In partnership with
In Today’s Issue:
🧠 Meta reads words off your brain
🇰🇷 South Korea's $880bn chip and AI bet
💧 Erin Brockovich vs. the data centers
📱 Cursor puts coding agents on your phone
✨ And more AI goodness…
⚡The Signal
The most science-fiction headline of the week is also the most grounded: Meta can now read words off your brain without opening your skull.
Brain2Qwerty v2 decodes the sentences a person is typing from their brain activity at 61% word accuracy, using a MEG scanner rather than a surgical implant, and Meta says accuracy keeps climbing with more data alone. The physical side of AI got louder elsewhere too: South Korea committed at least $880bn to chips, data centers and robots, while Erin Brockovich turned her data-center map into a national fight over water, power and secrecy. Coding agents advanced again, with Claude Opus 4.7 topping Epoch's new MirrorCode benchmark and Cursor shipping an app that runs agents from your phone. The thread: AI is moving off the screen and into brains, buildings and factory floors, which is exactly where the hard constraints and the real fights now live.
All the best,
Kim Isenberg
(Bloomberg via Getty Images)
South Korea Bets $880bn on Chips, AI and Robots
South Korea unveiled at least $880bn (£666bn) in planned investment to expand its chip, data-center and robotics capacity, calling the build-out a matter of national "survival."
President Lee Jae-myung announced the plan on Monday alongside the heads of Samsung and SK Hynix, naming semiconductors, physical AI and AI data centers the "triple axis" for a "great leap forward." The push aims to spread advanced factories beyond Seoul as Taiwan, China and Japan race to expand chip output; SK Hynix, a major Nvidia supplier, topped a $1tn valuation in May on AI-memory demand.
👉 tl;dr: Korea is matching the US hyperscalers' AI spending at the national level, treating chips and data centers as industrial policy.
💧 Erin Brockovich Takes On AI's Data Centers
Environmental campaigner Erin Brockovich is rallying communities against AI data centers, warning that residents are "up against forces that have all the money in the world."
Brockovich, whose 1990s fight against utility PG&E became a Hollywood film, built an open-source map after thousands of people wrote to her about projects near their homes. Her core complaint is secrecy: developers often sign nondisclosure agreements with local officials, so residents learn little about water use, power draw or round-the-clock generator noise before a project is approved. She says people in Australia, India, Scotland and Ireland have contacted her too.
👉 tl;dr: The data-center boom now has an organized grassroots opponent with a famous name and a national map.
📱 Cursor Puts Coding Agents in Your Pocket
Cursor launched a native iOS app in public beta that lets developers start, steer and review coding agents from their phone.
You can kick off cloud agents or remote-control ones running on your computer, direct them by voice or slash commands, and get push notifications and lock-screen Live Activities when work is ready. The app also lets you review diffs and merge pull requests on the go, pitched for on-call incidents and urgent bug fixes. It is free on every paid plan, with 75% off Composer 2.5 runs from mobile through July 5.
👉 tl;dr: Agentic coding is going asynchronous and mobile; you no longer need a laptop open to ship.
Before you trust an AI coding agent on real work, make it prove itself on a "rebuild it" test.
Why it helps: Epoch's new MirrorCode benchmark shows the best models can now reconstruct whole programs from scratch, but they still fail more often than they succeed, so a from-scratch rebuild quickly exposes where yours breaks.
Try this: "Without looking at the original source code, reimplement this command-line tool using only its README and test suite. Run the tests, list every case you cannot pass, and flag any behavior you had to guess."
🎬 Watch This
The Trump administration is letting Anthropic restore access to its Mythos 5 model for a select group of US companies and federal agencies, partly reversing an earlier clampdown. NYT reporter Sheera Frenkel breaks down on CBS News what the carve-out means, who gets access first, and why Washington's stance on frontier models keeps shifting.
"It takes a lot of energy to train a human."
(Sam Altman, OpenAI. Getty Images / TechCrunch)
– Sam Altman, CEO of OpenAI, at the AI Impact Summit in New Delhi (Fortune)
Newly surfaced Claude app strings appear to tie Fable 5 usage credits to identity verification, according to code spotted in the latest release. The strings suggest Fable 5 would sit behind a usage-credit system billed outside your plan, with one line reportedly reading "Your credits will be added once your identity is verified." That is notable because Anthropic had earlier described identity checks as unrelated to Fable and limited to flagged accounts. Decrypt says it verified the strings by extracting them from the shipped Claude Code package; Anthropic has not formally announced the change.
Meta Reads Words Off Your Brain,
No Implant Required
The Takeaway
👉 Meta's non-invasive Brain2Qwerty v2 decodes typed sentences from brain activity at 61% word accuracy, up from roughly 8% for earlier non-invasive methods.
👉 It reads the brain with a MEG scanner (magnetoencephalography), which senses magnetic fields from outside the head, so there is no surgical implant.
👉 Training used about 22,000 sentences typed by nine volunteers, each in the scanner for ~10 hours; the best participant hit 78%.
