Learning to Reason with LLMs(openai.com)
1573 points by fofoz | 1186 comments
OpenAI research post introducing o1, a large language model trained with reinforcement learning to reason using a long internal chain of thought before answering.
First screen, captured .
Looks similar
AI agents: Less capability, more reliability, please(sergey.fyi)
393 points by serjester | 230 comments
Blog post by Sergey Filimonov arguing AI agents should prioritize reliability over expanded capability, critiquing flight-booking demos.
I built non-autoregressive decision models with RL a year ago(laya.convaiinnovations.com)
1270 points by nandakishor_ml | 304 comments
Blog post by ConvAI Innovations' founder describing Laya, an open-weight non-autoregressive decision engine trained with RL, and claiming priority over a frontier lab's similar work.
How GLM built its own inference infrastructure(z.ai)
373 points by whiteros_e | 261 comments
Z.ai research blog post describing how its GLM model was used to build its own inference infrastructure, framed as a step toward recursive self-improvement.
Related stories
The Little Book of Deep Learning [pdf](fleuret.org)
420 points by dataminer | 29 comments
A PDF of François Fleuret's The Little Book of Deep Learning, a short introductory textbook on deep learning.
A new PNG spec(programmax.net)
610 points by bluedel | 553 comments
A blog post announcing the release of a new PNG specification after two decades of stalled progress and explaining why it matters.
Thinking fast and slow in AI: The role of metacognition (2021)(arxiv.org)
165 points by teleforce | 68 comments
An arXiv paper proposing a multi-agent AI architecture inspired by Kahneman's fast and slow thinking, with a metacognitive module choosing between system 1 and system 2 agents.
The HTML-in-Canvas API origin trial(developer.chrome.com)
110 points by andrewstuart | 54 comments
Chrome for Developers blog post introducing the HTML-in-Canvas API origin trial, which lets developers draw interactive DOM content into 2D canvas and WebGL/WebGPU textures.
TimeGPT-1(arxiv.org)
373 points by PaulHoule | 114 comments
arXiv paper introducing TimeGPT, a foundation model for time series forecasting that performs zero-shot predictions on unseen datasets.
TODOs aren't for doing(sophiebits.com)
380 points by todsacerdoti | 218 comments
A blog post by Sophie Alpert arguing that TODO comments are valuable as documentation of known issues even if never acted on, and shouldn't be auto-deleted or tracked.