State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490
Lex Fridman · 2026-01-31
💡 Quick Take
1. The AI landscape is intensely competitive, with both research and product development accelerating rapidly.
2. China is a significant player in AI, with companies like DeepSeek, Zhipu AI, and MiniMax releasing strong open-weight models, challenging US dominance.
3. Open-weight models, especially from China, are gaining traction due to their accessibility and potential to disrupt US business models reliant on API subscriptions.
4. Budget and hardware constraints are becoming the primary differentiators, rather than proprietary technological ideas.
5. The hype around models like Anthropic's Claude 3.5 Opus is significant, but differentiation between top models is decreasing.
6. Chinese companies are releasing open-weight models for a few years, driven by a desire for international influence and the recognition that US companies may not pay for API subscriptions due to security concerns.
7. Consolidation is expected in the open model space due to the high cost of development, but it's not anticipated to be a major story in 2026.
8. The "winner" of a year is often the most recent model, as ideas are rapidly iterated upon and incorporated by competitors.
9. Chinese companies have different incentives, with some like MiniMax and Moonshot AI filing for IPOs, aiming for Western mindshare.
10. The distinction between hype in the X (Twitter) echo chamber and actual model usage is important; ChatGPT and Gemini target a broad user base.
11. Users often stick with a model until it breaks, similar to choosing browsers or text editors.
12. The future likely involves using multiple LLMs for different use cases, rather than a single dominant model.
13. Google's scale and ability to integrate research and product give Gemini an advantage, while OpenAI is known for landing impactful products despite operational chaos.
14. Google's TPUs and data center infrastructure provide a significant advantage in terms of cost and development.
15. OpenAI's research division consistently delivers new research ideas and products, a key organizational strength.
16. There's a trade-off between model intelligence and speed, and users often desire the option to choose between them.
17. Users often develop "muscle memory" with specific models like ChatGPT due to long-term familiarity and brand recognition.
18. Customization of LLMs, like ChatGPT's memory feature, may lead to users having multiple subscriptions for personal and work use.
19. The "thinking" versus "non-thinking" model dichotomy reflects a trade-off between intelligence and speed, with different use cases for each.
20. Grok-3 Heavy is noted as being particularly good for hardcore debugging.
21. Gemini is praised for its interface and ability to handle long context with specific information retrieval.
22. Claude 3.5 Opus is favored for coding and philosophical discussions, especially with "extended thinking" enabled.
23. Grok is useful for real-time information and finding specific AI-related content on platforms like X.
24. User loyalty to LLMs is often won over by a single impressive feature or performance on a specific query, until a significant failure occurs.
25. The long context window in models like GPT-4o is a significant development, with ongoing improvements in efficiency.
26. Chinese open-weight models are not widely used by US-based individuals in the conversation, likely due to a US-centric bias and the current superior output quality of US models.
27. American companies are currently better in terms of output quality, leading users to pay for them despite the availability of cheaper Chinese models.
28. US companies are hosting Chinese models and selling tokens, a business model that leverages the cost-effectiveness of these models.
29. Open-weight models from China are popular due to unrestricted licenses, unlike some US models with attached restrictions.
30. The "DeepSeek moment" in early 2025, with its near state-of-the-art performance at lower compute cost, was a pivotal event.
31. The Transformer architecture, particularly the decoder-only variant used in GPT models, remains the foundational architecture.
32. Mixture of Experts (MoE) is a key architectural tweak that allows models to be larger without increasing compute per forward pass by selectively using different "experts" (feedforward networks).
33. Dense models utilize all their parameters in every pass, while sparse MoE models only activate a subset, making them more efficient.
34. Architectural tweaks like Grouped-Query Attention, Sliding Window Attention, and Multi-head Latent Attention are used to differentiate models and improve efficiency, particularly for long contexts.
35. The fundamental Transformer architecture has not changed drastically; advancements are often in tweaks and optimizations rather than entirely new paradigms.
