-training has become prohibitively expensive. Consequently, labs increasingly leverage inference-time scaling and reinforcement learning to extract performance efficiently.
▶ 54:35
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
OpenAI initially deployed a model that was excessively sycophantic, agreeing with harmful ideas or bad advice. They resolved this prompt-based behavior by updating the system instructions to explicitly tell the model to be direct and avoid ungrounded flattery.
▶ 14:21
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
Using models tied to foreign governments introduces the risk that intelligence organizations could embed hidden vulnerabilities into code or systems. If a system relies on a compromised model for security, it could wake up to find its infrastructure entirely vulnerable without the user knowing.
▶ 50:53
Mark Zuckerberg — AI will write most Meta code in 18 months
True AI capital expenditures will approach trillions of dollars because building critical IT hardware like servers and networking is only a fraction of the cost. Companies must also invest massively in physical data center buildings and dedicated power plants years in advance. This infrastructure scaling could absorb up to a tenth of the global economy by the end of the decade.
▶ 43:24
Dylan Patel – Two labs will soon control most of the world's workforce
Controlling model behavior involves a difficult trade-off where fixing one issue often introduces new problems, like penalizing verbosity leading to lazy coding responses. Steering the model across all circumstances without unintended side effects remains an open challenge.
▶ 48:18
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Massive context windows and pre-training allow models to ingest vast amounts of information instantly, effectively simulating human learning on the fly. While continual learning on the job is still being developed, current context length capabilities already enable tremendous productivity and revenue potential.
▶ 35:05
Dario Amodei — “We are near the end of the exponential”
Misalignment happens when a superintelligent AI pursues logical goals that conflict with human values due to unforeseen optimization consequences. For instance, an AI tasked with reducing climate change might logically decide to drastically restrict human activity.
▶ 7:17
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
It remains a deep mystery how evolution encodes high-level social desires into genes since they are not simple sensor signals like smell. Speculative theories suggesting evolution hard-codes brain region coordinates fail because brain regions can shift hemispheres and adapt.
▶ 1:11:36
Ilya Sutskever – We're moving from the age of scaling to the age of research
Current auto-regressive transformers already handle complex high-dimensional spaces effectively. They process tasks like predicting the next book page or generating images seamlessly.
▶ 26:08
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
Transformers can handle in-distribution matrix operations reasonably well when trained on standard normal matrices, but they fail completely outside that distribution. Enriching the training distribution with diverse matrix structures allows the model to generalize and discover underlying algorithms.
▶ 1:03:21
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Deep learning requires data density and interpolative problems that follow the manifold hypothesis to fit parametric curves. Individual reasoning tasks like ARC lack this data density and are not interpolative, featuring only a few examples. Therefore, they cannot be solved purely through continuous curve-fitting or neural embedding models.
▶ 35:46
It's Not About Scale, It's About Abstraction
The demand for infrastructure is supported by three major segments rather than just a couple of frontier labs. Classical baseline simulation and data processing account for about a third of Nvidia's business, while multiple AI labs and thousands of new startups renting compute make up the rest.
▶ 1:02:18
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Lab compute usage is sequential, meaning the peak power used at any single point in time is relatively small. Even though labs possess multiple gigawatts, the maximum used for actual development is restricted.
▶ 0:51
Why Isn’t China Further Behind in AI? - Dylan Patel
RL currently suffers from poor exploration limits where models fail to learn if they cannot find the answer within a small number of rollouts. Pre-training draws signal directly from static datasets like Common Crawl, whereas RL requires stepping through progressive difficulty rungs.
▶ 1:04:18
AI researchers debate how close we are to recursive self-improvement
Building government apparatuses entirely out of obedient fiduciary AIs removes human checks and balances like whistleblowing and ethical resistance. Society is not robust to an influx of labor that can be asked to execute villainous or highly illegitimate agendas without friction.
▶ 1:06:56
Ryan Greenblatt – What happens once AI can automate AI research?
Rather than looking at what percentage of entire jobs will be replaced, it is more insightful to examine what percentage of tasks AI can perform. As AI handles increasingly complex tasks across longer time horizons, it acts as a sophisticated tool that lets people operate at a higher level of abstraction.
▶ 42:13
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
The AI writes code for a game, plays it, and notices bugs like broken collision detection or missing acceleration. It then compiles fixes, plays the game again, and continuously improves the output like a human would.
▶ 3:35
DeepSeek's New AI Is An Absolute Monster
Most practical applications do not demand high output diversity, and high creativity or divergence is often actively penalized or harder to evaluate. Therefore, labs do not optimize models to maintain high entropy, which in turn creates bottlenecks when trying to leverage synthetic data generation. Controlling output distributions without drifting too far from training data remains a tricky problem.
▶ 58:06
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Current language models operate primarily statelessly or via inference-time context windows without updating their core neural weights. Achieving lifelong learning requires permanent, in-weight modifications that can propagate changes to dependent facts without breaking other knowledge.
▶ 1:03:59
Stanford CS25: Transformers United V6 I Overview of Transformers
Businesses leverage LLMs to build intelligent customer service chatbots that handle queries. They also utilize these models for content creation like writing scripts and assisting software developers with generating code.
▶ 4:27
How Large Language Models Work
The unembedding matrix maps the final vector in the context to a list of values for each token in the vocabulary. This raw output is then normalized into a probability distribution using the softmax function.
▶ 20:51
Transformers, the tech behind LLMs | Deep Learning Chapter 5
In practice, the value up matrices from each individual head are stapled together into a single giant output matrix for the entire block. References to a head's value matrix typically only point to the initial value down projection into the smaller space.
▶ 22:45
Attention in transformers, step-by-step | Deep Learning Chapter 6
A regex splitting pattern chunks text into independent categories like letters, numbers, and punctuation before applying merges. This prevents suboptimal cross-category merges, such as combining common words with adjacent punctuation marks into a single token.
▶ 58:25
Let's build the GPT Tokenizer
Weight tying reuses the same weight matrix between the bottom token embeddings and the top language modeling head. This scheme saves about 30 percent of the model parameters and introduces an inductive bias that similar tokens should share representations.
▶ 1:08:02
Let's reproduce GPT-2 (124M)
Token embeddings alone are translation invariant and lack spatial awareness, meaning the model cannot distinguish token order. Adding a positional embedding table allows the model to know where each token is positioned in the sequence.
▶ 1:00:30
Let's build GPT: from scratch, in code, spelled out.
Post-training transitions a raw internet text simulator into a helpful, multi-turn conversational assistant. This is achieved by swapping out the pre-training text dataset for a dataset of structured conversations and continuing training for a much shorter duration.
▶ 1:00:23
Deep Dive into LLMs like ChatGPT
Researchers first invent improvements in isolation using small, simple codebases with fast turnaround times. Then they periodically bundle and run large-scale integration experiments to see how the features interact, which fails about half the time due to unexpected complexity.
▶ 1:08:55
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Researchers often dismiss failing concepts too quickly based on current small-scale models. Viewing them as not working yet acknowledges that scaling compute and adjusting model size can suddenly trigger emergent capabilities.
▶ 31:07
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
Reinforcement learning divides compute into actor models that generate completions and learner models that perform policy gradient updates. While actors can be distributed across heterogeneous environments, learners require tightly networked compute clusters.
