News Archive
Pick a date to read that day's stories.
Days with a blue dot have stories.
- 01 of 10MODELS
OpenAI ships GPT-5.5 with native agentic reasoning
GPT-5.5 arrives with built-in tool use, long-horizon planning, and a new "deliberate mode" that lets it reason for minutes before answering. Benchmarks show large jumps on coding and math.
OUR TAKEAgents no longer need scaffolding, which just made the framework wars considerably harder for LangChain and CrewAI.
- 02 of 10REGULATION
EU AI Act enforcement begins for general-purpose models
The European Commission officially started enforcing the AI Act's GPAI obligations. Providers must publish training-data summaries, evaluate systemic risk, and report serious incidents.
OUR TAKEExpect a Brussels Effect: shipping two model versions is more painful than simply complying globally.
- 03 of 10STARTUPS
Anthropic raises $15B at $300B valuation
Anthropic closed a megaround led by Google and new sovereign-wealth investors, pushing its valuation past Stripe and SpaceX-era benchmarks. The funds go to compute and safety research.
OUR TAKEFoundation-model economics now look like aerospace: a handful of national-scale labs and everyone else fine-tuning on top.
- 04 of 10HARDWARE
Nvidia's Rubin chips begin sampling, 2.5x training throughput
The successor to Blackwell, Rubin, started shipping to hyperscalers. Early reports cite 2.5x training and 4x inference gains, with HBM4 and a new NVLink switch fabric.
OUR TAKEEvery 'GPU-poor' complaint from 2024 is about to repeat in 2026, with power as the bottleneck.
- 05 of 10RESEARCH
DeepMind paper claims continual learning without catastrophic forgetting
A new architecture mixes sparse mixture-of-experts with episodic memory, letting models keep learning post-deployment without overwriting prior skills.
OUR TAKEIf this replicates, the fine-tune-then-freeze era ends and safety becomes a living problem too.
- 06 of 10PRODUCT
Apple Intelligence finally adds on-device agents in iOS 20
iOS 20 lets Siri chain actions across apps using on-device models, with cloud fallback only for heavy tasks. Developers get a new AppIntents 3 framework.
OUR TAKEApple is betting users want privacy-first agents more than the smartest ones, probably right.
- 07 of 10SAFETY
Frontier labs sign joint commitment on bioweapon evals
OpenAI, Anthropic, Google DeepMind, xAI, and Meta agreed to share red-team results on CBRN risks with a new third-party evaluator before model release.
OUR TAKEFirst real coordination since Bletchley 2023, voluntary, but the optics of refusing now cost.
- 08 of 10OPEN SOURCE
Mistral releases Mixtral-Next under fully permissive license
A 240B-parameter MoE that benchmarks within 5% of GPT-5-mini, released Apache 2.0 with weights, training recipe, and a 12T-token dataset card.
OUR TAKEOpen source isn't catching up, 'good enough and free' is now its own product category.
- 09 of 10INFRASTRUCTURE
Microsoft and OpenAI break ground on 5GW Stargate campus
The first Stargate site in Wisconsin will draw 5 gigawatts at full build-out, with a dedicated nuclear PPA. Online by 2028.
OUR TAKEAI buildouts are now indistinguishable from energy policy, whoever permits faster, wins.
- 10 of 10CULTURE
Hollywood studios reach landmark AI licensing deal with SAG-AFTRA
The new contract sets per-use rates for digital replicas, requires consent for training data, and creates a royalty pool funded by streaming AI features.
OUR TAKEA template every creative industry will copy, the question shifts from fair use to rate card.