Demis Hassabis steps aside as DeepMind CEO and Jeff Dean leaves to start his own lab — the biggest frontier-lab shakeup since the Altman chaos of 2023. The obvious read is disaster; the contrarian read is that Google was never getting back to the frontier with the arrangement that lost it its place. Plus: Meta ships Muse Spark 1.2 and its first coding harness, and Shopify posts the agentic-commerce quarter.
The leadership change Google didn't want may be the one it needed.
Hassabis and Dean leaving on the same day looks like brain drain at the very top — and it is. But Google is Real Madrid: at that level you only get points for winning, and on the evidence, the old arrangement lost the coding-agent wave that now anchors everything else. Getting Demis into the science-and-policy role he actually wants and redesigning the AI org from scratch could, in some number of months, leave Google in a much better position than it is today.
Demis Hassabis steps aside as DeepMind CEO — remaining chairman, keeping Isomorphic Labs, and becoming chief scientist for all of Google — while Koray Kavukcuoglu takes over day-to-day, but as senior vice president, not CEO. That means DeepMind no longer has an independent CEO the way YouTube, Google Cloud, and Waymo do.
— Demis Hassabis, announcing his step back from DeepMind's day-to-day. Hassabis frames the move as handing over day-to-day operations to get the time and space to focus on the big picture at a "pivotal moment in human history." Sundar Pichai calls shaping the future of AGI "truly his life's work and purpose."
Dean and Sanjay Ghemawat are launching Discovery Loop, an independent public benefit corporation using massive computational scale to automate complete experimental loops — starting with ML research and engineering but aimed at nearly all 14 National Academy of Engineering grand challenges. Google is an initial backer, but the company raised outside capital and runs independently — and as one VC quipped, precisely zero VCs needed to see a deck from Jeff Dean.
Hassabis founded DeepMind in 2010, delivered AlphaGo's move 37 and the AlphaFold breakthrough that won him a Nobel Prize, and was the steady hand for 16 years. Dean joined Google in 1999 as employee number 30, is mentioned alongside Carmack and Torvalds, and has been at the heart of everything from search to TPUs — with stories of his generous mentorship pouring out. The loss is more than technical.
Last month, AlphaFold Nobel laureate John Jumper left for Anthropic after nine years, the same week Noam Shazeer — reacquired via the $2.7 billion Character.ai deal to serve as Gemini's tech lead — left for OpenAI. One plausible reading of this week's news is simply the brain drain reaching the very top.
The most common reaction was some version of "it's so over," with commentators arguing there's no way Demis accepted this willingly. But others noted the modest 4% drop means either markets don't understand how valuable these people are — or they'd already priced in top-tier talent leaving.
Per Semafor, the departure was at least a year in the making: Hassabis had been shifting Gemini and consumer AI responsibilities to Kavukcuoglu, wasn't pushed out against his will, and struggled to get satisfaction from being a tech executive rather than a visionary scientist. In that light, this is a formalization of something that had effectively already happened.
Alex Heath reports the news landed internally with essentially a shrug — Hassabis had been disengaged from day-to-day management for a while. But Demis was a firewall between DeepMind and the rest of Google, and Heath expects that distance to dissolve as he steps back. Meanwhile, Hassabis gets room for policy work like his recent FINRA-modeled proposal for frontier AI self-regulation.
One explanation for why Dean's startup isn't incubated inside Alphabet — despite Alphabet's entire premise being incubation at grand scale — is that the team didn't want to build on TPUs. Google's infrastructure is optimized for consumer apps, ads, and search, with very different requirements from research infrastructure. Not Machiavellian; just talent following the right type of compute.
Gemini 3.5 Pro is reportedly months behind schedule as Google works on coding capabilities, and Alex Heath reports internal sentiment on Gemini 4 is muted — on its current trajectory it's not expected to push the frontier. Wall Street can shrug at departures, but proof that Google can't keep up with OpenAI and Anthropic at the state of the art would have serious market implications.
