The AI Gilded Age Is Coming for the Human Mind

Two personnel announcements landed within days of each other this week. Viewed separately, they look like the usual churn of Silicon Valley’s executive class. Viewed together, they reveal where the artificial-intelligence race is heading next.

At Google, Demis Hassabis is surrendering day-to-day control of Google DeepMind to become chair of the division and chief scientist of Alphabet. He will spend more time thinking about artificial general intelligence, global strategy and the use of AI to discover medicines.

https://x.com/demishassabis/status/2085034334914769203

At OpenAI, researcher Naomi Bashkansky has walked away to join a young San Francisco company called Conduit, which is attempting to decode thoughts using non-invasive neural sensors. Its founders do not bother with timid language. They call the project “telepathy.”

https://x.com/NaomiBashkansky/status/2085043839589617918

The first Gilded Age was built by industrialists who controlled railways, oil fields, steel mills and the arteries of physical commerce. This one is being built by people attempting to control the movement of intelligence itself: from data centers into scientific laboratories, from algorithms into drugs, and eventually from human brains directly into machines. The chatbot was only the opening act.

Google separates the AI philosopher from the AI factory

In a company announcement published by Google, CEO Sundar Pichai said Hassabis would become chair of Google DeepMind and Alphabet’s chief scientist while continuing to lead Isomorphic Labs, the drug-discovery company spun out of DeepMind.

Koray Kavukcuoglu, DeepMind’s longtime chief technology officer and Google’s chief AI architect, will become senior vice president of Google DeepMind. He will oversee Gemini model development, frontier research, the Gemini app and Google’s developer-facing AI teams.

Kavukcuoglu will run the machinery: shipping models, working with product divisions, improving Gemini and turning research into revenue. Hassabis will become something closer to Alphabet’s philosopher-scientist in residence, considering AGI, scientific discovery and the long-term consequences of creating systems that may exceed human capabilities.

Hassabis wrote that humanity had reached “a pivotal moment” and that AGI now felt “close at hand.” Pichai said the new position would let Hassabis concentrate on shaping that future rather than managing the daily operations of a sprawling organization.

This is not necessarily a demotion disguised by a grander title. Hassabis has always appeared more interested in foundational breakthroughs than quarterly product road maps. He founded DeepMind in 2010, sold it to Google for roughly $650 million in 2014 and remained focused on the original mission: solving intelligence and then using intelligence to solve everything else.

The problem is that Google now needs to do two contradictory things simultaneously. It must conduct research that may define the next century while also selling AI products this quarter.

Pichai emphasized that the Gemini app has surpassed 950 million monthly users, while downloads of Google’s Gemma models have exceeded 900 million. Those are not laboratory curiosities. They are industrial-scale products requiring speed, reliability and relentless execution.

Putting Kavukcuoglu in operational control is therefore a rational division of labor. He has spent more than 13 years at DeepMind and contributed to systems including WaveNet and the deep-reinforcement-learning architecture behind DeepMind’s early Atari work. People inside Google Cloud reportedly welcomed the change, expecting it to pull DeepMind closer to the commercial side of Alphabet.

Yet the reshuffle comes amid a broader talent migration. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le — four of Google’s most consequential engineers and researchers — are leaving to establish Discovery Loop, a public-benefit company focused on automating scientific and engineering breakthroughs. Google will invest in the new organization and supply cloud infrastructure.

Alphabet shares fell around 4% following news of the restructuring and departures. Wall Street generally prefers its geniuses safely locked inside the corporate castle.

Hassabis wants AI to earn its place in history

Hassabis will now spend more time at Isomorphic Labs, where the ambition is not merely to make drug development cheaper but to transform it from an expensive process of trial and error into something closer to computational engineering.

Consumer AI is becoming brutally competitive. Models improve, features are copied, prices fall and supposedly unique capabilities become commodities within months. Scientific AI is different. A system that discovers a viable cancer treatment, predicts an important biological mechanism or shortens clinical development by several years creates defensibility, social legitimacy and enormous economic value.

Hassabis has repeatedly argued that improving health should be AI’s most important application. His work on AlphaFold helped demonstrate that machine learning could solve scientific problems rather than merely imitate language, generate pictures or help office workers produce more PowerPoint slides. He and DeepMind researcher John Jumper shared part of the 2024 Nobel Prize in Chemistry for protein-structure prediction.

The strategic calculation is obvious: another marginally better chatbot may impress a benchmark leaderboard, but a successful drug can change millions of lives and produce billions of dollars in revenue.

Google’s reorganization therefore looks less like Hassabis stepping away from the AI race than moving toward its highest-stakes section. Kavukcuoglu will fight the model war. Hassabis will pursue the argument that AI deserves civilization-level investment because it can produce civilization-level results.

OpenAI loses a researcher to the telepathy business

Naomi Bashkansky resigned from OpenAI on July 23 and began work the following day as a founding researcher at Conduit. She had spent approximately 18 months at OpenAI, working on alignment research and contributing to internal projects concerning AGI and resilience.

In an essay titled “Why I’m leaving OpenAI to build telepathy”, she describes Conduit’s mission without the customary layer of startup modesty: “We’re building telepathy: thought-to-text models, trained on non-invasive neural data.”

Her forecast is aggressive. By 2027, she imagines workers wearing neural headbands that continuously transmit intentions to AI coding agents. By 2030, AI models could communicate directly with representations of thoughts rather than waiting for those thoughts to be converted into language. By 2035, she believes AI may feel less like an external tool and more like an additional sense or limb.

