From Nicholas Carlini's Anthropic blog post:

I tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kernel. Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.

Bit more commentary on the capabilities benchmarking angle:

This project was designed as a capability benchmark. I am interested in stress-testing the limits of what LLMs can just barely achieve today in order to help us prepare for what models will reliably achieve in the future.

I’ve been using the C Compiler project as a benchmark across the entire Claude 4 model series. As I did with prior projects, I started by drafting what I wanted: a from-scratch optimizing compiler with no dependencies, GCC-compatible, able to compile the Linux kernel, and designed to support multiple backends. While I specified some aspects of the design (e.g., that it should have an SSA IR to enable multiple optimization passes) I did not go into any detail on how to do so.

Previous Opus 4 models were barely capable of producing a functional compiler. Opus 4.5 was the first to cross a threshold that allowed it to produce a functional compiler which could pass large test suites, but it was still incapable of compiling any real large projects. My goal with Opus 4.6 was to again test the limits.

Over nearly 2,000 Claude Code sessions across two weeks, Opus 4.6 consumed 2 billion input tokens and generated 140 million output tokens, a total cost just under $20,000. Compared to even the most expensive Claude Max plans, this was an extremely expensive project. But that total is a fraction of what it would cost me to produce this myself—let alone an entire team.

This was a clean-room implementation (Claude did not have internet access at any point during its development); it depends only on the Rust standard library. The 100,000-line compiler can build a bootable Linux 6.9 on x86, ARM, and RISC-V. It can also compile QEMU, FFmpeg, SQlite, postgres, redis, and has a 99% pass rate on most compiler test suites including the GCC torture test suite. It also passes the developer's ultimate litmus test: it can compile and run Doom.

This reminds me of a passage from L Rudolf L's history of the future:

By 2026, more code gets written in a week than the world wrote in 2020. Open source projects fork themselves into an endless orgy of abundance. Some high school students build functionally near-identical versions of Windows and Google Drive (and every video game in existence) from scratch in a month, because they can and they wanted one new feature on top of it. Everyone and their dog has a software product line. Big Tech unleashes a torrent of lawsuits against people cloning their products, echoing the Oracle v Google lawsuit about Java, but those lawsuits will take years to complete, and months feel like decades on the ground.

Back to Carlini on where Opus 4.6 fell short: 

The compiler, however, is not without limitations. These include:

• It lacks the 16-bit x86 compiler that is necessary to boot Linux out of real mode. For this, it calls out to GCC (the x86_32 and x86_64 compilers are its own).
• It does not have its own assembler and linker; these are the very last bits that Claude started automating and are still somewhat buggy. The demo video was produced with a GCC assembler and linker.
• The compiler successfully builds many projects, but not all. It's not yet a drop-in replacement for a real compiler.
• The generated code is not very efficient. Even with all optimizations enabled, it outputs less efficient code than GCC with all optimizations disabled.
• The Rust code quality is reasonable, but is nowhere near the quality of what an expert Rust programmer might produce.

The resulting compiler has nearly reached the limits of Opus’s abilities. I tried (hard!) to fix several of the above limitations but wasn’t fully successful. New features and bugfixes frequently broke existing functionality.

As one particularly challenging example, Opus was unable to implement a 16-bit x86 code generator needed to boot into 16-bit real mode. While the compiler can output correct 16-bit x86 via the 66/67 opcode prefixes, the resulting compiled output is over 60kb, far exceeding the 32k code limit enforced by Linux. Instead, Claude simply cheats here and calls out to GCC for this phase (This is only the case for x86. For ARM or RISC-V, Claude’s compiler can compile completely by itself.)

When Scott posted Does age bring wisdom? 8 years ago, I read it and thought "will this happen to me?" These passages got burned into my impressionable young-ish brain:

I turn 33 today. I can only hope that age brings wisdom.

We’ve been talking recently about the high-level frames and heuristics that organize other concepts. They’re hard to transmit, and you have to rediscover them on your own, sometimes with the help of lots of different explanations and viewpoints (or one very good one). They’re not obviously apparent when you’re missing them; if you’re not ready for them, they just sound like platitudes and boring things you’ve already internalized.

Wisdom seems like the accumulation of those, or changes in higher-level heuristics you get once you’ve had enough of those. I look back on myself now vs. ten years ago and notice I’ve become more cynical, more mellow, and more prone to believing things are complicated. For example:

1. Less excitement about radical utopian plans to fix everything in society at once
2. Less belief that I’m special and can change the world
3. Less trust in any specific system, more resignation to the idea that anything useful requires a grab bag of intuitions, heuristics, and almost-unteachable skills.
4. More willingness to assume that other people are competent in aggregate in certain ways, eg that academic fields aren’t making incredibly stupid mistakes or pointlessly circlejerking in ways I can easily detect.
5. More willingness to believe that power (as in “power structures” or “speak truth to power”) matters and infects everything.
6. More belief in Chesterton’s Fence.
7. More concern that I’m wrong about everything, even the things I’m right about, on the grounds that I’m missing important other paradigms that think about things completely differently.
8. Less hope that everyone would just get along if they understood each other a little better.
9. Less hope that anybody cares about truth (even though ten years ago I would have admitted that nobody cares about truth).

