The essay argues the recursive self-improvement feedback loop has moved from theoretical to measurable, citing public statements from Anthropic, OpenAI, and DeepMind researchers plus internal productivity signals suggesting 1.5x-2x throughput on research-engineering workflows. Tasks like writing CUDA kernels, paper reproductions, and triaging failed experiments that used to take a day now take minutes to hours.
Frontier-lab and ML practitioners in the thread confirm the essay matches their day-to-day experience — current models are completing engineering grunt work that previously consumed senior ML engineer time. Their testimony grounds the magnitude claims in lived workflow rather than speculation.
A faction of commenters accepts that models speed up engineering tasks but argues this conflates two different things: faster kernel-writing and sweep-running does not automatically translate into more novel insights or better theoretical breakthroughs. The bottleneck in research, they argue, has never been typing speed.
A smaller camp in the thread argues that essays like this — and the leaked internal estimates they rely on — are structurally useful to frontier labs seeking investment and regulatory positioning. The narrative that 'AI is accelerating AI' justifies capex and alignment urgency regardless of whether the underlying productivity gains are durable.
The essay is careful to hedge its own strongest claim: current models don't do research autonomously, they compress the time cost of research-engineering tasks for humans who direct them. That reframing — 'extremely fast interns who never sleep' — is itself a viewpoint, positioning the acceleration as significant but still human-gated rather than runaway.
A new essay on *Planned Obsolescence* (Helen Toner and co-authors' blog, which has become something of a required read for people who track frontier-lab behavior) argues that AI systems are now materially speeding up AI research itself. The post pulls together public statements from Anthropic, OpenAI, and Google DeepMind researchers, plus internal productivity signals that have leaked into talks and papers, and concludes that the feedback loop everyone has been theorizing about for a decade is no longer theoretical. It's happening, in measurable if messy ways, right now.
The specifics are worth reading carefully. Researchers at the frontier labs are reporting that Claude, GPT-5-class models, and Gemini are completing engineering tasks that used to eat a day of a senior ML engineer's time — writing CUDA kernels, implementing paper reproductions, running sweeps, triaging failed experiments — in minutes to hours. The post cites internal estimates (hedged, as they should be) of 1.5x to 2x throughput on certain research-engineering workflows, with some sub-tasks seeing much larger gains. That doesn't make the models autonomous researchers. It makes them extremely fast interns who never sleep, which is enough to matter.
The Hacker News thread (244 points) is more interesting than usual for a post of this type. The discussion splits roughly three ways: practitioners who say "yes, this matches what I see day-to-day"; skeptics who point out that research-engineering throughput is not the same as research progress; and a smaller camp arguing the whole framing serves lab PR. Notably, few of the top replies disagree on the underlying mechanism — the disagreement is about magnitude and what it implies.
The recursive self-improvement scenario has always been the uncomfortable base case in AI forecasting. Not the sci-fi version with glowing eyes, but the mundane one: better models make better tools for the people building the next models, which makes the next models better, which makes the next tools better. The interesting question was never whether this loop exists — it obviously does, humans have been doing it since the first compiler — but how tight the loop gets and how much of the work can be automated before the humans become the bottleneck.
What the post does well is separate the loop into pieces. Research has at least four distinct activities: generating hypotheses, designing experiments, implementing and running them, and interpreting results. AI is clearly accelerating implementation and running — the boring but massive middle. Interpretation is a mixed bag; models are good at summarizing results and bad at noticing when a result is actually weird. Hypothesis generation is where humans still dominate, and the post is honest about that. The frontier labs are not yet producing novel research directions via model prompts. They are producing papers faster because the engineering tax dropped.
That distinction matters because it predicts the shape of the next 18 months. If the bottleneck has moved from "write the training code" to "have the right research taste," then the labs with the deepest bench of senior researchers — not the ones with the most GPUs — pull ahead. This inverts a lot of recent conventional wisdom about compute moats. It also explains some of the frantic hiring of veteran scientists at eight-figure packages: not because they write more code, but because they have more defensible taste, and taste is now the gating input.
The skeptics in the HN thread raise a fair point: productivity gains in isolated tasks don't always compound into wall-clock research speedups, because real research is bottlenecked by compute, data, and waiting for experiments to finish. A 10x speedup in writing training code doesn't help if you still have to wait three weeks for a run. This is correct, and it's also why the labs are pouring money into infra: if coding is no longer the bottleneck, the next bottleneck is cycle time, and cycle time is a capex problem they know how to attack.
The deeper worry — barely mentioned in most reactions, which is itself notable — is evaluation. The pace of capability progress has already outrun the ability of public benchmarks to measure it; adding AI-accelerated research on top makes this gap wider, faster. The models that get released are evaluated on whatever was in vogue six months ago. The models being trained right now, with AI help, will arrive before the eval community has caught up. You can argue about whether this is dangerous; you can't really argue it isn't happening.
For practitioners outside the frontier labs, the implication is less about existential risk and more about tooling drift. The gap between what Anthropic-internal Claude-with-full-tooling can do and what you get from the public API is widening, because the internal version is being used to build the next internal version. If you build products on top of these APIs, assume the floor keeps rising faster than your release cycle, and architect accordingly: don't hard-code capability assumptions, keep the model layer swappable, and treat anything that works today as a baseline, not a ceiling.
Second, the research-engineering gains are leaking out, slowly. The same patterns that make internal ML engineers 2x faster — agentic coding harnesses, aggressive test-driven loops, model-driven debugging — are starting to show up in publicly available tools. If you haven't seriously tried running an agent on your own research or infra work in the last three months, your mental model of "what AI coding tools can do" is probably stale. The useful question isn't "can it write code," which was last year's question. It's "can it run my build, read the failure, and iterate without me." Increasingly, yes.
Third, hiring. If taste is the gating input at the frontier, it's also increasingly the gating input at every tier below. The developers who get leveraged up by these tools are the ones who already had strong opinions about what good code looks like; the ones who didn't are getting replaced by the tools, not amplified by them. This is uncomfortable to say out loud, and it's the thing every senior engineer I know has been quietly noticing for a year.
The honest forecast is not that AGI is six months away because of this loop — nobody serious is claiming that. The honest forecast is that the next model generation will arrive sooner than the previous cadence implied, that the generation after that will be weirder than any of us can currently predict, and that the labs themselves don't fully know what they're building because the thing building it is partially opaque to them. That is a reasonable thing to be unsettled by, and also a reasonable thing to keep shipping products against. Both can be true. Both are.
Every dot in these graphs is composed of multiple benchmarks smooshed together.Benchmarks are by definition a curated set of things that can be measured and evaluated.Things that can be measured and evaluated can be (expensively) trained on.The progress we are observing comes from figuring out how t
AI is accelerating my work...work which only exists because of AI.If I had to do my individual tasks without AI, I would be slower. But if I take a step back, I'm really not changing the world any faster than I was before.
Not trying to discredit the premise but I do wonder how much this is from vertical scaling of model intelligence in weights vs the sheer amount of capital being thrown at training and horizontal scaling of agent swarms etc.Anecdotally it seems like the majority of continued uplift comes from models
What exactly is driving all these capability improvements? Are they making more data? Using existing data more effectively? Modifying architecture? Changing the training somehow to pick up slack that was previously left behind?I’m kinda surprised that we expect the models to have infinite capacity t
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
As somebody who does ML research, LLMs certainly make ML research faster. Maybe even 10x in some cases, as the author suggests.But that’s just the parts of ML research that take place within your compute environment, namely experiments on model architecture, training strategy, UI/UX, etc. The a