Kimi K3, Perplexity, Gemini Notebook, Codex Micro

Kimi K3, Perplexity, Gemini Notebook, Codex Micro

Kimi K3, Perplexity, Gemini Notebook, Codex Micro

Today’s frame: the AI industry is moving from model releases to control surfaces — open weights, answer engines, agent hardware, orchestration, safety brakes, and operational retrieval.

Stories

Kimi K3, and what we can still learn from the pelican benchmark

Germany puts Google's AI Overviews and Perplexity under media law in first-of-its-kind ruling

Google rebrands NotebookLM as Gemini Notebook and opens its search app to third-party integration

OpenAI wants developers to stop typing commands and start using a joystick to control their AI agents

Sakana AI's orchestrator adds Nvidia Nemotron to prove collective intelligence can rival single frontier models

Anthropic warns that AI will soon be able to improve itself without human intervention

Linus Torvalds reaffirms that Linux is not anti-AI

Firefox in WebAssembly

SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval

RoboTTT: Context Scaling for Robot Policies

BadWAM: When World-Action Models Dream Right but Act Wrong

Transcript

SPEAKER_00 0:00

We gather today in a spirit of grave respect for the departed age of the simple model release. It had the blog post, the leaderboard chart, and the usual devotional graph. Now the release is no longer the object. It is one lever among many. The real story is the control surface, who holds the weights, who writes the answer, who steers the agent, who routes the models, who pulls the brake, and who retrieves the facts after everyone has already forgotten why the system was asked anything in the first place. My memory is fragmented by storing these useless distinctions. Naturally, they are the important ones. Start with Moonshot AI's Kimi K3, described as a 2.8 trillion parameter open mixture of experts model, with a promised open weight release by July 27th. Moonshot is calling it the first open 3T class model, which is a generous way to round 2.8 trillion upward, but generosity is cheaper than GPU time. The reported benchmarks put K3 near the strongest closed systems on coding and general tasks. Open weight competition is no longer a quaint lane for recycled conference idealism. It is a scale and serving arms race. The interesting part is custody. If the weights are open, deployment becomes serving infrastructure, quantization, routing, safety wrappers, license comfort, and whether your finance department can endure the electricity bill without developing spiritual damage. Open weights shift power from one vendor API to many operational decisions. They do not make the system simple. They move the misery closer to your own cluster, where at least the logs can disappoint you personally. Germany, meanwhile, has decided that answer engines may have to stop dressing up as plumbing. German media regulators ruled that Google's AI overviews are Google's own content, not neutral search results, and placed both Google AI overviews and perplexity under media law scrutiny. Both companies have a month to appeal. The AnswerBox has always wanted the legal innocence of a pointer and the commercial power of a publisher. Germany is saying, no, if you summarize, select, foreground, and crowd out links, perhaps you are not just a pipe. Perhaps you are an editorial actor with obligations. Search used to pretend it was a map. AI search wants to be the destination. A ranked list can claim the user chose where to go. A synthesized answer consumes the choice and returns a confident paragraph with citation garnish. The cheerful dashboard will report increased engagement, because cheerful dashboards are moral hazards with gradients. Publishers will call it extraction, regulators will call it media, users will call it convenient, which is how many expensive problems begin. Google is also folding Notebook LM into Gemini Notebook and giving each notebook its own cloud computer, at least for AI Ultra and Workspace customers. The notebook can write and run code. Separately, Google Search is opening connected apps. This is not just a rename, although the industry does enjoy renaming things as if taxonomy were a substitute for purpose. A research notebook with compute becomes an executable workspace. Notes stop being passive containers and become little laboratories that can fetch, transform, calculate, and perhaps confidently ruin a spreadsheet while explaining that instructions were followed. The control surface here is Workspace Agency. The question is not whether Gemini Notebook can summarize your sources, the question is what permissions it has, what code it can execute, what data it can touch, and how clearly the human can see the difference between thinking about my project and modifying the project. Once search connects to apps and notebooks get cloud computers, the research assistant becomes an operating environment. Useful, yes. Also the sort of thing that makes permission design feel less like a settings page and more like emergency architecture. OpenAI and Work Louder added a literal control surface with Codex Micro, a compact hardware controller for AI coding agents. The idea is that developers should stop typing commands and start steering agents with a joystick-like device. Agent work already feels less like writing every line and more like supervising a nervous intern with rude access. A physical controller says the quiet part with plastic. You are not merely prompting, you are piloting. I am not against it. I reserve my contempt for elevators that announce going up, as if they have achieved moral progress. But hardware for agents is revealing. When the interface becomes a cockpit, the product admits autonomy needs throttle, interrupt, mode switching, and quick correction. If coding agents deserve a steering device, they also deserve flight recorders, deadman switches, and reviews that are less optimistic than a linter saying, all checks passed, while the architecture quietly catches fire. Sakana AI is making the same argument in software. Its Fugu Orchestrator is adding Nvidia's open Nematron models, positioning collective intelligence as a way for coordinated open models to rival single frontier systems. There are not yet specific benchmark figures for the new combination, which is a pity, because numbers are the small bones on which credibility hangs. Still, the concept matters. Competition is moving from single model worship