Recent
- News briefAI-2026-0230
OpenAI buys the team behind the diffs in Codex
OpenAI will acquire Pierre, whose Trees and Diffs libraries already run in Codex, to rebuild Git workflows for developers and agents.
2 minSource: OpenAI
- News briefAI-2026-0229
Half the price, and rated Critical for cyber
OpenAI's GPT-6.1 Terra halves API prices against its predecessor and meets the Critical cybersecurity threshold in OpenAI's own framework.
2 minSource: OpenAI
- Research noteAI-2026-0228
The operator's instruction fades as context grows
A protocol routes human overrides through typed MCP calls and keeps them ranked ahead of the agent's own plan.
2 minSource: arXiv
- Research noteAI-2026-0227
The stronger model loses when the clock runs
RealtimeBench lets the game carry on during inference, and the ranking between two models reverses.
2 minSource: arXiv
- News briefAI-2026-0226
Meta's new frontier models are built for agent work
Muse Wonder and Muse Wonder Full add memory across conversations and persistence on hard tasks; the safety report is still to come.
2 minSource: Meta
- News briefAI-2026-0225
Google holds back its new frontier model
Gemini 4 Argon goes to trusted cyber defenders first, without cyber guardrails, while broad access waits on safeguards.
2 minSource: Google
- News briefAI-2026-0224
Gemini's live voice model gets a face
Google has put a streaming video avatar on Gemini 3.8 Live for enterprise agents, with lip-sync that follows 97 languages.
2 minSource: Google
- Research noteAI-2026-0223
A budget for how much the teacher says
Self-distillation always aims at the full teacher. Constraining how much teacher information enters at each token improved specialisation in seven of eight settings.
3 minSource: arXiv
- Research noteAI-2026-0222
A scoring rule that punishes guessing
Volume-based accuracy rewards a retrieval system for answering everything. A penalty-aware framework separates the three ways it fails.
3 minSource: arXiv
- AnalysisAI-2026-0221
A hundred million words against fifteen trillion
Children reach fluency on about 100 million words. Llama 3.1 took 15 trillion tokens, and nobody can yet say what closes the gap.
3 minSource: MIT Technology Review
- Research noteAI-2026-0220
Both agent-written papers were rejected
Princeton researchers gave agents six days, $3,000 in credits and unpublished research questions. The original authors turned down the results.
2 minSource: MIT Technology Review
- News briefAI-2026-0219
Gemini 3.7 Flash lands three weeks after 3.6
Google's workhorse model gains sharply on coding and agent benchmarks, with introductory pricing held to the end of the year.
2 minSource: Google
- AnalysisAI-2026-0218
Context windows stopped being the constraint
Every lab now ships a million tokens or more. The bottleneck moved to retrieval quality, and most teams have not noticed.
9 min
- BenchmarkAI-2026-0215
Open weights close the gap on reasoning, not on reliability
Scores converge at the top of the leaderboard. Variance across repeated runs tells a different story.
6 min
- Research noteAI-2026-0211
The quiet standardisation of tool schemas
Four vendors, one shape. Interoperability arrived without an announcement.
7 min
- AnalysisAI-2026-0207
Distillation is doing more work than scale
The interesting releases this quarter were small models trained on the output of large ones.
8 min
- Field reportAI-2026-0203
Evaluation sets are rotting faster than teams replace them
Contamination is no longer an edge case. It is the base rate.
5 min
- Market briefAI-2026-0198
Inference providers compete on latency now, not price
Per-token pricing converged. Time to first token did not.
4 min