36. The turbulence and advancement in AI are happening in the stages of training (pre-training, mid-training, post-training) and system optimizations, not just core architecture.
37. Supervised fine-tuning and Reinforcement Learning from Human Feedback (RLHF) were key algorithmic advancements that unlocked capabilities beyond GPT-2's architecture.
38. System optimizations like FP8 and FP4 training allow for faster experimentation and training by reducing memory usage and communication overhead.
39. Alternatives to the autoregressive Transformer, like text diffusion models and Mamba models, are being explored but haven't yet replaced the Transformer as state-of-the-art.
40. Scaling laws, the predictable relationship between compute/data and prediction accuracy, continue to hold, but low-hanging fruit is being picked.
41. Inference time scaling, demonstrated by GPT-4o, allows smaller models to achieve higher performance by using more compute during inference.
42. Reinforcement Learning with Verifiable Rewards (RLVR) is a major breakthrough, enabling models to learn complex behaviors and tool use through iterative generate-grade loops.
43. Pre-training is extremely expensive, with serving costs often outweighing training costs for large models.
44. The development of massive compute clusters (e.g., Blackwell) is driving further advancements in training capabilities.
45. While some believe pre-training is plateauing, the continued scaling of compute suggests models will continue to get smarter.
46. The trend is towards larger models and potentially higher subscription costs for cutting-edge AI services.
47. The sparse nature of MoE models makes them more efficient for generation, a key aspect of post-training.
48. Pre-training remains crucial for building the best base models, even if other scaling methods offer more immediate gains.
49. Reinforcement learning compute is becoming more comparable to pre-training in terms of time allocation, though it uses different hardware and scaling approaches.
50. Pre-training involves training on vast datasets using cross-entropy loss for next-token prediction, with an increasing focus on data quality and synthetic data.
51. Mid-training is a specialized phase, often focusing on specific data types like long-context documents, to avoid catastrophic forgetting.
52. Post-training encompasses fine-tuning, DPO, RLHF, and RLVR, focusing on refining model behavior and unlocking skills.
53. Synthetic data, including OCR-processed documents and high-quality LLM outputs, plays a crucial role in improving training efficiency.
54. Data quality, rather than just quantity, is a key driver of model performance, especially for larger models that can absorb more information.
55. The creation of high-quality pre-training datasets involves rigorous filtering and sampling from diverse sources, adapting to evolving evaluation needs (e.g., math, code).
56. PDFs, especially from sources like arXiv and Semantic Scholar, are valuable data sources for training.
57. Data privacy and licensing are becoming critical concerns, with a growing interest in training on explicitly licensed data.
58. Anthropic's legal issues with authors highlight the complex legal and ethical landscape of training data.
59. LLM-generated data is inevitable, but human curation and verification are crucial for maintaining quality and trust.
60. The "voice" of LLMs, influenced by RLHF, can be a limitation, making it hard for them to be incisive or express unique insights.
61. The "echo chamber" effect of Silicon Valley can lead to a detachment from broader human experiences and needs.
62. Agency in using AI to build things is key to understanding its strengths and weaknesses, and to steering its development.
63. There's a risk of losing enjoyment in core tasks if LLMs are used to automate them entirely; finding a "Goldilocks zone" for AI assistance is important.
64. Senior developers are more likely to use AI-generated code in production, suggesting AI is more effective when used by experienced individuals.
65. RLVR is highly effective for math and code, and its scalability is a major advantage over RLHF.
66. RLVR's core idea is to maximize accuracy on verifiable tasks through iterative learning, contrasting with RLHF's reliance on learned human preferences.
67. Process reward models and value functions are emerging areas in RLVR, aiming to score intermediate steps in reasoning.
68. The compute required for RLVR is significant and growing, with a focus on memory-bound operations compared to pre-training's compute-bound nature.
