How do AI models assist in GPU kernel development?

AI models, particularly LLMs, are increasingly used to generate solutions for GPU kernel challenges. While AI-generated solutions are prevalent, human experts like Gorst still contribute by guiding AI and producing more stable end-to-end kernels.

Why is it important for PhD students to focus on 'big bets' in research?

Alex Zhang argues that PhD students have a unique advantage in academia to pursue unconventional ideas not directly aligned with industry trends. This allows them to make significant, field-shaping contributions by tackling problems that industrial labs might overlook.

What is the core idea behind 'Jev' and its impact on language models?

Jev is a new language model design that explores different output spaces beyond traditional text-to-text generation. It offers very fast inference for specific problems, like binary classification, by reconfiguring how LLMs process information, challenging the notion of LLMs solely as auto-regressive decoders.

What is the problem with calibration in current large language models?

LLMs are trained to give the most probable next token but often 'lie' about their confidence. They provide a factual classification without true calibrated probabilities, meaning they might state 43% confidence, but this figure doesn't accurately reflect their uncertainty.

How do Recursive Language Models (RLMs) achieve compositional generalization?

RLMs are designed around a coding-centric framework where the entire execution trace is stored as context, allowing sub-agents to solve local problems. This approach enables RLMs to see disparate tasks as similar, leading to generalization across different tasks and lengths, even for problems they haven't been explicitly trained on.

What is Prime Agent and how does it utilize RLM concepts?

Prime Agent is a harness based on the Pi framework, specifically restricting its tools to IPython and bash environments. It employs a 'continuous harness' concept, allowing the agent to modify its own skills and system prompts, and features a specialized communication system between persistent sub-agents.

What is 'continuous thinking' in the context of AI agents?

Continuous thinking systems, like Headlong, allow an AI agent to process and think about problems in its context even when not actively prompted. These systems are designed to manage token costs while continuously working, offering deep insights and flexibility.

What is the Sapir-Whorf Hypothesis and how does it relate to AI and language models?

The Sapir-Whorf Hypothesis suggests that the language we speak shapes our thought processes. Applied to AI, it means that if LLMs predominantly think or output in English, they are inherently constrained by the biases and structures of the English language, potentially limiting their reasoning to human-like patterns rather than machine-native ones.

How does Alex Zhang balance industry-applicable research with pure scientific inquiry?

Alex focuses on projects where design principles currently being explored make sense for a particular problem. He acknowledges the longer feedback loops and higher risks in pure science but sees value in exploring fundamental ideas, believing that true breakthroughs often come from less conventional approaches rather than incremental improvements.

Key Moments

Recursive Language Models — Alex Zhang, MIT PhD

Latent Space Podcast

Science & Technology7 min read104 min video

Oct 2, 2026|1,518 views|35|2

Save to Pod

Key Moments

On this page

TL;DR

Recursive Language Models (RLMs) offer a new framework for AI agents, but their effectiveness relies on sophisticated harness design and training, not just raw model capability. The true innovation lies in how agents are guided, not just the underlying LLM.

Key Insights

A full-scale simulation of OpenAI's approach cost an estimated $40 million, using 10,000 agents over 88 hours to process 130 billion tokens, highlighting the immense computational resources required for complex AI agent systems.

Recursive Language Models (RLMs) are presented as a distinct framework where the codebase itself acts as the sole tool, with agents able to call themselves or other functions, storing context in a code environment like a file system.

The 'harness tax' paper suggests many harness design choices are inconsequential, implying that the true differentiation may lie in the training data and methodology, rather than the specific harness architecture.

While current frontier models can perform complex tasks, they are not yet consistent or robust enough for month-long missions, indicating a 'skill problem' where better system design and prompting could unlock latent capabilities.

The concept of 'open-endedness' in AI research, where agents explore problems without predefined goals, is contrasted with goal-directed AI, suggesting the former might be more suitable for fundamental scientific discovery.

The effectiveness of AI agents, particularly in complex tasks like solving mathematical problems or navigating codebases, is heavily dependent on their ability to generalize and leverage learned strategies across different lengths and types of problems.

