AI Actually
Issue No. 14 · Sunday · June 1, 2026
Something quietly significant happened this week while everyone was looking at the big splashy Anthropic headlines.
A stock-trading app handed an AI agent your brokerage account. A startup in New York started sending camera-wearing cleaners into people’s apartments — for free. Scientists dropped a tool that’s better at predicting cancer-related proteins than anything that came before it. And Apple is about to rebuild its most famous product from scratch, using a competitor’s brain.
We’ll get to Anthropic. It’s last.
Siri’s biggest makeover ever — and the quiet voice race happening underneath it
Apple’s WWDC developer conference is nine days away — June 8 — and Bloomberg’s Mark Gurman dropped the biggest preview yet of what’s coming.
The new Siri is being rebuilt around Google’s Gemini AI, gaining a standalone chatbot-style app with conversation history, a new “Search or Ask” panel triggered by swiping down from the top of the display, and integration directly into the Dynamic Island. Apple is also testing a system that lets users route queries to different AI services — including Gemini, ChatGPT, and Claude — directly from within Siri. Apple settled a $250 million class-action lawsuit over delayed Siri features just weeks before this WWDC — meaning promised AI upgrades that never shipped already cost the company real money. The stakes for June 8 are unusually high.
The Siri news is the headline. But beneath it, a parallel race has been quietly building: how you talk to AI in the first place.
Right now, most people still type into AI chatbots the same way they typed into Google in 2005. A new generation of tools is betting that’s already obsolete. Wispr Flow is an AI voice keyboard that works inside every app — press a hotkey, speak naturally, and it inserts clean, polished text wherever your cursor is sitting, whether that’s Gmail, Notion, Slack, or a code editor. It doesn’t just transcribe; it figures out what you meant to write and formats it for context. Slack messages come out casual. Emails come out professional. In May 2026, Wispr Flow raised $260 million at a $2 billion valuation, confirming that voice AI dictation has moved from niche utility to mainstream enterprise category.
Typeless, a Stanford-backed competitor, is taking the same bet. It targets professionals who need faster, more natural text capture — remote workers, writers, executives — with real-time filler removal, per-app tone adaptation, and a quoted dictation speed of around 220 words per minute.
Why it matters: Apple is finally building what voice input was always supposed to be — a real AI that knows what you mean, not just what you said. But Wispr Flow and Typeless are already living in that future, quietly. They’re betting the interface shift from “typing into AI” to “talking to AI” is happening whether Apple ships it or not. It probably is. For a company that invented the touchscreen, arriving second to a hotkey app is not a great look.
Someone put a camera on a cleaning person, and the robots are the point
A startup called Shift launched this week in New York City with a genuinely strange offer: a professional will come clean your apartment for free.
The catch: they arrive wearing a recording device.
Shift is an offshoot of a Germany-based company called Microagi, which already oversees data collection work in several countries. The data — footage of real humans cleaning real apartments, messy and uncontrolled — is anonymized and sold to AI labs and robotics companies trying to train the next generation of household robots. A robot can’t learn home cleaning from staged lab videos. Real homes have clutter, narrow corridors, misplaced objects, and the kind of chaos that a lab environment can’t replicate. Shift’s bet is that human demonstrations of real-world chores are becoming the most valuable training data in robotics.
Shift plans to expand beyond New York to San Francisco, London, Zurich, and Munich. The company is already operating in 15 countries with users who earn cash rewards for performing everyday chores — including washing dishes — that are recorded for training data.
Why it matters: Most AI conversations this year have been about white-collar jobs — coders, writers, lawyers. Physical trades have largely escaped that conversation because a body moving through a real space is still genuinely hard to automate. Shift is trying to close that gap, and they’ve figured out a clever way to fund it: make the training data collection the product. You get a clean apartment. They get closer to a robot that can replace the person who cleaned it.
Robinhood just gave an AI agent access to your brokerage account
AI agents have spent the past year asking for access to your calendar, your inbox, your spreadsheets. This week, Robinhood asked for something more awkward.
Robinhood launched “agentic trading” — a beta that lets users connect AI agents to a dedicated Robinhood account, set a budget, and let those agents trade stocks. Gold Card users also get an agentic virtual card that an AI assistant can spend within user-set limits. The agents can analyze portfolios, suggest strategies, and execute actual trades. The company plans to expand beyond stocks into options, crypto, futures, event contracts, and prediction markets.
There are controls: a dedicated account, a budget cap, and Robinhood says it will require approvals. But the direction of travel is clear.
