A new bias study requested thinking budgets from 0 to 8,192 tokens. GPT-5.1 used about 117 at every positive setting, and DeepSeek-R1 and Qwen3 thought even when told not to. Measured by what the models actually spent, more thinking did not reliably reduce bias.
A new Seoul National University paper finds LLMs repeat the original answer under negation 37 to 71% of the time, and more often when they're confident. The fix that looks best on standard metrics quietly teaches one stock wrong answer.
TasteVal finds AI experimental taste doubling every 3 months since Dec 2025. The same paper labeled 2,000+ model experiments and found zero invented ideas.
OpenAI’s Sep 30 writeup shows protected reasoning was extracted without breaking encryption: by replaying encrypted CoT across chats until it surfaced in the clear.
FTA freezes failed tool observations before generation. Across six models and 3,600 human-annotated responses, false-success falls from 22.8% baseline to 0.8% with a structured evidence contract, while useful responses rise from 74.9% to 98.8%.
Anthropic's Project Swap put Claude agents into a book barter market. Live efficiency hit ~0.55 vs a ~0.89 optimum; ~85% of the shortfall was preference representation, not the trading floor.
Salesforce AI Research’s ten-day agent campaign argues continual operation without forgetting is a harness problem, before continual learning can even start.
Anthropic's Claude sped up 30+ biomolecular models ~4× and cut binder-campaign GPU cost by ~100×. The kernels are the story.
Nature’s Paper2Agent converts papers and code into tested MCP servers that act like virtual corresponding authors.
By Venkata Anirudh Devireddy · Endoblog.dev For over a decade, people who wanted to get into the tech industry were told the best way to do so was to learn to code. This was sound advice back then when coding was the bottleneck in creating software. However, that bottleneck is
By Venkata Anirudh Devireddy · Endoblog.dev They're used interchangeably a lot, but they shouldn't be. The difference between an agent and an assistant tells you about what kind of actions it can perform. An assistant responds. An agent acts. Assistants just respond to your questions. If
By Venkata Anirudh Devireddy · Endoblog.dev There used to be a clear line between the best closed models and the best open weight models. However, the gap between these two categories is starting to shrink, which is something every AI developer should be aware of. What "open weight"