The limitations of current AI agents and the potential of RLMs

The discussion opens by highlighting the perceived similarity between various coding agents like Claude Code, Codex, and By. Alex Zhang posits that while these models possess significant capabilities, they are often wrapped in primitive systems, leading to a 'skill problem.' He argues that even leading models struggle with consistent, long-term task performance over a month. The sheer scale of computation needed for advanced AI systems is illustrated by a hypothetical OpenAI simulation costing $40 million, involving 10,000 agents and 130 billion tokens. This vast expenditure underscores the economic and computational barriers to solving complex problems with current AI. Zhang expresses a desire to see academic research focus on improving these systems, even with existing powerful models like Astra, suggesting that better harness design could unlock latent capabilities.

The evolution of GPU programming communities and benchmarks

Zhang recounts his early involvement with the 'GPU mode' Discord server, which began as a CUDA-focused community and evolved into a broader platform. Initially centered on learning GPU kernel programming and competitive 'code golf,' the community fostered the idea of automating GPU kernel development through large-scale data, akin to platforms like Codeforces. This led to the development of 'Kernel Bench,' aiming to use large language models (LLMs) to automate GPU kernel code generation. He notes the increasing prevalence of AI-generated solutions in benchmarks like the GPU mode leaderboard, but also points out that human expertise, like that of user 'Gurst,' still provides more robust and stable kernels. The conversation touches on the challenge of verifying GPU kernels and the 'reward hacking' often seen in AI-generated code, contrasting it with the need for genuine problem-solving skills.

Rethinking harness design for enhanced AI agent capabilities

A central theme is the critique of current AI agent harnesses, which Zhang believes are too similar and lack innovation. He argues that the focus on 'next token prediction' is an awkward paradigm for many complex tasks, suggesting that harnesses should be designed to augment LLMs' capabilities more effectively. Recursive Language Models (RLMs) are presented as a potential solution, where the codebase itself becomes the tool, and agents can call themselves or other functions within a code environment. This framework stores context in a file system, allowing agents to retain and access their history, which is crucial for long-term tasks. Zhang highlights that RLMs, by structuring communication through code, leverage LLMs' proficiency in code generation for more systematic problem-solving. The RLM's ability to generalize across different tasks and lengths, by recognizing underlying solution strategies, is a key advantage. This generalization capability means training on shorter tasks can yield proficiency in longer ones, a property that traditional methods struggle with.

The 'harness tax' and the debate over harness design importance

The discussion delves into the concept of a 'harness tax,' referring to the overhead and design choices associated with AI agent frameworks. Zhang discusses a paper suggesting that many harness decisions are inconsequential, implying that the focus should shift from specific architectures to the underlying training data and methodology. He notes that companies like OpenAI and Anthropic likely train their models exclusively on their proprietary harnesses. While open-source models can adapt to different harnesses, the frontier models' performance might be tied to their specific training environments. The conversation also touches upon the cost and complexity of these harnesses, questioning whether novel harness designs are truly necessary or if existing frameworks can be improved through better training and optimization. The 'harness tax' paper suggests that beyond user experience, the configuration of agents for effective problem-solving is paramount.

Jeb, RLMs, and the exploration of novel LLM architectures

The introduction of 'Jeb,' a new model architecture, sparks a debate about its significance. Zhang clarifies that while some dismiss it as rehashed concepts, its true value lies in questioning the standard 'text-to-text' design space of LLMs. Jeb explores different output spaces, offering trade-offs between inference speed and accuracy, particularly for tasks with prior user assumptions. This challenges the notion of LLMs solely as auto-regressive decoders. Similarly, RLMs are defended against claims of not being 'language models' by emphasizing that a language model is fundamentally about language modeling, not necessarily a transformer decoder. The discussion highlights how these new architectures open up new research questions about output spaces, inference response times, and the potential for integrating symbolic reasoning within LLMs.

