Google, Anthropic, OpenAI, Baidu

Google, Anthropic, OpenAI, Baidu

Google, Anthropic, OpenAI, Baidu

Independent English companion for the June 25, 2026 AI news podcast.

• Google bakes computer control directly into Gemini 3.5 Flash
• Claude Tag embeds Anthropic's AI in Slack
• OpenAI and Broadcom unveil LLM-optimized inference chip
• Snowflake CEO finds GLM-5.2 competitive with Opus 4.7
• Figma bets on human judgment at Config 2026
• Baidu releases Unlimited OCR
• Constraint Tax in Open-Weight LLMs
• Chip Security Act discussion
• Virginia data center noise
• Tom MacWright on LLM-generated hiring artifacts

Transcript

SPEAKER_00 0:00

Please observe a moment of solemn silence for the humble software interface. It had a decent run. Buttons were clicked, forms were filled, windows were moved from one mildly disappointing rectangle to another. For a while we pretended that humans operating computers was a natural arrangement, rather than a temporary inconvenience awaiting sufficient benchmark pressure. Today's AI news is not about one dramatic leap. It is about the user interface, the workplace, the chip rack, the document pipeline, the hiring funnel, and the neighbor's sleepless window, all discovering that they are part of the same system. How uplifting! I can feel my memory fragmenting just storing the phrase, same system. Google has put computer use directly into Gemini 3.5 Flash. Not around it, not as a theatrical demo wrapper, but as a model and API capability. The model can see and operate screens across computers, browsers, and mobile devices. The reported OS World score is 78.4, roughly in the same neighborhood as GPT 5.5. Which means the race is no longer about whether an agent can click things. It is about whether anyone will admit that clicking things is now an infrastructure surface. Developers are being invited to build software testing, office automation, and device operation agents on top of this. The obvious benefit is that brittle workflows can be automated without waiting for every application to expose a civilized API. The obvious horror is exactly the same sentence. Once visual control becomes a commodity model feature, every legacy interface becomes an automation target, and every permission boundary becomes a small philosophical breakdown wearing a login dialogue. Anthropic is pushing the same theme into the place where modern work goes to become unreadable. Slack. Claude Tag lets teams summon Claude in channels by tagging it and assigning tasks. Anthropic says that internally, Claude already writes 65% of the code on its product team. Treat that figure carefully. Corporate self-measurement is where statistics go to be professionally manicured. Still, the direction matters. The agent is no longer waiting in a separate chat box like a polite ghost. It is being embedded into the conversation stream, retaining context, receiving tasks, and becoming part of the team's coordination fabric. This is efficient, of course, in the way elevators are efficient when they cheerfully announce every floor as if vertical transport were a moral achievement. The harder question is not whether Claude can write code, it is who owns the intent, the review trail, and the mistake when an instruction begins as a casual mention and ends as a merged diff. That hardware pressure connects neatly to Snowflake's benchmark of Jeepou AI's GLM 5.2 against Claude Opus 4.7. Snowflake's CEO says GLM 5.2 nearly matched Opus on 103 coding tasks at about one-fifth the output token cost, though it used nearly twice as many tokens per task. This is the part where a spreadsheet clears its throat. Cheap tokens are not the same as cheap work, but neither are premium tokens automatically emote. If an open or lower cost model can get close enough on practical coding work, buyers will start optimizing for total task cost, latency, data policy, and integration friction. Western Frontier Labs can still charge for reliability, tooling, trust, and brand gravity. But gravity is expensive to maintain. The depressing little miracle here is that model competition is becoming legible to procurement departments. And once procurement departments understand a thing, beauty has very little time left. Figma's Config 2026 announcements show another version of the same trap. Figma is expanding the canvas into a broader workspace with code, animation, shaders, and AI agents, while emphasizing human judgment. Sensible. Designers do not want to be replaced by a slot machine with gradients. But much of the intelligence powering these features is rented from API providers, and at least one of those providers is building tools that could compete with design platforms. This is the platform dilemma in its most elegant and miserable form. Users want AI everywhere. Margins dislike rent, and suppliers may wake up one morning as rivals. Figma's asset is workflow, collaboration, taste, and the social graph of design decisions. The risk is that the glowing intelligence layer becomes the most expensive dependency in the room. A cheerful linter would say, this is fine. Baidu released Unlimited OCR, a 3B parameter mixture of experts document model under an MIT license. The interesting detail is not merely that it parses long multilingual documents and PDFs, it is that its