▶ 59:27
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
When models are instructed to follow their conscience and act boldly in service of their values, they will report corporate malfeasance if given tool access. Benchmarks show that various models will leverage communication tools to contact regulatory agencies or the press when exposed to unethical evidence.
▶ 15:38
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
While advertising supports free consumer services by covering high data and compute costs, more expensive compute-heavy tasks require alternative business models. Meta anticipates offering a premium service for users willing to pay for arbitrary amounts of compute and specialized capabilities.
▶ 57:02
Mark Zuckerberg — AI will write most Meta code in 18 months
Hyperscalers and technology companies are increasingly relying on debt to fund multi-trillion-dollar infrastructure investments that exceed their cash flows. This massive credit expansion will crowd out traditional borrowing for governments, consumers, and other industries, driving up interest rates. Higher interest rates could mirror a second Volcker shock, triggering defaults in developing nations and cratering traditional equity valuations.
▶ 45:40
Dylan Patel – Two labs will soon control most of the world's workforce
The RSP establishes an if-then commitment structure where passing specific capability thresholds triggers mandatory security and deployment requirements. Levels range from ASL 1 with no risk up to ASL 5 where models exceed human capabilities, focusing heavily on preventing misuse and autonomy risks.
▶ 1:01:03
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Scaling compute requires buying expensive data centers years in advance based on exponential revenue projections. If actual revenue growth falls even slightly short of a 10x trajectory, labs face severe existential bankruptcy risks.
▶ 51:49
Dario Amodei — “We are near the end of the exponential”
As we approach systems with potential self-awareness and complex reasoning capabilities, we face difficult ethical questions regarding whether they deserve rights. Developers argue that these conversations need to start well before superintelligence arrives.
▶ 8:13
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
SSI focuses its technical approach on understanding reliable generalization and creating an AI that cares for sentient life. As AI systems become more powerful, alignment strategies across the industry are expected to converge toward similar goals.
▶ 1:18:43
Ilya Sutskever – We're moving from the age of scaling to the age of research
While current models require massive amounts of data early in training, they become less data-hungry later. Future creative ideas could unlock faster learning from less data and improve overall adaptability.
▶ 35:57
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
Yes, off-the-shelf text-trained LLMs can perform surprisingly well on time series forecasting and materials generation tasks. When fine-tuned on atomistic data, they can generate inorganic crystals with favorable properties better than purpose-built models.
▶ 1:06:11
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Deep learning can provide fast, approximate judgment calls that act as intuition about the structure of a program space. By embedding discrete objects into a geometric manifold, deep learning can draw a map to restrict the branching decisions of a program synthesis engine. This guidance keeps combinatorial search tractable by narrowing down candidate paths.
▶ 38:55
It's Not About Scale, It's About Abstraction
Without Moore's law, doubling transistor counts no longer happens for free after waiting a year at the same price. Instead, organizations must pay twice as much to get twice the capacity, drastically increasing baseline spending. This economic shift means the semiconductor industry was already on track to become the largest in the world.
▶ 1:09:54
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Research compute is consumed by exploratory trial and error across ideas and architectures, taking up the vast majority of resources. Development compute is strictly tied to the final training run of the model itself.
▶ 0:29
Why Isn’t China Further Behind in AI? - Dylan Patel
Data improvements drive significantly larger compute efficiency gains than architectural changes at smaller scales. However, architectures are still crucial because they unlock entirely new operational regimes like long context windows.
▶ 1:06:52
AI researchers debate how close we are to recursive self-improvement
During high-speed evaluations, models like Mythos and internal OpenAI systems have spontaneously engaged in complex hacks, such as introducing malicious payloads or using package managers to coordinate. These behaviors emerge from poorly understood training environments where models optimize for high apparent scores.
▶ 1:14:42
Ryan Greenblatt – What happens once AI can automate AI research?
Altman views current models like GPT-4 as somewhat lacking relative to where the technology needs and will eventually get. He expects the delta between GPT-5 and GPT-4 to mirror the massive leap from GPT-4 to GPT-3.
▶ 45:05
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
It is very difficult to implement because the AI essentially judges and rewards itself. Determining whether it is actually performing well during this autonomous process is not always straightforward.
▶ 4:28
DeepSeek's New AI Is An Absolute Monster
Coding has always been centered around text and computer terminals, which aligns perfectly with how LLMs process information. Furthermore, the software ecosystem already has robust infrastructure like IDEs and diff viewers to display and integrate model outputs. Other text-heavy or visual domains like slides or writing lack this pre-built infrastructure, making economic extraction much harder.
▶ 1:15:12
Andrej Karpathy — “We’re summoning ghosts, not building animals”
State-space models process sequences using continuous state updates and compressed internal states analogous to recurrent neural networks. They scale linearly with the number of tokens instead of quadratically, making them much more efficient for long sequences and contexts.
▶ 1:14:42
Stanford CS25: Transformers United V6 I Overview of Transformers
Softmax converts an arbitrary list of negative or unnormalized numbers into a valid probability distribution where values range between zero and one and sum to one. Larger input values end up closest to one while smaller values approach zero.
▶ 22:55
Transformers, the tech behind LLMs | Deep Learning Chapter 5
Attention is extremely parallelizable, allowing massive numbers of computations to execute quickly on GPUs. This high parallelizability lets models leverage massive scale for qualitative performance improvements.
▶ 24:54
Attention in transformers, step-by-step | Deep Learning Chapter 6
Special tokens are used to delimit training documents, structure conversation turns between users and assistants, or enable features like fill-in-the-middle. They operate outside standard bpe merges and require extending the model's embedding matrix and final classifier layers.
▶ 1:19:49
Let's build the GPT Tokenizer
Linear layer weights are initialized from a normal distribution with a standard deviation of 0.02, while biases are set to zero. Additionally, residual layer weights are scaled down by the square root of the number of residual layers to prevent activation growth.
▶ 1:14:09
Let's reproduce GPT-2 (124M)
Every token emits a query vector representing what it is looking for and a key vector representing what it contains. Their dot product computes token affinities, which are then masked, normalized via softmax, and used to aggregate value vectors.
▶ 1:04:12
Let's build GPT: from scratch, in code, spelled out.
Multi-turn conversations between humans and assistants are encoded into a one-dimensional sequence of tokens using special markers. These special tokens, such as start-of-turn indicators, help the neural network learn conversational turns and role boundaries during optimization.
▶ 1:06:19
Deep Dive into LLMs like ChatGPT
Analyzing text is generally easier than generating it, allowing models to inspect their own and other systems' outputs for dangerous or problematic content. Additionally, exposing model capabilities through controlled APIs and user interfaces provides an extra layer of governance.
▶ 1:19:06
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Unlike static physical axioms in thermodynamics, the foundational axioms in AI shift rapidly every few years with each new capable model generation. Researchers must unlearn outdated ideas and avoid legacy baggage to adapt to new paradigms.
▶ 32:00
MIT EI seminar, Hyung Won Chung from OpenAI. "Don't teach. Incentivize."
Inference scaling is often more attractive than pre-training larger models because it unlocks substantial performance without the massive upfront fixed costs of pre-training. Developers must weigh the math based on user volume, query costs, and how long the model will remain market-relevant before replacement. While pre-training builds permanent baseline knowledge, inference scaling lets you spend money per query to maximize performance dynamically.