Why it matters: Up until recently, AI agents were mostly advisors — they could look at your finances and tell you what to do. Robinhood just made them actors. The shift from “AI that recommends” to “AI that executes” is arguably the most consequential UX decision in the technology right now. Once you hand an AI a budget, you need spending caps, audit logs, rollback buttons, and a panic switch. The design problems are no longer theoretical. Every company building an AI agent with real-world consequences — in finance, in procurement, in operations — is about to face the same question Robinhood just answered: how much trust, exactly, and with what guardrails?
A lab backed by Zuckerberg just released the most powerful protein model ever built
This one won’t trend. It should.
Mark Zuckerberg and Priscilla Chan’s nonprofit Biohub — which we covered back in Issue No. 4 when it committed $500 million to a Virtual Biology Initiative — just released the thing that money was building.
It’s called ESMFold2. It’s a model trained on 2.8 billion protein sequences, and it claims the top spot on every major benchmark for predicting how proteins fold and interact with each other — including outperforming Google DeepMind’s AlphaFold, which was itself considered a scientific breakthrough when it launched.
The part that matters for the rest of us: ESMFold2 is already being tested in actual labs against actual disease targets. The model designed proteins capable of binding to five different cancer and immune disease targets, with success rates ranging from 36% to 88%. In drug discovery terms, those numbers are remarkable. The system also includes ESM Atlas — a map of 6.8 billion protein sequences and 1.1 billion predicted structures. The whole thing is open source.
Why it matters: This is what Demis Hassabis, Google DeepMind’s CEO, means when he talks about AI ending disease — not as rhetoric, but as a literal engineering roadmap. The gap between “AI that can predict protein structure” and “AI that can design a drug” has been closing for three years. ESMFold2 closes it further. This isn’t a chatbot that gets better at writing emails. It’s infrastructure for rebuilding how medicine works. The fact that it’s open and free is almost as significant as what it can do.
And finally: Anthropic had a week
We’ve been here before. Every few months, the valuation goes up. This one went up a lot.
Anthropic raised $65 billion at a $965 billion valuation, making it the most valuable private startup on Earth — ahead of OpenAI, ahead of SpaceX, ahead of almost everything. The round was led by Greenoaks, Sequoia, Altimeter, and Dragoneer, with new strategic investors including Samsung, Micron, and SK Hynix. Revenue has reportedly crossed $47 billion annualized (versus OpenAI at $27 billion). The same week, it released Claude Opus 4.8 — top benchmarks on coding, reasoning, and financial analysis; four times fewer missed bugs than its predecessor; adjustable reasoning effort; and parallel AI subagents that can run simultaneously. It also teased something called “Mythos-class” models arriving in the coming weeks.
Why it matters: The number keeps going up because the revenue keeps going up. At $47 billion annualized, Anthropic isn’t a startup with a promising valuation anymore — it’s a large enterprise software company that hasn’t gone public yet. The Opus 4.8 release is legitimately good (early testing put it at the top of both engineering and writing benchmarks), but the model cadence is now so fast that any individual release matters less than the trajectory. The question worth sitting with isn’t “is Claude good?” — it’s “what does it look like when an AI company this profitable eventually goes public?” That’s a different kind of story.
What we’re skipping (and why)
China restricts AI researchers from traveling abroad — Genuinely interesting geopolitics, and probably worth a deeper look. The short version: Beijing is treating its top AI talent the way it treats semiconductor engineers — as a strategic asset that shouldn’t leave. We’re saving this for a future Wednesday issue where we can give it the framing it deserves.
Cognition raises $1B at a $26B valuation — Cognition makes Devin, the AI coding agent. The funding is large. The revenue ($492M annualized) suggests it’s not crazy. But it’s a VC round for a developer tool, and the mechanics don’t land differently than the dozen similar rounds we’ve tracked this year.
OpenAI Foundation pledges $250M for worker transition — The money is real. The initiatives aren’t announced yet. We’ll check back when there’s something to actually evaluate.
xAI tells employees to stop talking to Cursor workers — Extremely inside-baseball. Filed under: AI company drama that matters if you’re at one of those companies, and not much otherwise.
See you Wednesday morning.
— AI Actually
Apple's test of a system routing Siri queries to Gemini and Claude signals the death of closed ecosystems.
I analyze this platform dynamic at theaifounder.substack.com, because application builders lose their primary user relationship the moment hardware giants commandeer the routing layer.
What customer acquisition channels remain for standalone model developers after iOS intercepts every initial prompt?