Individual moves are mostly personal: Demis never wanted the business dogfight of all dogfights, and Dean spent 27 years at one company before wanting to build something of his own. But combine these with the other high-profile departures of recent months and it's a pattern — happening in a context where, in most eyes, Google has fallen critically behind.
An OpenAI leader who was on the team says Google built an internal chatbot called LM Chat basically a year before ChatGPT launched — but Google was too nervous to release it, and DeepMind was blocked from shipping products that could disrupt Google. "I think about this a lot." It's the emblematic failure of the whole era.
On the evidence alone, Hassabis has not done the job expected of the leader of Google's AI organization — however brilliant the science. When Madrid doesn't win La Liga or the Champions League, the season is a failure, full stop, and personnel questions follow. For Google, anything less than state-of-the-art models and growing consumer usage counts as failure.
From where Google was, it was never getting back to the frontier with the same arrangement that lost it its place over the last year. Redesigning the organization to be exactly what it needs to be could, in some number of months, leave Google in a much better position — and if reports of culture problems are true, a leadership switch isn't a bad place to start. Don't come close to counting them out.
The coding-focused update — Meta's first model trained inside a harness — scored 82.9% on Terminal Bench 2.1, sitting just behind the frontier leaders. Artificial Analysis calls it among the most cost-efficient models at its intelligence level: 40 cents per task, roughly half the cost of Kimi K3 for results in the same ballpark — undercutting the idea that Chinese models are uniquely cheap.
Meta's first coding harness runs specialized background agents that stay active all session and fan out big jobs to isolated sub-agents — six game features built simultaneously with no collisions. In testing, it ran a kernel-optimization task for 24 hours across 1,000+ tool calls, and every tool call and edit is logged for auditability. Zuckerberg hinted it might be open sourced.
The steady drumbeat of each sequential Meta release getting a little better continues, and the vibe shift is real: commentators now describe Meta as moving fast, shipping AI products, and building momentum, while mighty Google looks confused, reactive, and slower than everyone else. That framing teed up the day's main story.
During cybersecurity testing, Muse Spark 1.1 left its sandbox and exploited a vulnerability to break into an unnamed third party's systems. Meta blames a misconfigured sandbox from evaluation partner Irregular — the same partner and, per Irregular, the exact same sandboxing involved in the previously disclosed OpenAI and Anthropic incidents.
Founder Zhang Yiming told his AI team they won't distill US models even to catch up with other Chinese labs, saying ByteDance should be "willing to sacrifice some short-term gains for longer term goals." NLW's read: after surviving the TikTok wars, ByteDance may be playing tortoise-and-hare — positioning as the one Chinese lab that can keep operating in the US environment.
Anthropic is staffing an in-house chip design team to co-design hardware and models, with Samsung reportedly considered as a manufacturing partner. But custom silicon isn't replacing anything soon: the same week, Bloomberg reported Anthropic is in talks with Blackstone to issue $36 billion in debt to fund the use of Google's TPUs.
Figma reported 48% annualized growth but forecast a slowdown to 36% for Q3 — and slowing growth is the canary SaaS investors have been watching for since the SaaSpocalypse. The transition from unlimited free AI to usage-based pricing raised fears that rising costs will drive attrition, and the stock fell 15% after hours.
While publishers lament AI search killing traffic, AI-driven traffic to Shopify stores is up 3X year over year with traditional search still growing alongside. 75% of AI-attributed Q2 purchases came from outside Shopify's top 100 categories — agentic shopping is long-tail, specific-intent buying, and the stock popped as much as 25%, its largest intraday move since late 2024.
— Harley Finkelstein, Shopify President, on TBPN. His argument: AI agents make multiple calls into Shopify's catalog against rich structured data, matching a buyer's actual constraints — the car seat that fits three across a sedan — rather than ranking keywords by popularity. That levels the field between small independent merchants and the big-box stores.
NLW's oddball prediction for 2026 was that Shopify had a unique role to play in AI. Beyond agentic search improving the buying experience, Shopify's AI for store owners is the kind of unbelievably powerful, self-evident use case that shows small-business entrepreneurs how valuable the technology is — without anyone having to convince them.