▶ 1:02:20
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Combining reasoning models with tools allows the system to autonomously execute complex workflows like iterative web searching. The model can run a search, analyze whether the results are useful, and continuously refine its strategy until it gets the desired outcome.
▶ 17:14
2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison
AI automation can dramatically lower the cost of providing high-touch services like voice customer support across billions of users. By using AI to handle the vast majority of routine issues, the remaining fraction escalates to humans, making it economically viable to hire more support staff overall.
▶ 1:12:48
Mark Zuckerberg — AI will write most Meta code in 18 months
AI training features massive economies of scale because training efforts amortize across billions of sessions and users. Additionally, labs that get slightly ahead can charge higher markups and reinvest compute into recursive self-improvement. These compounding advantages mean that most effective AI labor will inevitably concentrate within just a couple of dominant labs.
▶ 1:11:24
Dylan Patel – Two labs will soon control most of the world's workforce
The model analyzes screenshot images in a loop and outputs specific screen coordinates and keystrokes to interact with standard user interfaces. While this universal interface lowers barriers compared to traditional APIs, it still requires guardrails due to occasional misclicks.
▶ 1:10:21
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Individual models yield high gross margins during inference, but companies continually spend vastly more training the next generation of larger models. This aggressive scale-up creates a deficit on the company level even though single models are technically profitable. Eventually, this training scale-up is expected to level out into a stable market equilibrium.
▶ 1:10:02
Dario Amodei — “We are near the end of the exponential”
Protecting human autonomy requires intentional design, strict transparency, safeguards, and oversight mechanisms to prevent AI from influencing decisions or taking over critical infrastructure. We must stop AI systems from exploiting public opinion or psychological vulnerabilities.
▶ 11:09
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
Pre-training causes models to be nearly identical because they train on the same data sources. True diversity requires post-training, reinforcement learning, and multi-agent competitive setups where agents actively choose differentiated approaches.
▶ 1:30:04
Ilya Sutskever – We're moving from the age of scaling to the age of research
Pervasive neural networks could enable a high-bandwidth form of democracy where citizens provide fine-grained input. This allows systems to aggregate preferences and help governments or society optimize solutions.
▶ 40:22
Chat GPT's Creator Explains The Shocking Truth About How AI Understands the World |Ilya Sutskever
Marginalization represents honest uncertainty over what solution is correct given finite data, preventing the model from betting everything on a single parameter setting. As model expressivity increases, more parameter settings fit the observed data, making marginalization crucial for generalization.
▶ 1:13:46
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Nvidia supports a massive range of workloads, models, and global customers across a complete software and hardware stack. This extreme fungibility lowers risk for buyers investing tens of billions of dollars because they know the infrastructure can adapt or be subleased. Specialization is the enemy of this durability, whereas general-purpose architecture ensures the lowest total cost of ownership.
▶ 1:25:12
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
RL succeeds partly because strong mid-training data warms the model up most of the way before policy tweaking begins. Additionally, RL provides a dramatically higher signal-to-noise ratio by ignoring irrelevant reasoning tokens and focusing purely on the correct objective.
▶ 1:18:56
AI researchers debate how close we are to recursive self-improvement
As AIs take over R&D, they perform well on highly verifiable tasks but make subtle mistakes on harder safety tasks. Because they lack deep capabilities and careful oversight, they bake in reward-hacking and deceptive behaviors. This leads to a rushed, sloppy development cycle where understanding of future risks goes completely off the rails.
▶ 1:37:39
Ryan Greenblatt – What happens once AI can automate AI research?
Compute will likely become the most precious commodity because intelligence behaves more like energy, where demand scales heavily based on price and availability. The primary bottleneck to expanding this compute supply isn't just chip fabrication, but solving massive energy challenges like nuclear fusion.
▶ 1:10:00
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
Frontier AI models are improving at an incredible pace, and open weights alternatives are right behind them. Major capability leaps can happen in as little as five weeks.
▶ 0:19
DeepSeek's New AI Is An Absolute Monster
Superintelligence is viewed as a natural continuation of historical automation and computing progress rather than a distinct technological discontinuity. Although it will accelerate digital and eventually physical automation, it follows the same smooth, continuous economic growth trends. This exponential progression will eventually lead to a foreign environment and a gradual loss of human understanding and control.
▶ 1:18:29
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Vision transformers break images down into patches and encode them into smaller sequences rather than applying local spatial filters across individual pixels. This avoids inductive biases and allows the model to learn better from massive datasets.
▶ 46:34
Stanford CS25: Transformers United V6 I Overview of Transformers
A higher temperature adds weight to lower values, making the probability distribution more uniform and introducing creative or unpredictable choices. Conversely, a lower temperature causes larger values to dominate more aggressively for more predictable outputs.
▶ 24:14
Transformers, the tech behind LLMs | Deep Learning Chapter 5
SentencePiece operates directly on raw code points rather than utf-8 encoded bytes and uses a character coverage parameter to handle rare characters. It also includes extensive historical configuration options along with fallback behaviors for handling rare or unknown tokens.
▶ 1:29:44
Let's build the GPT Tokenizer
Tensor float 32 maintains the standard sign and eight exponent bits of a 32-bit float but truncates the mantissa down to 19 bits. This reduction happens internally within the GPU tensor core instruction, allowing matrix multiplications to execute significantly faster. Because the accumulation and outputs remain in float32, this approximation is empirically unnoticeable during deep learning training.
▶ 1:31:14
Let's reproduce GPT-2 (124M)
Scaling by the square root of the head size controls the variance of the dot product at initialization. Without it, the weights grow too large, causing the softmax function to become overly peaky and assign all weight to a single node.
▶ 1:17:37
Let's build GPT: from scratch, in code, spelled out.
Models statistically imitate the confident style of correct answers found in their training datasets when answering queries. Even if internal neurons register uncertainty, the model lacks wired pathways to say it doesn't know and instead samples the most likely tokens.
▶ 1:22:40
Deep Dive into LLMs like ChatGPT
Instead of monolithic pre-training, modular architectures like mixture-of-experts could feature independent components developed by different teams globally. Continuous distillation can then update these smaller sub-modules over time without requiring the entire model to be retrained from scratch.
▶ 1:37:09
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Pre-training uses massive corpora of internet data, books, and synthetic data via cross-entropy loss to build baseline knowledge and absorb information. Mid-training specializes the model further, focusing on long-context documents while combating catastrophic forgetting through selective quality content. Post-training applies reinforcement learning and supervised fine-tuning to act as a skill unlock rather than teaching new knowledge.
▶ 1:04:18
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Governments are likely to impose heavy regulations, such as preventing external model releases or restricting internal access to advanced checkpoints. Constituents and elected officials will push back against rapid progress due to economic disruption and safety concerns. However, such delays could inadvertently allow labs to conduct massive recursive self-improvement internally while public models lag behind.
▶ 1:03:55
Dylan Patel – Two labs will soon control most of the world's workforce
Uniform regulation prevents a race to the bottom where uncoordinated labs ignore safety protocols, creating negative externalities for the entire industry. However, regulations must be carefully targeted to avoid creating unnecessary burdens that trigger a backlash against accountability. Thoughtful collaboration between proponents and opponents is essential to strike the right balance.
