All of Randaly's Comments + Replies

You seem to be assuming that there's not significant overhead or delays from negotiating leases, entering bankruptcy, or dealing with specialized hardware, which is very plausibly false.

If nobody is buying new datacenter GPU's, that will cut GPU progress to ~zero or negative (because production is halted and implicit knowledge is lost). (It will also probably damage broader semiconductor progress.)

This proportionally reduces cost of inference (and also of training).

This reduces the cost to rent a GPU-hour, but it doesn't reduce the cost to the owner. (Open... (read more)

Thanks for explaining. I now agree that the current cost of inference isn't a very good anchor for future costs in slowdown timelines.

I'm uncertain, but I still think OpenAI is likely to go bankrupt in slowdown timelines. Here are some related thoughts:

1. OpenAI probably won't pivot to the slowdown in time.
   1. They'd have < 3 years to do before running out of money.
   2. Budgets are set in advance. So they'd have even less time.
   3. All of the improvements you list cost time and money. So they'd need to continue spending on R&D, before that R&D has improved their

I agree that finances are important to consider. I've written my thoughts on them here; I disagree with you in a few places.

(1) Given Altman's successful ouster of the OpenAI board, his investors currently don't have much drive/desire/will to force him to stop racing. They don't have much time to do so on the current pace of increasing spending before OpenAI runs out of money.

(2) It's not clear what would boost revenue that they're not already doing; the main way to improve profits would just be to slash R&D spending. Much of R&D spending is spent ... (read more)

What prompts did you use? Can you share the chat? I see Sonnet 3.7 denying this knowledge when I try.

I want to clarify that I'm criticizing "AI 2027"'s projection of R&D spending, i.e. this table. If companies cut R&D spending, that falsifies the "AI 2027" forecast.

In particular, the comment I'm replying to proposed that while the current money would run out in ~2027, companies could raise more to continue expanding R&D spending. Raising money for 2028 R&D would need to occur in 2027; and it would need to occur on the basis of financial statements of at least a quarter before the raise. So in this scenario, they need to slash R&D spend... (read more)

My intuitions are more continuous here. If AGI is close in 2027 I think that will mean increased revenue and continued investment

Gotcha, I disagree. Lemme zoom on this part of my reasoning, to explain why I think profitability matters (and growth matters less):

(1) Investors always only terminally value profit; they never terminally value growth. Most of the economy doesn't focus much on growth compared to profitability, even instrumentally. However, one group of investors, VC's, do: software companies generally have high fixed costs and low marginal costs,... (read more)

Thanks for the response!

So maybe I should just ask whether you are conditioning on the capabilities progression or not with this disagreement? Do you think $140b in 2027 is implausible even if you condition on the AI 2027 capability progression?

I am conditioning on the capabilities progression.

Based on your later comments, I think you are expecting a much faster/stronger/more direct translation of capabilities into revenue than I am- such that conditioning on faster progress makes more of a difference.

The exact breakdown FutureSearch use seems relatively u

I'd be curious to hear more about what made you perceive our scenario as confident. We included caveats signaling uncertainty in a bunch of places, for example in "Why is it valuable?" and several expendables and footnotes. Interestingly, this popular YouTuber made a quip that it seemed like we were adding tons of caveats everywhere,

I was imprecise (ha ha) with my terminology here- I should have only talked about a precise forecast rather than a confident one, I meant solely the attempt to highlight a single story about a single year. My bad. Edited the post.

Typo: The description for table 2 states that "In total, 148 of our 169 tasks have human
baselines, but we rely on researcher estimates for 21 tasks in HCAST.". This is an incorrect sum; the right figure is 149 out of 170 tasks, per the table.

Footnote 18 of the timelines forecast appears to be missing its explanation.

Those were in fact some of the cases I had in mind, yes, thank you - I read the news too. And what one learns from reading about them is how those are exceptional cases, newsworthy precisely because they reached any verdict rather than settling, driven by external politics and often third-party funding, and highly unusual until recently post-2016/Trump. It is certainly the case that sometimes villains like Alex Jones get smacked down properly by libel lawsuits; but note how wildly incomparable these cases are to the blog post that Spartz is threatening to

This is a combative comment which fails to back up its claims.

how surely only noble and good people ever sue over libel

if you really believe lawsuits are so awesome and wonderful

He did not say this. This is not reasonable for you to write.

you can count on one hand the sort of libel lawsuit which follows this beautiful fantasy.

