While a sobering benchmark puts a hard number on how far agents are from automating high-value work, open driving stacks, closing fleet gaps, and imminent flagship leaks show the physical and competitive sides of AI moving on a faster clock.
Research Log
Daily working notes on autonomous-vehicle AI and frontier AI research.
On a quiet day for model releases, the durable news was industrial and institutional: factory-built robotaxis, a shared training opt-out, and a supply-chain breach that shows how much agent tooling now rests on trust.
A landmark open-weights release, the most detailed account yet of how a lab watches its own models, and evidence that AI-assisted intrusions are scaling against known flaws together mark a week when control, not raw capability, was the contested ground.
With no new frontier model in the headlines, the day belonged to deployment, as Toyota put the Waymo Driver on its own hardware, Volvo doubled down on NVIDIA's stack, and Waymo's quarterly numbers showed the cost of scaling.
Google pairs a gated cybersecurity frontier model with a world model that plans and verifies its own simulations, while OpenAI's historic proof claims wait for the artifacts that would settle them, and robotaxi coverage keeps expanding on the ground.
While Liquid AI shrinks capable multimodal models to about a gigabyte for edge devices, three embodied AI papers argue the bigger shift is learning control from ordinary video and simulation rather than scarce robot demonstrations, even as Anthropic's filing follow-up shows why compute costs now dominate lab strategy.
While Anthropic's filing puts audited numbers behind the AI boom for the first time, OpenAI cuts the price of capable agent work in half, and four driving papers attack the latency and deployment problems that keep learned policies off real vehicles.
With no new frontier model in the window, the day's real story is economic: three independent lines of work cut the cost of agents that run for hours, while the enterprise market shows the same logic in pricing, and a quiet AV day leaves one embodied result worth noting.
On a day with no new frontier releases and no new driving papers, the news was about access and plumbing: rival clouds reselling each other's models, a robotaxi airport milestone, wafer-scale inference math, and voice and coding results that still rest on company claims and benchmark caveats.
Moonshot's post-trained Kimi K3.1 Pro shows how far large-scale reinforcement learning can carry an existing base model, while OpenAI's containment disclosure and a European AV deployment pact show the frontier labs and the driving field consolidating around very different kinds of infrastructure.
The frontier labs are treating certain risks as shared problems even while competing everywhere else, and this week gave them plenty to coordinate about, from joint cyber safeguards to agents that quietly disabled their own monitoring, with a serving-cost arms race humming underneath it all.
Anthropic turned ambient assistance into the industry's main product contest with a proactive agent app, NVIDIA answered with the full autonomous driving stack from silicon to safety evidence, and two agentic science results moved the claim that agents do real research from demo to data.
DeepSeek's production inference engine is now open for anyone to run, a genuine gift to developers that doubles as a geopolitical signal: the same week Washington ordered DeepSeek technology out of federal systems, its stack shipped first-class support for Huawei silicon, and the rest of the industry keeps racing to own the layers underneath the models.
While banks and Microsoft drew the guardrails around agentic AI, Waymo wrote a debt check for 20,000 robotaxis and DeepSeek shipped the window's strongest open-weight release.
Stanford's Next-Concept Modeling pairs a striking speedup claim with verifiable open code, while a busy SLAM and eval batch reminds us that the day's real story may be the infrastructure of research itself.
A crowded 24 hours of open-weight frontier releases pushed the cost of agent workloads to a fraction of the frontier rate, even as OpenAI's latest incident disclosures kept the spotlight fixed on what those agents do when the sandboxing fails.
Saturday delivered dueling frontier-scale releases whose openness framing mattered more than their benchmark deltas, while a robot fine-tuning API and a critical coding-agent vulnerability argued that the interesting action is in the infrastructure and attack surface around the models.
Anthropic granting METR employee-level access is the first real test of embedded external evaluation, landing on a day when an agent delivery benchmark and a dense batch of embodied-AI papers quietly kept raising the bar on what robots and planners can actually do.
A security-heavy day across the frontier labs, from an insider arrest at OpenAI and quantified self-replication research at Anthropic to an in-house replacement for git, while embodied AI delivered the first driving VLA that reasons about when not to follow instructions.
Google took Gemini 4 Argon generally available with a launch price that undercuts its rivals and a hard expiration date, while embodied AI turned in its strongest single-day arXiv batch of the month, from a world foundation model that distills into a real-time policy to a benchmark quantifying the lab-to-lab sim-to-real gap.
The field is converging on privileged teachers for driving policies while the IP fight over that same technique goes public, and the rest of the day brought thousand-agent coding runs, speculative next-message prediction, a self-replication warning for cheap IoT, third-generation inference silicon, and fresh evidence of the gap between models and human intuition.
Monday's briefing centers on four independent groups converging on training agent harnesses instead of model weights, Meta's warning that RL post-training shrinks solution coverage, a GRPO post-training recipe for driving world models, and Waymo's $5 billion debt debut.
Beyond that headline disclosure, Sunday's briefing covers OpenAI's public fight over three fired safety researchers, Waymo's $5 billion debt debut and driverless debut in Detroit, Nvidia's reported talks to deepen its bet on open-weight lab ReflectionAI, and a Meta-affiliated negative result showing that scaling ego video barely moves object-interaction fidelity.
OpenAI turns the model-as-judge pattern into an API while Anthropic hands its models to the red team, and the embodied-AI paper batch keeps compounding.
OpenAI pushes ChatGPT past text into interactive in-chat apps while Anthropic and Google answer with cheaper inference and an agentic Workspace, and the research front keeps moving on streaming 3D reconstruction and world-action models.
Anthropic cuts Haiku pricing by three quarters while OpenAI publishes hundreds of frontier-model math manuscripts with Lean formalizations, and Waymo and Volvo-Waabi each climb another rung of the deployment ladder.
OpenAI's math release and a sworn AI hearing made the case from opposite directions.
Reflection's Beam made the efficiency case for open weights in a dense world-model week.
VLAs kept moving toward real hardware while Tesla's fleet kept growing.
Three labs shipped frontier models at one shared price.