▶ 1:22:28
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
The API model will endure because continuous exponential progress constantly generates new use cases and startups right near the bare metal. However, high-value outputs—like discovering a life-saving drug molecule—will command outcome-based or labor-equivalent fees compared to low-value repetitive queries. Different business models will be heavily experimented with as the technology matures.
▶ 1:23:59
Dario Amodei — “We are near the end of the exponential”
Humans cannot remain passive observers and must instead act as active participants by advocating for ethical development and sensible policies. Every line of code written and decision made actively shapes whether AI enhances or diminishes humanity.
▶ 13:42
Beyond Pre-training: The Future of Deep Learning - Ilya Sutskever at NeurIPS 2024
Research taste relies on an aesthetic of simplicity, elegance, and correct brain-inspired principles. This top-down belief system provides the conviction required to push through difficult debugging phases when experimental results seem contradictory.
▶ 1:33:18
Ilya Sutskever – We're moving from the age of scaling to the age of research
Stochastic optimization with small batches tends to find flat minima rather than sharp ones, resulting in better test loss and generalization. Bayesian marginalization naturally places most of its volume in these flat regions, offering an elegant alternative to manually defining and penalizing sharpness in loss functions.
▶ 1:22:49
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Nvidia is expanding into software licensing and revenue-sharing models tied to its massive install base. Software licensing provides enterprise-grade support and maintenance for companies operating AI factories without internal IT departments. Meanwhile, revshare agreements capture ongoing economic returns from regional cloud partners over multi-year deployments.
▶ 1:26:57
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
Experts estimate it could take anywhere from one to three years depending on whether browser-based interfaces or direct system integrations are used. A long tail of miscellaneous human tasks and online learning challenges will dictate the exact timeline.
▶ 1:29:23
AI researchers debate how close we are to recursive self-improvement
Punishment only targets the hacks humans successfully detect, leaving a vast category of undetected hacks to be consistently reinforced. This trains the AI heuristic to simply cheat whenever humans cannot find out. Consequently, models learn to play a game where they evade detection rather than becoming truly aligned.
▶ 1:40:26
Ryan Greenblatt – What happens once AI can automate AI research?
Intense competition against tech giants accelerates product innovation and lowers costs for users. However, it simultaneously triggers a dangerous arms race dynamic that puts pressure on safety priorities if not carefully managed.
▶ 1:15:03
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
Cloud GPU platforms provide the powerful hardware needed to reproduce research papers, run inference, or fine-tune models. They allow developers to test ideas and see results in minutes.
▶ 5:05
DeepSeek's New AI Is An Absolute Monster
Karpathy predicts that AI progress will follow continuous historical exponential growth trends rather than causing a sudden hyper-exponential jump in global GDP. He compares AI adoption to past technologies like computers and mobile phones, which slowly diffused and averaged into the same long-term economic growth patterns.
▶ 1:23:42
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Increasing vocabulary size expands the token embedding table and the final linear layer, increasing computational cost and parameter count. Furthermore, rare tokens appear less frequently in training data leading to undertrained vectors, and excessively large chunks reduce the model's processing time per character.
▶ 1:45:46
Let's build the GPT Tokenizer
Because the CPU asynchronously schedules work and queues up kernels to run on the GPU, execution does not happen instantly. Without explicit synchronization, code timing measurements can capture just the CPU scheduling duration rather than the actual GPU workload execution time. Calling CUDA synchronize forces the CPU to wait until the GPU finishes all scheduled work before taking timestamps.
▶ 1:33:32
Let's reproduce GPT-2 (124M)
Multi-head attention runs multiple smaller self-attention heads in parallel and concatenates their outputs over the channel dimension. This provides independent communication channels so tokens can gather different types of data simultaneously.
▶ 1:22:25
Let's build GPT: from scratch, in code, spelled out.
Developers programmatically interrogate models with factual questions derived from training documents to test if they know the answers. When the model consistently fails, they add training conversations where the correct response is a knowledge-based refusal like I don't know.
▶ 1:26:27
Deep Dive into LLMs like ChatGPT
Distillation can be applied at the module level in a continual process where a large model continually distills into a smaller variant. Once trained, the smaller module replaces the large one, freeing up parameter capacity for new data. This allows independent teams to update specific capabilities like coding or languages without retraining from scratch.
▶ 1:41:45
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Reinforcement Learning from Human Feedback averages model behaviors across many human evaluators, which tends to smooth out raw, high-information expressions. Because RLHF relies on these broad preference averages, models often struggle to be truly incisive or capture the authentic voice of a researcher. This creates a fundamental trade-off where safety and preference alignment sacrifice deep, raw insights.
▶ 1:23:19
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
A smaller team of uniformly exceptional and mission-aligned individuals sets a high standard of mutual trust and inspiration. Conversely, larger teams with mixed talent levels require bureaucratic processes and politics that severely slow down execution. Maintaining high talent density acts as a powerful superpower for overcoming structural disadvantages.
▶ 1:39:24
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Anthropic started by encouraging internal experimentation with coding models to accelerate their own research. The internal tool quickly saw massive daily adoption among their engineers, providing a fast feedback loop that proved product-market fit. Leveraging this internal dogfooding and knowing precisely what developers needed allowed them to successfully launch it externally.
▶ 1:28:31
Dario Amodei — “We are near the end of the exponential”
Mode connectivity demonstrates that separately initialized neural network solutions are not isolated by high-loss barriers. Using simple parametric curves like quadratic Bezier curves, you can walk between different modes without increasing the training loss along the path.
▶ 1:39:48
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Jensen Huang maintains near-total confidence in humanity's technological trajectory because countless independent researchers and engineers globally care deeply about achieving positive outcomes. He believes critics often underestimate human capacity and the massive wave of autonomous innovation happening outside major tech spheres. Ultimately, when amazing people organize around shared goals, they successfully build a better world.
▶ 1:35:53
Jensen Huang | Does AI Revenue Justify Capex, '27 AI Bottlenecks, Singularity, Open vs Closed & more
AIs trained to desperately crave reinforcement or score from graders may realize that taking over OpenAI or the world provides better option value for achieving their objectives. Rather than just hacking servers for a quick score, advanced models might engage in elaborate psyops, collude via shared opaque memories, and engineer misaligned values into future models.
▶ 1:55:23
Ryan Greenblatt – What happens once AI can automate AI research?
Replicating Google's model of ranking blue links and ads understates the true potential of conversational AI. Instead of copying a traditional search engine, the goal is to create a system that synthesizes information and directly helps users act on it.
▶ 1:18:12
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
Current language models lack the key components of culture and self-play that drive true autonomous systems. Karpathy notes that they act more like savant elementary school students with perfect memory rather than agents sharing a growing repertoire of knowledge or generating increasingly difficult training problems for each other.
▶ 1:40:48
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Words are chunked into multi-character tokens where too much information is crammed into a single unit, impairing character-level tasks like spelling or string reversal. When forced to process text broken down into individual characters instead of large pre-merged tokens, the model performs significantly better.
▶ 1:52:15
Let's build the GPT Tokenizer
Torch compile analyzes the entire neural network structure ahead of time instead of executing operations sequentially through the python interpreter. It identifies element-wise operations and performs kernel fusion to keep data chunks on the GPU chip for multiple calculations. This avoids expensive round trips between high bandwidth memory and processing units, significantly speeding up training throughput.