This is not true. This is obviously not true. A successful and important libel case (against Giuliani) was literally headline news this week. You can exceed five such cases just looking at similar cases: Dominion v Fox; Sma... (read more)

IMO, both U.S. and UK libel suits should both be very strongly discouraged, since I know of dozens of cases where organizations and individuals have successfully used them to prevent highly important information from being propagated, and I think approximately no case where they did something good (instead organizations that frequently have to deal with libel suits mostly just leverage loopholes in libel law that give them approximate immunity, even when making very strong and false accusations, usually with the clarity of the arguments and the transparenc

I'm not disputing that specific people at Skunk Works believed that their tech was disliked for being good; but that's a totally insane belief that you should reject immediately, it's obviously self-serving, none of those people present any evidence for it, and the DoD did try to acquire similar technology in all these cases.

Again, this is a direct quote on procurement incentives from a guy who was involved on both the buy and sell side of the SR-71 back in the day.

This is quote from, per you, somebody from the CIA. The CIA and Air Force are different orga... (read more)

Your discussion of Skunk Works is significantly wrong throughout. (I am not familiar with the other examples.)

For example, in 1943 the Skunk Works both designed and built America’s first fighter jet, the P80 Shooting Star, in just 5 months. Chief engineer Kelly Johnson worked with a scrappy team of, at its peak, 23 designers and 105 fabricators. Nonetheless, the resulting plane ended up being operationally used by the air force for 40 years.

The P80 was introduced in 1945; the US almost immediately decided to replace it with the F-86, introduced in 1949. Th... (read more)

No.

I'm pretty negative on how you fail to discuss any specific claim or link to any specific evidence, but you spend your longest paragraph speculating about the supposed bias of unnamed people.

You haven't really written enough to be clear, but I suspect that you have confused concentration camps with death or extermination camps? Regardless, the recent UN report did pretty specifically support claims of concentration camps- see points 37-57

I also found that, controlling for rents, the partisanship of a state did not predict homelessness (using the Partisan Voting Index)

This is not a useful way of looking at this; homelessness would be almost entirely controlled by city, not state, policies. State partisanship in large part measures not how blue or red the states' cities are, but rather how urban or rural the state as a whole is.

This, and the Bahrain/UAE cases, seem more likely to be driven by concerns about whether/how well the Chinese vaccines work?

On the other hand, look at the US wars in Vietnam, Iraq or Afghanistan. The outcomes of these wars were determined much more by political forces (in both of the relevant countries) than by overwhelming force.

Insurgencies aren't a good comparison for conventional wars like the Nagorno-Karabakh war.

The overall thrust here seems like an application of Clausewitz's maxim that "war is the extension of politics by other means". However, the specific politics suggested seem very unrealistic.

• You suggest ways to impact Azerbaijan's internal politics by targeting harm to specific groups. I see no reason to believe that Armenia had any substantial ability to deal much harm to Azerbaijan at all, so this isn't relevant. In general, it would be much harder for Armenia to advance to deal significant damage to Azerbaijan's homeland than it would be to defend.
• Assas

Try vaxstandby.com

The source article is here. The numbers are not how much of the total the subgroups make up, they are how quickly each subgroup is growing. The text continues:

The number critically ill with covid-19 in that age group grew by about 30% in the week before January 2nd, and also in the following week—but by just 7% in the week after that (see chart 2). By contrast, among those aged between 40 and 55 (who were vaccinated at a much lower rate at the time) the weekly change in the number of critically ill remained constant, with a 20-30% increase in each of those three weeks.

I have no idea why Dr. Moncef Slaoui, the head of Operation Warp Speed, was asked to resign and transition things over to someone else. Seems like if someone does their one job this effectively you’d want to keep them around.

While it's possible that Moncef Slaoui's resignation was caused by the Biden transition's request, he'd been publicly clear for months that he would resign in late 2020 or early 2021, as soon as 2 vaccines were approved. Here's a news article of him saying this from November.

Plausibly the Biden transition just wanted him to resign at a... (read more)

Blade Runner 2045 movie

2049, not 2045.

