From Hinton to Dean: Google's AI Talent Diaspora Jeff Dean is one of the true legends in the field. When people of that caliber leave to start something new, talent follows. Discovery Loop (Dean + Sanjay Ghemawat + Oriol Vinyals + Quoc Le) is already proving the point — reports this week put them in talks for ~$1B at a ~$10B valuation while still in the earliest stage and just beginning to hire. This is not a new pattern. It is the acceleration of one I covered in AI L4IR. Geoffrey Hinton left Google in 2023 so he could speak freely about AI risks. The original Transformer authors all eventually left. In the more recent wave we have seen clear directional flows: multiple high-profile researchers (including Nobel laureate John Jumper and key Gemini/AlphaFold contributors) to Anthropic; Noam Shazeer and others to OpenAI; and researchers such as Denny Zhou and Dawn Song to Meta's Superintelligence efforts. Some have even moved beyond the first destination — Shazeer being the clearest serial case (Google → Character.AI → back to Google → OpenAI). Every major lab has now either lost or gained from the concentrated pool Google spent years building. Pure-play labs, companies like Cohere, and Chinese efforts all benefit when deep expertise becomes more distributed. The result is faster diffusion of systems knowledge, research taste, and experimental methods — which compresses gaps even as competition intensifies. For Google, this can still be a useful crisis if treated that way. Visible high-profile exits force the hard questions: Are we fast enough? Is ownership and focus real enough inside the walls? Is compute allocation clear? The organizations that treat talent loss as pure damage usually keep losing. The ones that use it as a forcing function to raise the internal bar often come out sharper on velocity and retention. Google still has residual depth, infrastructure, and the ability to back the people who leave (as it is doing with Discovery Loop). The test is whether the new leadership structure and the pressure from these exits produce a genuine step-up — or whether the pattern simply continues. Legends leaving doesn't just create competitors. It raises the competitive temperature for everyone, including the place they left.

**Distillation, Efficiency, and State Power: How China Is Closing the AI Gap — and Why the US Is Moving to Restrict Chinese Models** Caught between two superpowers, many scientists and business leaders today are carrying a burden that rarely gets discussed publicly. As the United States and China compete for leadership in artificial intelligence, the people actually building these systems are finding themselves in an increasingly difficult position — one where technical excellence, commercial pressures, ethical responsibility, and national interests are pulling in different directions at once. In my latest essay, I explore how Chinese AI labs have managed to close the performance gap with the West despite facing significant restrictions on advanced chips. Through highly efficient architectures like Mixture-of-Experts, aggressive quantization, and deep hardware-software optimization for domestic accelerators, they are delivering frontier-level capabilities at a fraction of the cost. At the same time, close coordination between the Chinese government and its leading AI companies has created structural advantages in data access, strategic focus, and speed of execution. This progress has not gone unnoticed in Washington. Concerns about large-scale model distillation, the stripping of safety guardrails, and the potential national security implications of widely adopted Chinese models have led to growing calls for restrictions on how U.S. organizations use these systems. The debate is no longer just about chips and compute — it now extends to which models American companies and institutions should be allowed to rely on. What often gets lost in these strategic discussions is the human dimension. Engineers, researchers, and executives are being asked to navigate an environment where every technical decision can carry geopolitical weight. The pressure to move fast, stay competitive, and deliver value to customers exists alongside growing expectations around security, ethics, and alignment with national interests. For many, this creates a quiet but real strain. The image attached tries to capture something of this tension. I wrote this piece because I believe we need a more honest conversation about the real trade-offs involved — not just in terms of technology and policy, but in terms of the people doing the work. If we want responsible AI development, we need to create conditions where those building it can operate with clarity rather than constant conflict. You can read the full essay here: https://kentkaufman.substack.com/p/distillation-efficiency-and-state I'd be interested in your thoughts — especially from those working directly in AI development or policy.

