Fork just before the model changes its mind
Why fork placement decides step level credit in tree structured reinforcement learning, and how belief shift branching offers a practical answer.
Core topic
AI infrastructure
Inference and GPU infrastructure that turns models into services
The performance and reliability decisions behind distributed inference, GPU scheduling, KV caches, MoE serving, and on-premises operations.
Related stories
Asking about consciousness with calculus, what Gradland says about experience
Following the hypothesis that Jacobian structure shapes phenomenal experience through the differentiable world of Gradland, and how effective rank and cohesion explain vividness and duration.
Should SFT Learn Every Token, What Trimming Teaches Us About Math Reasoning
A walkthrough of TrimSFT, which replaces uniform SFT loss with logit gap weighting for mathematical reasoning. Why trimming both extremes helps, and what the results across six models show.
What changes when AI learns entangled instead of smart alone
Walking through a paper that models entangled collective behavior as probability waves and tests it on Chinese intraday trading data, and asking why AGI might need entangled game modules, and where the limits sit.
Rewriting the statistical mechanics of learning in function space
Reading a theory that treats functions and their operators as macroscopic variables for learning dynamics, and tracing why error and fluctuations settle along smooth directions.
Reading images and text with one encoder
An editorial commentary on NeoMME, a single transformer that handles images and multilingual text without a separate vision tower and delivers strong document retrieval at compact sizes
IFM Releases K2 Horizon, Six Apache 2.0 Models From 0.9B to 375B
IFM, the frontier lab launched by MBZUAI, released six models from 0.9B to 375B under Apache 2.0 with the pretraining corpus. Here is what the fleet structure means.
The uncomfortable diagnosis that DNS spreads scams
Following the measurement that more than 10 percent of new gTLDs look like scams, a walk through the abuse structure and what engineers should watch.