I note that that chart doesn't contain data for 2026. It also doesn't contain data for 2030.
Unless something catastrophic happens, [0] construction four years out is effectively locked in. When what's being constructed is expected to increase CO2 emissions by ~50% after it's built, it's quite reasonable to assert that the LLM Industry has reversed our collective work on cutting CO2 emissions. It's exceedingly reasonable to bitch bitterly about it when those very same entities driving the construction today were screaming left and right about all of the initiatives they'd been championing to reduce power consumption and Save The Planet and whatnot just a few years ago.
[0] ..such as a fuckoff-huge bubble that's being used to justify said construction messily exploding...
The construction plans are nuts and it looks very much like a real estate bubble.
Chip efficiency gains will mean that by 2030 you’ll be getting 100X the inference throughput on half the power. All this construction is not necessary.
They're in line with what's needed to meet the revenue projections of the LLM companies. Your "100x the inference" assertion ignores the most expensive part of the work of creating SOTA LLMs... the "training". That is what the DCs are primarily being built to do.
If this is about training, why does it have to be so geographically distributed? Why put data centers in cities? Why not just put the data centers near clean power sources?
In fact, using AI requires also inference power that needs to be closer to the consumers.
> ...why does it have to be so geographically distributed?
Probably because of economics, politics, and logistics. Much easier and safer to have a bunch of relatively small projects running in a large variety of jurisdictions than one fuckoff-huge project that's the size of all of the smaller projects combined, don't you think? There's probably some folksy saying about that sort of thing...
> Why not just put the data centers near clean power sources?
a) Unless you live in Texas, folks in the US are connected to a nationwide electricity-distribution grid. Folks in nearly all other other countries are in similar situations. Proximity to generation is nice, but not mandatory.
b) If there was grid capacity available to run the new datacenters that are in various stages of project completion, why would they be running on-site natural-gas fired turbines? Those are way less efficient and way more expensive to run than grid-scale generation facilities. That'd be a huge waste of money if the datacenter operators could hook into the grid, yeah?
> ...that needs to be closer to the consumers to reduce latency.
Needs? No. RTT all the way across the globe is right around 250 milliseconds; I checked. For this part of these sorts of programs, servers in the same region as their clients are nice to have but not at all required.
Depends on how big the efficiency gains are. There is not infinite demand. Jevon’s paradox occurs when there is latent demand that is released at a lower price point.
I also found that these projections are 15 years ahead and assume nearly all power is from gas when battery and solar costs are still falling geometrically and data center demand has reignited modular nuclear interest.
I’m generally skeptical of all predictions of the form “look at this terrifying thing that happens if we extrapolate the current trends indefinitely!” See: “The Population Bomb” by Paul Erlich and countless other doom predictions of a similar nature.
Yeah my biggest concern is we have already tipped the climate and it’s currently ratcheting up to the next stable slot… one where we maybe lose Miami and experience huge agricultural region shifts that will destabilize governments and societies.
But hey. FAFO. We will live and adapt. Humans are like cockroaches, as a bio professor said to me once, and he explained that this is not an insult.
There's something I find difficult when reading this piece. The thing is, I've been making a living as a programmer and I'm now in my seventh year, but programming wasn't actually my major. I majored in computational physics in university, jumped into programming on that basis, and taught myself programming. So does that mean I have no right to make up programming problems and ask questions? I mean, couldn't Bill Gates have studied biology hard and understood things outside his own field too?
Or is the theme of the piece that one ought to have epistemic humility?
But the thing is, this piece itself also makes claims about psychology, power theory, and human nature to some degree, which is ironic. But that can't be it. Because this piece has no epistemic humility in it.
Then, if we demand according to what this piece says, in the end it comes down to "no one can test AI outside the major they studied in university" — but honestly, isn't it just a matter of buying the textbooks and papers and testing AI with them?
> does that mean I have no right to make up programming problems and ask questions?
Of course not, but that's not what's being argued. If you believed you could authoritatively declare what AI is or is not, that would be different, though.