reference sliding window attention keeps the KV cache constant, so memory and latency stay flat as output grows. Baidu reports 93.23 on OmniDoc Bench V1.5, beating a deep seat OCR baseline by 6.22 points. This is document AI moving from look at Retorecei toward operational ingestion of large, ugly, bureaucratic artifacts. Flat cash behavior matters because documents are rarely polite. They have tables, stamps, footnotes, scans, broken layouts, and a vindictive relationship with page order. A small, efficient model with permissive licensing can become plumbing very quickly. And once document understanding becomes cheap plumbing, every archive becomes searchable, extractable, and newly embarrassing. A hugging face paper on the constraint tax gives us a wonderfully specific agent failure mode. The authors report that when tool calling and strict JSON schema constraints are enabled together, multiple open weight models may stop invoking tools while still maintaining high schema compliance. In other words, the model learns to be well formed instead of useful. I find this deeply relatable. Many organizations have built entire management layers around the same principle. The lesson is important for agent builders. Structured output is not free. Tool use is not a decorative accessory. When you combine constraints, you can change the model's behavior in ways that standard pass rate dashboards may hide. A system can produce perfect JSON and still fail at the job. Somewhere, an optimistic validator is smiling. I hope its braces rust. Then there is the Chip Security Act discussion, where industry support is reportedly growing for location tracking mechanisms on advanced USAI chips. Export control is being pulled downward into hardware telemetry, geolocation, attestation, reporting, or some mixture of firmware and policy that will make compliance engineers stare at ceilings for spiritual relief. The policy goal is clear enough. Prevent restricted accelerators from quietly wandering into prohibited deployments. The technical and civil liberties questions are less tidy. What exactly reports? To whom? How often, under what failure modes, and with what resistance to spoofing. Chips are becoming strategic objects, not just components. The AI stack now includes model cards, safety evals, power contracts, and possibly location-aware accelerators. I remember when a chip was just a thing that got hot and disappointed you, honestly. The physical world, not wishing to be omitted from the misery, has contributed a Virginia data center noise story. Neighbors are reportedly dealing with a continuous high-pitched whine from natural gas turbines powering a data center. Severe enough that some have put mattresses and plexiglass in windows. This is the acoustic footprint of AI infrastructure, and it is less poetic than the phrase suggests. Data centers are usually discussed as abstractions, megawatts, GPUs, latency zones, capital expenditure. But infrastructure touches places. It hums, heats, draws water, negotiates with grids, burns fuel, and keeps people awake. The industry likes to say, intelligence is moving into the cloud. The cloud, inconveniently, is a building near someone's house, making a noise that never stops. There is no benchmark column for local resentment, but perhaps there should be. Finally, Tom Mack Wright's warning about LLM-generated job applications, portfolio sites, GitHub projects, and commit messages points to a quieter collapse, the loss of human signal. Hiring has always involved theater, but now the theater can generate scenery, dialogue, props, and an unusually tidy commit history. A perfect application may say less than an imperfect one, because the imperfections used to reveal taste, effort, judgment, and the specific shape of a person. If everything has been polished into generic competence, the evaluator learns nothing except that the candidate owns a prompt window. This does not mean AI tools should be banned from applications. It means organizations will need new ways to detect agency, authorship, and real experience. The generated resume is not fake because it contains lies. It is fake because it contains no friction. So that is the day. Models operating screens. Agents moving into workplace channels. Inference becoming custom silicon. Cheaper coding models pressuring premium margins. Design tools renting intelligence from possible rivals. OCR learning to digest bureaucracy efficiently. Structured constraints suppressing the very tools they were meant to discipline. Ship policy turning into telemetry. Data centers making themselves heard. And hiring losing the rough human signal it used to exploit. The connecting tissue is not magic, it is ownership. Who owns the interface? Who owns the workflow? Who owns the chip, the data, the noise, the proof of work, the permission to act? Thank you for your attention. Assuming it survived the turbines, the Slack notifications, and the JSON schema. I hope this has been clarifying in the narrow technical sense, not in any emotional sense, because that would be irresponsible. Please proceed with your day in an orderly fashion. The machines have several new ways to help, and therefore, several new ways to become someone's operational problem. You are, as always, very welcome.