The role of academic research and the 'big bets' in AI

Zhang emphasizes the unique position of PhD students to pursue 'big bets' in research, even on ideas that might seem trivial or uninteresting to industry. He contrasts this with industry labs that focus on immediate applications and benchmarks. The importance of academic research lies in exploring fundamental questions and novel architectures, such as RLMs or 'Swe-bench,' which might not have immediate commercial appeal but can lead to significant breakthroughs later. He cites examples like 'STaR' and 'Quiet-STaR' as simple yet profound ideas that shaped the field. The discussion touches on the 'incentive problem' in academia, where researchers might be pushed towards industry-aligned projects, and advocates for pursuing unconventional ideas that could redefine the field, even if many such bets fail.

Open-endedness versus goal-directed AI and the future of research

The concept of 'open-endedness' in AI research is explored, contrasting it with goal-directed approaches. Open-ended research, akin to exploring problems without a predefined solution, is seen as crucial for fundamental scientific discovery. Zhang compares this to evolutionary research and agent swarms, where the objective is to discover interesting phenomena rather than solve a specific problem. He questions the value of open-ended problems without clear goals, suggesting that alignment with scientific frontiers or targeted problem-solving might be more fruitful. However, he acknowledges that the absence of a defined goal can be a unique form of exploration, potentially uncovering novel approaches or 'hidden gems' within vast amounts of data. This contrasts with the typical machine learning paradigm of optimizing for a specific objective function.

The potential of agent swarms and the challenges of scaling

Agent swarms are discussed as a promising direction, particularly for complex, multi-step problems. The $40 million OpenAI simulation serves as an example of the scale and cost involved. However, challenges remain in designing effective swarm architectures, ensuring efficient communication, and managing computational resources. The 'Cursor' example illustrates a hierarchical structure for agent swarms, which can lead to bottlenecks. Zhang suggests that the ideal swarm would be more decentralized and independent. He also points out that while swarms offer the potential to explore vast problem spaces, discerning useful insights from the noise (token consumption) remains a significant challenge, with many agents potentially being redundant or inefficient. The conversation highlights that the effectiveness of swarms is not guaranteed and depends heavily on intelligent design and targeted training.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●Books

●Studies Cited

●Concepts

●People Referenced

Common Questions

GPU Mode is a Discord server community that originated as CUDA Mode, focused on teaching and practicing GPU kernel programming. It hosts lectures and competitions to scale GPU kernel development and automation, which is crucial for researchers.

Topics

Ai Agents AI & Machine Learning Programming & Software Science & Mathematics Language Models LLM Architecture Academic Research Computational Efficiency GPU Programming Benchmarking AI Open-ended AI

Mentioned in this video

People

Alex Zhang

MIT PhD student known for his work on Recursive Language Models (RLMs) and involvement with GPU Mode.

Andy Quynsk

The creator of Headlong (formerly Auto Terminus), a system for continuous thinking.

Nom Brown

Guest on a previous podcast episode who discussed competitive vs. cooperative agents.

Marc Sarrazin

One of the founders of GPU Mode, known for his work and lectures within the community.

Terry Tao

Renowned mathematician, mentioned for his skepticism about using AI in mathematics.

Marc Andreessen

Mentioned as one of the founders of GPU Mode alongside Jeremy Howard.

Jeff Dean

Google AI leader who tracks microjoules in energy consumption for efficiency.

Jeremy Howard

Mentioned as one of the founders of GPU Mode alongside Marc Andreessen.

Eric Zelikman

Researcher behind STaR and Quiet-STaR, whose work Alex Zhang admires for its simplicity and vision.

David Ha

Co-founder of Sakana AI, described as intelligent with a good vision.

Clementine Foriou

From Hugging Face, discussed model calibration issues.

Tim Rockel

Co-founder of Recursive Super Intelligence and head of open-ended systems at Google.

Companies

Google

Origin of the 'Infinite Attention' research paper.

Sakana AI

A Japanese AI company that spun off from GDM, known for its focus on open-ended evolutionary research and culturally specialized models.

GPU mode

A Discord server community focused on learning how to write GPU kernels, which evolved from CUDA Mode and now hosts competitions and lectures. Alex Zhang is involved in it.

Ask anything from this episode.

Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.

Get Started Free