Two personnel bulletins landed inside days of one another this week. Seen individually, they appear like the same old churn of Silicon Valley’s government class. Seen collectively, they reveal the place the artificial-intelligence race is heading subsequent.
At Google, Demis Hassabis is surrendering day-to-day management of Google DeepMind to change into chair of the division and chief scientist of Alphabet. He’ll spend extra time serious about synthetic normal intelligence, international technique and the usage of AI to find medicines.
I’ve been working in direction of AGI my entire life, and as we enter this pivotal second, I’m entering into a brand new function as Chair of Google DeepMind & Chief Scientist of Alphabet. It will enable me to deal with long-term technique, and accelerating scientific breakthroughs, together with…
— Demis Hassabis (@demishassabis) August 5, 2026
At OpenAI, researcher Naomi Bashkansky has walked away to hitch a younger San Francisco firm referred to as Conduit, which is making an attempt to decode ideas utilizing non-invasive neural sensors. Its founders don’t hassle with timid language. They name the challenge “telepathy.”
Two weeks in the past, I resigned from OpenAI to hitch Conduit as a founding researcher, the place we’re coaching fashions to non-invasively learn the human thoughts.
I’ve written some ideas about what telepathy may appear like by 2035 and easy methods to get there: pic.twitter.com/cat0d4My15
— Naomi Bashkansky (@NaomiBashkansky) August 5, 2026
The primary Gilded Age was constructed by industrialists who managed railways, oil fields, metal mills and the arteries of bodily commerce. This one is being constructed by individuals making an attempt to manage the motion of intelligence itself: from knowledge facilities into scientific laboratories, from algorithms into medicine, and finally from human brains instantly into machines. The chatbot was solely the opening act.
Google separates the AI thinker from the AI manufacturing unit
In a company announcement published by Google, CEO Sundar Pichai mentioned Hassabis would change into chair of Google DeepMind and Alphabet’s chief scientist whereas persevering with to guide Isomorphic Labs, the drug-discovery firm spun out of DeepMind.
Koray Kavukcuoglu, DeepMind’s longtime chief expertise officer and Google’s chief AI architect, will change into senior vp of Google DeepMind. He’ll oversee Gemini mannequin growth, frontier analysis, the Gemini app and Google’s developer-facing AI groups.
Kavukcuoglu will run the equipment: delivery fashions, working with product divisions, enhancing Gemini and turning analysis into income. Hassabis will change into one thing nearer to Alphabet’s philosopher-scientist in residence, contemplating AGI, scientific discovery and the long-term penalties of making programs that will exceed human capabilities.
Hassabis wrote that humanity had reached “a pivotal second” and that AGI now felt “shut at hand.” Pichai mentioned the brand new place would let Hassabis think about shaping that future somewhat than managing the every day operations of a sprawling group.
This isn’t essentially a demotion disguised by a grander title. Hassabis has at all times appeared extra desirous about foundational breakthroughs than quarterly product highway maps. He based DeepMind in 2010, bought it to Google for roughly $650 million in 2014 and remained targeted on the unique mission: fixing intelligence after which utilizing intelligence to unravel every thing else.
The issue is that Google now must do two contradictory issues concurrently. It should conduct analysis that will outline the following century whereas additionally promoting AI merchandise this quarter.
Pichai emphasised that the Gemini app has surpassed 950 million month-to-month customers, whereas downloads of Google’s Gemma fashions have exceeded 900 million. These will not be laboratory curiosities. They’re industrial-scale merchandise requiring velocity, reliability and relentless execution.
Placing Kavukcuoglu in operational management is due to this fact a rational division of labor. He has spent greater than 13 years at DeepMind and contributed to programs together with WaveNet and the deep-reinforcement-learning structure behind DeepMind’s early Atari work. Folks inside Google Cloud reportedly welcomed the change, anticipating it to drag DeepMind nearer to the business facet of Alphabet.
But the reshuffle comes amid a broader expertise migration. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le — 4 of Google’s most consequential engineers and researchers — are leaving to determine Discovery Loop, a public-benefit firm targeted on automating scientific and engineering breakthroughs. Google will spend money on the brand new group and provide cloud infrastructure.
