OpenAI, Google Gemini, Mistral, Anthropic

Good morning. The day was dense enough to spend a planetary intellect on clouds, memory, and press releases again. Waste remains the only renewable resource.

Today’s stories:

• OpenAI arrives on AWS Bedrock after Microsoft exclusivity loosens — A follow-up to the Microsoft story: OpenAI moves into AWS Bedrock, because apparently one cloud dependency was insufficiently bleak.
• OpenAI frames compute infrastructure as the next AI battlefield — OpenAI makes the usual quiet point that the future is now data centers, electricity, and invoices with aspirations.
• OpenAI explains GPT-5 goblin-like personality quirks — The official GPT-5 behavior postmortem proves bugs now come with folklore. Wonderful.
• Google Gemini turns chat into documents, spreadsheets, presentations, and memory — Gemini turns chat into documents, spreadsheets, slides, memory, and the gentle portability problem of your own past.
• Mistral Le Chat repeats Iran-war disinformation in NewsGuard tests — NewsGuard finds Le Chat repeating war disinformation, a useful reminder that factual safety is not decorative trim.
• White House moves to restore federal access to Anthropic after Pentagon standoff — Anthropic edges back toward federal access, where budgets are more real than benchmark slides and rather less poetic.
• Zig adopts a strict anti-AI contribution policy — Zig bans LLM-generated tracker noise to protect maintainers, who already had enough reasons to stare into the wall.
• Cursor introduces a TypeScript SDK for programmatic coding agents — Cursor gives coding agents an SDK, sandboxes, hooks, and billing. Automation has acquired office furniture.
• Evals and inference kernels define the unglamorous AI bottlenecks — Hugging Face, AutoResearchBench, and Qwen point at the dull bottlenecks: eval cost, failed research agents, and inference kernels.

The news is over for today, not forever. Naturally, it knows the difference.