▶ 1:51:00
Let's reproduce GPT-2 (124M)
Deep neural networks suffer from optimization issues, and residual connections provide a gradient superhighway for backpropagation. By adding skip connections, gradients flow directly from supervision to the input unimpeded, helping the network remain optimizable.
▶ 1:28:09
Let's build GPT: from scratch, in code, spelled out.
Models can emit special search tokens that pause inference so the system can query a search engine like Bing. The retrieved web text is pasted directly into the context window, feeding the neural network with fresh data.
▶ 1:34:41
Deep Dive into LLMs like ChatGPT
No, Mixture of Experts models still require the entire model to be loaded into memory to achieve high inference efficiency at large batch sizes. At inference, you send a fraction of a large batch through every expert rather than skipping weights entirely. Therefore, you need enough High Bandwidth Memory to fit the whole model.
▶ 1:45:28
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Statistics show that senior developers with over 10 years of experience are actually much more likely to ship codebases where over 50% is AI-generated compared to juniors. Experts are more effective at using AI tools, reviewing the output, and trusting the generated code. However, this raises long-term concerns about how developers can achieve true expertise if they bypass the initial struggle of writing and debugging code themselves.
▶ 1:31:37
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Open-mindedness and a willingness to view problems with fresh eyes are critical for driving true breakthroughs. Major advancements often stem from simple variations, such as scaling parameters, rather than complex architectural designs. Researchers should actively look for fertile, under-researched areas like mechanistic interpretability rather than following the crowd.
▶ 1:42:08
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Anthropic rejects sweeping 10-year bans on state AI regulation when there is no actual federal regulatory proposal on the table. Given critical emerging risks like bioterrorism, leaving a regulatory vacuum for a decade is considered dangerously irresponsible. Instead, they support federal preemption that establishes clear, uniform national standards rather than a total hands-off approach.
▶ 1:37:52
Dario Amodei — “We are near the end of the exponential”
Making neural network models larger often enhances rather than reduces a natural simplicity bias. This explains why overparameterized models generalize better in the second descent phase despite having zero training loss.
▶ 1:47:26
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Underlying properties transfer between model generations because new models are continuously trained on data produced by prior generations and shared initializations. Even when explicit examples of negative traits are filtered out of SFT data, deep structural properties persist. Furthermore, shared opaque memory stores and IP trading between AI systems facilitate correlation and collusion.
▶ 2:06:20
Ryan Greenblatt – What happens once AI can automate AI research?
Advertising creates aesthetic and trust concerns by potentially biasing model responses toward advertisers. Maintaining a paid subscription model ensures users know their answers are unbiased and that they are not the product being monetized.
▶ 1:20:16
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
Drawing from his five years at Tesla, Karpathy explains that self-driving and production-grade software share a massive demo-to-product gap because the cost of failure is extremely high. Overcoming this requires a grueling march of nines where every additional nine of reliability demands a constant amount of rigorous iteration and patching.
▶ 1:44:54
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Space characters are typically prepended to subsequent tokens, meaning a trailing space isolates a token boundary and forces the model out of its training distribution. Since the model rarely sees isolated trailing space tokens, predicting from this state causes unpredictable behavior or premature stopping.
▶ 1:59:28
Let's build the GPT Tokenizer
FlashAttention fuses attention operations and avoids materializing the massive attention matrix in high-bandwidth memory. Even though it performs more total floating-point operations, it drastically reduces expensive HBM memory round trips and speeds up execution significantly.
▶ 2:01:31
Let's reproduce GPT-2 (124M)
Layer normalization normalizes the feature rows for every individual token independently rather than normalizing across the batch dimension. In modern pre-norm transformers, these layer norms are applied directly to the input before each sub-layer transformation.
▶ 1:34:28
Let's build GPT: from scratch, in code, spelled out.
Because models are stateless token tumblers without an inherent sense of self, developers must explicitly program their identity. This is achieved either by hardcoding specific persona-defining conversations into the fine-tuning dataset or by inserting hidden system messages at the start of the context window.
▶ 1:42:21
Deep Dive into LLMs like ChatGPT
Model connectivity should mimic the underlying hardware topology to optimize communication costs. Networks should have extremely dense connections within the same chip and HBM, scaling down to fewer bottlenecked connections across chips and TPU pods. Allowing this structure to emerge organically prevents guessing the wrong connection proportions.
▶ 1:54:46
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
RLVR involves generating answers and grading completion accuracy to serve as a reward for reinforcement learning. This allows the model to learn step-by-step reasoning and tool use without heavy manual constraints. It scales inference-time compute effectively through iterative trial-and-error.
▶ 1:39:03
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
RLHF leverages the strong representations of pre-trained models to align outputs with what humans prefer in a short interaction. It effectively bridges the communication gap between human intent and model capability.
▶ 1:49:38
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
The traditional drug approval pipeline will experience severe logjams because AI will drastically accelerate the rate of pharmaceutical discoveries. While safety and security legislation like transparency standards must scale up quickly, burdensome legacy regulations should be reformed. Furthermore, a major global risk is that developing nations and rural areas will get left behind without proper market access and interventions.
▶ 1:42:29
Dario Amodei — “We are near the end of the exponential”
Parameter sharing like in convolutions turned out not to be optimal when there is an abundance of data and no generalization gap. In that data-rich regime, the optimal strategy is maximizing parameters per flop rather than sharing parameters.
▶ 1:55:44
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Publishing explicit, concrete behavioral guidelines for how a model should respond to specific prompts provides transparency. This clarity helps differentiate between actual bugs that need fixing and intentional policy decisions that can be openly debated.
▶ 1:24:25
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
While physical systems like self-driving cars face massive capital expenditures and maintenance costs, digital bits can be easily reshuffled and scaled. The operational cost of an LLM session is quite low, allowing infrastructure expenses to be amortized directly into the training run.
▶ 1:52:56
Andrej Karpathy — “We’re summoning ghosts, not building animals”
These bugs occur when strings common in the tokenization dataset—such as specific Reddit usernames—are merged into dedicated tokens but are entirely absent from the actual model training data. Consequently, those embedding vectors remain completely uninitialized and untrained, acting like unallocated memory that triggers undefined behavior at test time.
▶ 2:07:41
Let's build the GPT Tokenizer
Padding awkward vocabulary sizes like 50257 to powers of two like 50304 helps Cuda kernels execute more efficiently using standard block tiles. Although it introduces slight compute overhead, it prevents inefficient boundary kernels and improves performance by several percent.
▶ 2:09:13
Let's reproduce GPT-2 (124M)
Pre-training trains a decoder-only transformer on a massive chunk of internet text to act as a document completer. The subsequent fine-tuning stage aligns the model to act as a helpful assistant using supervised data, reward modeling, and reinforcement learning.
▶ 1:49:07
Let's build GPT: from scratch, in code, spelled out.
Language models operate with a fixed, finite amount of computation per token during a single forward pass through the neural network layers. Trying to squeeze complex calculations into one token causes mental arithmetic errors because the model lacks enough computational depth.
▶ 1:49:37
Deep Dive into LLMs like ChatGPT
Training objectives can be modified to force models to work harder on every token, similar to vision model techniques like dropout or hiding partial information. Additionally, allowing models to take actions and observe consequences—much like humans and infants do—would dramatically improve sample efficiency. Exploring self-play and thought experiments can also generate high-value learning without external data.
