I think the linked tweet is possibly just misinterpreting what the authors meant by "transistor operations"?  My reading is that "1000" binds to "operations"; the actual number of transistors in each operation is unspecified. That's how they get the 10,000x number - if a CPU runs at 1 GHz, neurons run at 100 Hz, then even if it takes 1000 clock cycles to do the work of neuron, the CPU can still do it 10,000x faster. 

(IDK what the rationale was in the editorial process for using "transistor operations" instead of a more standard term like "clock cycles", but a priori it seems defensible. Speculating, "transistors" was already introduced in the sentence immediately prior, so maybe the thinking was that the meaning and binding of "transistor operations" would be self-evident in context. Whereas if you use "clock cycles" you have to spend a sentence explaining what that means. So using "transistor operations" reduces the total number of new jargon-y / technical terms in the paragraph by one, and also saves a sentence of explanation.)

Anyway, depending on the architecture, precision, etc. a single floating point multiplication can take around 8 clock cycles. So even if a single neuron spike is doing something complicated that requires several high-precision multiply + accumulate operations in serial to replicate, that can easily fit into 1000 clock cycles on a normal CPU, and much fewer if you use specialized hardware.

As for the actual number of transistors themselves needed to do the work of a neuron spike, it again depends on exactly what the neuron spike is doing and how much precision etc. you need to capture the actual work, but "billions" seems too high by a few OOM at least. Some reference points: a single NAND gate is 4 transistors, and a general-purpose 16-bit floating point multiplier unit is ~5k NAND gates.

The passage seems fine to me; I commented on Erdil's post and other brain efficiency discussions at the time, and I still think that power consumption is a more objective way of comparing performance characteristics of the brain vs. silicon, and that various kinds of FLOP/s comparisons favored by critics of the clock speed argument in the IAB passage are much more fraught ([1], [2]).

It's true that clock speed (and neuron firing speed) aren't straightforwardly / directly translatable to "speed of thought", but both of them are direct proxies for energy consumption and power density. And a very rough BOTEC shows that ~10,000x is a reasonable estimate for the difference in power density between the brain and silicon.

Essentially, the brain is massively underclocked because of design-space restrictions imposed by biology and evolution, whereas silicon-based processing has been running up against fundamental physical limits on component size, clock speed, and power density for a while now. So once AIs can run whatever cognitive algorithms that the brain implements (or algorithms that match the brain in terms of high-level quality of the actual thoughts) at any speed, the already-existing power density difference implies they'll immediately have a much higher performance ceiling in terms of the throughput and latency that they can run those algorithms at. It's not a coincidence that making this argument via clock speeds leads to basically the same conclusion as making the same argument via power density.

• Tricky hypothesis 1: ASI will in fact be developed in a world that looks very similar to today's (e.g. because sub-ASI AIs will have negligible effect on the world; this could also be because ASI will be developed very soon).

• Tricky hypothesis 2: But the differences between the world of today and the world where ASI will be developed don't matter for the prognosis.

Both of these hypotheses look relatively more plausible than they did 4y ago, don't they? Looking back at this section from the 2021 takeoff speed conversation gives a sense of how people were thinking about this kind of thing at the time.

AI-related investment and market caps are exploding, but not really due to actual revenue being "in the trillions" - it's mostly speculation and investment in compute and research.

Deployed AI systems can already provide a noticeable speed-up to software engineering and other white-collar work broadly, but it's not clear that this is having much of an impact on AI research (and especially a differential impact on alignment research) specifically.

Maybe we will still get widely deployed / transformative robotics, biotech, research tools etc. due to AI that could make a difference in some way prior to ASI, but SoTA AIs of today are routinely blowing through tougher and tougher benchmarks before they have widespread economic effects due to actual deployment.

I think most people in 2021 would have been pretty surprised by the fact we have widely available LLMs in 2025 with gold medal-level performance on the IMO, but which aren't yet having much larger economic effects. But in relative terms it seems like you and Christiano should be more surprised than Yudkowsky and Soares.