The AI Frontier Race Just Got Rewritten — Grok 4.5 Forces a New Tradeoff Between Intelligence, Cost, and Speed The rules of the frontier AI race are changing in real time — and this week made that unmistakably clear. For the past two years, the prevailing assumption was straightforward: the company with the highest benchmark scores would eventually dominate. That assumption is breaking down. We're moving from a world where raw intelligence was the primary differentiator to one where the real competition is about who can best balance intelligence, cost, and speed at scale. Grok 4.5 is the clearest signal of this shift. While it doesn't lead every benchmark, it delivers near-frontier performance with significantly better token efficiency and much lower pricing — proving more economically attractive than pricier, higher-scoring alternatives in many real-world agentic and coding workflows. xAI is betting the winning formula is no longer "maximum intelligence at any cost," but strong intelligence at the lowest cost per successful outcome. OpenAI reached a similar conclusion. Instead of releasing GPT-5.6 as a single model, they launched it as a three-tier family (Sol, Terra, and Luna) — an acknowledgment that different customers need different points on the intelligence-cost-speed spectrum. This reflects a deeper shift: the AI TAM is stratifying into distinct sub-markets — high-intelligence/low-volume use cases, cost-efficient/high-volume agentic work, specialized coding agents, knowledge work automation, robotics, and more. Companies that fail to define which sub-markets they're targeting risk misallocating resources in an increasingly differentiated landscape. Structural advantages matter more too. xAI's integration into the broader SpaceX ecosystem — the Colossus supercluster, in-house chip plans, orbital data center ambitions — gives it a path to lower long-term compute costs pure software labs will find hard to match. Three distinct approaches are emerging among the leading U.S. labs: - Anthropic: maximum capability and safety, under greater regulatory constraints. - OpenAI: serving multiple segments through tiering while navigating government review. - xAI: fastest on public release, leaning hardest into efficiency, cost, and vertical integration. The winners of this next phase won't necessarily have the smartest model on every benchmark — they'll be the ones who best combine intelligence, cost, and speed while building the infrastructure to deliver it economically at scale. The game has moved from a simple 2D board to 3D chess. Full analysis and benchmark comparison (Grok 4.5, Fable 5, GPT-5.6, Gemini, and others): https://kentkaufman.substack.com/p/openais-gpt-56-strikes-back-against How do you see this playing out over the next 12–18 months?

Anthropic's Fable 5 & Mythos 5 – Still Disabled, But the Story Is More Nuanced Than Headlines Suggest - Update 6/19/2026: On June 9, Anthropic released Claude Fable 5 — its first publicly available "Mythos-class" model — along with the more powerful Mythos 5. Just three days later, on June 12, the U.S. government issued an export control directive requiring Anthropic to block access for all foreign nationals. To comply, Anthropic disabled both models globally for every user. As of today (June 19), access remains fully disabled. Anthropic has described the situation as a misunderstanding and is actively engaged with the Commerce Department to restore availability. The company has also floated proposals for closer ongoing coordination with the administration on security matters. The triggering event was research — reportedly shared by Amazon — showing a prompting technique that could bypass some of Fable 5's guardrails. However, cybersecurity experts who reviewed the findings (including former government advisor Katie Moussouris) have clarified that this was NOT a traditional high-risk jailbreak. It was a narrow "Defense Oriented Prompting" method — essentially asking the model to "fix this code." The technique surfaced only minor, previously known vulnerabilities and did not demonstrate novel or catastrophic misuse potential. This episode highlights the growing tension between frontier AI capability, export controls, and responsible deployment — with real implications for companies building or using advanced AI systems. I've published a deeper analysis covering the timeline, the technical details behind the reported technique, the government response, and what this means for AI governance and organizations navigating these risks. Full piece here: https://kentkaufman.substack.com/p/from-mythos-to-fable-and-now-to-government What are your thoughts on how export controls should evolve for frontier models? I'd value perspectives from those working in AI, security, or regulated industries.

Frontier Models