Calls Anthropic's justification 'hideous' — argues that inevitability arguments are a coward's logic when the people deploying them are precisely the ones with the most agency over whether the dangerous thing happens. The framing offloads moral responsibility onto a hypothetical competitor while Anthropic itself accelerates the very dynamic it claims to fear.
Reads the document as investor-facing positioning timed to Anthropic's public-offering preparations. The 8× productivity headline and the 'frontier lab' framing function as growth-narrative inputs for institutional buyers, regardless of how the safety content is dressed.
Argues the LOC caveat is a non-answer: the central problem isn't the magnitude of AI-generated code but its quality distribution. Rounding down a flawed proxy still leaves you measuring the wrong thing, and the quality question is precisely what RSI claims hinge on.
Frames the LOC critique as the 'easy take' that's already been made, and pushes past it to ask what RSI actually means inside a frontier lab. Treats the metric debate as a distraction from the more important question of how inevitability framing shapes policy.
Reports his own 'agentic iterative optimization' experiments — asking models to speed up real benchmarks under strict regression and quality constraints (MSE, output equivalence). The loop produces modest, real improvements, but the work is being done by the eval harness keeping the model honest, not by the model getting smarter across iterations. This undercuts the implicit RSI narrative that the model itself is the engine of acceleration.
The memo's core argument: frontier recursive self-improvement is going to be pursued by someone, and a safety-focused lab being at the frontier is preferable to ceding that ground. The 8× engineer productivity claim is offered as evidence that the loop is already underway internally, justifying continued investment despite acknowledged metric weaknesses.
Anthropic published a memo titled *When AI Builds Itself: Our progress toward recursive self-improvement*, framing the company's internal use of Claude to accelerate Claude's own development. The headline number — an estimated 8× lines of code per engineer per day in Q2 2026 — comes with a footnote conceding that LOC is a bad proxy for productivity, and that the number is therefore 'rounded down.' The HN thread (473 points) latched onto the metric, but the more substantive content of the memo is the *justification*: that frontier RSI work is going to happen, and Anthropic would rather be the lab doing it.
The top community response, from user *ivraatiems*, called the moral framing 'hideous': "We must blast forwards into making this dangerous thing because if we don't, someone else surely will" is a coward's argument when the people making it are the ones with the most control over whether it happens at all. Another commenter, *jameson*, read the entire document as IPO-roadshow material, noting the unusual timing as Anthropic prepares to go public. A third, *mrandish*, pushed on the LOC caveat: rounding down doesn't fix the metric, because the *quality* distribution of AI-generated code is the entire question.
Meanwhile, *minimaxir* — one of the more rigorous practitioners on HN — described his own experiments with what he calls agentic iterative optimization: instructing a model to speed up real benchmarks by X% without regressing tests or output quality (MSE for statistical code, etc.). His results are interesting precisely because they're modest: the loop works, but most of the gain comes from the eval harness being honest, not from the model getting cleverer between iterations.
The LOC critique is the easy take, and @Top10_dev already ran it. The harder question is what 'recursive self-improvement' actually means inside a frontier lab in 2026, and whether the public framing of *inevitability* — which Anthropic has now made a load-bearing pillar of its own narrative — is the framing the rest of the industry should adopt.
There are three substantive claims layered in the memo, and they should be evaluated separately. The first is empirical: Claude is meaningfully accelerating Anthropic's own engineering output. Probably true, with the caveat that *every* frontier lab can say this about its internal coding assistant. The second is methodological: that this acceleration compounds — that each generation of model makes the next one cheaper and faster to build. This is the actual RSI claim, and the memo doesn't quite show it. An 8× LOC number in a single quarter is a level shift, not a compounding curve. To demonstrate recursion, you'd need to show that Q3 is faster than Q2 *because* Q2's model was used to build it, not just because tooling improved.
The third claim is the political one: that responsible labs should accelerate, because the alternative is ceding the frontier to less careful actors. This is the Manhattan Project frame, and it has the same load-bearing flaw it had in 1942: 'someone else will' is an empirical claim that should be tested, not a moral premise that ends the argument. Anthropic's competitive landscape is real, but it is also small enough that a public coordination signal — 'we are slowing down here, and we challenge our peers to do the same' — would be observable. The memo's authors chose the other rhetorical move.
The practitioner contrast is instructive. *minimaxir*'s agentic optimization loops — run by individual engineers on individual benchmarks — produce real gains, but the binding constraint is *the eval*, not the model. If you can't write a benchmark that the model can't game, the loop optimizes for the wrong thing within three iterations. Most teams discover this the hard way: the model 'speeds up' the code by deleting branches that the benchmark doesn't exercise. The same risk applies at frontier scale, only the deleted branches are alignment properties.
If you're running agentic optimization loops in your own codebase — and many of us now are, whether we call it that or not — the takeaway from both the memo and the community pushback is the same: invest in the eval before you invest in the loop. A loop with a soft eval is a confident regression generator. A loop with a sharp eval is genuinely useful, but the marginal return on a better model is much smaller than the marginal return on a better benchmark.
For team leads, the LOC framing is also worth pushing back on internally. If your AI-tooling rollout is being measured in lines added, you will get more lines, and most of them will be bad. Output velocity is the wrong KPI for the same reason it was the wrong KPI in 2005: it rewards verbose refactors and punishes deletions. The honest metric is throughput on a fixed problem set — closed tickets per engineer, with a constant difficulty distribution — and almost nobody measures it that way because it requires actual editorial work on the backlog.
And for anyone evaluating frontier labs as vendors: the race-logic framing is now load-bearing for Anthropic's public narrative. That has implications for how the company will react to safety incidents, regulatory friction, and capability disclosures over the next 18 months. Buy the model on its merits; price the policy posture separately.
The interesting question isn't whether Claude is writing more of Claude — it almost certainly is, at Anthropic and at every other frontier lab. The question is whether the *compounding* claim survives the next two quarters. If Anthropic publishes a Q4 2026 update showing the acceleration rate itself accelerating, the RSI thesis has teeth. If the next memo quietly drops the LOC framing and pivots to a different metric, you'll know which one didn't hold up.
I don't quite understand the intent of such article other than to promote themselves given an odd timing that the company is planning on going public, so I can only conclude that this is just part of the IPO roadshow.LLMs certainly have made significant changes to our lives, but I haven't
I have been doing more experiments with what I have now been calling agentic iterative optimization: telling the LLM to optimize code such that it speeds up all real-world-representative benchmarks by X% without cheating or causing regressions in both tests and performance metrics (e.g. MSE for stat
Whether or not Anthropic is right about what AI can accomplish, whether these performance gains are real or not, their moral stance here is absolutely hideous to me."We must blast forwards into making this dangerous thing because if we don't, someone else surely will," is a coward
> "A caveat: Lines of code is an imperfect measure"I'm pleased they at least included this. However, they address the caveat by 'rounding down' the estimated multiple of the gain. I'm not sure that is the correct adjustment, especially once we understand the range is
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
>A caveat: Lines of code is an imperfect measure, as it measures quantity over quality. So 8× lines of code/engineer/day in the second quarter of 2026 is almost certainly an overstatement of the true productivity gain. Nonetheless, it indicates an acceleration. At Anthropic, we don’t re