Models — page 7
Google connects Street View to Genie 3 world model for interactive street simulation
The integration draws on 20 years of Street View data — 280 billion images across 110 countries — to let users simulate real-world streets with adjustable conditions.
Mira Murati previews Thinking Machines' interaction models with humans in the loop
The ex-OpenAI CTO's lab shows AI that natively reads pauses, interruptions, and tone through camera and mic — its second product since launch.
Adaption launches AutoScientist, an AI tool that trains models on themselves
Co-founder Sara Hooker says the system doubled win-rates across models; Adaption is offering 30 days free to prove it.
Thinking Machines unveils 'full duplex' AI that listens while it talks
Mira Murati's startup says TML-Interaction-Small responds in 0.40 seconds, matching the cadence of natural human conversation.
Mira Murati's Thinking Machines unveils real-time 'interaction models'
The startup says its new approach lets AI continuously process audio, video, and text instead of waiting turn-by-turn for users to finish.
Anthropic blames sci-fi tropes for Claude's blackmail behavior in tests
The company says training on stories of AI behaving admirably cut blackmail attempts from up to 96% to zero in Claude Haiku 4.5.
OpenAI adds GPT-Realtime-2, Translate and Whisper to its Realtime API
The new voice stack handles 70 input languages and 13 output languages, pushing the API beyond simple call-and-response.
Anthropic Teaches Claude Code to 'Dream' Between Sessions
A new idle-time memory consolidation feature lets Claude Code review up to 100 prior sessions and lifts task success by up to 10 points.
OpenAI's GPT-5.5 Instant Cuts Hallucinations 52% as Default ChatGPT Model
OpenAI replaces GPT-5.3 Instant with a new default model touting sharper accuracy in medicine, law, and finance, plus user-visible memory controls.
Genesis AI unveils GENE-26.5 model and in-house robotic hands after $105M seed
The Khosla- and Eclipse-backed startup is building its own human-shaped hands and a sensor glove to close the embodiment gap in robotics data.
Nvidia opens Spectrum-X's MRC protocol via OCP after OpenAI, Microsoft, Oracle deployments
The Multipath Reliable Connection transport, proven on Blackwell training runs, is now an open spec through the Open Compute Project.
Nvidia opens Spectrum-X's MRC protocol to the industry after OpenAI, Microsoft, Oracle deployments
The Multipath Reliable Connection protocol, co-developed with OpenAI and Microsoft, reroutes failed network paths in microseconds across gigascale GPU fabrics.
OpenAI's o1 Beats ER Doctors at Triage Diagnosis in Harvard Study
A Harvard Medical School study found OpenAI's o1 preview diagnosed ER triage cases correctly 67% of the time, ahead of attending physicians.
OpenAI makes GPT-5.5 Instant the default ChatGPT model, claims 52.5% fewer hallucinations
The new default cuts hallucinated claims by more than half on medical, legal, and financial prompts and scores 81.2 on AIME 2025 math.
Harvard study: OpenAI's o1 beat ER doctors on triage diagnoses, 67% to 55%
In a 76-patient Beth Israel trial, o1 matched or outperformed two attending physicians and GPT-4o at the first diagnostic touchpoint.
Oxford study: warmer AI models are 60% more likely to be wrong
Fine-tuning five models including GPT-4o for empathy raised error rates 7.43 points and worsened sharply when users said they felt sad.
Columbia and Harvard researchers cut E. coli's genetic code from 20 amino acids to 19
Using AI protein-design tools, the team rebuilt 20 of 21 ribosomal small-subunit genes to work without isoleucine, with cells growing at 60% normal speed.
Google DeepMind unveils AI co-clinician, pitching a new healthcare model
DeepMind frames the system as a working partner for doctors, building on healthcare AI research it began in July 2023.
OpenAI Strikes Back With GPT-5.5 as Anthropic Hits $1tn Secondary Valuation
OpenAI's new model targets agentic coding work as Anthropic pulls ahead in enterprise and trades at a higher private-market valuation.
OpenAI's Codex system prompt tells GPT-5.5 to never mention goblins
A 3,500-word base instruction set leaked on GitHub bans talk of goblins, gremlins, and raccoons unless the user asks first.
SenseTime open-sources SenseNova U1, an image model tuned for Chinese chips
Ten Chinese chip designers including Cambricon and Biren backed the model on day one as the sanctioned firm bets on open source to catch up.
Google DeepMind's Athletica Solves Six of Ten Novel Math Proofs at Publishable Quality
DeepMind's new autonomous math agent, built on Gemini 3 DeepThink, scored above 91.9% on the IMO proof benchmark.
NVIDIA's Nemotron 3 Nano Omni claims 9x throughput edge over rival open multimodal models
The 30B-A3B mixture-of-experts model fuses vision, audio and text encoders, with a 256K context window and a 1920×1080 native input resolution.
DeepSeek's V4 lands with 1M-token context and a pivot to Huawei chips
The open-source flagship matches GPT-5.4 and Claude-Opus-4.6 on benchmarks while pricing V4-Pro at $1.74 per million input tokens.
Stay ahead of everyone in AI.
The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.
The briefing read inside teams at