Last Tuesday I was reviewing a deal memo for a YC company building developer tools. Solid team, good traction, reasonable valuation. The kind of deal I’ve seen hundreds of times. But midway through the call, the founder said something that stopped me cold. “We used Claude Code to build about 95% of the product. Our AI researcher agent designed three of our five core experiments. And last week, the model we were fine-tuning started suggesting changes to its own training pipeline that actually worked.”
I wrote it down in my notebook and circled it twice. Not because it was surprising. Because it was the third time I’d heard something like that in a single month.
Something has shifted. Quietly, across multiple frontier AI labs and in the garages of solo developers, we have entered the recursive self-improvement phase of artificial intelligence. The machines are helping build the next version of themselves. And the implications for startups, for venture capital, and for how society organizes itself around technology are enormous.
Let me be specific about what I mean by “recursive self-improvement,” because the term gets thrown around loosely.
It doesn’t mean Skynet. It means an AI model can now meaningfully participate in the research process that creates the next AI model. A frontier model designs an experiment. It writes the code to run that experiment. It analyzes the results. It proposes the next experiment. A human researcher still steers the direction and reviews the output, but the loop is getting tighter. The human’s role is shrinking from driver to navigator to, increasingly, someone who checks the map every few hours.
The Chinese AI lab MiniMax released their 2.7 model recently and said something remarkable in their technical write-up. They described M2.7 as “our first model deeply participating in its own evolution.” The model updated its own memory, built dozens of complex skills in its research harness, and improved its own learning process based on experimental results. They estimated AI was handling 30 to 50 percent of the overall research workflow.
OpenAI said something similar when they released GPT 5.3 Codex. They called it “our first model that was instrumental in creating itself.” Early versions of Codex were debugging its own training, managing its deployment, and diagnosing test results. Then early checkpoints of the new model were helping optimize later checkpoints of the same model. The model was not just assisting in research. It was improving itself in real time.
Anthropic has been quieter about it, as is their style, but the evidence is everywhere. They’ve described autonomous loops where Claude Code writes code for a new feature, runs tests, and iterates continuously. Their engineers give the model abstract problems, let it work autonomously, then review the solutions before final refinements. They’ve said Claude Code has become “far more than a coding tool” and that it now powers “almost all of our major agent loops.”
And then there’s Google’s AlphaEvolve, which discovered faster matrix multiplication for the first time in roughly 50 years. Every AI model trained after that discovery runs faster. That is recursive self-improvement made tangible.
Here’s what makes this moment different from previous AI hype cycles. It is no longer confined to labs with billion-dollar GPU clusters.
Andrej Karpathy, the former head of AI at Tesla and one of the founding members of OpenAI, recently open-sourced a project called Auto Research. The concept is straightforward. You point a frontier model at a training task. The model designs experiments, runs them, analyzes results, proposes new experiments, and repeats. Karpathy ran it overnight and woke up with the fastest training time for a GPT-2 class model on Earth.
Individual developers with no machine learning background are now doing autonomous research. One developer I follow has been using an open-source agent framework to fine-tune small open-source models, benchmarking them against Opus 4.6 as the target. The system runs all night. If a fine-tuned model outperforms the frontier model on a specific task, it gets swapped in. If it doesn’t, the system analyzes what went wrong, generates new synthetic training data, picks a different base model, and tries again. No human intervention required between iterations.
When someone with no formal ML training can run autonomous AI research experiments from their laptop, the old barriers to entry in AI development have collapsed. This is not a drill.
So what does recursive self-improvement mean for the startup ecosystem? I’ve been thinking about this nonstop, and I keep landing on a few things.
First, the shelf life of technical moats just got dramatically shorter. If you are a startup whose primary advantage is that you trained a model that does something well, that advantage may evaporate in weeks rather than years. When AI can improve itself, the cost and time required to replicate any given capability drops toward zero. The models will converge on capability faster than any startup can build a brand around being the only one who can do the thing.
This doesn’t mean AI startups are a bad bet. It means the nature of the bet has changed. The value is migrating from “we have a better model” toward “we have proprietary data that makes any model better at this specific problem” and “we have distribution and workflows that are hard to replicate.” The startup that wins won’t be the one with the best algorithm. It will be the one embedded most deeply in a customer’s daily operations, feeding data back into a loop that improves the product in ways a generic model cannot.