> couldn't Bill Gates have studied biology hard and understood things outside his own field too?
Yes but he would still be a dilettante. Belief to the contrary is more likely to be an indication of engineer's disease.
To be honest with you, I don't understand why you rate this piece so highly. This piece basically shows a strong self-contradiction. If it's due to human writing having self-contradiction, then this piece would be human. If that's the case, I think AI writing is better.
The point of the matter is that the intellectual humility the author of this piece advocates isn't present in this piece. What papers are cited for the psychological claims the author of this piece makes? Of course, there's no need to do that in a blog essay. But the target of this piece's criticism is the absolutization of one's own standards and epistemic humility, and the problem is that this piece also commits that contradiction. This is a matter of the piece's level. I respect the intellectual aesthetics of people who like this kind of writing, but it's not my taste.
Well said. I don't believe the author has any credentials or accolades in psychology or sociology, but they certainly have strong opinions on how other people should form opinions.
What I am seeing in the article is an unchecked case of cognitive dissonance. That's just the overwhelming sense I get. You identified it as self-contradiction. The reason I label it cognitive dissonance is that the author is arguing against something that they themselves are projecting; that subject matter experts in one domain are overconfident in another.
This article's self-fulfilling prophecy is that it will one day be used as an argument from authority, if it hasn't already.
The overconfidence effect is well documented and a known thing in light aircraft private aviation. People with education and good paying professional careers (doctors, lawyers, dentists, etc) who have the disposable income to buy a $300,000 aircraft and maintain it can easily become overconfident in their abilities. Because they're good at one thing in their life that requires a lot of education and training, they think it automatically extends to every other domain of knowledge that they set themselves to.
The original v-tailed Beechcraft Bonanza used to be known as the "v-tailed doctor killer".
Every IT project I've worked on for the last 25years have been successful even though usually 30% of IT project fail.
I'm smart enough to know that I can't make a distinction between my expertise or luck for the success of those project. But my ego wants it to be my expertise because I'm the only constant in those projects
> Anything approaching my skill level is superhuman
Or someone has the impression that Ai can code as well as them (and find errors) but at 10x the speed or more. They then conclude that human ability and superhuman speed is superhuman in total.
This article first makes a point that Gates isn’t expert enough in biology to assess AIs capability to answer complex questions on the topic as being a breakthrough.
Then immediately goes on to say that his comments about AI coding are arrogant despite being an expert in that field with more experience than nearly anyone.
The funny thing is, even before AI staff engineers (especially in larger companies) often didn’t get to write code; most of the work was supervising other engineers who actually wrote the code for the architectures they designed… which was mostly docs. I used to joke that my favorite IDE’s were “IntelliJ, Google Docs and other engineers.”
I would consider the average staff engineer, if appropriately titled, to be an expert in their field. If not a staff engineer then when exactly would one be considered an expert?
> being an expert in that field with more experience than nearly anyone
Nothing annoys me more than business folks who try to claim street cred from their halcyon days as a programmer.
These folks are consistently the ones who make the worst decisions because they falsely believe their intuition of how software worked 20 years ago is valuable insight for technical decisions being made today.
I always get the "ick" when I see someone successful in one domain assume they're automatically brilliant in all others. Like someone who made a fortune selling B2B SaaS going on a podcast talking nonsense about "quantum" or whatever.
I agree with you generally but this was a poor example of it. He wasn't claiming to be an expert in biology or that the AI would be if it did well on the rest. He was saying it was proving it could read and use information usefully. (And showing unbelievable hubris he said he could write a high school biology problem without a PhD in biology education)
he presents the idea of Bill Gates observing an OpenAI model a. doing "complex" biology problems, b. some coding problems, and then based on his assumption that Bill Gates knows rudimentary biology, and used to also write code, the fact that Bill Gates is impressed with the OpenAI model means Bill Gates thinks he has superhuman biology and coding skills. This simply does not follow.
