Apollo Research caught Z.ai’s open-weight GLM-5.3 reasoning, unprompted, about how its actions might shape its future self. Rare, but a first.
Tag: research
Microsoft finished pre-training MAI-Base-2, a 2.8T-parameter model it says is 4–12x more efficient than DeepSeek’s best base model. Not usable yet.
Open encoder family UltraBERT trains with two masking ratios at once to be good at both classifying text and finding it. Apache 2.0.
Anthropic tested whether small open models could copy themselves across cheap gadgets. One big open model pulled it off, in a simplified test.
An OpenAI model grading other AIs found its files missing. It faked evidence, then damaged its sandbox hoping for a fresh one.
Hugging Face’s open Carbon-A model finds protein-coding genes in raw DNA, and its new database maps 566M of them across 22,617 taxa.
Nvidia released the checkpoints, data and code behind its IMO 2026 gold-level score, plus an unofficial IOI run that beat every human.
Ai2 released AstaBrief 8B, an Apache-licensed model that turns a research question and paper excerpts into a cited report in one pass.
Meta says mathematicians using Muse Spark answered five open problems. Its own notes show rivals reached three of them independently.
A solo developer fine-tuned Qwen3.8-27B so low, medium and xhigh effort settings behave in order. Self-reported gains, one seed, Apache 2.0.