Trump continues to promise a vaccine by late October. The head of the CDC says that’s not going to happen. Trump says the head of the CDC is ‘confused.’ The CDC walks the comments back. On net, this showed some attempt by the CDC to not kowtow to Trump, but then a kowtow, so on net seems like a wash.

This is missing the last step, which is that the CDC then walked back its walk back (?!?). See here:

The CDC scrambled to explain; by about 6 p.m., the agency was claiming Redfield had misunderstood the original question and

I don't really have a great answer to that, except that empirically in this specific case, Spain was indeed able to extract very large amounts of resources from America within a single generation. (The Spanish government directly spent very little on America; the flow of money was overwhelming towards Europe, to the point where it caused notable inflation in Spain and in Europe as a whole.) I don't disagree that running a state is expensive, but I don't see why the expense would necessarily be higher than the extracted resources?

(1) Local support doesn't end after the first stages of the war, or after the war ends. I mentioned having favored local elites within one society/ethnicity continue to do most of the direct work in (2); colonizers also set up some groups as favored identities who did much of the work of local governance. For example, after the Spanish conquest, the Tlaxcala had a favored status and better treatment.

(2) Not sure why you'd expect low fidelity control to imply that it ends up as a wash in terms of extracting resources, can you clarify?

I feel like there's two points causing the confusion:

(1) The assumption that natives are an undifferentiated mass. There were a variety of mutually hostile indigenous peoples, who themselves sough out allies against each other; and, in particular, who sought to balance the strongest local powers. Seven Myths of the Spanish Conquest, page 48:

The search for native allies was one of the standard procedures or routines of Spanish conquest activity throughout the Americas. Pedro de Alvarado entered highland Guatemala in 1524 not only with thousands of Nahua all

[Like, it's not for nothing that the Aztecs told the Conquistadors that they thought the latter group were gods!]

It is unlikely that the Aztecs actually believed that the Conquistadors were gods.  (No primary sources state this; the original source for the gods claim was Francisco Lopez de Gomara, writing based on interviews with conquistadors who returned to Spain decades later; his writing contains many other known inaccuracies.)

Claims that are related to, but distinct from, the Aztecs believing that the Conquistadors were gods:

• The Aztecs, and other

On Diamond and writing, see previous discussion here. It is highly unlikely that writing was critical:

• Pizarro was illiterate
• The Aztecs had writing, yet didn't beat the Spaniards (or avoid having their leader kidnapped)
• Cortes' conquests were only a decade or so before- a short enough period that writing wasn't necessary to communicate the lessons. Pizarro was physically present in the Americas at the time.
• There's not actually any clear pathway from "have writing" -> "Atahualpa refuses to leave his army to meet with Pizarro". Writing did not make all

The specific evidence you’ve cited is weak. (1) You write that “The argument that we should be listening to experts and not random people would make a lot of sense if the "armchair" folks didn't keep being right.” It is extremely easy to be right on a binary question (react more vs less). That many non-experts were right is therefore more-or-less meaningless. (I can also cite many, many examples of non-experts being wrong. I think what we want is the fraction of experts vs non-experts who were right, but that seems both vague and unobtainable.)

(Note that t

That quote seems to provide no evidence that the 'literate tradition' mattered. Cortes' conquest was only 14 years before; Pizarro had arrived in the New World 10 years before that; Cortes' conquest involved many people and was a big/important deal; even if the Spanish had no writing at all, Pizarro would likely have known the general outline of Cortes' actions.

It's strictly speaking impossible to rule out Pizarro indirectly being influenced by writing; but I don't think it would be possible for stronger evidence against the importance of writing in this specific case to exist.

The Portuguese presumably were reasonably educated

Pizarro was illiterate.

That is not true; the CSA had worse railroads, but they were still important throughout the war. Some of the most important Union offensives late in the war- the Atlanta campaign and the siege of Petersburg- were intended to sever the South's railroads; and the war ended almost immediately after the Union cut off the railroad routes to the CSA capital of Richmond at the Battle of Five Forks. Both sides were heavily reliant on railroads for supply, and also used railroads to move troops (for the CSA, e.g. moving Longstreet's corps to fight at Chickamagua).