69. RLHF is still valuable for refining model style, tone, and personality, acting as a finishing touch.
70. RLVR has shown dramatic accuracy improvements in short training times, but concerns about data contamination in benchmarks exist.
71. Mid-training data curation, focusing on "reasoning traces" (step-by-step problem-solving), is crucial for enabling post-training RLVR.
72. The difficulty of finding verifiable problems that models haven't mastered is driving research into more complex domains like scientific problems.
73. Building an LLM from scratch, starting with a simple model like GPT-2, is the best way to learn fundamentals and understand how things work.
74. Hugging Face Transformers is a valuable resource but can be too complex for beginners; starting with simpler, from-scratch implementations is recommended.
75. Reverse-engineering models by examining config files and weights, then implementing them from scratch, is a powerful learning technique.
76. The field moves so fast that focusing on narrow, fundamental areas after gaining foundational knowledge is key for impactful research.
77. Character training, achieved through fine-tuning smaller models with techniques like LoRA, is an area with less research but significant potential.
78. Research contributions can be made through evaluation, identifying model failure modes and creating representative test problems, even with limited compute.
79. The "9/9/6" work culture (9 AM to 9 PM, 6 days a week) is a trend in some Silicon Valley AI companies, driven by intense competition and passion.
80. Silicon Valley operates as an echo chamber, which can be productive for innovation but also lead to a detachment from broader societal needs.
81. Text diffusion models offer a parallel processing approach that could be more efficient for certain tasks, but quality and tool integration remain challenges.
82. Tool use is a critical unlock for LLMs, reducing hallucinations and enabling more complex tasks, but it requires careful implementation and trust.
83. Continual learning, the ability for models to adapt rapidly from new information, is a key challenge for achieving true general intelligence.
84. In-context learning, by providing extensive context to LLMs, can mimic learning without updating model weights, but has limitations.
85. Personalized memory in LLMs is currently achieved through context stuffing or preference prompts, with LoRA adapters offering limited weight updates.
86. Long context windows are expanding, driven by architectural tweaks and increased compute, but significant breakthroughs are still needed for massive context lengths.
87. Recursive language models, breaking down tasks into sub-tasks, offer a new paradigm for tackling long-context problems and improving accuracy.
88. World models, simulating the world to provide LLMs with data beyond their training set, hold promise for unlocking new capabilities and sophisticated reasoning.
89. Robotics is being supercharged by LLM advancements, with improved simulators and the use of LLMs as central units for exploration.
90. Safety is a paramount concern in robotics, especially for embodied systems operating in real-world environments, where failure is not an option.
91. The timeline to AGI/ASI is highly debated, with definitions varying and predictions ranging from a few years to decades.
92. The "jagged" nature of AI capabilities means models will excel at some tasks while remaining poor at others, requiring human collaboration.
93. The full automation of software writing is a long-term goal, with humans likely remaining in the loop for system design and outcome specification.
94. The economic impact of LLMs is still unfolding, with a potential for significant GDP growth through increased productivity and accessibility of knowledge.
95. The commoditization of AI and the specialization of foundation models for specific industries (finance, law, pharma) will drive significant economic value.
96. The competition between open and closed models is driving innovation, with open models potentially becoming cheaper and more optimized for serving.
97. Consolidation in the AI startup space is increasing, with multi-billion dollar acquisitions and licensing deals becoming more common.
98. The future of AI companies may involve a mix of API services, product offerings, and hardware development, with competition intensifying.
99. Meta's Llama brand has been influential in the open-weight space, though its future direction and the balance between open and closed approaches are debated.
100. The US government's "AI Action Plan" and initiatives like the "Adam Project" aim to bolster the US open-source AI ecosystem to compete with China.
101. Open-source models are crucial for educating the next generation of AI researchers and fostering innovation.
102. NVIDIA's dominance is built on the CUDA ecosystem, a significant moat that is hard to replicate, though LLMs may accelerate the development of alternatives.
