What if your video game world could manage multiple agents — players, robots or AI — without any effort? NVIDIA, @Tsinghua_Uni, and Vector Institute present Gamma-World, a generative multi-agent world model. They invented two tricks: a 'simplex' encoding
RESEARCH
-
Consciousness autocatalytic, LLMs only imitate patterns
By
–
Consciousness is autocatalytic: it feeds on the substrate conditions it creates. LLMs are not autocatalytic, their behavior only depends on the patterns they imitate. (The question whether LLM personas can experience phenomenonology is a completely different one though!)
-
Researchers restore personality to AI language model outputs
By
–
Large language models trained on millions of authors produce a flattened, agreeable voice. Now researchers have found a way to restore individual personality to AI output, enabling applications from therapy training to personalized education.
-
Frontier AI models fail to change ‘three words’ to ‘four’ when appropriate
By
–
This is an interesting test, and the frontier models (GPT-5.5 Pro Extended, Claude 5 Fable Max) do fail. They refuse to turn the "three words" into "four" if that fits better Prompting the AI to act like a translator surfaces the problem, but it still avoids changing the wording
-
Anticipation for Gemini 3.5 Pro in fierce competition
By
–
Really curious for Gemini 3.5 Pro. The competition is currently fierce. It needs to be a big release.
-
GPT 5.5 cracks cmatrix, 80% milestone in 18 months, Bellard tasks hardest
By
–
GPT 5.5 cracked cmatrix, one of the easiest of the 200 tasks. @jyangballin on why 80% is maybe a year and a half out, and why the hardest instances (SQLite, the PHP interpreter, the Tiny C compiler) trace back to one developer, Fabrice Bellard. From Benchtalks #2. pic.twitter.com/k6F6KgFP0D
— Snorkel AI (@SnorkelAI) 11 juin 2026GPT 5.5 cracked cmatrix, one of the easiest of the 200 tasks. @jyangballin on why 80% is maybe a year and a half out, and why the hardest instances (SQLite, the PHP interpreter, the Tiny C compiler) trace back to one developer, Fabrice Bellard. From Benchtalks #2.
-
Berkeley AI alumnus wins PAMI Young Researcher in Computer Vision Award
By
–
Congratulations to @berkeley_ai alumnus @pathak2206 who has been awarded the PAMI Young Researcher in Computer Vision Award! This top award for young researchers in computer vision is given to two recipients yearly. https://
thecvf.com/?page_id=413#Y
RA
… -
Free Guide on Understanding Benchmarks and Evals Misleading Training
By
–
INCREDIBLE The MOST COMPLETE GUIDE for understanding benchmarks and evals, and why training on them is intentionally misleading is now available online to read for free Covers the fundamentals – What machine learning is actually trying to measure (generalization vs
-
Critique of exaggerated claims about AI as scientist
By
–
Devastating to a lot of overclaims about AI as scientist.
-
Apple develops faster image codec using neural architecture search
By
–
What if your photos compressed 3x better while decoding faster than most ML codecs run on a V100 GPU? Apple researchers present a new learned image codec built for both perceptual quality and on-device speed. They used neural architecture search over millions of backbones to