This would be like if I went in my garage, spent several weeks with the 3d printer and CNC machine to build a robot that can "dance", then I made a video of me dancing about as well as your average dad at a wedding along with my dancing robot and I said "this thing can dance better than me!", therefore: "Ha, your robot sucks! You must think you're Baryshnikov if you call that dancing". I mean this is kind of sad.
Slightly more than a couple of years ago, just observing an LLM coherently doing work at the academic level of a sixth grader was brand new to all of us. Which was objectively impressive. It doesnt matter "so what? anyone can do 6th grade math"....this..is not..the point...
edit: the point they want to make, that these people are not qualified to make statements about "superintelligence", and that they see themselves fit to determine "superintelligence" based on watching the robots be smarter than them, is one I generally agree with, and this is a kind of hubris. but it's not a hubris you can assign to the general LLM-using population as this seems to be doing.
I find the authors point not very compelling.. Bill Gates was by all accounts an exceptional programmer at one point and then spent a long time evaluating software engineers. He seems like a natural person to gauge AI progress in coding?
Around when AI was getting much better at SQL I was writing SQL everyday at work and had been for years. I don’t think it would’ve been hubris for me to judge that it was getting really good and would shortly be super human at SQL. It would’ve been correct anyway.
He wrote software for controlling hydroelectric dams when he was still in high school. At a time where the tooling was incredibly worse than it is now.
What's your definition of exceptional? Because that fits mine.
Computers, programming languages, and software in general were simple back then. It didn't take much to program them. You couldn't do much with a few KB of memory and a single-core single-thread CPU with basic instructions.
It took some amount of talent to squeeze something useful into computers with those constraints, but the upper limit on complexity also put an upper limit on how exceptional anything could be.
Bill Gates is/was an exceptional businessman. His technical abilities were nothing special.
By your definition, is John Carmack an exceptional programmer? After all, Wolfenstein3d is also simple, if you ignore the challenges of limited hardware and of being an early mover who has to invent lots of things from scratch. Who from that time period would qualify?
Exceptional is a relative term not an absolute one. I dislike Gates but you can't simply argue that he was not technically exceptional because the primitive tools he used capped how complex the software he wrote could be from a modern perspective of software complexity. It is ridiculous. By the same token, any software anyone living can possibly produce pales by the standards of a 2090 dev.
The reasonable take is that Gates was an exceptional dev for his era for the things he did with the things he had.
Gates and Allen wrote an 8080 emulator on the Harvard PDP-10, then wrote Altair BASIC for the simulated 8080, which worked on the first try on a real Altair without their ever having seen one, after laboriously entering it with toggle switches at MITS in Albuquerque. <https://news.ycombinator.com/item?id=43579471>
The author talked about programming specifically after biology. Even so, if you got a 5 on say AP Bio back in the day and have kept up with biology (professionally in Bills case or as a parent).. it seems like you would be a natural person to gauge how good AI is at AP exam level questions.
The author's point is that "better than you" doesn't mean "superhuman", it just means "usefully good".
If we assume that you're 95th percentile at writing SQL, it's going to be amazing when companies can just copy-paste the 0.95 SQL dev instead of having to pay for that expertise -- but it doesn't really mean that AI improvement is going exponential, or that we've reach the singularity, or that AI will cure all disease and end human suffering or any of the other nonsense you see AI boosters predicting on this website every day.
Yeah I think the authors just mistaken. If you are top 5% or so in some skill and you see AI over a few years go from median to surpassing your skill level, you can use common sense and realize it’s going to get past the next 5%, and it has.
It feels like the author is making a bit of a big deal about bill gates declaring superhuman in January when it might take a year or so before it’s undeniable?
> Bill Gates was by all accounts an exceptional programmer at one point
Perhaps, but that point was in the 1970s, 50 years before he was evaluating OpenAI, and there were vastly fewer programmers in the 1970s, so "exceptional" then isn't necessarily exceptional now.
A footnote in the article mentions a crucial way he was exceptional: "Bill Gates went to one of the only high schools in the world that had a computer".