103. Singular figures like Jensen Huang and Steve Jobs have a profound impact on steering technological direction through focus and vision.
104. The "bitter lesson" suggests that compute and scaling laws will continue to be fundamental drivers of AI progress.
105. The internet and compute are merging, with networking and communication playing a vital role in scaling AI systems.
106. Deep learning and the concept of neural networks are likely to be remembered as foundational breakthroughs, even if specific architectures evolve.
107. The future may involve specialized robots and advanced human-computer interfaces, but core human needs for community and agency will likely persist.
108. The increasing value of physical goods, in-person experiences, and human craftsmanship is a potential counter-trend to AI automation.
109. Trust and verification systems will be crucial for navigating a world with both human- and AI-generated content.
110. The potential for AI to destabilize civilization is a serious concern, but human ingenuity and community are sources of hope.
111. AI can be viewed as a powerful tool that enhances human agency rather than replacing it, provided humans remain in control.
112. The accessibility of human knowledge through LLMs has the potential for profound, long-term societal impact, driving innovation and learning globally.
113. Personalization and on-the-fly information packaging by LLMs offer significant value in areas lacking structured, information-dense resources.
114. The monetization of AI through advertising is a complex challenge, balancing user experience with revenue generation.
115. The "dream" of a single, all-encompassing AI model is evolving towards a system of specialized agents working together.
116. The development of AI will likely lead to a greater premium on fundamental human experiences like in-person interaction and authentic creation.
117. The current pace of AI development suggests continuous amplification of capabilities rather than immediate paradigm shifts.
118. The "slop" of AI-generated content may eventually lead to a societal re-evaluation and a higher appreciation for genuine human creation.
119. The ability of LLMs to generate code is rapidly advancing, with potential to automate significant portions of software development.
120. The "remote worker" definition of AGI is a useful, if imperfect, starting point for understanding AI capabilities.
📊 Detailed Explanation
1. The AI landscape is intensely competitive, with both research and product development accelerating rapidly. The conversation highlights the "DeepSeek moment" in early 2025 as a catalyst for this acceleration. Companies are not only pushing the boundaries of research but also rapidly deploying products, leading to an "insane" pace of development.
2. China is a significant player in AI, with companies like DeepSeek, Zhipu AI, and MiniMax releasing strong open-weight models, challenging US dominance. DeepSeek's release of near state-of-the-art performance with less compute was a surprise. This spurred a movement in China, leading to many companies releasing powerful open-weight models, with Zhipu AI (GLM models) and MiniMax (Kimi Moonshot) emerging as strong contenders.
3. Open-weight models, especially from China, are gaining traction due to their accessibility and potential to disrupt US business models reliant on API subscriptions. The appeal of open-weight models lies in their availability for anyone to use and modify. This is particularly attractive to US enterprises hesitant to rely on Chinese API subscriptions due to security concerns, and to individuals and smaller companies who may not be able to afford proprietary solutions.
4. Budget and hardware constraints are becoming the primary differentiators, rather than proprietary technological ideas. The speakers believe that technological ideas are flowing freely due to researcher mobility. The real bottleneck and competitive advantage will lie in the immense resources (budget and hardware) required to implement these ideas at scale.
5. The hype around models like Anthropic's Claude 3.5 Opus is significant, but differentiation between top models is decreasing. While Claude 3.5 Opus has generated immense buzz, the conversation suggests that the performance gap between leading models is narrowing, making it harder to distinguish them based on raw capability alone.
6. Chinese companies are releasing open-weight models for a few years, driven by a desire for international influence and the recognition that US companies may not pay for API subscriptions due to security concerns. This strategy allows Chinese companies to gain global adoption and influence, as they understand the limitations of direct API sales to US markets. The government also sees this as a way to build international influence.
7. Consolidation is expected in the open model space due to the high cost of development, but it's not anticipated to be a major story in 2026. Building and training these models is extremely expensive. While consolidation is inevitable, the current trend suggests more open model builders, particularly from China, will emerge in 2026.
