In partnership with
In Todayโs Issue:
๐ฉบ ChatGPT opens your medical records to every US user
๐ Oracle cuts 21,000 jobs as its AI bet bites
๐ค Washington moves to keep Chinese robots out of the military
โ๏ธ Europe's AI law starts fining before it's finished
โจ And more AI goodnessโฆ
โก The Signal
Two of the biggest AI labs spent the same day teaching their assistants to listen.
Dear Readers,
On July 23, OpenAI and Anthropic both shipped voice upgrades within hours of each other, and the timing matters. OpenAI pushed its full-duplex GPT-Live model into the ChatGPT desktop app and Codex, so engineers can steer coding jobs without touching the keyboard. Anthropic moved Claude's voice mode onto its stronger models and let it reach the tools you have already connected, from Gmail to Notion. Voice is turning from a party trick into the control layer for real work, which sharpens an old question: once talking is the fastest way to run software, who is really steering?
All the best,
Kim Isenberg
Whoโs actually reading this?
Weโre planning next yearโs coverage and building it around you.
Three taps, no typing.
Why it matters: we need to know who our readers are; itโs the first question any serious sponsor asks and we canโt answer it right now. It also gives us a real click in every issue, which helps our numbers.
(ChatGPT app / Getty Images via TechCrunch)
๐ฉบ ChatGPT Wants to Read Your Medical Records
OpenAI just handed ChatGPT the keys to your medical file. Health in ChatGPT is now live for every logged-in US user over 18, on the free, Go, Plus, and Pro tiers. It links Apple Health and records from providers like Epic and Oracle Health, then answers questions about labs, medications, and sleep in plain language. OpenAI says it already fields 300 million health questions a week, and warns the tool is "not intended for use in the diagnosis or treatment of any health condition."
๐ tl;dr: Real convenience meets your most sensitive data, and the privacy trade-off is now everyone's.
(Oracle chairman Larry Ellison / The Jerusalem Post)
๐ Oracle's AI Bet Costs 21,000 Jobs
Oracle is cutting about 21,000 jobs, roughly 13% of staff, even as it goes all-in on AI. Headcount falls from around 162,000 to 141,000 while the firm funds a reported $300 billion compute deal with OpenAI. A credit downgrade from BBB to BBB- has triggered fresh collateral demands, including a $7 billion guarantee sought by Wisconsin regulators for a data center.
๐ tl;dr: The AI boom's backbone is brutally expensive, and Oracle is trimming people to pay for machines.
(An AgiBot humanoid robot / South China Morning Post)
๐ค The US Moves to Keep Chinese Robots Out of the Military
Congress wants Chinese humanoid robots nowhere near the US military. A provision in the House-passed NDAA would bar the Pentagon from buying, leasing, or operating humanoid robots made by "foreign adversaries" including China, Russia, and Iran. It is far narrower than the sweeping GUARD Act floated last year by Representative John Moolenaar: commercial sales to US businesses and consumers stay untouched, for now.
๐ tl;dr: Washington's first hard line on physical AI starts at the barracks door.
๐ฌ Watch This
Elon Musk sat down with Zanny Minton Beddoes, editor-in-chief of The Economist, for a wide-ranging talk on where AI is headed. He argues that AI will surpass the combined intelligence of humanity within about five years, and that people may lose control within ten. Along the way he sketches an "age of amazing abundance," floats a scheme for rival labs to cross-check each other's frontier models, and explains why he thinks the industry's leaders can still cooperate.
An unverified leak circulating among AI watchers claims OpenAI's next model, tagged GPT-5.7 or GPT-6, is aiming for an August launch, built on a fresh pre-training run with a 1.5M-token context window and a big step up in agentic reasoning. With no model card, price, or date from OpenAI, treat the details as rumour, not roadmap.
Two Labs, One Day: Voice Becomes How You Run Your AI
The Takeaway
๐ On July 23, OpenAI brought its full-duplex GPT-Live voice model to the ChatGPT desktop app and Codex, so you can dictate, steer, and debug coding jobs hands-free.
๐ The same day, Anthropic upgraded Claude's voice mode to run on Opus and Sonnet, up from the small Haiku model, and to reach connected apps like Gmail and Notion mid-conversation.
๐ Both push voice past dictation into a control layer: you run software and orchestrate agents by speaking, the way you would brief a colleague.
๐ The catch: Anthropic left its underlying voice model untouched, so latency still trails OpenAI, and hands-free tool access raises real oversight questions.
For three years, using an AI assistant mostly meant typing at it and waiting your turn. This week, two of the biggest labs decided that was the slow way to work.