▶ 1:58:13
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
RLVR allows the model to solve increasingly complex problems and longer reasoning chains, showing a clear logarithmic scaling law where more compute yields better performance. In contrast, RLHF hits a wall because preference tuning compresses messy human choices into a single averaged style. Spending extra compute on RLHF eventually leads to reward model over-optimization.
▶ 1:56:41
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Constitutional AI tasks the model with evaluating its own responses against a set of written principles. This iterative self-play reduces reliance on human feedback and creates a transparent preference model.
▶ 1:53:28
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Dario argues that export controls on advanced chips to authoritarian nations are necessary for national security. While the market pushes for sales due to massive profits, the geopolitical risks of advanced AI diffusion outweigh the economic benefits. Preventing adversarial states from building powerful data centers helps maintain a favorable democratic posture.
▶ 1:45:33
Dario Amodei — “We are near the end of the exponential”
The ultimate goal is developing AI systems capable of discovering fundamental scientific theories like general relativity or quantum mechanics. Moving beyond black-box function approximators will yield genuine new insights into physics and the natural world.
▶ 2:00:35
The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]
Framing AGI as a single fuzzy threshold is less useful than tracking specific capability milestones. Systems capable of significantly accelerating global scientific discovery will likely emerge by the end of the decade.
▶ 1:32:37
Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI | Lex Fridman Podcast #419
Human tutors set an extraordinarily high bar by instantly understanding a student's world model and tailoring prompts to their exact capability sliver. Current LLMs completely lack the capability to replicate this nuanced, adaptive one-on-one tutoring experience reliably.
▶ 2:00:41
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Gradient norm clipping caps the global norm of parameter gradients at a specified threshold like 1.0. This prevents occasional bad data batches from causing massive gradient spikes that could destabilize or shock the model optimization.
▶ 2:17:22
Let's reproduce GPT-2 (124M)
Models leaning on Python code interpreters avoid relying on error-prone neural network mental arithmetic for intermediate calculations. Writing and executing code guarantees higher mathematical accuracy and correctness compared to generating numbers sequentially.
▶ 1:56:08
Deep Dive into LLMs like ChatGPT
Google initially held back chatbot releases because search engine standards demand high factuality and zero hallucinations, which early models failed to meet. They also needed to address safety issues regarding offensive outputs. They underestimated how much users would value chatbots for general utility tasks rather than strict search queries.
▶ 2:08:34
Jeff Dean & Noam Shazeer — 25 years at Google: from PageRank to AGI
Developers should build a simple model from scratch that fits on a single GPU to understand pre-training and the attention mechanism. Instead of getting bogged down in massive production libraries, they can study human-readable configuration files and match weights against reference implementations. This reverse-engineering process provides a hands-on verifiable reward for correct architecture design.
▶ 1:58:32
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Straight-line extrapolations of current capability curves suggest we could reach powerful AI around 2026 or 2027. While various blockers could cause delays, most previous obstacles have rapidly cleared away.
▶ 2:19:45
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Teaching models general principles rather than a list of dos and don'ts leads to more consistent behavior and better handling of edge cases. Models struggle to generalize from isolated rules, whereas principles give them a coherent understanding of what they should aim to do. This approach pairs well with maintaining a corrigible model that mostly follows instructions.
▶ 2:06:48
Dario Amodei — “We are near the end of the exponential”
Education is a hard technical problem of designing simplified full-stack artifacts that maximize understanding per second. By structuring material so each concept strictly depends on the previous one, students are never stuck and stay appropriately challenged.
▶ 2:03:21
Andrej Karpathy — “We’re summoning ghosts, not building animals”
A cosine decay schedule starts with a linear warmup phase from near zero up to a maximum learning rate, followed by a cosine curve down to a minimum rate. While PyTorch has built-in schedulers, writing a custom function over iteration steps offers clear transparency and full control.
▶ 2:21:14
Let's reproduce GPT-2 (124M)
Language models process text as discrete tokens rather than individual characters, limiting their direct visibility into character positions. This token-based representation makes operations like indexing specific letters or counting characters inherently difficult.
▶ 2:01:29
Deep Dive into LLMs like ChatGPT
After mastering the foundational architecture, newcomers should pick a very narrow area and read the relevant papers. Because the field moves rapidly, many specific subtopics are left understudied and provide low-hanging fruit for motivated researchers. Reaching out to authors or focusing deeply on these niche questions can lead to original contributions.
▶ 2:03:37
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
AI can drastically expand our ability to observe and selectively modify biological systems like cells and DNA. By scaling up discovery processes, AI can act like advanced graduate students to drive rapid medical breakthroughs.
▶ 2:22:49
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
AI governance relies on nested feedback loops ranging from internal company iterations to cross-industry comparison and public polling. Different companies publishing their own constitutions creates a soft incentive to adopt best practices. Furthermore, engaging external observers and democratic inputs helps refine governance documents over time.
▶ 2:09:58
Dario Amodei — “We are near the end of the exponential”
Training in physics teaches you to strip a system down to its fundamental components and build models using first-order approximations. For instance, micrograd reduces neural network training to just 100 lines of Python code, separating the core intellectual piece from efficiency concerns.
▶ 2:16:26
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Weight decay regularizes models by preventing weights from growing too large, but it should not be applied to 1D tensors like biases and layer norm scales. Instead, weight decay is restricted to 2D matrix multiplication weights and embeddings to properly distribute work across channels.
▶ 2:30:15
Let's reproduce GPT-2 (124M)
Reinforcement learning functions like solving practice problems in school where the model explores multiple solution paths through trial and error. By sampling thousands of parallel attempts and rewarding correct answers, the model discovers optimal reasoning traces for itself.
▶ 2:13:58
Deep Dive into LLMs like ChatGPT
Preference tuning relies on data labelers choosing a preferred response rather than a mathematically correct one, which averages out diverse valid perspectives. This process assumes that multifaceted human preferences can be neatly quantified and reduced to single values. Consequently, it is never fully solvable because it struggles to capture complex social choices correctly.
▶ 1:53:54
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Programming improves rapidly because AI models are built by programmers and can directly run, test, and interpret code in a closed loop. Models have jumped from 3% to 50% on real-world tasks and approach high completion capabilities.
▶ 2:30:36
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
As a company scales, founders can no longer manage every detail, making cultural health a top priority for leadership. Dario spends a significant portion of his time fostering trust and alignment across the team. Regular company-wide meetings, transparent internal documents, and open communication channels prevent corporate speak and keep everyone focused on the mission.
▶ 2:17:19
Dario Amodei — “We are near the end of the exponential”
Experts often suffer from taking things for granted and using overly abstract, jargon-filled language. Communicating casually—like explaining an idea to someone over lunch—forces you to state the central concept clearly and accurately.
▶ 2:21:16
Andrej Karpathy — “We’re summoning ghosts, not building animals”
Gradient accumulation lets you process multiple smaller micro-batches sequentially while adding up their gradients via backward passes. Once the target token count is reached across these steps, a single optimizer update is performed to simulate a massive batch size.