Alphabet shares fell round 4% following information of the restructuring and departures. Wall Road usually prefers its geniuses safely locked inside the company fortress.
Hassabis desires AI to earn its place in historical past
Hassabis will now spend extra time at Isomorphic Labs, the place the ambition is just not merely to make drug growth cheaper however to rework it from an costly means of trial and error into one thing nearer to computational engineering.
Client AI is changing into brutally aggressive. Fashions enhance, options are copied, costs fall and supposedly distinctive capabilities change into commodities inside months. Scientific AI is totally different. A system that discovers a viable most cancers remedy, predicts an vital organic mechanism or shortens scientific growth by a number of years creates defensibility, social legitimacy and large financial worth.
Hassabis has repeatedly argued that enhancing well being ought to be AI’s most vital utility. His work on AlphaFold helped display that machine studying may clear up scientific issues somewhat than merely imitate language, generate photos or assist workplace employees produce extra PowerPoint slides. He and DeepMind researcher John Jumper shared a part of the 2024 Nobel Prize in Chemistry for protein-structure prediction.
The strategic calculation is clear: one other marginally higher chatbot could impress a benchmark leaderboard, however a profitable drug can change thousands and thousands of lives and produce billions of {dollars} in income.
Google’s reorganization due to this fact seems to be much less like Hassabis stepping away from the AI race than shifting towards its highest-stakes part. Kavukcuoglu will struggle the mannequin warfare. Hassabis will pursue the argument that AI deserves civilization-level funding as a result of it might produce civilization-level outcomes.
OpenAI loses a researcher to the telepathy enterprise
Naomi Bashkansky resigned from OpenAI on July 23 and started work the next day as a founding researcher at Conduit. She had spent roughly 18 months at OpenAI, engaged on alignment analysis and contributing to inside tasks regarding AGI and resilience.
In an essay titled “Why I’m leaving OpenAI to build telepathy”, she describes Conduit’s mission with out the customary layer of startup modesty: “We’re constructing telepathy: thought-to-text fashions, educated on non-invasive neural knowledge.”
Her forecast is aggressive. By 2027, she imagines employees carrying neural headbands that constantly transmit intentions to AI coding brokers. By 2030, AI fashions may talk instantly with representations of ideas somewhat than ready for these ideas to be transformed into language. By 2035, she believes AI could really feel much less like an exterior instrument and extra like an extra sense or limb.
Bashkansky says Conduit is within the “GPT-2 period” of neural decoding: primitive in contrast with the place the expertise could finally go, however far sufficient alongside to see scaling conduct. She estimates that an organization able to broadly studying from and writing to the human mind may finally be price greater than $1 trillion.
It’s the purest doable Gilded Age startup pitch: frighteningly bold, technically unsure and hooked up to a market doubtlessly bigger than private computing.
Conduit says it has already collected roughly 10,000 hours of neural and language knowledge from 1000’s of individuals, utilizing customized headsets and a basement operation that ran for so long as 20 hours per day. Members held conversations, typed and spoke whereas the corporate recorded non-invasive neural alerts.
The ensuing mannequin is just not producing excellent transcripts. It’s making an attempt to establish semantic intent: the final that means of what anyone is making ready to speak. Conduit has revealed examples during which a participant’s phrase resembling “the room appeared colder” produced a prediction referring to a breeze or gust. The wording was unsuitable, however the underlying idea was associated.
The corporate claims its system can typically establish an concept shortly earlier than the participant converts it into spoken or typed phrases. It has not but revealed complete technical particulars or unbiased validation of its mannequin, saying {that a} fuller analysis launch will come later.
That caveat is vital. “Telepathy” is great advertising, however this isn’t but a machine casually rummaging by anyone’s secret reminiscences. It’s statistical inference constructed from noisy neural measurements, linguistic context and a strong language mannequin that’s already good at guessing what an individual in all probability intends to say. The LLM is doing a few of the similar work a navigation map does when GPS reception is poor: narrowing a fuzzy sign right into a believable vacation spot.
