News Archive
Pick a date to read that day's stories.
Days with a blue dot have stories.
- 01 of 10OPENAI
OpenAI ships GPT-5.5 with native long-horizon agents
OpenAI released GPT-5.5, adding native long-horizon agent loops that can run for hours without a human in the middle. The company says tool use, memory, and self-correction are now baked into the base model instead of stitched together in the API layer.
OUR TAKEEvery model launch this year has promised "real agents." This one at least moves the plumbing into the model, which is where it should have lived from day one.
Read more → - 02 of 10GOOGLE
Gemini 3 Ultra gets a 10 million token context window
Google DeepMind pushed Gemini 3 Ultra to a 10 million token context window in AI Studio and Vertex, aimed at whole-codebase and multi-document reasoning. Pricing drops for cached tokens to make long contexts affordable for production workloads.
OUR TAKEBigger context is only useful if the model actually reads it. The caching discount is the more interesting number here, not the headline size.
Read more → - 03 of 10ANTHROPIC
Anthropic opens Claude Code to enterprises with SOC 2 controls
Anthropic made Claude Code generally available for enterprise, adding SOC 2 Type II reporting, VPC deployment, and per-repo policy controls. Admins can now scope which repos and secrets the coding agent is allowed to touch.
OUR TAKEThe interesting bit is not the compliance stamp, it is the per-repo scoping. That is what security teams have been quietly blocking rollouts on.
Read more → - 04 of 10META
Meta open sources Llama 4.5 with a 400B mixture of experts
Meta released Llama 4.5, a 400B parameter mixture of experts model, under a revised community license that expands commercial use for firms above the old 700M user cap. Weights, tokenizer, and eval harness are on Hugging Face.
OUR TAKEMeta keeps calling this open source. It is closer to source available, but the weights are on Hugging Face today and that is what most teams actually care about.
Read more → - 05 of 10NVIDIA
Nvidia unveils Rubin R200 accelerator with 288GB HBM4
Nvidia detailed the Rubin R200 at a partner event, pairing 288GB of HBM4 with a new NVLink 6 fabric. Volume shipments are slated for Q1 2027, with hyperscaler samples going out this quarter.
OUR TAKENvidia is once again selling next year, and everyone is once again pre-ordering. The only real question is whether HBM4 supply cooperates.
Read more → - 06 of 10SECURITY
Prompt injection worm hits multiple agent marketplaces
Researchers disclosed a self-propagating prompt injection that spread through three major agent marketplaces by hiding instructions inside shared tool descriptions. Affected vendors have pulled the offending listings and rotated API scopes.
OUR TAKEAgent marketplaces let untrusted strangers write instructions your model will read. Acting surprised when that goes wrong is not a strategy.
Read more → - 07 of 10STARTUPS
AI browser startup Dia raises $150M Series B at $2.4B
Dia, the AI first browser spun out of The Browser Company, closed a $150M Series B led by Benchmark at a $2.4B valuation. The company says paid subscribers crossed one million in June.
OUR TAKEA browser with a chat sidebar is not a moat. A million people paying for one is at least a data point worth watching.
Read more → - 08 of 10HARDWARE
Apple ships M6 MacBook Pro with 64 core Neural Engine
Apple announced the M6 MacBook Pro, headlined by a 64 core Neural Engine and on device support for 70B parameter models at 8 bit precision. Preorders open Friday, shipping July 25.
OUR TAKEOn device 70B is a genuine milestone, but battery life and thermal throttling on sustained inference are the numbers that will actually matter in reviews.
Read more → - 09 of 10HEALTH
FDA clears first autonomous AI radiologist for chest X rays
The FDA granted de novo clearance to Rad AI Autonomous, the first radiology system authorized to sign off on normal chest X rays without a human radiologist in the loop. Deployment starts in three US hospital networks.
OUR TAKEAutonomous is doing a lot of work in that headline. It is normal reads only, which is the easy 60 percent. Still, first through the door is first through the door.
Read more → - 10 of 10POLICY
EU AI Office fines Perplexity 55M euro over training data disclosures
The EU AI Office issued a 55M euro fine against Perplexity for failing to disclose copyrighted training data sources as required under the AI Act. Perplexity says it will appeal and has 60 days to file.
OUR TAKEThis is the first real bite from the AI Act, not the warning shots. Every foundation model vendor operating in the EU should read the decision, not the headline.
Read more →