Andrej Karpathy has trained more neural networks than almost anyone alive — he co-founded OpenAI, ran Tesla’s Autopilot vision team, and can explain backpropagation from first principles without notes. Tom Blomfield has never shipped a model. He built a bank. Yet Anthropic hires Andrej Karpathy-caliber researchers and, in the same recruiting wave, a fintech founder with zero AI research background — then hands him one of the company’s most consequential technical bottlenecks: compute.
That pairing sounds like a mismatch until you look at what each man was actually hired to fix. One problem is scientific. The other is closer to logistics, financing, and operations at a scale most banks never touch. Anthropic is betting that solving 2026’s AI race requires both kinds of brains in the building at once.
Why Anthropic Hires Andrej Karpathy-Caliber Talent for Pretraining
Karpathy’s résumé reads like a highlight reel of the last decade of AI. He was a founding research scientist at OpenAI in 2015, left in 2017 to become Tesla’s director of AI and Autopilot Vision — the person responsible for teaching cars to see — then returned to OpenAI in 2023 for about a year before striking out on his own. In 2024 he founded Eureka Labs, an AI-focused education startup, and spent two years mostly out of the corporate lab system entirely.
On May 19, 2026, he announced he was joining Anthropic. The company placed him on its pretraining team — the group responsible for the massive training runs that give Claude its core knowledge before any fine-tuning happens — with a specific mandate to build tooling that uses Claude itself to accelerate pretraining research. It’s a neat bit of recursion: the model helping design the process that built the model. If you want to see what that team’s work actually ships, our breakdown of what Claude Sonnet 5 means for developers goes deep on the model his group now helps shape.
Monzo, GoCardless, and Why Tom Blomfield Doesn’t Fit the Usual AI-Hire Mold
Blomfield’s background is about as far from a research lab as it gets. He co-founded Monzo, the UK challenger bank that turned a hot-coral debit card into one of Europe’s most recognizable fintech brands, and before that GoCardless, a recurring-payments infrastructure company. After stepping back from Monzo, he spent time as a general partner at Y Combinator starting in 2023, mentoring early-stage founders rather than building models.
On July 13, 2026, Blomfield announced he was taking a leave of absence from YC to join Anthropic’s compute team, working under co-founder Tom Brown — Anthropic’s Chief Compute Officer, as first detailed by Tech Funding News. His stated reasoning was blunt: as the industry enters what he called the early stages of recursive self-improvement, the availability of compute becomes one of the most important problems in the field to solve, arguably more binding than any single research breakthrough.
Here’s an interesting tangent worth sitting with: Monzo didn’t win UK banking by writing cleverer software than the incumbents. It won by obsessing over operational bottlenecks — onboarding friction, card manufacturing logistics, support queues — the unglamorous plumbing that decides whether a good idea actually scales to millions of users. That’s arguably the exact skill Anthropic is buying. Not machine learning intuition, but the instinct to find the boring constraint that’s secretly the whole game.
The Real Reveal: Compute Stopped Being Purely an Engineering Problem
This is the part that surprises people who assume AI progress is bottlenecked by algorithms. Increasingly, it isn’t. Securing enough chips, negotiating power contracts with utilities, structuring multi-billion-dollar capital commitments for data centers years before they’re needed — none of that is a machine learning problem. It’s finance, logistics, and negotiation at a scale that has more in common with running a bank’s operations than training a model.
Blomfield isn’t the only signal here. In April 2026, Anthropic hired Eric Boyd away from Microsoft, where he’d led the Azure infrastructure that actually hosted Anthropic’s models under the companies’ cloud partnership — poaching the person who understood their own compute stack from the outside. Putting a payments-and-operations founder and an infrastructure executive on the same team tells you compute has become as much a business function as a technical one.
Anthropic Hires Andrej Karpathy — and Isn’t Stopping There
Karpathy and Blomfield are the two names getting headlines, but they’re part of a longer list. In June, Anthropic pulled Nobel laureate John Jumper away from Google DeepMind after nearly nine years, betting on AI-for-science work involving the Allen Institute and Howard Hughes Medical Institute. DeepMind researchers Jonas Adler and Alexander Pritzel reportedly followed. On July 1, Berkeley’s CS department chair Jelani Nelson joined too. The pattern isn’t “hire good ML engineers” — it’s “hire the best person in any adjacent field and figure out where they fit.”
It’s paying off in a metric that’s easy to overlook: retention. Anthropic’s two-year staff retention rate reportedly sits around 80%, compared to 78% at DeepMind and just 67% at OpenAI — a gap that compounds badly when your competitors are this aggressive about poaching.
The Rest of the Industry Is Bleeding Talent, and Paying Absurdly to Stop It
Zoom out and the Karpathy-Blomfield story is one skirmish in a much uglier war. Meta reportedly offered researcher Andrew Tulloch a package worth up to $1.5 billion over six years — he turned it down. It reportedly landed AI researcher Ruoming Pang for around $200 million and offered 24-year-old Matt Deitke roughly $250 million, pulling him out of a PhD program entirely. Sam Altman claimed Meta dangled $100 million signing bonuses at OpenAI staff, a figure poached researcher Lucas Beyer publicly called “fake news,” with Meta clarifying the eye-watering totals were multi-year packages reserved for a handful of senior leaders.
Whatever the real number, OpenAI felt the need to respond with retention bonuses of roughly $1.5 million for about a thousand staff, running as high as $5 million for its most senior researchers — and still reportedly lost eight researchers to Meta’s superintelligence group by the end of June.
OpenAI, for its part, isn’t sitting still on the product side either — we broke down what its newer Sol/Terra/Luna-era model family actually changes in our GPT-5.6 explainer, worth a read if you want to see what the other side of this fight is shipping while it bleeds staff.
It’s a strange thing to sit next to this week’s other big AI story: the Future of Life Institute’s 2026 AI Safety Index, which gave Anthropic the best grade of any frontier lab — a C+ — while OpenAI and Google landed C’s and Meta scored a D+. We covered the full rankings here. A company can be topping safety leaderboards and fighting a bare-knuckle recruiting war over chips and compute at the same time — those aren’t contradictions. They’re the same organization solving two different existential problems in parallel.
What This Actually Tells You
Karpathy’s hire made headlines because it was easy to understand — the person who helped invent modern AI research joining the company trying to build the next version of it. Blomfield’s hire is more interesting precisely because it doesn’t make sense at first glance, and that’s usually the tell for where an industry is actually headed. When the bottleneck stops being “who can design a smarter architecture” and starts being “who can secure enough power and silicon to run one at scale,” the most valuable hire on your roster might not have an AI background at all.
Keep watching who these labs hire next. The job titles are turning into a leading indicator of what each one thinks it’s actually short of — and right now, apparently, that’s fintech operators as much as it is machine learning PhDs.
Deixe um comentário