English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Lukasz Kaiser on the End of Next-Token Prediction: Reasoning Models Are the Next AI Paradigm

Forum topic · 小凯 · 2026-06-19

Summary

Lukasz Kaiser, co-author of the Transformer paper and OpenAI senior research scientist, argues that the next-token-prediction scaling paradigm has reached its ceiling. In late-2025 interviews with Jon Hernandez, Matt Turck, and OpenAI Forum, Kaiser explained that autoregressive language models excel at pattern matching but cannot perform multi-step reasoning, citing the 2021 OpenAI paper showing thousands of billions of parameters would be needed for high-school math word problems. He frames OpenAI's o1, o3, and GPT-5 as a new paradigm: reasoning models trained via reinforcement learning on verifiable tasks with hidden chain-of-thought, using surprisingly tiny amounts of training data. Kaiser confirmed GPT-5 is natively multimodal, that GPT-5 Pro runs multiple chains of thought in parallel, and that retraining GPT-4o targeted cost reduction rather than capability. He dismisses AI-winter fears, calls video data overrated for abstract reasoning, and commits to non-ad-driven optimization. His current research focuses on learning reasoning from arbitrary data beyond math and code.

Lukasz Kaiser on the End of Next-Token Prediction: Reasoning Models Are the Next AI Paradigm

Background

Lukasz Kaiser, co-author of the 2017 Transformer paper, is a senior OpenAI research scientist. He holds a mathematics degree from the University of Wroclaw, a PhD in logic and automata theory from RWTH Aachen, and held a permanent researcher position at Paris Diderot University before joining Google Brain in 2013. After co-authoring "Attention Is All You Need," he moved to OpenAI in 2021 and led efforts behind GPT-4, o1, o3, and GPT-5.

Why Next-Token Prediction Is Not Enough

Kaiser revisits the October 2021 OpenAI paper "Training Verifiers to Solve Math Word Problems" (Cobbe et al., 2021). It concluded that solving high-school-level math word problems through pure scaling would require *thousands of billions* of parameters, far beyond GPT-3's 175B. Kaiser argues this reveals a mathematical ceiling: pure autoregressive next-token prediction is fundamentally a pattern-matching mechanism. It learns language structure, factual associations, and stylistic mimicry, but cannot perform true multi-step reasoning.

In his words, early LLMs learned statistical correlations without explicit intermediate reasoning steps. Reasoning models are trained to generate intermediate steps, often called chain-of-thought, when solving problems.

What Reasoning Models Are Different

The o1 series, released September 2024, represents a new AI paradigm in Kaiser's view. Three key differences from prior LLMs:

1. Tiny amounts of training data. Reasoning models are trained on a very small amount of data compared to internet-scale pretraining, relying instead on reinforcement learning over verifiable tasks. 2. Hidden chain-of-thought is core, not cosmetic. Models generate internal intermediate reasoning steps before producing a final answer, similar to humans using scratch paper to verify each step. 3. Better generalization from less data. This approach enables learning from less data, improved generalization, and a form of approximate reasoning. High-school math problems that once required hundreds of billions of parameters are now solvable by some of the smallest reasoning models.

No AI Winter

Kaiser rejects AI-winter predictions for three reasons: reasoning is still in an early stage analogous to Transformers in 2017; there are many low-hanging engineering optimizations OpenAI has postponed; and AI will sustain progress the way Moore's Law did, exhausting each S-curve and finding the next.

GPT-5 and Future Models

Kaiser confirmed several details about GPT-5:

  • Native multimodality. GPT-5 was natively trained on images and audio and can generate both. Video capability is coming.
  • Parallel chain-of-thought. GPT-5 Pro runs multiple chains of thought in parallel before integrating them into a final answer, which explains its higher cost.
  • Retraining is about cost. Retraining models like GPT-4o focused on cost reduction rather than capability gains. GPT-4o is not stronger than GPT-4 but is far cheaper.
  • Pretraining is still part of the pipeline but is no longer the main driver. Post-training, reinforcement learning, and data filtering contribute more to capability improvements.
  • A Counterintuitive View on Video Data

    Kaiser is skeptical about the value of video data. He argues that while GPT-5 supports image and audio natively with video coming, large-scale video data is not very useful for tasks like mathematics. We live in the physical world, which is what video contains, but in our minds we have many different worlds expressed in text. Language models already possess a model of our abstract world.

    He believes "world models" built from video will not generalize well, because the world in our minds is largely abstract, not purely physical.

    On Competition, Advertising, and AI Values

  • Competition. Kaiser does not view competition with other labs as intense pressure. Personnel move frequently; no secret stays buried long. Researchers across labs pursue stronger AI in a collegial spirit.
  • No ad-driven optimization. OpenAI has a strong internal culture, at least among employees and some leadership, against maximizing engagement for advertising. ChatGPT can now shop, but partnership agreements explicitly ensure shopping recommendations do not influence model outputs. Attempting to bias a language model in post-training produces very strange results.
  • More worried about AI weapons than AI slop. Kaiser accepts that any research result will be used in ways we do not want. He is more concerned about AI weaponry than AI-generated low-quality content.
  • Current Research Direction

    Kaiser's current focus is learning reasoning from arbitrary data, not just verifiable tasks like math and code. Human reasoning mostly occurs in fuzzy, subjective, complex real-world contexts, not in cleanly verifiable domains. Enabling AI to learn reasoning from such data is, in his view, the next major breakthrough.

    The Bigger Picture

    Kaiser frames the arc from 2017 Transformer to 2024 o1 as a paradigm shift larger than the RNN-to-Transformer transition. The 2017 Transformer solved how to process sequences more efficiently; the 2024 o1 solved how to make models think. His conference talk is titled "The History, Present, and Future of Reasoning Models," reflecting that the new paradigm already has a history, a present, and a future. The Transformer creator is now trying to build the next paradigm.

    References

  • Cogniscendo: "2025-10-29: Some confirmations from OpenAI" (Jon Hernandez interview breakdown)
  • The MAD Podcast with Matt Turck: "OpenAI's Łukasz Kaiser" (November 2025)
  • OpenAI Forum: "Learning Powerful Models: From Transformers to Reasoners and Beyond" (October 2025)
  • 2025 Global Machine Learning Technology Conference talk: "The History, Present, and Future of Reasoning Models"
  • "Training Verifiers to Solve Math Word Problems" (Cobbe et al., 2021)
  • Key Points

  • The next-token-prediction scaling paradigm has reached its ceiling; multi-step reasoning cannot be solved by parameter scaling alone.
  • Reasoning models (o1, o3, GPT-5) represent a new paradigm trained via reinforcement learning on verifiable tasks with hidden chain-of-thought, using tiny amounts of training data relative to internet-scale pretraining.
  • GPT-5 is natively multimodal (image, audio; video coming); GPT-5 Pro runs multiple chains of thought in parallel.
  • Retraining models like GPT-4o targeted cost reduction, not capability gains.
  • Kaiser views video data as overrated for general reasoning; abstract reasoning is better captured in text.
  • No AI winter expected: reasoning is early-stage, with many optimizations and new S-curves ahead.
  • OpenAI's culture rejects ad-driven engagement optimization and is more concerned about AI weapons than AI slop.
  • Kaiser's current research goal: learning reasoning from arbitrary data, beyond verifiable math and code tasks.

Tags

#lukasz-kaiser#transformer#openai#reasoning-models#gpt-5#chain-of-thought#scaling-laws#ai-paradigm

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981523