The "you're sued" part is part of what ensures that the forms get filled out honestly and comprehensively.

Depending on the kind of audit you do, the actual deliverable you give your auditor may just be a spreadsheet with a bunch of Y/N answers to hundreds of questions like "Do all workstations have endpoint protection software", "Do all servers have intrusion detection software", etc. with screenshots of dashboards as supporting evidence for some of them.

But regardless of how much evidence an external auditor asks for, at large companies doing important audits, every single thing you say to the auditor will be backed internally with supporting evidence and justification for each answer you give. 

At a bank you might have an "internal audit" department that has lots of meetings and back-and-forth with your IT department; at an airline it might be a consulting firm that you bring in to modernize your IT and help you handle the audit, or, depending on your relationship with your auditor and the nature of the audit, it might be someone from the audit firm itself that is advising you. In each case, their purpose is to make sure that every machine across your firm really does have correctly configured EDR, fully up to date security patches, firewalled, etc. before you claim that officially to an auditor.

Maybe you have some random box used to show news headlines on TVs in the hallways - turns out these are technically in-scope for having EDR and all sorts of other endpoint controls, but they're not compatible with or not correctly configured to run Microsoft Defender, or something. Your IT department will say that there are various compensating / mitigating controls or justifications for why they're out of scope, e.g. the firewall blocks all network access except the one website they need to show the news, the hardware itself is in a locked IT closet, they don't even have a mouse / keyboard plugged in, etc. These justifications will usually  be accepted unless you get a real stickler (or have an obstinate "internal auditor"). But it's a lot easier to just say "they all run CrowdStrike" than it is to keep track of all these rationales and compensating controls, and indeed ease-of-deployment is literally the first bullet in CrowdStrike's marketing vs. Microsoft Defender:
 

CrowdStrike: Deploy instantly with a single, lightweight agent — no OS prerequisites, complex configuration, or fine tuning required.
Microsoft: Complicated deployment hinders security. All endpoints require the premium edition of the latest version of Windows, requiring upfront OS and hardware upgrades for full security functionality.

You wrote in a sibling reply:

Further, the larger implication of the above tweet is that companies use Crowdstrike because of regulatory failure, and this is also simply untrue. There are lots of reasons people sort of unthinkingly go with the name brand option in security, but that's a normal enterprise software thing and not anything specific to compliance.
 

I agree that this has little to do with "regulatory failure" and don't know / don't have an opinion on whether that's what the original tweet author was actually trying to communicate. But my point is that firms absolutely do make purchasing decisions about security software for compliance reasons, and a selling point of CrowdStrike (and Carbon Black, and SentinelOne) is that they make 100% compliance easier to achieve and demonstrate vs. alternative solutions. That's not a regulatory failure or even necessarily problematic, but it does result in somewhat different outcomes compared to a decision process of "unthinkingly going with the name brand option" or "carefully evaluate and consider only which solutions provide the best actual security vs. which are theater".

The audit requirements Mark is talking about don't exist. He just completely made them up.

The screenshotted tweet says that you're required to install something like Crowdstrike, which is correct and also seems consistent with the ChatGPT dialogue you linked?

There are long lists of computer security practices and procedures needed to pass an audit for compliance with a standard like ISO27001, PCI DSS, SOC 2, etc. that many firms large and small are subject to (sometimes but not necessarily by law - e.g. companies often need to pass an  SOC 2 audit because their customers ask for it).

As you say, none of these standards name specific software or vendors that you have to use in order to satisfy an auditor, but it's often much less of a headache to use a "best in class" off-the-shelf product (like CrowdStrike) that is marketed specifically as satisfying specific requirements in these standards, vs. trying to cobble together a complete compliance posture using tools or products that were not designed specifically to satisfy those requirements. 

A big part of the marketing for a product like CrowdStrike is that it has specific features which precisely and unambiguously satisfy more items in various auditor checklists than competitors.