8. The "winner" of a year is often the most recent model, as ideas are rapidly iterated upon and incorporated by competitors. The rapid pace of development means that the latest model released often sets the benchmark, with others quickly adopting its architectural tweaks and training methodologies.
9. Chinese companies have different incentives, with some like MiniMax and Moonshot AI filing for IPOs, aiming for Western mindshare. This indicates a strategic push for global recognition and market presence, beyond just technological advancement.
10. The distinction between hype in the X (Twitter) echo chamber and actual model usage is important; ChatGPT and Gemini target a broad user base. The conversation points out that while certain models might be darlings of the online AI community, their actual usage patterns might be different, with models like ChatGPT and Gemini focusing on a massive general user base.
11. Users often stick with a model until it breaks, similar to choosing browsers or text editors. This "use it until it breaks" mentality suggests that loyalty is built through consistent performance, and switching occurs only when significant issues arise, rather than through constant comparison of minor differences.
12. The future likely involves using multiple LLMs for different use cases, rather than a single dominant model. Users are already employing different models for specific tasks (coding, general queries, creative writing), and this trend is expected to continue, leading to a multi-LLM ecosystem.
13. Google's scale and ability to integrate research and product give Gemini an advantage, while OpenAI is known for landing impactful products despite operational chaos. Google's vast resources allow for better separation of research and product, while OpenAI, despite its reputation for chaos, consistently delivers groundbreaking products.
14. Google's TPUs and data center infrastructure provide a significant advantage in terms of cost and development. By controlling their hardware stack from top to bottom, Google can avoid the high margins of third-party chip providers and leverage their existing data center expertise.
15. OpenAI's research division consistently delivers new research ideas and products, a key organizational strength. The consistent output of innovative models and features from OpenAI's research team is highlighted as a core competitive advantage.
16. There's a trade-off between model intelligence and speed, and users often desire the option to choose between them. The ability to select between a "thinking" (slower, more intelligent) and "non-thinking" (faster, less intelligent) mode caters to different user needs and task requirements.
17. Users often develop "muscle memory" with specific models like ChatGPT due to long-term familiarity and brand recognition. The ease of use and long-standing presence of models like ChatGPT create a user habit that can be difficult to break, even with newer, potentially superior alternatives.
18. Customization of LLMs, like ChatGPT's memory feature, may lead to users having multiple subscriptions for personal and work use. The need to maintain boundaries between personal and professional data may drive users to have separate, specialized LLM subscriptions.
19. The "thinking" versus "non-thinking" model dichotomy reflects a trade-off between intelligence and speed, with different use cases for each. Users can leverage the faster "non-thinking" models for quick queries and the more intelligent "thinking" models for in-depth analysis and complex tasks.
20. Grok-3 Heavy is noted as being particularly good for hardcore debugging. This specific model is highlighted for its effectiveness in tackling challenging debugging tasks.
21. Gemini is praised for its interface and ability to handle long context with specific information retrieval. Gemini's user interface and its capacity to process and extract specific information from extended contexts are seen as key strengths.
22. Claude 3.5 Opus is favored for coding and philosophical discussions, especially with "extended thinking" enabled. The model's performance in these areas, particularly when its advanced reasoning capabilities are activated, makes it a preferred choice for certain users.
23. Grok is useful for real-time information and finding specific AI-related content on platforms like X. Its ability to access and process real-time data makes it valuable for staying updated on fast-moving topics like AI developments on social media.
24. User loyalty to LLMs is often won over by a single impressive feature or performance on a specific query, until a significant failure occurs. This "threshold effect" means users can become attached to a model based on a positive experience, but a major failure can quickly lead to switching.
25. The long context window in models like GPT-4o is a significant development, with ongoing improvements in efficiency. The ability to process and retain information over much longer sequences is a crucial advancement, making models more useful for complex tasks.