Bashkansky says Conduit is in the “GPT-2 era” of neural decoding: primitive compared with where the technology may eventually go, but far enough along to see scaling behavior. She estimates that a company capable of broadly reading from and writing to the human brain could eventually be worth more than $1 trillion.

It is the purest possible Gilded Age startup pitch: frighteningly ambitious, technically uncertain and attached to a market potentially larger than personal computing.

Conduit says it has already collected approximately 10,000 hours of neural and language data from thousands of participants, using custom headsets and a basement operation that ran for as long as 20 hours per day. Participants held conversations, typed and spoke while the company recorded non-invasive neural signals.

The resulting model is not producing perfect transcripts. It is attempting to identify semantic intent: the general meaning of what somebody is preparing to communicate. Conduit has published examples in which a participant’s phrase such as “the room seemed colder” produced a prediction referring to a breeze or gust. The wording was wrong, but the underlying concept was related.

The company claims its system can sometimes identify an idea shortly before the participant converts it into spoken or typed words. It has not yet published comprehensive technical details or independent validation of its model, saying that a fuller research release will come later.

That caveat is important. “Telepathy” is excellent marketing, but this is not yet a machine casually rummaging through somebody’s secret memories. It is statistical inference built from noisy neural measurements, linguistic context and a powerful language model that is already good at guessing what a person probably intends to say. The LLM is doing some of the same work a navigation map does when GPS reception is poor: narrowing a fuzzy signal into a plausible destination.

The science is real, but the hype is running ahead

Conduit is not alone in believing that scale will unlock non-invasive brain interfaces. Meta’s Brain2Qwerty v2 research used magnetoencephalography, or MEG, to decode sentences from the brain activity of nine participants as they typed. Each person contributed about ten hours of recordings. The system achieved an average word-error rate of 39%, while its best participant produced many sentences containing no more than one incorrect word.

That is impressive, but it is still a controlled experiment involving participants actively typing inside specialized equipment. MEG machines are hardly consumer headbands. They are large, expensive devices designed for laboratories and hospitals.

Portable EEG systems are more practical but produce far noisier signals. A recent Scientific Reports analysis found that several supposedly successful EEG-to-text systems performed similarly when supplied with random noise, suggesting that models were sometimes memorizing linguistic patterns rather than genuinely decoding brain activity.

So the honest verdict is neither “telepathy has arrived” nor “this is science fiction.” Non-invasive neural decoding is advancing. Large language models are particularly useful because they can transform incomplete signals into coherent possibilities. But the gap between recognizing broad intent under controlled conditions and continuously reading unrestricted thoughts in everyday life remains enormous.

The important question is whether massive datasets can close that gap. Conduit is betting that neuroscience will follow the same bitter lesson that reshaped AI: sophisticated handcrafted theories will eventually lose to enough data, enough compute and a sufficiently large model.

It is a bold bet. History suggests we should not dismiss bold bets merely because they initially sound ridiculous. History also suggests that founders describing trillion-dollar markets are not neutral observers.

Neural data could become the most valuable data on Earth

If Conduit’s approach works, the consequences will reach far beyond faster typing.

Today’s computers force humans to compress complex intentions into clicks, gestures and sentences. A brain interface could bypass that bottleneck. Instead of instructing an AI agent through carefully constructed prompts, a user might communicate uncertainty, curiosity, visual ideas and half-formed intentions directly.

That could make AI dramatically more useful. It could also produce the most invasive surveillance technology ever created.

Search histories reveal what people investigate. Social-media activity reveals what they publicly engage with. Neural information could reveal what they considered before deciding not to speak.

The governance problems are obvious. Who owns the raw signals? Can they be retained for model training? Can an employer require a productivity headset? Can an insurance company infer cognitive or emotional conditions? Can a platform use neural responses to optimize advertising before the user has consciously articulated a preference?

The most dangerous version of brain-computer technology may not be a government machine that perfectly reads minds. It may be a convenient consumer product that reads them imperfectly, stores the results indefinitely and buries consent inside a 40-page terms-of-service agreement.

That is how power usually enters society: not wearing jackboots, but offering free shipping and a better user experience.

The next platform war is over the human operating system

Hassabis moving toward AGI and computational biology, Google veterans launching an automated-discovery company and an OpenAI researcher defecting to build neural interfaces are not unrelated developments.

They indicate that the center of gravity is moving beyond general-purpose language models.

The largest laboratories have already trained systems capable of producing text, code, images, audio and video. The new competition is over what those systems connect to. Google wants AI connected to scientific discovery and medicine. Conduit wants it connected to intention. Meta is researching the neural machinery of language. Other companies are pursuing robotics, autonomous laboratories and synthetic biology.

The next monopolies may not be built around websites or smartphone operating systems. They may be built around the interface between intelligence and the physical human being.

That is what makes this an AI Gilded Age. The technology is genuinely extraordinary, and its benefits could be enormous. It is also being developed by a small number of extraordinarily wealthy institutions competing to own the infrastructure through which people work, think, discover and perhaps eventually experience reality.

The original Gilded Age laid steel tracks across continents. This one is laying invisible tracks between the mind, the machine and the molecule.

Whoever owns those tracks will not merely own the next great technology platform. They may own the route by which human intention enters the world.