On July 23, OpenAI extended GPT-Live, the full-duplex model it launched July 8 that can listen and speak at once, into the ChatGPT desktop app on macOS and Windows, wired straight into Codex, its coding agent, and ChatGPT Work. In one demo, two engineers spoke to the same session in the same room, each giving separate instructions while the model kept working. The pitch is hands-free agentic coding: you start jobs, review pull requests, and debug by talking, while the model hands the heavy reasoning to GPT-5.5 in the background.
(OpenAI CEO Sam Altman / Getty Images via Fortune)
Almost at the same time, Anthropic shipped its own upgrade. Claude's voice mode, once stuck on the small Haiku model, can now run on Opus or Sonnet and talk problems through in ten languages. More important, voice reaches the tools you have already connected, Gmail, Google Calendar, Slack, Canva, and Notion, so you can reschedule a meeting or draft a document by speaking. Free users get Haiku and one app; paid plans unlock the rest.
(Claude voice mode / Anthropic via TechCrunch)
The symmetry is the story: two rivals, same day, moved voice from a novelty you demo once into the layer you use to operate the machine. They took different roads: Anthropic admits it did not change its underlying voice model, so its latency still feels clumsier than OpenAI's continuous audio. And giving a talkative agent live access to your inbox and codebase is convenient right up until it does the wrong thing fast. The quicker you can tell software what to do, the quicker a misheard instruction becomes an action you never wanted.
Why it matters: When talking becomes the fastest way to run your tools and your code, the interface stops being something you look at and becomes something you converse with. That shifts who does the typing, and it puts a premium on the machine hearing you right.
Sources:
๐ https://venturebeat.com/orchestration/agentic-coding-goes-hands-free-as-openai-brings-gpt-lives-full-duplex-voice-control-to-codex-and-chatgpt-on-the-desktop
๐ https://x.com/claudeai/status/2080376094939603366
๐ https://techcrunch.com/2026/07/23/anthropic-updates-claude-voice-mode-with-more-capable-models/
๐ https://fortune.com/2026/07/23/openai-launches-chatgpt-voice-desktop-programming/
Win AI Search Without a Big Team
92% of VCs use AI to find companies. 58% of buyers start there, too. If you're not showing up in AI answers, you're invisible before the conversation even starts. Join HubSpot for Startups, Anthropic, and Marketing Against the Grain on July 30 (11 am ET) for a live AEO teardown. Real startup. Real recs. Register and unlock the free Startup Visibility Bundle.
The chart: Oracle's stock over the last six months, the other half of the layoffs story. Shares closed at $120.04 on July 23, down about 34% in six months after peaking near $250 in early summer. A 2% after-hours bounce barely dented the slide.
The lesson: Wall Street has soured on the pay-now, earn-later math of the AI infrastructure race. Oracle staked its future on a reported $300 billion compute commitment to OpenAI and a wave of data centers; investors have roughly halved the value they briefly gave that bet.
The caveat: One stock is not the whole trade. Oracle carries its own baggage, from collateral demands to heavy debt, so the drop is partly company-specific. Still, it is a clear warning that the market will punish the gap between huge outlays and revenue that has not arrived.
โ๏ธ Europe Fines First, Builds Later
โก Bottom line: On August 2 the EU AI Act starts handing regulators the power to fine AI providers up to โฌ35M or 7% of global revenue.
๐ก Why it matters: Any company serving EU users now carries the compliance load, even as Brussels delays the law's hardest parts.
๐ What it means: Europe is quick to regulate and slow to build, and that gap is becoming its defining AI problem.
On August 2, the European Union starts writing fines. That is when the EU AI Act, the world's first comprehensive AI law, hands regulators real enforcement power over providers of general-purpose AI models, the systems behind ChatGPT, Claude, and Gemini. The penalties are steep: up to โฌ35 million or 7% of worldwide revenue, a ceiling above the GDPR's, and every one of the 27 member states has to stand up its own AI regulator.
(The EU Artificial Intelligence Act)
What actually arrives is transparency paperwork. Under Article 50, chatbots must announce they are machines, and AI-generated images, audio, and video must be labelled as synthetic. Useful in theory, though it lands the same week Black Forest Labs shows off a model that spins up video, audio, and action in one shot, a reminder of how far the tech has run ahead of the rules.
(European Commission President Ursula von der Leyen)
The awkward part is what Brussels quietly dropped. A late amendment, the Digital Omnibus, has pushed the heavier high-risk rules, covering hiring, education, and biometrics, more than a year down the road to late 2027, after industry and Washington complained. So Europe gets the fines and the forms now and the substance later, all while it struggles to field a single frontier lab of its own. Writing the rulebook was never the hard part; building something worth regulating is.
Think You Know What AI Does Next?
Which model leads the next benchmark? Which AI lab ships the next major breakthrough?
Kalshi lets you trade on real-world AI and technology events as the industry moves. If you follow launches, model updates, and benchmarks closely, put that knowledge to work.
Bonus credit varies from $15 to $500. Terms apply.