The sentence "X was an exceptional programmer" surely indicates that the exceptionality was in relation to his time period.
Are you going to say that someone like Salieri and Chopin were not technically exceptional because not that many people in the planet had a piano available?
> Are you going to say that someone like Salieri and Chopin were not technically exceptional because not that many people in the planet had a piano available?
There's a crucial difference: music has not become significantly more complex over the past 200 years. As far as I'm aware, pianos themselves haven't changed significantly since then. In contrast, computers and programming have changed dramatically since the 1970s. Whatever Gates did in the 1970s, it's not clear how 2020s Bill Gates, who stopped programming full-time decades ago, is qualified to evaluate contemporary programming.
1. Can someone _who haven't programmed for 50 years_ correctly evaluate current coding AI, and
2. Is what made a good programmer in the 70s the same as what make is good today to evaluate current coding AI.
If someone did competitive bike racing in 1970 and stopped since then, he probably forgot a lot of things and is not a good judge on current cycling. But if he were to be magically transported from that time, I don't see what he would lack. Sure, bikes are different (and better), nutrition is better understood, and so on. But the core job is the same.
> You conflate two different arguments: 1. Can someone _who haven't programmed for 50 years_ correctly evaluate current coding AI, and 2. Is what made a good programmer in the 70s the same as what make is good today to evaluate current coding AI.
I actually don't. See my other comment: "You're conflating two different things, (1) difficulty in principle with (2) practical knowledge and experience." https://news.ycombinator.com/item?id=50498280
> competitive bike racing
You cherry-picked an activity that depends primarily on the physical limits of the human body, which as far as I'm aware hasn't evolved significantly since the 1970s.
Still, training and sports science has evolved, as you note: "nutrition is better understood". Thus, I suspect that an exceptional racer magically transported from the 1970s might not be as exceptional now. Moreover, the Earth's population is larger now, and bike racing has become popularized by TV coverage, attracting more enthusiasts, not to mention more money, so the competition is now tougher. You can see this even more clearly in sports that were traditionally very white, such as golf and tennis: more non-white people are now competing in those sports, opening to a global population. At the top, the bar has been raised.
> Do you think programers in the 70s would be considered unskilled nowadays?
Yes. C++ was created in 1985. JavaScript was created in 1995. I could go on and on. If a 70s programmer frozen in ice was unthawed today, that person would be unable to get a programming job.
The funny thing is that contemporary professional programmers who rely on Claude are already complaining that they've forgotten how to program, so how much could Bill Gates possibly remember 50 years later?
Sure as we all know writing those Javascript CRUD apps like what the majority of today's programmers do is so much more difficult than what those losers in the 70s did.
> Sure as we all know writing those Javascript CRUD apps like what the majority of today's programmers do is so much more difficult than what those losers in the 70s did.
You're conflating two different things, (1) difficulty in principle with (2) practical knowledge and experience.
Could Bill Gates learn contemporary programming if he put in the time and effort? Perhaps. The point is that he hasn't. Although he was a 1970s programming expert, he's a 2020s programming n00b.
Any competent and experienced programmer who has learned a few diverse programming languages, even if those are only languages of the 60s, not even of the 70s, e.g. ALGOL 68, IBM PL/I, APL, LISP, SNOBOL, SIMULA 67, will be able to switch in a week or so to any other programming language of today, and reach the same level of competence as a programmer who has programmed only in that language for 10 years.
The job postings that require things like "5 years of experience with C++" or "5 years of experience with JavaScript" are really dumb, because anyone who really understands programming can do that in any programming language, after a few days needed to memorize its quirks, to avoid checking its specification continuously.
The number of innovations in programming languages that have been done after 1980 is extremely small in comparison with those done before 1980.
There were vastly fewer programmers in the 1970s, but there is no evidence whatsoever that the median skill or the top skill were lower than today.
The extra programmers of today have filled the voids in the distribution of programmers vs. skill, but in the best case for the present programmers they have not shifted the distribution towards a higher skill. If there was any shift in the distribution, it seems more likely that it was towards a lower skill.
