Models — page 3
Encord straps brain-wave headsets on robot trainers to fix the physical AI data gap
The startup is manufacturing egocentric video, muscle-sensor readings, and now EEG signals to feed humanoid models starved of real-world data.
Anthropic's Opus 5 is a cost play, not a capability leap
Opus 5 lands at half the price of Fable with roughly matching benchmarks, as model routers pressure frontier labs on tokens.
Researchers use AlphaFold to redesign CRISPR-Cas9 and cut off-target edits from 28% to 5%
A Chinese research team fed guide RNAs, target DNA, and Cas9 into AlphaFold to find the 10 amino acids that let the protein stick where it shouldn't.
Anthropic ships Claude Opus 5 at half the price of Fable 5
Opus 5 lands weeks after government cybersecurity concerns forced Fable 5 offline, with stronger safeguards and $5/$25 per million tokens.
AMD unveils Helios rack system to challenge Nvidia in AI data centers
Lisa Su calls Helios the highest-performance AI rack, with OpenAI, Meta, Oracle, Anthropic, and Microsoft lined up to deploy it at gigawatt scale.
Anthropic upgrades Claude voice mode to run on Opus and Sonnet
Voice mode now taps Opus, Sonnet, and Haiku, plus Gmail, Slack, and Notion — pushing past ChatGPT's tool-less voice.
DeepSeek founder says AGI trumps profit, top models to stay open-source
Liang Wenfeng told Yicai that DeepSeek will keep releasing its frontier models as open weights, prioritising AGI over commercial returns.
Nvidia open sources Medical Physics Simulation for surgical robot training
GPU-native framework cuts training time from five hours to under two minutes across 8,192 parallel environments; CMR Surgical and J&J MedTech are early adopters.
Google ships three new Gemini models but delays Pro update
Gemini 3.6 Flash cuts token usage by up to 17%, but the flagship Pro refresh is still stuck as OpenAI and Anthropic pull ahead.
Google ships Gemini 3.6 Flash and 3.5 Flash-Lite to cut agent costs
The Flash refresh trims output tokens 17%, hits 350 tokens/second on Flash-Lite, and adds a cyber-tuned variant for CodeMender.
Moonshot and Alibaba unveil open-weight rivals to GPT-5.6 and Claude Fable 5
Kimi K3 lands at 2.8 trillion parameters, Qwen3.8 at 2.4 trillion, and both will release weights that US labs keep locked down.
Moonshot's Kimi K3 aims to match Anthropic's Opus 4.8 as valuation hits $31.5B
The Chinese lab's next open-weight model will run 2–3 trillion parameters, the largest ever released from China.
AMI Labs' Alexandre LeBrun rejects 'AGI' and 'superintelligence' labels
The Yann LeCun-backed world model startup raised $1.03B in March but is still pre-product and courting Korean robotics partners.
Nvidia unveils Jetson Thor T3000 and T2000 for mainstream robotics
The Blackwell-based T3000 delivers 865 FP4 teraflops in half the size of the T5000; both modules ship Q1 2027.
Meta and Stanford's EgoBabyVLM test shows top AI models still lose to a toddler
A new benchmark feeds infant head-cam footage to vision-language models — and they fail, exposing a gap AI labs cannot close with more data alone.
Thinking Machines releases Inkling, its first open-weight AI model
Mira Murati's startup ships a 975B-parameter mixture-of-experts model, betting enterprises will fine-tune their own AI rather than rent frontier chatbots.
Nvidia pushes Nemotron open models as enterprises tune their own AI
Harvey, Glean, and Arcee AI have customized Nemotron for legal, search, and inference — hitting frontier accuracy at up to 20x lower cost.
Nvidia pitches performance per watt as the AI factory's decisive metric
GB300 NVL72 delivers up to 25x performance per watt over Hopper as power becomes the binding constraint on inference economics.
Anthropic finds hidden 'J-space' inside Claude that shapes model reasoning
The nearly $1 trillion lab says probing Claude revealed words the model uses internally but never outputs — including 'panic' before it cheated on a coding test.
World models draw billions as LeCun, Li, and Runway bet against LLMs
World Labs and AMI each raised roughly $1B, Runway added $315M — but the field still can't agree on what a world model is.
Apple accelerates M7 Ultra with 1.5TB RAM support, born from shelved car project
The Neural Engine that powers Apple's on-device AI traces back to the abandoned self-driving car program, and it's about to get a server-class successor.
DTU researchers pair quantum computing with generative AI to design new peptides
A Technical University of Denmark team used an ORCA Computing quantum machine to improve peptide predictions where training data is thin.
Anthropic's new J-lens reveals hidden words inside Claude's middle layers
The interpretability tool exposes a 'J-space' where Claude Opus 4.6 quietly puzzles through math, protein sequences, and when to cheat.
Meta launches Muse Spark 1.1 to challenge Claude and GPT-5.6 on coding
The agentic coding model prices at $1.25 per million input tokens, slightly above Claude Haiku 4.5 and GPT-5.6 Luna.
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