So "opens up an expensive new chapter of his book" is colorful and somewhat exaggerated, but I wouldn't describe it as "misinformation" - it's definitely pointing at something real, which is that a lot of enterprise security software is sold and bought as an exercise in checking off specific checklist items in various kinds of audits, and how easy / convenient / comprehensive a solution makes box-checking is often a bigger selling point than how much actual security it provides, or what the end user experience is actually like.

I would just evaluate your argument on my own and I would evaluate the counterargument in the comment on my own.

The precise issue is that a sizable fraction of the audience will predictably not do this, or will do it lazily or incorrectly.

On LessWrong, this shows up in voting patterns, for example, a controversial post will sometimes get some initial upvotes and then the karma / trend will swing around based on the comments and who had the last word. Or, a long back-and-forth ends up getting far fewer votes (and presumably, eyeballs) than the top-level post / comment.

My impression is that most authors aren't that sensitive to karma per se but they are sensitive to a mental model of the audience that this swinging implies, namely that many onlookers are letting the author and their interlocutor(s) do their thinking for them, with varying levels of attention span, and where "highly upvoted" is often a proxy for "onlookers believe this is worth responding to (but won't necessarily read the response)". So responding often feels both high stakes and unrewarding for someone who cares about communicating something to their audience as a whole.

Anyway, I like Duncan's post as a way of making the point about effort / implied obligation to both onlookers and interlocutors, but something else that might help is some kind of guide / reminder / explanation about principles of being a good / high-effort onlooker.

What specifically do you think is obviously wrong about the village idiot <-> Einstein gap?  This post from 2008 which uses the original chart makes some valid points that hold up well today, and rebuts some real misconceptions that were common at the time.

The original chart doesn't have any kind of labels or axes, but here are two ways you could plausibly view it as "wrong" in light of recent developments with LLMs:

• Duration: the chart could be read as a claim that the gap between the development of village idiot and Einstein-level AI in wall-clock time would be more like hours or days rather than months or years.
• Size and dimensionality of mind-space below the superintelligence level. The chart could be read as a claim that the size of mindspace between village idiot and Einstein is relatively small, so it's surprising to Eliezer-200x that there are lots of current AIs landing in between them, and staying there for a while.

I think it's debatable how much Eliezer was actually making the stronger versions of the claims above circa 2008, and also remains to be seen how wrong they actually are, when applied to actual superintelligence instead of whatever you want to call the AI models of today.

OTOH, here are a couple of ways that the village idiot <-> Einstein post looks prescient:

• Qualitative differences between the current best AI models and second-to-third tier models are small. Most AI models today are all roughly similar to each other in terms of overall architecture and training regime, but there are various tweaks and special sauce that e.g. Opus and GPT-5 have that Llama 4 doesn't. So you have something like: Llama 4: GPT-5 :: Village idiot : Einstein, which is predicted by:

Maybe Einstein has some minor genetic differences from the village idiot, engine tweaks.  But the brain-design-distance between Einstein and the village idiot is nothing remotely like the brain-design-distance between the village idiot and a chimpanzee.  A chimp couldn't tell the difference between Einstein and the village idiot, and our descendants may not see much of a difference either.

(and something like a 4B parameter open-weights model is analogous to the chimpanzee)

Whereas I expect that e.g. Robin Hanson in 2008 would have been quite surprised by the similarity and non-specialization among different models of today.

• Implications for scaling. Here's a claim on which I think the Eliezer-200x Einstein chart makes a prediction that is likely to outperform other mental models of 2008, as well as various contemporary predictions based on scaling "laws" or things like METR task time horizon graphs:
  
  "The rough number of resources, in terms of GPUs, energy, wall clock time, lines of Python code, etc. needed to train and run best models today (e.g. o4, GPT-5), are sufficient (or more than sufficient) to train and run a superintelligence (without superhuman / AI-driven levels of optimization / engineering / insight)."
  