All these seem like convincing insights. But most of them are in the direction of elite opinion. There’s an innocent explanation for this: intellectual elites are pretty wise, so as I grow wiser I converge to their position. But the non-innocent explanation is that I’m not getting wiser, I’m just getting better socialized. ...

... eight years ago I was in a place where having Richard Dawkins style hyperrationalism was a useful brand, and now I’m (for some reason) in a place where having James C. Scott style intellectual conservativism is a useful brand. A lot of the “wisdom” I’ve “gained” with age is the kind of wisdom that helps me channel James C. Scott instead of Richard Dawkins; how sure am I that this is the right path?

Sometimes I can almost feel this happening. First I believe something is true, and say so. Then I realize it’s considered low-status and cringeworthy. Then I make a principled decision to avoid saying it – or say it only in a very careful way – in order to protect my reputation and ability to participate in society. Then when other people say it, I start looking down on them for being bad at public relations. Then I start looking down on them just for being low-status or cringeworthy. Finally the idea of “low-status” and “bad and wrong” have merged so fully in my mind that the idea seems terrible and ridiculous to me, and I only remember it’s true if I force myself to explicitly consider the question. And even then, it’s in a condescending way, where I feel like the people who say it’s true deserve low status for not being smart enough to remember not to say it. This is endemic, and I try to quash it when I notice it, but I don’t know how many times it’s slipped my notice all the way to the point where I can no longer remember the truth of the original statement. ... 

There’s one more possibility that bothers me even worse than the socialization or traumatization theory. I’m going to use science-y sounding terms just as an example, but I don’t actually think it’s this in particular – we know that the genes for liberal-conservative differences are mostly NMDA receptors in the brain. And we know that NMDA receptor function changes with aging. It would be pretty awkward if everything we thought was “gaining wisdom with age” was just “brain receptors consistently functioning differently with age”. If we were to find that were true – and furthermore, that the young version was intact and the older version was just the result of some kind of decay or oxidation or something – could I trust those results? Intuitively, going back to earlier habits of mind would feel inherently regressive, like going back to drawing on the wall with crayons. But I don’t have any proof.

Wisdom is like that.

Looking at Scott's list now that I'll also turn 33 this year:

• I do have a lot more high-level organizing frames than I did 8 years ago, but most of them don't sound like platitudes, maybe because I know how to decompose them into specific non-platitudinous concepts I've been saving in my various PKMs over the years (thanks gwern)
• No change on "radical utopian plans have zero chance of fixing everything at once" 
• Interestingly I went an epsilon in the opposite direction from Scott re: "I'm special and can change the world" due to a zero-chance skeptical baseline (likely due to low self-esteem) followed by a once-in-a-lifetime stroke of luck 
• I have in fact trended towards "anything useful requires a grab bag of intuitions etc"
• (skipping a few out of disinterest)
• re: hoping that more people care about truth, also trended in the opposite direction to my surprise, also maybe due to zero-chance skeptical baseline followed by repeated counterevidence
• the passage on "I believe X is true -> but it's low-status to say in public -> when others say it I start looking down on them for being bad at PR -> later I start looking down on them for being low-status -> "low-status" merges with "bad/wrong" -> X is instinctively bad/wrong unless I force myself to explicitly consider if X is true" was burned into my brain as such a terrible failure mode I've been on guard against it ever since, even now that I work in public health policy where there's a steep incentive gradient to warp reasoning in this direction. One thing I've noticed about myself is that when someone says low-status-but-true-X in public, what I find cringe isn't that they said X so much as how they said it 
• I've always wondered about the "NMDA receptor function changes with aging" thing, not so much that specific mechanism (which isn't what Scott believed anyway), but more generally how I'd be able to tell if this happens, and whether this is at least temporarily reversible or modulatable somehow 

FWIW the coauthor of the paper you linked provides more nuance here.

