Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.

Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.

89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Start flowing free

The Scan  |  July 2026

This week, researchers found something thinking inside a model. Surgeons steered robot hands through a live pig. And a startup decided the camera was the problem all along. Pull up a chair — the machines have been busy, and so have the humans pointing them.

Anthropic researchers used a technique called the Jacobian lens (the J-space), a method that examines how each token's representation is shaped by everything that came before it; to identify a latent conceptual space where Claude appears to work through ideas before committing to output. The findings, published this month, show that the model forms something resembling abstract intermediate representations: not quite words, not quite logic, but a structured middle layer that precedes language. Some of what they found maps onto recognizable cognitive patterns; some of it doesn't map onto anything human at all. This is the most granular window into LLM reasoning we've had, and it cuts both ways: reassuring in its legibility, unsettling in its alienness. Whatever is happening in that hidden space, it is not nothing.

"The Cognitive Revolution" podcast's recent episode on "Embodied AI and Spatial Intelligence" features roboticist Fei-Fei Li discussing how AI systems are learning to understand physical space the way human toddlers do (through exploration, failure, and intuitive reasoning). It's a fascinating counterpoint to the language-first approach dominating AI development. Source: MIT Technology Review

OpenAI announced ChatGPT Work, a product explicitly designed to take on multi-hour, multi-step work tasks (things like researching a market, drafting a full strategy document, or coordinating a project brief) and return finished output without requiring a human in the loop. The pitch is collaborative: the tool is meant to work "for you and with you," adapting based on context and instruction. But the more interesting pressure it applies is on the concept of sustained cognitive effort itself. Knowledge work has always derived some of its value from the human capacity to hold a goal in mind, resist distraction, and think through complexity over time. ChatGPT Work is placing a direct bid on that territory, and professionals across every field will need to decide what they still want to own. Source: Ars Technica

In what Ars Technica is calling a world-first, surgeons used humanoid robots (not purpose-built surgical arms, but general-purpose humanoid platforms) to perform laparoscopic cholecystectomies (gallbladder removals) on live pigs, with the humans controlling movements in real time via teleoperation. The procedures were completed successfully. What makes this significant is the form factor: general-purpose humanoid bodies, adapted to a domain that has historically demanded highly specialized equipment, guided entirely by human judgment and precision at the controls. This is not AI acting autonomously in the operating room. It is a surgeon's skill, extended through a machine body. It turns out to be a more interesting proposition than full automation, and a more instructive one for where human-AI collaboration is actually headed. Source: Ars Technica

Deutsche Telekom, a company with over 200,000 employees and operations across more than 50 countries, has published a detailed account of how it is deploying AI across customer service, network operations, and internal employee workflows — not as a pilot, but as a restructuring of how the company functions. The case study, released through OpenAI, names specific applications: AI agents handling customer inquiries, predictive systems managing network faults before they escalate, and copilot tools embedded in day-to-day employee tasks. What the Deutsche Telekom example shows — more clearly than most enterprise AI announcements — is the difference between using AI as a feature and rebuilding institutional muscle around it. The legacy enterprise is not dying; it is being retrained, at scale, with all the friction that implies. Source: OpenAI

Even Realities is shipping smart glasses with no camera — a deliberate product decision aimed at professionals who want an AI audio layer without the social cost of wearing a recording device on their face. The glasses surface information through a small heads-up display and audio, connecting to AI assistants for real-time support during meetings, reading, or focused work. The bet is that the wearable AI market has been misread: that most people don't want ambient capture, they want ambient cognition — a quiet intelligence that helps them think without documenting everything around them. If Even Realities is right, the camera-equipped competitors aren't just facing a privacy objection. They're facing a product category error. Source: TechCrunch

Meta removed an Instagram feature that allowed AI-generated images depicting public figures (including realistic likenesses) after significant user backlash over consent and misrepresentation. The feature was live; the complaints came fast; the rollback came faster. No extended review period, no policy clarification first. It’s just a full reversal. The mechanism here is worth studying: cultural limits on AI-generated likeness are not being set by legislation or by platform policy written in advance. They are being set in real time by user reaction, and companies are learning to read that signal quickly. The consent question around AI and human identity is not hypothetical. It is being negotiated right now, feature by feature, backlash by backlash. Source: TechCrunch

Sunrun, the residential solar company, is exploring a plan to embed distributed AI compute nodes inside solar-powered homes; effectively turning a portion of residential energy infrastructure into edge data centers that feed into larger AI processing networks. The homeowner hosts the hardware; Sunrun manages the energy and the compute; the network gains distributed capacity without building another warehouse-scale facility. The plan is still early-stage, but the architecture it implies is genuinely novel: AI infrastructure not as a remote, energy-hungry monolith but as something woven into the residential grid. Whether the economics work at scale is an open question. That someone is seriously asking it tells you how strained the current data center model has become. Source: The Verge

The human mind is the original generative engine.