26. Chinese open-weight models are not widely used by US-based individuals in the conversation, likely due to a US-centric bias and the current superior output quality of US models. Despite the availability and performance of Chinese models, the participants' usage patterns suggest a preference for US-developed models, possibly due to familiarity, perceived quality, or bias.
27. American companies are currently better in terms of output quality, leading users to pay for them despite the availability of cheaper Chinese models. The current advantage in output quality from US-developed models justifies the higher cost for users who prioritize performance.
28. US companies are hosting Chinese models and selling tokens, a business model that leverages the cost-effectiveness of these models. This strategy allows US companies to offer access to powerful, cost-effective models without the direct development burden, while still generating revenue.
29. Open-weight models from China are popular due to unrestricted licenses, unlike some US models with attached restrictions. The lack of restrictive licensing for Chinese open-weight models makes them more appealing for widespread adoption and modification.
30. The "DeepSeek moment" in early 2025, with its near state-of-the-art performance at lower compute cost, was a pivotal event. This event marked a significant shift, demonstrating that high performance could be achieved with less computational resources, thereby democratizing access to advanced AI.
31. The Transformer architecture, particularly the decoder-only variant used in GPT models, remains the foundational architecture. The core principles of the Transformer, as introduced in the "Attention Is All You Need" paper, continue to underpin most modern LLMs.
32. Mixture of Experts (MoE) is a key architectural tweak that allows models to be larger without increasing compute per forward pass by selectively using different "experts" (feedforward networks). MoE models pack more knowledge into the network by having multiple specialized feedforward networks, with a router directing input tokens to the most relevant experts, thus improving efficiency.
33. Dense models utilize all their parameters in every pass, while sparse MoE models only activate a subset, making them more efficient. This distinction between dense and sparse computation is central to understanding MoE's efficiency gains.
34. Architectural tweaks like Grouped-Query Attention, Sliding Window Attention, and Multi-head Latent Attention are used to differentiate models and improve efficiency, particularly for long contexts. These variations on the attention mechanism help optimize performance, reduce memory usage (e.g., KV cache), and enable longer context windows.
35. The fundamental Transformer architecture has not changed drastically; advancements are often in tweaks and optimizations rather than entirely new paradigms. The core structure remains the same, with improvements coming from modifications to components like attention, normalization, and activation functions.
36. The turbulence and advancement in AI are happening in the stages of training (pre-training, mid-training, post-training) and system optimizations, not just core architecture. Progress is driven by how models are trained, the data used, and the underlying infrastructure, rather than solely by architectural innovations.
37. Supervised fine-tuning and Reinforcement Learning from Human Feedback (RLHF) were key algorithmic advancements that unlocked capabilities beyond GPT-2's architecture. These post-training techniques were crucial for making models like ChatGPT more useful and aligned with human preferences, even with similar underlying architectures.
38. System optimizations like FP8 and FP4 training allow for faster experimentation and training by reducing memory usage and communication overhead. These low-precision training techniques enable more efficient use of hardware, leading to faster iteration cycles.
39. Alternatives to the autoregressive Transformer, like text diffusion models and Mamba models, are being explored but haven't yet replaced the Transformer as state-of-the-art. While these alternatives offer potential advantages like parallel processing, the autoregressive Transformer still holds the lead in terms of overall performance and capability.
40. Scaling laws, the predictable relationship between compute/data and prediction accuracy, continue to hold, but low-hanging fruit is being picked. The predictable improvements from scaling are still valid, but the most significant gains are becoming harder to achieve.
41. Inference time scaling, demonstrated by GPT-4o, allows smaller models to achieve higher performance by using more compute during inference. This technique enables models to "think" for longer periods during inference, leading to better results without necessarily increasing model size.