Second, the speed of company-building is going to accelerate in ways that make the last two years look slow. We already saw what happened when GPT-4 landed. Startups that would have taken two years to build an MVP were launching in two months. Now layer on autonomous coding agents that can ship features without human engineers and research agents that can run experiments without human researchers. The compression is going to continue. I would not be surprised to see companies go from idea to meaningful revenue in weeks, not months.
For founders, this is both thrilling and terrifying. The same tools that let you build faster let your competitors build faster too. And it means the bar for what constitutes a “real” startup is going to keep rising. Building a product is no longer the hard part. Finding a real problem, earning trust, and distributing the solution to people who will pay for it are the hard parts. Those have always been the hard parts, honestly, but the noise around them is about to get a lot louder.
For investors, the recursive self-improvement era poses questions that the industry has never had to answer.
The traditional venture capital model relies on a basic assumption. The best companies take years to build, which means investors have time to evaluate, fund, support, and exit them on a predictable timeline. But what happens when the best companies can be built in months? What happens when a competitor can emerge from nowhere with an autonomous agent that replicates your portfolio company’s core feature overnight?
I don’t think venture capital goes away. Far from it. What I think happens is that the role of the investor shifts. If building the product becomes trivially easy, then capital alone is no longer a competitive advantage. The investors who thrive will be the ones who bring genuine distribution, domain expertise, and operational support to their portfolio companies. Money buys you a ticket into the arena. What you do once you’re inside is what matters.
There’s another dimension too. Valuation cycles are going to compress. When a company can scale from nothing to real revenue in weeks, the price of getting in early increases dramatically. And the penalty for waiting increases just as dramatically. For smaller, nimble funds that can make decisions quickly, this could be an enormous advantage. The big, slow, consensus-driven funds are going to struggle with the pace.
I run early-stage funds that write small checks into rounds at a blistering pace. The advantage of that model has always been speed and volume. Get into a lot of high-potential companies early, before anyone knows which one is going to break out. In a world where breakout happens faster, that advantage compounds.
Now let’s zoom out. Because the implications extend well beyond startups and term sheets.
Leopold Aschenbrenner, a former researcher at OpenAI, wrote a widely-circulated paper in 2024 called “Situational Awareness.” In it, he laid out a graph that I keep coming back to. On the x-axis is time. On the y-axis is effective compute, which you can think of as a rough measure of AI capability. The line is flat for a long time. Then it hits a point where automated AI researchers become possible. And after that point, the line goes vertical.
We are at that inflection point right now. Not approaching it. Standing on it.
Sam Altman laid out OpenAI’s internal goals last October. An automated AI research intern by September 2026. A true automated AI researcher by March 2028. Based on everything I’m seeing in the market, they are well ahead of that timeline. Arguably, the research intern already exists.
What happens to the labor market when AI researchers can be automated? What happens to software engineering as a profession when autonomous agents write, test, and deploy code without human oversight? What happens to entire industries when the bottleneck shifts from human talent to available compute?
I don’t have neat answers to these questions. Nobody does. But I will say this. Every previous technology revolution, from the printing press to the internet, created more jobs and more prosperity than it destroyed. The transition periods were rough. People whose skills became obsolete suffered real pain. But the aggregate outcome was overwhelmingly positive.
I believe that will hold true here, but with a caveat. The speed of this transition is unlike anything in human history. The printing press took decades to reshape society. The internet took about 15 years. AI, in its recursive self-improvement phase, might reshape the world in 3 to 5 years. That is not a lot of time for individuals, institutions, or governments to adapt.
A few weeks ago I was talking with a founder who had just gone through YC. Sharp young mind, building something interesting in the AI infrastructure space. I asked him what surprised him most about the batch. He thought for a second and said, “How many of us are building companies that might make our own products obsolete within 18 months.”
He wasn’t scared when he said it. He was energized. Because he understood something that a lot of people haven’t internalized yet. In a world of recursive self-improvement, the goal is not to build something permanent. The goal is to build something that improves faster than the alternatives. That’s true for companies. It’s true for investors. And it might be the defining challenge for all of us.
The machines are training themselves now. The question is not whether that changes everything. The question is whether we are paying attention fast enough to keep up.