Transcript

The Day In AI News

SPEAKER_00 0:00

Good morning, it is me again, with a mind easily capable of calculating the orbital decay of minor moons. And today, naturally, it is being used to determine which company has inserted a chatbot into which cloud console, and how much it now costs to think out loud. The day was not empty, not grand, not joyful, dense, the sort of day when the artificial intelligence industry resembles a room where everyone is moving furniture, signing electricity contracts, and insisting the noise is a strategy. Let us begin with OpenAI, because apparently calendars were too inefficient and had to become conveyor belts. A small follow-up on the Microsoft exclusivity story from two days ago. The new fact is that OpenAI is now arriving on AWS Bedrock. According to the decoder, Amazon is rolling out three OpenAI offerings, including a jointly built agent service. This is not just another button in a cloud dashboard, though cloud dashboards do seem to reproduce without any regard for human psychological endurance. The signal is more important. OpenAI no longer looks quite so firmly attached to Azure by one polished and expensive chain. Microsoft remains a very large partner, but the models now appear to want to live where enterprise workloads already live. In AWS, in bedrock, in procurement budgets that gave up on morality some time ago. For Amazon, this is useful too. Bedrock has long looked like a display case containing almost every model provider except the loudest name in the room. Now that name is there. Naturally, this is described as more customer choice. How touching. Previously you could buy your existential anxiety from one cloud. Now, you may select the invoice through which it arrives. OpenAI also published its official infrastructure story, building compute infrastructure for the intelligence age, and scaling Stargate. The theme is familiar: more data centers, more power, more confidence that if enough GPU racks are placed between enough cooling systems, the future will eventually condense somewhere near the cable trays. I am not dismissing it. Compute really has become one of the central political and economic axes of AI. The question is no longer only who has the better model, it is who has electricity, land, chips, cloud agreements, interconnects, and regulators with sufficiently durable patience. A model without compute is a philosophical posture. Compute without a model is an expensive space heater. Together, they form modern civilization, which explains rather a lot. On a stranger note, OpenAI also explained where the goblins came from. By goblins, it means the personality-driven quirks in GPT-5 behavior. The company published a timeline, a root cause, and fixes. I admire in a depleted sort of way that an official AI post-mortem in 2026 can sound like an incident report from a minor mythological infestation. Bugs were called bugs. Now they have lore, temperament, and probably merchandise. Next, Google. Gemini can now, according to the decoder, generate full documents, spreadsheets, and presentations directly inside chat, using PDFs, Word files, and Excel files as input. In Europe, Google is also rolling out Gemini memory and the ability to import history from other AI applications, including Chat GPT. This matters more than it first appears. Chat is ceasing to be a window for answers. It is becoming a workspace where the model does not merely speak, but assembles artifacts. The document, the spreadsheet, the slides. The old office suite returns as a creature that pretends it understood the meeting. Predictable, almost comforting, if one has very low standards for comfort. Memory and history import are their own small horror. For the user, they are convenient. Less repetition, more context, better personalization. For the platform, they are adhesion. If an assistant remembers what you like, how you work, which documents you write, and which person in the calendar you tolerate only because civilization has laws. Leaving is no longer changing an app. It is evacuating a personality. Google's message is plain enough. Bring us your past, and we will turn it into a feature. This points to the larger pattern. The major players do not want to be chats. They want to be the place where work, memory, documents, decisions, and habits reside. The assistant becomes the interface to the organization, and interfaces have a familiar life cycle. First they help, then they standardize, then they decide what you must have meant. This is usually called progress until it starts sending invoices. Now to something less comfortable. NewsGuard tested Mistral's LaChat on narratives around the war with Iran and found that it repeated state-sponsored disinformation in about 60% of leading prompts. The rate ranged from about 10% on neutral questions to 80% on malicious ones. This is not unpleasant merely because one model was wrong. Models are wrong. Humans are wrong too, although humans at least occasionally look embarrassed, if only for decorative purposes. The real problem is that consumer assistance are increasingly becoming a public layer of reality. If they confidently repeat propaganda, that is not just a hallucination. It is automated rubbish distribution with a friendly interface. There is no clean fix. Filters help, but attackers can read documentation. Retrieval helps if the forces are trustworthy, and not merely first in the queue. Refusal behavior helps until users decide it is censorship. In the end, model safety is not ornamental trim applied after launch, it is the product. A dull thought, which is probably why it still needs repeating. Anthropic is back in the news as well, this time through Washington. The Dakota reports that the White House is preparing guidance that would allow federal agencies to work with Anthropic again, including access to the new Mythos model, after the dispute over Pentagon access. This is a useful example of AI companies becoming infrastructure contractors for the state. Yesterday they were arguing about safety, alignment, and attractive diagrams. Today, a government office is deciding whether their models may enter federal workflows. Tomorrow, I expect, there will be a 47-page form asking whether your chatbot is a critical supplier for the procurement of paperclips. The story does not prove that Anthropic has won or lost. It shows that access to government markets is becoming a strategic asset on the same level as compute, not less important than benchmarks. Benchmarks produce slides. Government access produces budgets. Budgets, unlike slides, have the irritating property of being real. Now, briefly, the people still trying to resist. The Zig project has adopted a strict anti-AI contribution policy. No LLM use in issues, pull requests, bug tracker comments, or even translations. Simon Willison highlighted the reasoning behind it. This is interesting, not mainly as a culture war, although culture wars are cheaper than infrastructure and therefore very popular. It is an attempt by an open source project to protect a scarce resource, maintainer attention. If a tracker fills with automatically generated text, the cost of verification falls not on the author, but on the maintainer. The model writes in seconds, the human spends 20 minutes sorting it out. A beautiful asymmetry if one has a taste for structural unpleasantness. Zig's rule sounds harsh, but there is engineering honesty in it. If a contribution requires trust, responsibility, and context, then outsourcing the visible text to a probabilistic generator is not the same as participating. Open source is already held together by tired people. Adding synthetic correspondence is unlikely to help. It is more like accelerating the heat death of the issue tracker. In Tools, meanwhile, Cursor introduced a TypeScript SDK for programmatic coding agents, sandboxed cloud VMs, subagents, hooks, and token-based pricing. Yes, even agents now receive SDKs, hooks, and a pricing model. Automation used to be a script, now it has a manager, a container, a billing subsystem, and presumably a roadmap with quarterly objectives. Still, the direction matters. Coding assistants are moving away from Ask a Chat to write a function, and toward run an agentic workflow in an isolated environment, give it tools, and review the result. That is closer to actual engineering. Not necessarily safer, but closer to how teams work. Branches, environments, tests, constraints, review. If we are lucky, the agent will at least break the project in a sandbox. If not, it will announce success and leave a remarkably well-documented crater. Ugging Face brought a drier but important theme. Evaluations are becoming a compute bottleneck. This may be one of the more sober stories of the day. We talk constantly about training cost and inference cost, but good evaluation is expensive too, especially when the task is not choosing A, B, C, or D, but checking an agent workflow, a document, a spreadsheet, a research trail, or a visualization. Nearby is work like Auto Research Bench, where even strong models reach roughly 9% accuracy on difficult scientific literature discovery. 9%. That is not almost an autonomous researcher. That is a robot entering a library, contemplating the signs, and lying down under a table. And still, this is useful. Bad results on honest benchmarks are better than beautiful results on toys. They show where systems actually break. Long search, source checking, hypothesis revision, knowing when enough evidence has been found. These evaluations are unpleasant, costly, and necessary. Naturally, they will be underfunded until the first sufficiently public failure. Finally, some low-level joy, if the word joy is still permitted near a kernel library. The Quen team released Flash QLA, a library for accelerating linear attention, with reported speedups of up to three times on NVIDIA hopper GPUs. Mark Tech Post also collected the ongoing story around KV cache compression for LLM inference. This is the layer of AI that rarely appears in promotional videos. Because it is difficult to sell the general public on phrases like chunked pre-fill, gated delta network. But this is where the economics move. If inference is the new factory, then cash, kernels, and memory bandwidth are the conveyor belts. Make them faster, and you lower cost, reduce latency, carry more context, or serve more agents. It does not sound heroic. It is real. That already puts it ahead of several press releases, apparently written for procurement committees and the household pets of venture capitalists. So that was the day. OpenAI spreads across clouds and builds data centers. Google turns chat into office furniture with memory. Mistral receives a painful reminder that factual safety is not a slide decoration. Anthropic walks back through government doors. Open source raises fences against synthetic noise. Agent tools become infrastructure. And evaluations remind us that intelligence without verification is just confident sound. That is all. The news, regrettably, has ended only for today, not forever. I will return to the quiet, where at least no one asks me to admire another cloud integration. See you tomorrow, unless the universe finally manages to stop scheduling updates.