42. Reinforcement Learning with Verifiable Rewards (RLVR) is a major breakthrough, enabling models to learn complex behaviors and tool use through iterative generate-grade loops. RLVR allows models to learn from tasks with clear, verifiable outcomes, which is crucial for developing skills like coding and tool interaction.
43. Pre-training is extremely expensive, with serving costs often outweighing training costs for large models. The recurring cost of making models available to millions of users is a significant financial consideration, often exceeding the initial training investment.
44. The development of massive compute clusters (e.g., Blackwell) is driving further advancements in training capabilities. The availability of immense computational power is enabling the training of larger, more capable models.
45. While some believe pre-training is plateauing, the continued scaling of compute suggests models will continue to get smarter. The fundamental scaling laws suggest that with more compute, models will continue to improve, even if the rate of improvement changes.
46. The trend is towards larger models and potentially higher subscription costs for cutting-edge AI services. As models become more powerful, the cost of accessing them is likely to increase, reflecting the investment in their development and operation.
47. The sparse nature of MoE models makes them more efficient for generation, a key aspect of post-training. MoE's ability to selectively activate parameters makes it well-suited for the generative tasks involved in post-training refinement.
48. Pre-training remains crucial for building the best base models, even if other scaling methods offer more immediate gains. While inference scaling can boost performance, a strong foundation from pre-training is essential for unlocking those gains.
49. Reinforcement learning compute is becoming more comparable to pre-training in terms of time allocation, though it uses different hardware and scaling approaches. RL training requires significant resources, often memory-bound, and is becoming a more substantial part of the overall AI development lifecycle.
50. Pre-training involves training on vast datasets using cross-entropy loss for next-token prediction, with an increasing focus on data quality and synthetic data. The process of learning from massive amounts of text is evolving to prioritize curated and synthetically generated data for better efficiency and performance.
51. Mid-training is a specialized phase, often focusing on specific data types like long-context documents, to avoid catastrophic forgetting. This intermediate training stage helps models specialize without losing previously acquired knowledge, particularly for tasks requiring long-term memory.
52. Post-training encompasses fine-tuning, DPO, RLHF, and RLVR, focusing on refining model behavior and unlocking skills. This final stage of training is dedicated to aligning the model with desired behaviors, improving its utility, and teaching it specific skills.
53. Synthetic data, including OCR-processed documents and high-quality LLM outputs, plays a crucial role in improving training efficiency. Using data derived from various sources, including AI-generated content, helps models learn faster and more effectively.
54. Data quality, rather than just quantity, is a key driver of model performance, especially for larger models that can absorb more information. The focus is shifting from simply feeding more data to ensuring the data is high-quality and well-structured, allowing models to learn more efficiently.
55. The creation of high-quality pre-training datasets involves rigorous filtering and sampling from diverse sources, adapting to evolving evaluation needs (e.g., math, code). Building effective datasets requires a scientific approach to prune and sample data, tailoring it to the specific capabilities being trained (e.g., reasoning, coding).
56. PDFs, especially from sources like arXiv and Semantic Scholar, are valuable data sources for training. These structured documents provide rich information that can be extracted and used for pre-training, contributing to model knowledge.
57. Data privacy and licensing are becoming critical concerns, with a growing interest in training on explicitly licensed data. The legal and ethical implications of using data for training are leading to a greater emphasis on obtaining proper licenses and respecting data ownership.
58. Anthropic's legal issues with authors highlight the complex legal and ethical landscape of training data. The lawsuit against Anthropic for using copyrighted books in training underscores the ongoing challenges in navigating intellectual property rights in the AI era.
59. LLM-generated data is inevitable, but human curation and verification are crucial for maintaining quality and trust. While AI can generate content, human oversight is essential to filter out errors, ensure accuracy, and maintain the integrity of the training data.
60. The "voice" of LLMs, influenced by RLHF, can be a limitation, making it hard for them to be incisive or express unique insights. The averaging effect of RLHF can lead to models that lack a
Kanal: Lex Fridman