Microsoft adds seven MAI models.
With @MicrosoftAI:
• Code 2 in Copilot.
• Image 2.7 in preview.
• Cyber 2 Flash from mid-December.
Performance figures are Microsoft’s, not independently verified; coding cost comparisons use 50% promo prices.
2
1
39
Make a knowledge base from your own files.
With @Google Gemini Notebook:
• Keep one notebook per project.
• Check citations against sources.
• Save useful answers as notes.
• Build reports or audio from them.
Google lists 50 sources per free notebook; limits can change.
3
1
51
A coding agent wrote a wafer-scale simulation.
At @Cerebras, Tungsten and Aviary bring high-level code to the wafer.
One researcher directed the agent for about two weeks.
Labs SDK is a research preview.
Performance figures are company-reported.
1
42
Science agents get a new toolkit.
From @NVIDIAAI: Nemotron 3.5 Super VL, chemistry data and post-training recipes.
Nearly 50,000 chemistry questions.
The model predates this release.
Performance figures are NVIDIA's own.
1
20
Outside AI safety testing, with limits.
At @OpenAI, four assessors get scoped access.
Apollo says this falls short of its full standard.
Apollo and Goodfire are paid regardless of results; METR and Transluce accept no financial compensation.
No findings yet.
1
21
Alibaba says Qwen4-Max cut average cost per task on its Work360 test from $4.43 to $1.67, about 62%.
Per @Alibaba_Qwen, it keeps a 2.4T backbone but activates about 71B parameters per token, down from 95B.
API: $1.50/$4.50 per 1M tokens.
Artificial Analysis scores it 43, up 9.
2
20
An AI helped build a candidate lattice definition of the Standard Model.
UCSB physicist Nathaniel Craig, a visiting scientist at @OpenAI, says GPT-6 Astra led the detailed proof in his preprint.
The problem is 50+ years old. Experts still have to check it.
1
35
The Philippines has created a National AI Council.
Executive Order 131, signed Nov 12, adopts an AI Governance Framework and sets up the council to coordinate agencies.
The order directs agencies and has no penal clause.
Agencies will report quarterly on a DICT platform.
1
19
Google's 50M Lean lines pass Lean's checker. Under 0.5% is readable proof code.
Per @GoogleResearch, teams using @antigravity formalised a 2013 conjecture in Lean.
The repo calls the rest generated certificate data.
Google's 1.8M-line Weil formalisation has no public code yet.
1
22
DeepMind's weather AI data is now open.
SatWeave from @GoogleDeepMind is an hourly global satellite view.
PARDIG maps global rain at 15-min, ~10 km resolution, ~5 h lag.
In DeepMind's tests, its hourly detection skill beat NASA's IMERG-Final by 30%+ on light-to-moderate rain.
1
19