I am pretty certain that any programmer who was exceptional in the 1970s would still be exceptional today.
Several decades ago, when I was very young, I reverse engineered a few Microsoft programs for CP/M, including the BASIC interpreter, the FORTRAN compiler and the FORTRAN run-time library.
At least the BASIC interpreter is likely to have included code written by Bill Gates. Given the hardware constraints, those programs were well written. They also included a lot of assembly language tricks that would not be used today, but which were justified in a program that had to run on very weak CPUs and using very little memory.
At that time, I considered those programs very instructive.
1) Although I think there is timeless inborn ability, I don't think there is timeless programming skill, irrespective of advances in technology. We clearly disagree about that, though: https://news.ycombinator.com/item?id=50498336
2) With regard to timeless inborn ability, I think it's obvious that you will end up with higher achievers if you start with a much larger population all working on the same activity.
He also tries to dunk on Gates by saying he claimed he produced novel problems in science to test AI.
I don’t think the author realizes that “AP” is an exam for American high schoolers. Gates is only claiming he made new test questions at a high school+ level that wouldn’t have been in the training set for AI. I don’t see how it’s so unbelievable he could do that.
> Gates is only claiming he made new test questions at a high school+ level that wouldn’t have been in the training set for AI. I don’t see how it’s so unbelievable he could do that.
It’s unbelievable because Gates is neither a biology exam author nor a biology textbook author, so how would he even know what‘s in the training set for AI?
> Gates is only claiming he made new test questions at a high school+ level that wouldn’t have been in the training set for AI. I don’t see how it’s so unbelievable he could do that.
It’s unbelievable because Gates is neither a biology exam author nor a biology textbook author, so how would he even know what’s in the training set for AI? To know, you would need to be an experienced biology teacher and survey the pedagogical literature.
If you did well on AP Bio in the distant past and have kept up with the field of biology like Gates it is normal you could ask some high school exam level questions.
When I was studying for AP Calculus or Literature etc I’m sure I could’ve made decent sample questions.
Nobody is disputing whether Gates could come up with some exam questions. They're disputing whether Gates could come up with novel exam questions that weren't in the AI training data.
Especially because good questions are more than simply asking about factoid X in Book Y. Good questions test if you really understood something and can transfer knowledge.
Coming up with such questions is not easy, and just because you did well on a test or "kept up with the field a bit" does not mean you can ASK such questions very well.
This is the crux though, isn't it? Are people as overconfident as the article claims, or are there many 'hobbyist' experts milling around? I did a Bachelor's in philosophy before switching to CS, and have been fascinated by IR/political theory since high school. While I could not hope to learn it all, I can hold a good conversation with a researcher in either philosophy or IR, especially if it intersects with a topic of my core expertise.
To flip it around, I have friends who I would consider (hobbyist) experts in CS. The geneticist is much better at C++ and OOP than me, the geologist knows a lot about 3d modelling and the whole GIS world, and the economist knows way more about DS, AI and statistics. Who am I to say they aren't experts? This kind of rigid division of field and expertise made no sense in university, and it only breaks apart more as we get older and more experienced.
> Are people as overconfident as the article claims, or are there many 'hobbyist' experts milling around? I did a Bachelor's in philosophy before switching to CS, and have been fascinated by IR/political theory since high school.
Heh, well, by coincidence, before I became a programmer I spent (too many) years as a PhD candidate in philosophy, and I'm skeptical about the extent to which undergraduates can be considered experts. ;-)
That was a long time ago, though, and despite my experience, I've undoubtedly fallen far behind the current work in the field.
Surely even undergrads know something, and knowing more than high-school curriculum - which is the bar we're discussing - isn't some particularly difficult achievement.
What if we flip this question around. Given what we’ve seen LLMs achieve today, if Bill Gates actually got an accurate read on their capabilities back then, what does that say about his expertise in biology? Maybe he got Dunning-Kreuger’d and made a lucky guess… or his assessment was accurate.