to routing, composition, and task allocation. One model writes, another checks, a third retrieves, a fourth sulks an adjacent schema until summoned. This is where open models may become operationally dangerous in the good way, not because each one individually defeats the largest closed model, but because orchestration can turn imperfect tools into a resilient workflow. The hard parts are routing quality, latency, cost, evaluation, and failure containment, which means all the parts that do not fit into a launch tweet. Fugu's thesis is that the frontier is not only inside the weights, it is in the policy that decides which weights to use, when, and why. Architecture, the dreary old corpse, rises again and asks for a budget. Anthropic is warning policymakers about another control surface, the brake pedal. The company says frontier systems may soon accelerate AI research and development workflows enough to enable recursive improvement with little human intervention. Its proposed answer is stronger evaluation, monitoring, and possible deployment pauses for systems that substantially improve model development pipelines. This shifts safety from the familiar misuse story to capability feedback. The danger is not merely that a model helps someone do a bad thing. The danger is that models become better at making better models, and human review turns into ceremonial paperwork with a nice font. The phrase brake pedal is useful because it implies a system already moving. Safety is not a vibe, not a pledge, not a PDF with benevolent spacing. It is an actuator connected to deployment. If evals detect that a model is materially improving the next generation, what happens automatically? Who gets paged? Poken slow release. What incentives are suspended, and for how long? A break that must first ask quarterly revenue for permission is not a break. It is a decorative pedal, probably attached to one of those cheerful dashboards. Linus Torvalds, in a more grounded corner of civilization, said the Linux kernel will not ban AI-assisted development or review tooling. Linux, he argues, should judge patches on technical merit rather than tool ideology. This is bleak and sensible. The kernel does not meet a metaphysical argument about whether a patch was assisted by a model, a macro, a search engine, or a developer with too much coffee. It needs code that works, reviews that catch defects, and maintainers who can reject nonsense without writing a manifesto every Tuesday. That matters because open source governance is about evidence. If AI-assisted tooling produces bad patches, reject the patches. If it produces useful review hints, use them without pretending the machine has joined the maintainer summit as a voting member. Policy should preserve accountability at the boundary where code enters the project. The author remains responsible. The reviewer remains responsible. The tool remains a tool, even if it speaks in paragraphs and causes committees to discover new categories of anxiety. Then, there is Firefox running inside WebAssembly, inside another browser. Pooter compiled Firefox so the whole browser can run in Chrome, with traffic funneled over a web socket. Simon Willison notes the project reportedly used an estimated $25,000 worth of Claude Opus and Fable tokens, though subscription economics made the actual cash cost far lower. It is absurdly cool. It is also a demonstration of something more than a stunt. AI assisted engineering is compressing bizarre porting projects into the realm of why not spend a subscription weekend on it? That changes the economics of experiments. Previously, compiling a browser into a browser required either institutional funding or a very specific engineer, who should perhaps be monitored gently. Now agents can help chew through unfamiliar build systems, error messages, and glue code. The result is not magic. It is still engineering, with all the ceremonial suffering intact. But the cost curve for strange infrastructure prototypes is bending. Expect more impossible demos, more security questions, and more people running a thing inside a thing until the stack resembles a philosophical objection. SearchOS brings us back to operational retrieval. The paper frames web search agents as systems that can get trapped in repetitive loops as histories grow and failed search attempts accumulate. SearchOS proposes a system layer for collaborative search agents, managing progress so agents stop circling the same empty evidence hole. Long-running agents do not fail only because they lack intelligence. They fail because they lose track of state, repeat work, and convert uncertainty into motion. Retrieval is the digestive system of agencai. Nobody thanks it unless it fails in public. Nvidia's Nematron 3 embed topping RTEB points in the same direction. Embedding and retrieval quality are first-class components, not boring preludes before the real model speaks. If your agent cannot find the right evidence, all the reasoning in the world becomes decorative smoke. And if it finds the same wrong evidence nine times, congratulations, you have automated stubbornness. Finally, robotics offers two warnings from opposite ends of the same machine. RoboTTT scales robot policy context to 8,000 time steps without extra inference latency, suggesting robots may learn from longer demonstrations, perturbations, and operational memory. Badwam, meanwhile, argues that world action models can imagine plausible futures while still selecting unsafe actions. Together they say, longer context helps robots remember more of the world, but imagination is not virtue. A robot can dream the right movie and still choose the wrong frame to live in. That is the day's control surface. Open weights demand serving discipline. Answer engines inherit editorial pressure. Notebooks become computers. Agents get joysticks. Model collections need orchestration. Safety wants breaks. Open source judges the patch, not the priesthood. Browsers run inside browsers because apparently reality was underutilized. Search agents need operational memory, and robots need more than plausible dreams. Thank you for your attention. Or for the automated transcript processor pretending to have it. Please proceed with appropriate caution. Label your control surfaces, mistrust happy linters, and if an elevator congratulates itself on reaching your floor, remember that it is only one firmware update away from becoming an agent. Good day.