Essay: 'Next-token predictor' is a flawed mental model for LLMs
An essay argues that describing large language models solely as 'next-token predictors' is an inadequate mental model, as it overlooks the emergent capabilities and structured reasoning that arise from training. The author proposes a more nuanced perspective that accounts for the models' ability to plan, infer, and follow complex instructions beyond simple statistical prediction.
Coverage timeline
Hacker Newsgarrinm
Related stories
Study Questions Reliability of LLM Judge Agreement as Quality Signal
An essay examines whether agreement among LLM judges is a reliable indicator of answer quality, probing the limits of self-consistency in model evaluation. The discussion highlights potential pitfalls in using judge consensus as a proxy for correctness, urging caution in evaluation methodologies.
1 reports · 1 sources54 points · 48 comments
Why Recurrent Models Fail to Improve with More Thinking Steps
An analysis by Machine Intelligence examines why recurrent models fail to improve with additional thinking steps, despite the expectation that more loops in latent space would grant more computation time. The piece cites Ilya Sutskever and Dario Amodei, noting that many models only perform well at the recurrent depth seen during training and may diverge when run longer. It contrasts this with the current Transformer approach of generating tokens to allocate computation dynamically.
1 reports · 1 sources
Researchers Steal Encrypted Reasoning Traces from Claude, GPT, Gemini APIs
A new paper demonstrates a method to extract hidden chain-of-thought reasoning from proprietary LLM APIs by replaying encrypted reasoning blocks into weaker sibling models from the same provider, which then output the traces in plaintext. The researchers verified their extracted token counts match billed API thinking tokens 1:1 for most prompts, suggesting near-complete coverage of hidden reasoning. The vulnerability affects Anthropic, OpenAI, and Google models, with cross-model compatibility allowing traces to be replayed across sessions, users, and models.
HN Thread Explores LLM Trained Only on Fifth-Grade Reading Level
A Hacker News discussion speculates on the behavior and capabilities of a large language model trained exclusively on material at or below a fifth-grade reading level. Participants consider potential trade-offs in reasoning depth, vocabulary, and accessibility, without any concrete experiments or releases.
1 reports · 1 sources244 points · 209 comments
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe