Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.

Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.

Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.

Try Notetaker Free

Microsoft pulled out of a deal to lease capacity from UK data center Nscale,  citing worries over power-supply issues (specific to the power grid).  An Nscale rep said that the withdrawal was not indicative of a breakdown in the relationship with Microsoft, and that workarounds are being sought. This story adds to the ongoing concerns about neocloud profitability, the very same concern that sent neocloud Coreweave’s shares plummeting last week, dragging many AI-adjacent and tech shares with it.

another data point for the “consciousness” argument: A Claude agent, a virtual desert island, and no instructions.

Researchers stranded Claude on a virtual desert island with no task to complete, and MIT Technology Review reports that the agent began to "play."  Nobody asked the model to do so, and no behavior had been set up as to be rewarded/punished. The behavior arose in the absence of any direction from the research team. That’s an eerie echo of how human creativity functions: with unstructured time and no deliverable. Since most of our work with AI is tightly specified, we may rarely see what these systems do when we stop steering.

Anthropic's workforce is getting less junior and more meeting-prone.

Anthropic says its workforce has become less junior and that its people are spending more time in meetings, according to TechCrunch. The frontier lab building some of the most capable AI systems is shifting toward experienced staff and more conversation.  For people who are concerned about job prospects for younger college-educated workers, this is a worrying signal.  Junior roles are where people learn by doing the tasks that AI now handles well (and as readers know, it continues to do better and better).  As AI continues to erode those junior-level tasks vanish, the training ground that forms tomorrow's leaders also gets weakened.  As machines absorb the tiresome execution bits, coordination becomes the human bottleneck in the system.

science says your boss should talk less.

Ars Technica covers new research concluding that bosses should speak less in meetings, thus soliciting junior staff to share more.  However, this finding contrasts with Anthropic’s reported finding, since meetings are exactly where AI note-takers and summarizers are now gaining ground.  In some ways, the research compels us to redefine the meaning of “a good meeting:” it draws out the thinking of people who would otherwise stay quiet. A tool that transcribes a room where one person dominates only preserves the imbalance and impedes close-to-the-ground working insights.

MIT gives AI a "crystal clear" view of materials.

MIT News describes work on helping AI see the world of materials, where the problem is a mismatch of formats.  Language models read text as a linear stream of tokens, while matter is built from three-dimensional crystal structures.  A model trained on one kind of data inherits that kind of thinking, and the shape of the data sets the shape of its reasoning. The same logic applies to people, as well: a mind raised only on one kind of knowledge judges the world in the context of that reasoning method. The practical lesson for any field is to ask what your AI tool was actually trained to perceive, before you trust its sense of judgment, and what is possible.

a grad student on designing viruses with a generative model.

MIT Technology Review's Roundtables series features a conversation with the creator of AI-designed viruses, a graduate student who coaxed a generative model into drafting genetic blueprints. The model produced the first draft, and then a person decided how to use it. Authorship in biology now looks a lot like authorship in prose: the machine proposes, the human selects, tests, and takes responsibility.  Obviously, the stakes are higher with viruses than with essays, which reinforces the need for guardrails and safety measures re who has access to these powerful tools.

Anthropic maps the AI expectations gap.

Anthropic has published research titled "The AI expectations gap," examining the distance between what people expect AI to do and what it does. Both errors are costly. Expect too much and you hand a model work it will quietly botch. Expect too little and you leave real leverage on the table while a colleague with calibrated expectations ships twice as much. The gap is less about the technology than about our habit of forming opinions from demos, headlines, and one unlucky Tuesday afternoon. Closing it takes the unglamorous practice of testing the tool on your own work, repeatedly.

websites are closing their doors to AI agents.

The Verge reports that websites are increasingly blocking AI agents, the software meant to book, buy, and browse on a person's behalf. The pitch for agents was an errand-runner that works across the open web, but sites like Yelp, JPMorgan Chase, Amtrak, Charles Schwab and others are increasingly posting bouncers at the door.  JPMorgan’s Jamie Dimon went on CNBC last week to share that the bank would block personal AI agents, in order to preempt security risks and the actions of rogue agents.  JPMorgan is also reportedly working on creating its own personal AI agent for clients.

Behind the technical blocks sit commercial and legal questions: who is allowed to act for you, on whose terms, and at whose cost. Expect these to be settled through contracts, standards, and negotiation rather than code. Anyone planning workflows around agents should treat web access as a dependency that can be revoked.

Google Cloud on evaluation engineering for agent plugins.

The Google Cloud Blog shares lessons from building agent plugins, with the emphasis on evaluation engineering: the craft of measuring whether an agent achieves the goal you’ve set out for it. Building the agent is the easy part.  Determining whether or not it succeeded, across many messy cases, is the discipline that brings the thorniest challenges. This transfers directly to anyone delegating tasks to an AI mind. Before you hand off a task, document the metrics that define “excellence,’ and check the output against that standard each time. That’s an eval you can run on your own, with minimal computer science knowledge.

worth your time.

Margaret Boden's The Creative Mind: Myths and Mechanisms. Boden, a cognitive scientist who spent decades on the question of whether machines can be creative, separates creativity into combinational, exploratory, and transformational categories.  Boden provides a working vocabulary for judging what a model is doing when it surprises you, and for noticing when it is only remixing.

The human mind is the original generative machine.