Maybe one of Logan Strohl's Fucking Goddamn Basics of Rationalist Discourse? Actually the whole thing is short enough to repost here in its entirety to save the trouble of clicking through:

1. Don't say false shit omg this one's so basic what are you even doing. And to be perfectly fucking clear "false shit" includes exaggeration for dramatic effect. Exaggeration is just another way for shit to be false.
2. You do NOT (necessarily) know what you fucking saw. What you saw and what you thought about it are two different things. Keep them the fuck straight.
3. Performative overconfidence can go suck a bag of dicks. Tell us how sure you are, and don't pretend to know shit you don't.
4. If you're going to talk unfalsifiable twaddle out of your ass, at least fucking warn us first.
5. Try to find the actual factual goddamn truth together with whatever assholes you're talking to. Be a Chad scout, not a Virgin soldier.
6. One hypothesis is not e-fucking-nough. You need at least two, AT LEAST, or you'll just end up rehearsing the same dumb shit the whole time instead of actually thinking.
7. One great way to fuck shit up fast is to conflate the antecedent, the consequent, and the implication. DO NOT.
8. Don't be all like "nuh-UH, nuh-UH, you SAID!" Just let people correct themselves. Fuck.
9. That motte-and-bailey bullshit does not fly here.
10. Whatever the fuck else you do, for fucksake do not fucking ignore these guidelines when talking about the insides of other people's heads, unless you mainly wanna light some fucking trash fires, in which case GTFO.

Yes please to the longer stab. 

You also made me wonder what a shortlist of crucial considerations for digital minds ranked by something like instantiable mind-seconds "swing factor" (or whatever the more sophisticated version should be) would look like, where reversible computation's swing factor is 10^54 by Claude's lights.

Meta: consider reposting on the EA Forum? 

Julia Wise's 2013 Giving now vs. later: a summary still seems good today:

Reasons to give now:

• You may get less altruistic as you age, so if you wait you may never actually donate.
• Estimates of the returns on investment may be over-optimistic.
• Giving to charities that can demonstrate their effectiveness provides an incentive for charities to get better at demonstrating that they're effective.  We can't just wait for charities to improve — it takes donations to make that happen.
• Having an active culture of giving encourages other people to give, too.
• Better to eliminate problems as soon as possible. E.g. if we had eliminated smallpox in 1967 instead of 1977, many people would have been spared.
• Giving to particular organizations can accelerate our learning about which causes are best to support. (Note: this wasn't in Julia's post, it's from Luke Muehlhauser's comment under it as to "most important reason missing from" this section)

Reasons to give later:

• As time passes, we'll probably have better information about which interventions work best.  Even in a few years, we may know a lot more than we do now and be able to give to better causes.
• Investing money may yield more money to eventually donate.
• When you're young, you should invest in developing yourself and your career, which will let you help more later.
• You can put donations in a donor-advised fund to ensure they will someday be given, even if you haven't yet figured out where you want them to go.

But it’s a topic that deserves more depth than that summary. ... 

Besides the links listed after that, you can also check out the patient altruism tag and "related entries" there, as well as the cause prioritization wiki's donating now vs later. 

a skill which I respect in other people and which I aspire towards is noticing when other people are experiencing suffering due to violations of positive narratives, or fulfillment of negative narratives, and comforting them and helping nudge them back into a good narrative.

interesting, any examples online?

I learned about ALARA from this post, so I wanted to note this here -- Reading List 01/17/2026 by Brian Potter at Construction Physics notes that ALARA may be taken down, maybe, not sure, can't find the memo to confirm:

Now it looks like the US Department of Energy is removing ALARA. Via E&E News:

Energy Secretary Chris Wright killed the Department of Energy’s decades-old radiation safety standard Friday.

Wright ended the department’s use of the As Low As Reasonably Achievable — or “ALARA” — principle, which has long been a staple of nuclear regulation. ALARA is rooted in the idea that any radiation exposure carries risks, but low doses can be justified by practical considerations. Critics in the nuclear power and health fields argue that the standard is overly burdensome with no real safety benefits.

The move could lower operational costs and accelerate projects using nuclear material, but it will alter an established safety-first culture. The change in safety standards may impact DOE’s ongoing advanced nuclear reactor pilot program and high-stakes radiation cleanups, like the Hanford site in Washington state that has been dubbed the most contaminated place in the Western Hemisphere.

I looked for the DOE memo that initiated this change but wasn’t able to find it, and no one other than E&E news seems to have picked up the story yet, so it’s not clear to me how “real” this is.

Another whimsical "benchmark": Terry Tao wrote on Mathstodon that

There are now sufficiently many different examples of Erdos problems that have been resolved with various amounts of AI assistance and formal verification (see https://github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erd%C5%91s-problems for a summary) that one can start to discern general trends.

Broadly speaking, we now see an empirical tradeoff between the level of AI involvement in the solution, and the difficulty or novelty of that solution.  In particular, the recent solutions have spanned a spectrum roughly describable as follows:

1.  Completely autonomous AI solutions to Erdos problems that are short and largely follow a standard technique.  (In many, but not all, of these cases, some existing literature was found that proved a very similar result by a similar method.) 