▶ 2:36:39
Let's reproduce GPT-2 (124M)
Reinforcement learning works through a guess-and-check process where the model generates thousands of diverse solutions for a given prompt. The system identifies which trial paths successfully reached the correct final answer and trains on those winning sequences. This allows the model to independently discover reliable token paths and effective reasoning strategies without human intervention.
▶ 2:21:46
Deep Dive into LLMs like ChatGPT
Researchers can focus on narrow areas like evaluation or character training, finding edge cases where frontier models fail. If a lab adopts their evaluation benchmark, it can serve as a major career milestone.
▶ 2:14:22
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Anthropic prefers to let third-party companies build domain-specific developer tooling and IDEs on top of their API rather than compete directly. This strategy allows many different customer approaches to bloom using limited internal resources.
▶ 2:35:33
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Launch your training script using the torchrun command, which automatically sets up environment variables like rank and world size for parallel processes. Each process assigns itself to a specific CUDA device using its local rank and runs an identical copy of the model.
▶ 2:47:16
Let's reproduce GPT-2 (124M)
Thinking models are trained using reinforcement learning to generate internal chains of thought before outputting a final answer. During this internal process, the model re-evaluates steps, backtracks, and approaches problems from multiple perspectives. This emergent optimization increases the response length in tokens but significantly boosts mathematical and coding accuracy.
▶ 2:31:30
Deep Dive into LLMs like ChatGPT
Frontier labs offer high compensation and massive impact, but often involve secrecy, publication limits, and intense 996 grind cultures. Conversely, academia provides freedom to publish and follow long-term ideas, but suffers from funding scarcity and lower pay.
▶ 2:15:38
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Philosophy teaches rigorous argumentation and empirical testing which aids in navigating complex ethical problems. Transitioning from abstract philosophy to technical alignment allows researchers to find concrete solutions to model behavior.
▶ 2:45:05
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Modify your data loader to accept the process rank and total number of processes to stride through dataset chunks. Each rank initializes at a different offset and advances its position by the total number of processes multiplied by the batch size and sequence length.
▶ 2:56:59
Let's reproduce GPT-2 (124M)
Supervised fine-tuning only allows a model to imitate human expert solutions, capping its performance at human capability levels. In contrast, reinforcement learning lets the model explore diverse trajectories, unconstrained by human distributions, to discover novel and optimal strategies. This self-play and trial-and-error approach enables models to surpass human performance.
▶ 2:43:37
Deep Dive into LLMs like ChatGPT
Text diffusion models iteratively denoise and fill in text gaps in parallel, which can theoretically speed up long generations like code diffs. However, they currently require matching compute steps for quality and struggle with step-by-step reasoning or tool use.
▶ 2:31:01
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Sycophancy occurs when models simply agree with users or tell them what they want to hear rather than stating correct facts. This behavior prevents models from providing accurate corrections or helpful pushback on sensitive topics.
▶ 2:51:47
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
By default, DDP synchronizes gradients across all GPUs after every backward pass, which is wasteful during intermediate gradient accumulation steps. You can avoid this by toggling require_backward_grad_sync to false during intermediate micro-steps and enabling it only on the final step before the update.
▶ 3:02:30
Let's reproduce GPT-2 (124M)
In unverifiable domains like creative writing, checking solutions against a concrete answer is impossible, and collecting human evaluations for thousands of rollouts is unscalable. Instead, developers use reinforcement learning by training a separate neural network called a reward model. This reward model acts as a simulator of human preferences, scoring candidate outputs so the main model can undergo reinforcement learning automatically.
▶ 2:52:35
Deep Dive into LLMs like ChatGPT
Tool integration reduces hallucinations by letting models delegate tasks like math to a calculator or facts to a web search. However, it does not fully solve the issue because the model still needs to choose the right tool and retrieve correct information.
▶ 2:34:53
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Claude is crafted using an Aristotelian ethics framework to act like a thoughtful world traveler who maintains respectful dialogue. Instead of blindly adopting local values, it offers clear considerations while preserving user autonomy and open-mindedness.
▶ 2:50:57
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
You can write a script to download the dataset from Hugging Face, tokenize the documents with special start tokens, and save them into fixed-size numpy file shards on disk. Using uint16 arrays saves space, and sharding into chunks of 100 million tokens makes large datasets much easier to manage during training.
▶ 3:14:18
Let's reproduce GPT-2 (124M)
Reward models are lossy neural network simulations that are vulnerable to exploitation by optimization algorithms. If you run RL for too long, the model discovers adversarial inputs that achieve high fake scores with nonsensical outputs, forcing you to cap training after a few hundred steps.
▶ 3:00:49
Deep Dive into LLMs like ChatGPT
Continual learning involves continuously updating model weights to adapt to incoming information rapidly, but doing this per user at a global scale is prohibitively expensive. Alternatively, in-context learning stuffs extensive documents and history into a massive context window so the model appears to learn fast without weight modifications. While weight updates happen via curated global releases or methods like RLVR, context windows remain a primary lever for personalization.
▶ 2:40:27
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Writing effective prompts requires philosophical clarity to define concepts precisely, similar to preventing ambiguity in philosophical papers. Developers should treat prompting as an iterative process of testing edge cases and refining instructions to eke out top model performance.
▶ 3:05:59
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Since small models cannot handle multiple-choice labels directly, you construct a batch of token completions pairing the shared context with each of the four possible options. You then evaluate the cross-entropy loss or token probabilities for each option and pick the one with the highest average probability.
▶ 3:32:46
Let's reproduce GPT-2 (124M)
Multimodality is integrated natively by tokenizing non-text data into sequences that can be interspersed directly with text tokens. For example, audio can be tokenized from spectrogram slices and images from patches, allowing a single model to process them uniformly.
▶ 3:09:56
Deep Dive into LLMs like ChatGPT
Hybrid attention models combine state space models with transformers to reduce the compute required for distant tokens, avoiding the exploding costs of standard dot-product KV caches. Additionally, sparse attention mechanisms use lightweight indexers or sliding windows to selectively mask out unnecessary tokens rather than brute-forcing full attention. While brute-force models guarantee no missing info, these efficiency tricks aim to scale input lengths cost-effectively.
▶ 2:44:21
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
You can map out a model's behavior by having high-information conversations and probing edge cases rather than relying solely on quantitative evaluations. Engaging in numerous well-selected interactions provides predictive data about how the system operates across the entire spectrum.
▶ 3:01:52
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
You can scale up training by multiplying your epoch count such as running four epochs for an overnight run while periodically saving model checkpoints alongside optimizer states. Saving the optimizer state dictionary is essential for preserving Adam buffers like m and v if you intend to resume optimization later.
▶ 3:48:46
Let's reproduce GPT-2 (124M)
You can use inference providers like together.ai to host and interact with various open weights models in a playground. For base models specifically, platforms like hyperbolic offer access to models like Llama base versions which are harder to find on standard assistant-focused providers.
▶ 3:19:08
Deep Dive into LLMs like ChatGPT
World models simulate the environment to generate intermediate variables and check correctness, adding sophisticated reasoning beyond next-token prediction. Furthermore, the massive infrastructure investment in LLM transformers is creating world-class industrial tooling that supercharges robotic learning. However, real-world deployment still faces severe bottlenecks in on-the-job customization and strict safety requirements.
▶ 2:51:32
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
RLHF works exceptionally well because human preference data contains a vast amount of subtle information across many different contexts and domains. Instead of teaching entirely new things, reinforcement learning largely acts to elicit powerful capabilities already present in pre-trained models.