Homepage seems to lack links to the last two books.

Now, imagine you’re a diplomat, at a diplomatic conference. You see a group of diplomats, including someone representing one of your allies, in an intense conversation. They’re asking the allied diplomat questions, and your ally obviously has to think hard to answer them. Your intuition is going to be that something bad is happening here, and you want to derail it at all costs.

Source? I feel very, very confident that this is false. You would only want to break things up if you felt very confident that your ally would screw up answering the questions; otherwise, having lots of people paying careful attention to your side's proposals would be a very good sign.

In, now.

Literally every sentence you wrote is wrong.

The worst crimes of the holocaust were a conspiracy within the Nazi government.

This is not true. The Holocaust was ordered by the popular leader of the German government; they were executed by a very large number of people, probably >90% of whom actively cooperated and almost none of whom tried to stop the Holocaust. (see e.g. Christopher Browning's Ordinary Men) German society as a whole knew that their government was attempting genocide; see e.g. What We Knew for supporting details, or Wikipedia for a s... (read more)

All of these are plausibly true of art departments at universities as well. (The first two are a bit iffy.)

Thanks, this helped me!

As I understand it, the mainstream interpretation of that document is not that Bin Laden is attacking America for its freedom; rather, AQ's war aims were the following:

• End US support of Israel (also, Russia and India)
• End the presence of US troops in the Middle East (especially Israel)
• End US support for Muslim apostate dictators

See, e.g., this wikipedia article, or The Looming Tower. Eliezer is correct that AQ's attacks were not caused by AQ's hted of American freedoms.

The argument doesn't understand what the moral uncertainty is over; it's taking moral uncertainty over whether fetuses are people in the standard quasi-deontological framework and trying to translate it into a total utilitarian framework, which winds up with fairly silly math (what could the 70% possibly refer to? Not to the value of the future person's life years- nobody disputes that once a person is alive, their life has normal, 100% value.)

No I'm not. The Fizzbuzz article cited above is a wiki article. It is not based on original research, and draws from other articles. You will find the article I linked to linked to in a quote at the top of the first article in the 'articles' section of the wiki article; it is indeed the original source for the claim.

The quote does not claim there has been no filtering done before the interview stage. If you read the original source it explicitly states that it is considering all aplicants, not only those who make it to the interview stage: "We get between 100 and 200 [resumes] per opening."

You seem to be confusing applicants with people who are given interviews. Typically less than half of applicants even make it to the interview stage- sometimes much, much less than half.

There's also enough evidence out there to say that this level of applicants is common. Starbucks had over a hundred applicants for each position it offered recently; Proctor and Gamble had around 500. This guy also says it's common for programmers.

unless you believe more than 100 people on the average get interviewed before anyone is hired

This is accurate for the top companies- as of 2011, Google interviewed over 300 people for each spot filled. Many of these people were plausibly interviewed multiple times, or for multiple positions.

Maybe, but this is the exact opposite of polymath's claim- not that fighting a modern state is so difficult as to be impossible, but that fighting one is sufficiently simple that starting out without any weapons is not a significant handicap.

(The proposed causal impact of gun ownership on rebellion is more guns -> more willingness to actually fight against a dictator (acquiring a weapon is step that will stop many people who would otherwise rebel from doing so) -> more likelihood that government allies defect -> more likelihood that the government... (read more)

The Syrians and Libyans seem to have done OK for themselves. Iraq and likely Afghanistan were technically wins for our nuclear and drone-armed state, but both were only marginal victories, Iraq was a fairly near run thing, and in neither case were significant defections from the US military a plausible scenario.

SlateStarCodex.

We can know that other amphibious assaults probably had lower or neglible friendly fire rates, because some other landings (some opposed) had absolutely lower rates of casulaties- e.g here, here, and here.

Thanks for your response!

1) Hmmm. OK, this is pretty counter-intuitive to me.

2) I'm not totally sure what you mean here. But, to give a concrete example, suppose that the most moral thing to do would be to tile the universe with very happy kittens (or something). CEV, as I understand, would create as many of these as possible, with its finite resources; whereas g/g* would try to create much more complicated structures than kittens.

3) Sorry, I don't think I was very clear. To clarify: once you've specified h, a superset of human essence, why would you apply... (read more)