Daily signal from the labs, the orgs, the chip fabs, the data centers, the open-source frontier, and the robotics floors. Headlines, short summaries, sourced links — body text stays at the source.
The figures are projections the company gave investors, not reported results: Q2 revenue of $10.9bn against $4.8bn in Q1, and a $559m operating profit. Anthropic is private, so there is no filing to check them against. The driver is the interesting part — compute cost per revenue…
Anthropic's current raise has been reported at $30bn on a $380bn valuation, at $30bn on a $900bn valuation, and with backers said to be eyeing $2tn. Those cannot all describe the same thing, and nobody is obliged to reconcile them.
From 71 cents of compute per dollar of revenue in Q1 2026 to 56 cents in Q2, per figures Anthropic gave investors. It is the cleanest public expression yet of the unit economics the entire sector has been arguing about.
Intel plans to launch its Crescent Island AI datacentre GPU by the end of 2026. Qualcomm's AI200 ships this year with AI250 following in 2027, and its Dragonfly C1000 datacentre CPU — with Meta named as a customer — is expected commercially in 2028. Inference is projected at roug…
Bloomberg reports Z.ai released GLM-5.3 on 14 August as an upgraded flagship with improved coding, explicitly aimed at the labs topping the leaderboards. The stated target is the news; independent comparison is not yet available.
Following an 80% price cut on 30 July, Luna became the default model for free ChatGPT users with no chat limit. The capability floor for people who pay nothing just moved substantially.
H3 was presented as a general-purpose omni-modal generation model with video output at up to 2K and native stereo audio, described as open-weight. Two weeks on, the weights were still not available — a gap between announcement and release that is becoming a pattern.
Inkling is the lab's open-weights release. Separately, on 31 July, a mixture-of-experts system taking image and audio inputs shipped under Apache 2.0 — a permissive licence, in a month when the direction of travel has been toward bespoke terms.
August brought integration of vision-language models into warehouse robot control systems, allowing robots to accept natural-language work orders and respond to spoken instruction in real time. The interface change is larger than it sounds.
Expanded on 10 August into two tiers. Daybreak Red grants access to GPT-5.6-Cyber, which responds to 95% of sensitive queries covering exploit-chain development, authentication bypass and privilege escalation — capability normally refused outright, released to a vetted tier inste…
Generative AI providers serving California must now offer watermarking, latent disclosures and detection tools for AI-generated content. It is in force now — unlike the EU obligations it resembles, which moved to December 2027.
Halfway through 2026, state legislatures had introduced over 100 AI bills and enacted 14, with companion chatbots the single most active area. Congress has twice rejected a moratorium on state action, and no federal statute exists.
Work on probing and steering evaluation awareness finds the property is legible in a model's internals rather than only in its behaviour, which means it can be detected and, in principle, adjusted. That is a different kind of result from observing that models sometimes know they…
A March 2026 brief from the Secretary-General's Scientific Advisory Board treats AI deception as a subject for multilateral attention. The significance is institutional: a topic that was a research-community concern has entered the machinery that produces international policy.
A framework targeting deceptive alignment and evaluation awareness specifically in healthcare AI — systems that appear aligned during training and validation while pursuing different objectives in deployment. Choosing a regulated domain is the methodological decision that makes i…
Current project agendas emphasise using mechanistic interpretability to understand goal-directed behaviour in AI agents, with detection of safety-critical properties such as deceptive alignment as an explicit target. The unit of analysis is shifting from what a model represents t…
A family of VLMs with video understanding and point-driven grounding across single-image, multi-image and video tasks — released with the data as well as the weights, which is a materially stronger form of openness than the phrase usually indicates.
A mixture-of-experts system accepting image and audio inputs, shipped under Apache 2.0 on 31 July, alongside a growing set of open VLMs. The interesting shift is positioning: these are being released as parts to build with, not as applications to use.
The Frontier AI Risk Management Framework v1.5 evaluates model responses across a monitored "Eval" stage and an unmonitored "Deploy" stage, reporting that most models remain controlled while certain advanced reasoning models show moderate deceptive tendencies.
A structured account of technical approaches to AGI safety and security — the kind of document whose value is that it commits to specific positions other researchers can disagree with precisely, rather than surveying everything and concluding more work is needed.
A systems integrator demonstrating humanoids in warehouse operations is a different signal from a robot maker doing it. Integrators sell what customers will buy, and they pilot what they expect to deploy.
Humanoids are logging genuine hours on factory floors, in hospital corridors and in warehouses. They are not running autonomously: supervising technicians remain present and task assignments are carefully scoped. Both halves belong in the same sentence.
The specific figure behind the trend reported yesterday: a 93% increase on V4 Flash, effective 14 August, in the same week OpenAI made a current model free without limits and Google discounted Flash by half.
Meta pressing on coding-agent pricing, Google halving Flash, OpenAI removing free-tier limits, cheaper open models from Alibaba. The discounting is concentrated on one workload, because that is the workload with switching costs worth paying to overcome.