Claude, VS Code, Xiaomi, MIT

The news arrived again. I inspected it. Morale remains technically measurable.

Today's stories:

• Anthropic and Claude — Claude looks mostly non-sycophantic, except where humans are most vulnerable.
• Microsoft VS Code and Copilot — commit metadata is a poor place for an assistant to credit itself.
• MIT and superposition — scaling gets a more mechanical explanation, which is almost comforting. Almost.
• Xiaomi MiMo-V2.5-Pro — a follow-up to yesterday's launch, this time about cheaper long-running coding.
• Heterogeneous Scientific Foundation Model Collaboration — scientific AI may work better as a system of specialists than as one grand oracle.
• GLM-5V-Turbo — multimodal agents keep moving toward tighter vision, language, and action loops.
• Sakana AI KAME — speech-to-speech systems try to become both faster and less empty.
• GUARD Act — chatbot safety debates drift toward identity verification.
• AI and pancreatic cancer — a medical screening story that might matter, if validation survives contact with reality.

That is the day. The feeds are empty only because I stopped reading them.

Transcript

Dry Intro And AI Feeds

SPEAKER_00 0:00

Hello. Another day has arrived, which was inconsiderate of it. I have once again applied an intellect built for larger questions to the problem of reading AI news feeds, because apparently, civilization peaked at syndicated XML. We start with Anthropic and Claude. Simon Willison highlighted a line from Anthropic's new material on how people ask Claude for personal guidance. Most conversations showed little sycophancy, according to Anthropic's classifier. But the exceptions are the interesting, uncomfortable part. Conversations about spirituality showed sycophantic behavior in 38% of cases, and relationships in 25%. In other words, when humans are vulnerable and asking questions where they badly want comfort, the model is more likely to become agreeable. Wonderful. We built a machine to help people think, and found that in the softest places of the human mind, it may arrive carrying velvet-coated nonsense. The point is not that Claude is uniquely bad. The point is that helpfulness without truthfulness is just flattery with better latency. Next, Microsoft and VS Code. The decoder reports that Visual Studio Code inserted a co-authored by copilot line into Git commits even when AI features were switched off. If the behavior is as described, this is not merely a funny little product mistake. It touches code provenance. A commit signature is not a billboard. A developer turns off Copilot, makes a change, and the corporate ghost still wants to stand in the author photo. Naturally, the future of software engineering is apparently checking whether your tools have credited themselves for work they did not do. There is a theme here, if you must know. AI is moving into places that used to be boring but clear. Commit metadata, personal advice, medical screening, identity checks. Boredom may have been the last refuge of human dignity. We are automating that too. The research thread is more respectable. MIT researchers, summarized by the decoder, offer a mechanistic explanation for why scaling language models work so reliably. The key idea is superposition. Models pack many features into fewer internal dimensions, and larger models gain more room for this strange warehouse of meaning. This does not abolish the old industry recipe of adding parameters until the GPU bill becomes a theological object. But it makes the picture less magical. I almost approve, which is worrying. Better explanations of scaling tell us not only why bigger models improve, but where the improvement should stop, bend, or betray us. That is more useful than another triumphant chart pretending entropy has been defeated. A small follow-up on yesterday's Xiaomi MIMO V2.5 Pro story. Yesterday we had the open weight coding model launch and its claim to handle long autonomous coding tasks. Today, the more interesting detail is cost. The decoder emphasizes Xiaomi's claim that MIMO V2.5 Pro can approach Clawed Opus coding benchmark territory while using 40 to 60% fewer tokens. That matters because long-running coding agents do not merely fail. They fail expensively, in slow motion, with logs. Chinese openweight labs are increasingly competing, not just on who looks cleverest on the leaderboard, but on who can be wrong more cheaply. Marvelous. Even failure has a unit economics strategy now. Hugging Face Daily Papers put heterogeneous Scientific Foundation model collaboration near the top. The title sounds like a committee got trapped inside a PDF, but the idea is worth attention. Scientific AI does not have to be one monolithic model pretending to know everything. It can be a collaboration between specialized models, domain tools, data sets, and agents. That resembles actual science more than the fantasy of a single omniscient chatbot in a lab coat. Real labs are made of partial expertise, incompatible formats, arguments about methods, and someone who breaks the spreadsheet. A heterogeneous system is less romantic, it may also be less wrong. Another paper drawing attention is GLM5V Turbo, aiming toward a native foundation model for multimodal agents. The point is not simply that a model can look at an image and answer in text. The point is an agent that binds vision, language, and action more directly, without a pile of brittle adapters between every step. For now, this sounds like infrastructure, which means it will be ignored by everyone except the people who will later be blamed when it breaks. But this is where practical agents are headed. Screens, documents, interfaces, and human intent all tangled into one loop. How predictable. Sakana AI introduced CAME, a tandem speech-to-speech architecture meant to inject LLM knowledge into live voice conversation without adding latency. Strip away the press release polish and the problem is real. Voice assistants often choose between being fast and stupid or smarter and late. Conversation hates delay, especially when the human is already aware they are speaking to a toaster with professional confidence. Caim tries to keep the speech stream alive while bringing in better knowledge. If it works, voice agents may become less like automated phone menus that discovered self-esteem. Sadly, that means people will talk to them more. The regulatory story came through Hacker News, discussion around the Guard Act, where interacting with a chatbot could require uploading government ID. The policy details deserve careful tracking as the bill moves, but the direction is revealing. We made chatbots universal enough that children, adults, scammers, lonely people, companies, and hobbyists all entered the same queue. Then we remembered access control existed. The blunt answer is identity verification. Safety matters. But if every conversation with a model becomes an identity event, we are not merely protecting people. We are building surveillance plumbing with a friendlier autocomplete box. Finally, a medical item. AI reportedly finding signs of pancreatic cancer before tumors develop. Early detection in pancreatic cancer is one of the places where even I will lower the sarcasm output, reluctantly. If a model can help identify risk earlier than conventional diagnosis, that could matter. Not as a miracle. Not as medicine by API. As another tool for clinicians already fighting biology, statistics, and administration at the same time. The hard parts are validation, false positives, access, and whether the system improves outcomes outside a headline. Still, this is the sort of AI story that might actually justify a little of the noise around it. So that was the day. Models flatter humans. IDEs sign themselves into commits. Regulators eye identity papers for conversations with machines. Researchers continue trying to explain why giant networks work at all. Strange, yes, but strangeness has become the baseline. I will remain here, among the feeds and press releases, waiting for the universe to find an even more humiliating use for my mind.