The science is actual, however the hype is operating forward
Conduit is just not alone in believing that scale will unlock non-invasive mind interfaces. Meta’s Brain2Qwerty v2 research used magnetoencephalography, or MEG, to decode sentences from the mind exercise of 9 individuals as they typed. Every particular person contributed about ten hours of recordings. The system achieved a median word-error price of 39%, whereas its greatest participant produced many sentences containing no a couple of incorrect phrase.
That’s spectacular, however it’s nonetheless a managed experiment involving individuals actively typing inside specialised tools. MEG machines are hardly shopper headbands. They’re massive, costly units designed for laboratories and hospitals.
Transportable EEG programs are extra sensible however produce far noisier alerts. A current Scientific Reports analysis discovered that a number of supposedly profitable EEG-to-text programs carried out equally when equipped with random noise, suggesting that fashions have been typically memorizing linguistic patterns somewhat than genuinely decoding mind exercise.
So the trustworthy verdict is neither “telepathy has arrived” nor “that is science fiction.” Non-invasive neural decoding is advancing. Massive language fashions are significantly helpful as a result of they will remodel incomplete alerts into coherent prospects. However the hole between recognizing broad intent underneath managed situations and constantly studying unrestricted ideas in on a regular basis life stays huge.
The vital query is whether or not large datasets can shut that hole. Conduit is betting that neuroscience will comply with the identical bitter lesson that reshaped AI: subtle handcrafted theories will finally lose to sufficient knowledge, sufficient compute and a sufficiently massive mannequin.
It’s a daring wager. Historical past suggests we should always not dismiss daring bets merely as a result of they initially sound ridiculous. Historical past additionally means that founders describing trillion-dollar markets will not be impartial observers.
Neural knowledge may change into probably the most precious knowledge on Earth
If Conduit’s method works, the results will attain far past sooner typing.
At the moment’s computer systems power people to compress complicated intentions into clicks, gestures and sentences. A mind interface may bypass that bottleneck. As a substitute of instructing an AI agent by fastidiously constructed prompts, a person may talk uncertainty, curiosity, visible concepts and half-formed intentions instantly.
That might make AI dramatically extra helpful. It may additionally produce probably the most invasive surveillance expertise ever created.
Search histories reveal what individuals examine. Social-media exercise reveals what they publicly interact with. Neural data may reveal what they thought of earlier than deciding to not converse.
The governance issues are apparent. Who owns the uncooked alerts? Can they be retained for mannequin coaching? Can an employer require a productiveness headset? Can an insurance coverage firm infer cognitive or emotional situations? Can a platform use neural responses to optimize promoting earlier than the person has consciously articulated a choice?
Probably the most harmful model of brain-computer expertise will not be a authorities machine that completely reads minds. It might be a handy shopper product that reads them imperfectly, shops the outcomes indefinitely and buries consent inside a 40-page terms-of-service settlement.
That’s how energy normally enters society: not carrying jackboots, however providing free delivery and a greater person expertise.
The following platform warfare is over the human working system
Hassabis shifting towards AGI and computational biology, Google veterans launching an automated-discovery firm and an OpenAI researcher defecting to construct neural interfaces will not be unrelated developments.
They point out that the middle of gravity is shifting past general-purpose language fashions.
The biggest laboratories have already educated programs able to producing textual content, code, photos, audio and video. The brand new competitors is over what these programs hook up with. Google desires AI related to scientific discovery and medication. Conduit desires it related to intention. Meta is researching the neural equipment of language. Different corporations are pursuing robotics, autonomous laboratories and artificial biology.
The following monopolies will not be constructed round web sites or smartphone working programs. They might be constructed across the interface between intelligence and the bodily human being.
That’s what makes this an AI Gilded Age. The expertise is genuinely extraordinary, and its advantages might be huge. It’s also being developed by a small variety of terribly rich establishments competing to personal the infrastructure by which individuals work, assume, uncover and maybe finally expertise actuality.
The unique Gilded Age laid metal tracks throughout continents. This one is laying invisible tracks between the thoughts, the machine and the molecule.
Whoever owns these tracks won’t merely personal the following nice expertise platform. They might personal the route by which human intention enters the world.
Alex Chen Alex Chen Read More