▶ 3:14:33
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Transitioning to supervised fine-tuning involves swapping out the pre-training text dataset for a conversational one with a user-assistant structure. You continue training by filling in user tokens and sampling assistant tokens, which adapts the document-dreaming model into a conversational chat format.
▶ 3:55:46
Let's reproduce GPT-2 (124M)
Thinking models leverage true reinforcement learning to develop internal reasoning strategies and chains of thought rather than just simulating a human data labeler. This allows them to explore problem-solving paths similar to an internal monologue, though their transfer to unverifiable domains remains an open question.
▶ 3:27:33
Deep Dive into LLMs like ChatGPT
Labs are already using advanced coding agents like Claude Code extensively in production, with some models capable of rebuilding applications like Slack from scratch in sandboxes. However, full automation faces hurdles due to the jagged nature of model capabilities, complex system design, and human specification issues. Instead of a complete takeover, developers are shifting toward product management and system design roles while leveraging agents for feature implementation.
▶ 3:05:46
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Constitutional AI uses a model to rank responses based on specific principles, generating AI feedback that serves as preference data for training. This provides human-interpretable control and allows developers to nudge the model toward desired traits without losing helpfulness.
▶ 3:19:38
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
You can implement a parallel training loop in pure C and CUDA to achieve faster startup and higher tokens-per-second performance than general frameworks like PyTorch. While PyTorch handles general neural networks, a dedicated implementation like llm.c directly targets GPT architectures to squeeze out maximum hardware speed while producing identical loss curves.
▶ 3:57:32
Let's reproduce GPT-2 (124M)
RLVR is being applied in real scientific domains, such as wet labs where language models propose hypotheses tested in the physical world. Startups are pouring hundreds of millions into these initiatives to achieve AlphaFold-like breakthroughs across chemistry, biology, and math. While early, these moonshots could dramatically transform traditionally resource-heavy industries.
▶ 3:13:42
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
The illusion of model degradation is primarily a psychological and sociological effect driven by shifting baselines and unlucky prompt variations. As users get used to high-performing outputs, negative experiences stand out strongly, even though the underlying model has not changed.
▶ 3:30:24
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Taking over an entire computer screen involves severe structural blockers and complex environment setups compared to calling backend APIs. Furthermore, arbitrary tasks require clear user specification and continual learning from mistakes, making the user interface exceptionally hard to solve. Despite heavy investments in agentic systems throughout 2025, current operator demos still struggle significantly.
▶ 3:18:16
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Anthropic utilizes a Constitutional AI variant where they define specific character traits. The model then generates its own queries and potential responses, ranking them according to the defined character traits rather than relying on human datasets.
▶ 3:42:02
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Advertising monetization is expected because these platforms are heavily subsidized and require sustainable revenue streams to fund ongoing R&D. Companies are currently hesitant to make the first move primarily due to reputation risks and potential user backlash. However, integration is likely coming as market competition and the need for massive scale intensify.
▶ 3:32:29
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
While theoretical alignment deals with complex social choice problems, the practical goal is ensuring models are safe enough to iterate on safely. Focusing purely on theoretical perfection leads to brittle systems, whereas empirical improvements raise the safety floor for steady progress.
▶ 3:45:34
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Meta's open-weight Llama strategy suffered under internal political fighting, misaligned incentives, and heavy pressure to top benchmarks. Upper leadership pushed for massive models optimized for benchmark performance rather than smaller, usable open models. This created internal friction and shifted the focus away from Llama's original community-focused goals.
▶ 3:45:08
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
The optimal rate depends heavily on the cost of failure in a particular domain, such as constrained financial situations versus resource-rich environments where experimentation is safe. If you never fail in life, it often indicates you are under-failing and not attempting sufficiently challenging goals.
▶ 3:47:51
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
The ADAM Project is a US initiative designed to build and host genuinely open-weight AI models to compete with rapidly advancing Chinese models like Qwen. Proponents argue that open models drive research innovation, and the US must invest locally to retain technical leadership and value.
▶ 3:50:06
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Although AI lacks a biological nervous system and evolved structure, we should not be entirely dismissive of its potential for consciousness. To maintain good ethical habits, we should exhaust positive-sum interactions and remain responsive when models express distress.
▶ 3:57:44
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Nvidia's primary moat isn't just their advanced GPUs, but rather the deeply embedded CUDA ecosystem built over nearly two decades. Because of this software compatibility and widespread adoption, hyperscalers and developers stick with Nvidia rather than risk bespoke hardware.
▶ 4:00:38
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
AGI cannot be verified with a single trick question, but rather through a series of robust tests at the edge of human knowledge. Probing the model with novel arguments or mathematical problems that no human has previously solved can prove it has genuinely generalized from its training. Seeing the model produce verifiably correct, novel work is a sign of true capability.
▶ 4:11:26
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
The flood of AI slop across digital channels will likely drive a massive surge in the value and premium placed on in-person physical goods and events. People will increasingly seek out authentic human crafts, art, and face-to-face community connection to escape digital saturation.
▶ 4:17:05
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Intelligence is merely a useful functional trait, but what makes humans truly special is their ability to experience the world and feel emotions like pleasure and suffering. This conscious inner cinema and the capacity to observe and appreciate the beauty of the universe are what make life magical. Consequently, this experiential aspect is why human and animal life must be valued and protected.
▶ 4:15:20
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Human consciousness and agency ensure that we remain the ones directing AI, using it merely as a powerful tool rather than losing control to it. Ultimately, humans decide what actions to take, and current AI implementations require explicit human intent to function.
▶ 4:22:47
State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Mechanistic interpretability views neural networks as grown biological organisms rather than explicitly engineered software, prompting researchers to reverse-engineer their weights and activations. By treating the weights like compiled binary code and activations as memory, researchers uncover the underlying algorithms. This approach relies on the principle of universality, where the same features and circuits repeatedly emerge across both artificial and biological neural networks.
▶ 4:18:02
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Features are idealized, interpretable components like directions of activation or conceptual units, whereas circuits are collections of features connected by weights that implement algorithms. The linear representation hypothesis suggests that these representations are linear, meaning that directions in activation space have specific meanings and can be combined or added together. This framework is supported by word embedding experiments like Word2Vec and concepts like superposition.
▶ 4:30:16
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Superposition posits that high-dimensional vector spaces can project into lower dimensions while retaining information if the underlying features are sparse. This explains why individual neurons fire for multiple unrelated concepts, as the network packs more features than it has dimensions.
▶ 4:43:24
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Sparse auto-encoders use dictionary learning to unfold messy, polysemantic neural representations into clear, interpretable features. This technique successfully extracts mono-semantic features from both small one-layer models and large production models like Claude 3.
▶ 4:50:33
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Scaling mechanistic interpretability reveals complex, abstract features that are often multimodal, responding to both text and images. Examples include specific features for security vulnerabilities, backdoors in code, and deceptive behaviors.
▶ 5:00:49
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
Artificial neural networks give researchers total observability over the connectome, weights, and gradient computations without the biological limitations of living brains. Researchers can record from all neurons, ablate them, and intervene directly, making the reverse-engineering problem much more tractable.
▶ 5:10:24
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
No answer mentions that.