  My read of task-time-horizon and scaling law-based models of AI progress is that they more strongly predict that further AI progress will basically require more GPUs. It might be that the first Einstein+ level AGI is in fact developed mostly through scaling, but these models of progress are also more surprised than Eliezer-2008 when it turns out that (ordinary, human-developed) algorithmic improvements and optimizations allow for the training of e.g. a GPT-4-level model with many fewer resources than it took to train the original GPT-4 just a few years ago.

Thanks for the report, should be fixed now.

The issue was that the LW GraphQL API has changed slightly, apparently. The user query suggested here no longer works, but something like:

{
      GetUserBySlug(slug: "max-h") {
          _id
          slug
          displayName
          pageUrl
          postCount
          commentCount
          createdAt
      }
}

works fine.

I prefer (classical / bedrock) liberalism as a frame for confronting societal issues with AGI, and am concerned by the degree to which recent right-wing populism has moved away from those tenets.

Liberalism isn't perfect, but it's the only framework I know of that even has a chance of resulting in a stable consensus. Other frames, left or right, have elements of coercion and / or majoritarianism that inevitably lead to legitimacy crises and instability as stakes get higher and disagreements wider.

My understanding is that a common take on both the left and right these days is that, well, liberalism actually hasn't worked out so great for the masses recently, so everyone is looking for something else. But to me every "something else" on both the left and right just seems worse - Scott Alexander wrote a bunch of essays like 10y ago on various aspects of liberalism and why they're good, and I'm not aware of any comprehensive rebuttal that includes an actually workable alternative.

Liberalism doesn't imply that everyone needs to live under liberalism (especially my own preferred version / implementation of it), but it does provide a kind of framework for disagreement and settling differences in a way that is more peaceful and stable than any other proposal I've seen.

So for example on protectionism, I think most forms of protectionism (especially economic protectionism) are bad and counterproductive economic policy. But even well-implemented protectionism requires a justification beyond just "it actually is in the national interest to do this", because it infringes on standard individual rights and freedoms. These freedoms aren't necessarily absolute, but they're important enough that it requires strong and ongoing justification for why a government is even allowed to do that kind of thing. AGI might be a pretty strong justification!

But at the least, I think anyone proposing a framework or policy position which deviates from a standard liberal position should acknowledge liberalism as a kind of starting point / default, and be able to say why the tradeoff of any individual freedom or right is worth making, each and every time it is made. (And I do not think right-wing frameworks and their standard bearers are even trying to do this, and that is very bad.)

I think it was fine for Nate to delete your comment and block you, and fine for you to repost it as a short form.

But my anecdote is a valid report of the historical consequences of talking with Nate – just as valid as the e/acc co-founder's tweet.

"just as valid" [where validity here = topical] seems like an overclaim here. And at the time of your comment, Nate had already commented in other threads, which are now linked in a footnote in the OP:

By "cowardice" here I mean the content, not the tone or demeanor. I acknowledge that perceived arrogance and overconfidence can annoy people in communication, and can cause backlash. For more on what I mean by courageous vs cowardly content, see this comment. I also spell out the argument more explicitly in this thread.

So it's a bit of a stretch to say that any AI safety-related discussion or interpersonal interaction that Nate has ever had in any context is automatically topical.

I also think your description of Nate's decision to delete your comment as "not ... allowing people to read negative truths about his own behavior" is somewhat overwrought. Both of the comment threads you linked were widely read and discussed at the time, and this shortform will probably also get lots of eyeballs and attention. 

At the very least, there is an alternate interpretation, which is that the comment really was off-topic in Nate's view, and given the history between the two of you, he chose to block + delete instead of re-litigating or engaging in a back-and-forth that both of you would probably find unpleasant and unproductive. Maybe it would have been more noble or more wise of him to simply let your comment stand without direct engagement, but that can also feel unpleasant (for Nate or others).