👉 Accuracy improves log-linearly with data, and Meta open-sourced the code and dataset, hinting the gap to implants could close without surgery.
Meta has shown an AI system that turns brain activity into text without touching the brain. Brain2Qwerty v2, from Meta's research lab, decodes the sentences a person is typing by reading their brain with a magnetoencephalography (MEG) scanner, a helmet-like device that picks up the tiny magnetic fields neurons produce. No electrodes are implanted; the volunteer simply sits under the scanner and types. An end-to-end deep-learning model reads the raw signals, and a fine-tuned language model uses context to fill in what the noisy signal misses.
(A volunteer types under a MEG scanner. Meta AI)
The headline number is 61% word accuracy on average, with the best participant reaching 78% and more than half of their sentences decoded with one word error or fewer. That is a big jump from the roughly 8% managed by previous non-invasive methods. The model learned from about 22,000 sentences typed by nine volunteers, each wearing the MEG scanner for around 10 hours.
(Brain2Qwerty's benchmark: word and character error rate by model, EEG vs MEG, where Brain2Qwerty leads. Meta AI / arXiv)
The catch is practical: a MEG scanner is a room-sized, shielded machine, not a wearable, so this is a lab result, not a consumer product. But Meta reports that accuracy improves log-linearly with data, meaning more recordings alone keep pushing the numbers up, and it open-sourced the training code and dataset. The company frames the work as a path toward communication for people with brain lesions or paralysis.
Why it matters: Non-invasive brain-to-text has trailed far behind surgical implants like Neuralink; closing that gap without surgery would make the technology dramatically safer and easier to scale, even while the hardware is still stuck in the lab.
Sources:
🔗 https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication
🔗 https://arxiv.org/abs/2502.17480
🔗 https://decrypt.co/372338/meta-brain2qwerty-ai-brain-activity-text
HubSpot AEO
Picture this. A buyer opens ChatGPT and asks for a recommendation in your category. Your competitor's name comes up. Yours doesn't. And that buyer never makes it to your website.
That's happening right now in markets everywhere. And most teams don't know it's happening because it never shows up in their analytics.
HubSpot AEO shows you exactly where your brand stands in AI search, where competitors are getting recommended instead of you, and tells you specifically what to fix. No expertise needed.
Try it free for 28 days. Just $50 a month after.
The chart: Epoch AI's new MirrorCode benchmark (built with METR) measures the solve@100% rate, the share of attempts where a model rebuilds a target program from scratch and passes 100% of its end-to-end tests. Claude Opus 4.7 leads at ~56%, ahead of GPT-5.5 (~44%) and Gemini 3.1 Pro Preview (~31%); the whiskers show ±1 standard error.
The lesson: Long-horizon autonomous coding is real. Even failed attempts usually pass 90%+ of tests, and Opus 4.7 reimplemented a ~16,000-line Go toolkit in 14 hours for about $251.
The caveat: These are open-source programs the models likely saw during pretraining, so scores may be inflated; the set is small (25 targets), and one hard task ran 19 days and cost $2,600.
🤖 Europe's Big Robot Bet: NEURA Raises Up to $1.4B
⚡ Bottom line: German firm NEURA Robotics raised up to $1.4 billion, the largest Series C ever for a full-stack robotics company.
💡 Why it matters: At a ~$7 billion valuation, it is Europe's loudest entry in the US-China humanoid race.
🔎 What it means: The funding is racing ahead of revenue, a bet that cognitive robots scale this decade.
Most of the money and hype in humanoid robots has flowed to the US and China. NEURA Robotics, based near Stuttgart in Germany, just changed that. On June 10 the company announced a Series C of up to $1.4 billion that it calls the largest ever for a full-stack robotics firm, at a valuation of around $7 billion. The backers read like a who's-who: Nvidia, Amazon, Qualcomm, German industrial giants Bosch and Schaeffler, the European Investment Bank, and crypto firm Tether as lead investor.
(Founder David Reger with NEURA's robot lineup. NEURA Robotics)
NEURA builds what it calls "cognitive" robots that learn tasks rather than follow fixed scripts, the same shift toward physical AI driving the whole field. Its range runs from industrial arms to a transport robot, topped by 4NE-1, a humanoid aimed at factories and, eventually, homes. The cash is meant to push production toward millions of robots by 2030 and fund "NEURA Gyms," real-world training grounds where robots learn by doing.
(NEURA's 4NE-1 humanoid robot. NEURA Robotics)
The caveat: the $1.4 billion is a ceiling tied to performance milestones, not a guaranteed lump sum, and NEURA's order book of over $1 billion is still mostly promises rather than shipped robots. Even so, the raise hands Europe a humanoid contender that can finally spend at US-and-China scale.
Viktor writes the code, builds the tool, ships the brief, all from inside Slack and Microsoft Teams. 20,000+ teams run their company through him. Hire on day zero. Get Started for Free