2.  AI-powered modifications of existing solutions (which could be either human-generated or AI-generated) that managed to improve or modify these solutions in various ways, for instance by upgrading a partial solution to a full solution, or optimizing the parameters of the proof.

3.  Complex interactions between humans and AI tools in which the AI tools provided crucial calculations, or proofs of key steps, allowing the collaboration to achieve moderately complicated and novel solutions to open problems.

4.  Difficult research-level papers solving one or more Erdos problems by mostly traditional human means, but for which AI tools were useful for secondary tasks such as generation of code, numerics, references, or pictures.

The seeming negative correlation between the amount of AI involvement and the depth of result is somewhat reminiscent of statistical paradoxes such as Berkson's paradox https://en.wikipedia.org/wiki/Berkson%27s_paradox or Simpson's paradox https://en.wikipedia.org/wiki/Simpson%27s_paradox .  One key confounding factor is that highly autonomous AI workflows are much more scaleable than human-intensive workflows, and are thus better suited for being systematically applied to the "long tail" of obscure Erdos problems, many of which actually have straightforward solutions.  As such, many of these easier Erdos problems are now more likely to be solved by purely AI-based methods than by human or hybrid means.

Given the level of recent publicity given to these problems, I expect that over the next few weeks, pretty much all of the outstanding Erdos problems will be quietly attempted by various people using their preferred AI tool.  Most of the time, these tools will not lead to any noteworthy result, but such failures are unlikely to be reported on any public site.  It will be interesting to see what (verified) successes do emerge from this, which should soon give a reasonably accurate picture of what proportion of currently outstanding Erdos problems are simple enough to be amenable to current AI tools operated with minimal human intervention. (My guess is that this proportion is on the order of 1-2%.)  Assessing the viability of more hybridized human-AI approaches will take significantly longer though, as human expert attention will remain a significant bottleneck.

So I'll whimsically define the "Erdos problems benchmark" to be "the proportion of currently outstanding Erdos problems amenable to current AI tools operated with minimal human intervention", and the current "SOTA" to be Tao's guess of 1-2% as of Jan 2026. My guess is it won't be saturated in ~2 years like every other benchmark because open math problems can be unboundedly hard, but who knows?

I bet that Steve Byrnes can point out a bunch of specific sensory evidence that the brain uses to construct the status concept (stuff like gaze length of conspecifics, or something?), but the human motivation system isn't just optimizing for those physical proxy measures, or people wouldn't be motivated to get prestige on internet forums where people have reputations but never see each other's faces.

Curious to see what Steven Byrnes would actually say here. I fed your comment and Byrnes' two posts on social status to Opus 4.5, it thought for 3m 40s (!) and ended up arguing he'd disagree with your social status example:

Byrnes explicitly argues the opposite position in §2.2.2 of the second post. He writes:

"I don't currently think there's an innate drive to 'mostly lead' per se. Rather, I think there's an innate drive that we might loosely describe as 'a drive to feel liked / admired'... and also an innate drive that we might loosely describe as 'a drive to feel feared'. These drives are just upstream of gaining an ability to 'mostly lead'."

And more pointedly:

"I'm avoiding a common thing that evolutionary psychologists do (e.g. Secret of Our Success by Henrich), which is to point to particular human behaviors and just say that they're evolved—for example, they might say there's an 'innate drive to be a leader', or 'innate drive to be dominant', or 'innate drive to imitate successful people', and so on. I think those are basically all 'at the wrong level' to be neuroscientifically plausible."

So Byrnes is explicitly rejecting the claim that evolution installed "status-seeking" as a goal at the level Eli describes.

Byrnes proposes a three-layer architecture: first, there are primitive innate drives—"feel liked/admired" and "feel feared"—which are still at the feeling level, not the abstract-concept level. Second, there's a very general learning mechanism (which he discusses extensively in his valence series, particularly §4.5–4.6) that figures out what actions and situations produce those feelings in the local environment. Third, there are some low-level sensory adaptations (like "an innate brainstem reflex to look at people's faces") that feed into the learning system. Status-seeking behavior emerges from this combination, but "status" itself isn't the installed goal.

Why does this matter? Eli presents something like a dichotomy: either (A) evolution can only do sensory-level proxies that break in novel contexts (like male preferences for big breasts, which don't update when you learn a woman is infertile), or (B) evolution can install abstract concepts as goals (like status). Byrnes' model offers a third option: evolution installs feeling-level drives plus a general learning mechanism. The learning mechanism explains why status-seeking transfers to internet forums—the primitive drive to "feel liked/admired" still triggers when you get upvotes, and the learning system figures out how to get more of that—without requiring that "status" itself be the installed goal. This third option actually supports the original claim Eli is arguing against. Evolution didn't need to install "status" as a concept; it installed feelings + learning, and the abstract behavior emerged.

(mods, let me know if this is slop and I'll take it down)