Hot take: Modern AI models could probably pass an interview for a junior AI researcher position, get put on a performance improvement plan within their first month or two, and be fired after a couple more months.
This observation might sound like a criticism of current AI capabilities, but it's actually the key to understanding why recursive self-improvement could be much closer than most people expect.
If an AI can pass the interview and survive a few weeks before struggling, that puts it squarely in the range of "actual but mediocre human researcher" - not in some completely different category. We're talking about performance that's recognizably in the same distribution as human junior researchers, just toward the bottom end.
This matters enormously for recursive self-improvement timelines. The traditional framing often assumes we need AI to reach some much higher capability threshold - like "superhuman researcher" level - before recursive self-improvement kicks in. But if current models are already in the range where they could plausibly contribute to AI research teams, even if they need supervision and produce lower-quality work, that changes the timeline math significantly.
Recursive self-improvement might not require waiting for AI to become the world's best researcher. It might just require AI to become good enough to be a net positive contributor to AI research - and then deploying that capability at massive scale.
Consider this: even if each AI researcher is notably worse than the median human researcher, 10,000 of them working simultaneously would dwarf the research capacity of OpenAI, Anthropic, DeepMind, and all the academic labs combined. And they'd be working 24/7 with no coordination overhead, salary costs, or human limitations.
At that point, it's not about individual AI researchers being particularly brilliant. It's about having an overwhelming research apparatus that can try thousands of different approaches in parallel, rapidly iterate on promising directions, and scale up successful experiments immediately.
The difference between "can kinda fake their way through basic AI research" and "half-way decent junior AI researcher" is not that big - probably smaller than the difference between GPT-2 and GPT-4.
GPT-2 could barely maintain coherence over a few paragraphs, while GPT-4 can engage in complex reasoning, coding, and analysis. If current models are already at "can kinda fake their way through basic AI research," then we're potentially talking about one or two more major model generations to reach "half-way decent junior AI researcher."
The beautiful thing is that this approach plays to AI's strengths while working around its weaknesses. You could assign different groups to:
Current AI already participates in these activities - it just does them poorly. It can debug a distributed training run that's not making progress, suggest learning rate adjustments, hypothesize about data issues. The problem isn't that it can't participate; it's that it's inefficient and often wrong.
But going from "participates poorly" to "participates adequately" is typically much easier than going from "can't participate at all" to "participates poorly." We've already crossed the harder threshold.
This suggests that at most in two generations, humans will be mostly coordinating swarms of automated AI researchers, not doing research themselves. The economic incentives become overwhelming - why hire a team of 20 human researchers when you can deploy 1,000 AI researchers for a fraction of the cost?
Humans would naturally shift into orchestration roles: setting research priorities, evaluating promising directions, making high-level strategic decisions. The actual grunt work of implementation, experimentation, and systematic exploration gets handled by the swarms.
This timeline helps explain why figures like Dario Amodei are staking their reputations on "powerful AI" arriving by 2027. If you can see that current models are already in the "can kinda do AI research" range, and you know roughly what the next two generations of improvements will bring, then "competent AI researcher swarms" by 2027 becomes a very reasonable prediction.
Once you have AI systems that can meaningfully accelerate AI development - even if they start out mediocre - you've created a feedback loop that becomes very hard to predict. Each improvement makes the AI researchers more capable, which accelerates the next round of improvements, which makes them even more capable.
The compounding effects could be dramatic. Instead of AI progress being limited by human researcher bandwidth, you'd have 24/7 research cycles, massive parallel exploration, and rapid iteration at scales impossible for humans.
We might not be waiting for some distant breakthrough, but for an economic tipping point where you can afford to run thousands of AI researchers on your problem. And given how fast compute costs are dropping and AI deployment is scaling up, that tipping point might be surprisingly close.
The threshold for recursive self-improvement might not be "AI becomes superhuman at research." It might just be "AI becomes cheap enough to deploy at massive scale for research tasks." And we may already be much closer to that threshold than most people realize.