Last week, Andrej Karpathy, the former head of AI at Tesla, published something that stopped me mid-scroll. His new open-source project, autoresearch, gives a coding agent a single GPU and a 630-line Python training file, then steps aside. The agent modifies the code, runs a five-minute training job, checks the score, commits the improvement, and loops, completely autonomously, indefinitely, without anyone in the room.
In its first run, the agent found roughly 20 real optimizations that Karpathy hadn’t caught manually. Broken attention scaling. Missing regularization. Each fix stacked on the last. The result: an 11% improvement on the “Time to GPT-2” benchmark. The chart shows 83 experiments, 15 kept improvements, a curve trending steadily down. Karpathy’s job, as he describes it, reduced to writing a program.md file that describes the research direction.
The researcher’s new job is writing the spec. Not the code, not the experiment, not the analysis – those are now delegated. The valuable work shifts upstream: defining the problem clearly enough that an agent can run with it. That’s a different skill set, and a different kind of leverage, than anything researchers have had before.
This Is Happening Everywhere at Once
What makes this moment different isn’t any single announcement, it’s the simultaneity. Just consider the developments over the past two months: in the same week that autoresearch shipped, Google quietly launched Workspace Studio to general availability. It lets any employee describe an automation in plain English and deploy an agent that reasons through their Gmail, Drive, and Calendar. No code.
24 hours before, Microsoft announced its own frontier suite built on M365 Copilot the same week. Anthropic shipped Claude Cowork, and OpenAI launched ChatGPT 5.4, their smartest model yet. Perplexity built a computer agent. 60 days ago, OpenClaw went live and very quickly there’s an entire ecosystem around the project with new Macbook mini being launched by Apple, a collaboration between Dell and Nvidia for a Mini PC created for agents, “ClawCon” meetups all over the world…
The “AI assistant” era, where AI helped you write things, is being replaced with the agentic era, where AI does things. And the infrastructure layer is no longer the frontier. The application layer is.
Why This Is the Founding Moment
The pattern that follows infrastructure commoditisation is consistent: the people who build the right applications at the right time create the most durable companies. It happened with cloud in 2008, mobile in 2010, and APIs in 2014. The window between “infrastructure mature enough to build on” and “application layer saturated” is where the best pre-seed bets are made.
We are inside that window right now. Foundation models have matured faster than anyone predicted. Agent frameworks are converging. The cost of running inference is falling monthly. But most of the value in enterprise workflows, vertical markets, and consumer behaviour hasn’t been captured yet, because the people who understand both the AI capability and the domain problem deeply enough to build the right thing are only just going stealth.
Where Pre-Seed Investors Fit
At pre-seed, we’re not buying traction. We’re backing the clarity of a founder’s thinking about a problem before the market has confirmed it. The most useful thing an early investor can do, and the thing we try to do at Remagine Ventures is to engage when the thesis is still forming. Before the architecture is set. Before the first enterprise conversation. When the question is still “what, exactly, are we building and for whom?”
That conversation is worth having three months before you think you’re ready to raise. The best founders we’ve backed came in with a sharp point of view and an open question. The investor’s job, in that moment, is to help stress-test the question, not to answer it.
We are actively looking for teams with technical depth plus operational instinct and willingness to talk to the market early in their journey. If you’re building something at the intersection of AI and a hard domain problem, and you’re six months into stealth wondering whether it’s too early to talk to investors: it isn’t.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- Weekly Firgun Newsletter – April 24 2026 - April 24, 2026
- Israel’s most promising startups in 2026 - April 22, 2026
- Bubble, or Super-Cycle? What the AI Boom Means for Founders Right Now - April 20, 2026