Given his track record and the state of AI today, I know which one I find more likely.
The article is a case of choosing and using an example so poorly, at length, that it casts severe doubt on the reasoning.
Take the claim of superhuman for example. If an AI is as good as Bill Gates at coding, but is relentless and tireless, then that is fairly called superhuman, the same way a drilling machine was superhuman relative to John Henry.
Ah yes, the professor from Hoboken is here to teach us that tech bros are delusional, and it’s really professors from Hoboken who know what is actually going on.
Biases and heuristics have become somewhat mainstream “intellectualism”, but just like people using the words prolific and myriad incorrectly, is leveraged here to a false conclusion.
The anti-AI narrative continues, to the detriment of those who listen to its self-destructive guidance that equates excessive enthusiasm with delusion.
Both the author's and your perspectives can be wrong at the same time, yet have elements of truth. Certainly the phrase "coastal elites" did not come from nothing. At the same time, the critiques have their own form of intellectual elitism.
The narrative of AI being of illusory value, only appealing to the delusional tech elite, is one that serves only to soothe the emotions of those who are unwilling to embrace reality.
That's not the narrative that I read in the post. It was people who excel in on area believing their expertise translates to other things. The leaders of Ai are doing this
None of this matters. By the time truth out itself, the Kushner's, Gate's, Andreeson's, and Altman's of the world will have walked off with hundred of billions of our 401k money.
I am no fan of Bill Gates. He's pedophile adjacent, in my *opinion.
But I was there. He worked really, really fucking hard. As did most of the early crowd. If you're asking yourself how people back then worked so hard, just realize that there are drugs that can help you do that.
> I picked on Bill Gates above to illustrate the overconfidence effect
After having read your blog i can assert you've demonstrated that effect on yourself. Bill has all capability and credentials proven so we know this guy is an authority on software engineering topics, you have zero if i'm not mistaken.
Gates doesn't have the biology credentials. He's a pedophile, to top it off, so most probably that's where the article should have started with. It's telling that for most of the last 30 or so years the most wealthy person on this planet has been a pedophile, come to think of it maybe we (the human species) do deserve what's coming.
There's this thing people with strong opinions about something love to do: appeal to authority when it suites their agenda. This guy was once, a very long time ago in technical terms, very accomplished, so therefore everything he says now is more credible.
It's absolutely infuriating because it allows people with a lot of money tell you they know what they're talking about without proving it. Your money is your perceived effort, no matter how you acquired it.
I would bet his spectacularly good calls vastly outnumber his spectacularly bad calls. After stepping down as CEO, he was the chief architect and remains a technology advisor of the tech company that is currently one of the most valuable companies ever. That basically implies a long rarely-broken string of good calls. He has led software engineering at a scale only a handful of people ever have.
I would say he has more authority on software engineering topics than most people on earth, but that’s just me.
You are just focusing on the tech bit. Bill gates never was a very good person to begin with and then he did some stuff that pretty much offset whatever good he did.
As for software engineering: if you think Bill Gates is an authority in the field I would like to introduce you to Donald Knuth, Peter Norvig, Fabrice Bellard, Dennis Ritchie, Linus Torvalds and many others. Bill Gates wouldn't make the top 100.
I think you're conflating computer science with software engineering. Other than maybe Peter Norvig and Linus none of them would top the list for "software engineering" per se. And talking of computer science, consider that as an undergrad student Bill Gates found and published an algorithmic result that was not beaten for 30 years: https://www.npr.org/2008/07/04/92236781/before-microsoft-gat...
> I would bet they don't.
I mean, how many "But but but antitrust!" and "Nobody got fired for buying Microsoft!" excuses can be made for Microsoft being where it is today?
I am focusing on the tech bit because the subthread was about his technical skills. As an engineer and a tech executive, everything I've read about him places him far above most people you could name. I also think he ranks very highly as a general scientific and systems thinker. As an example, he was harping on about pandemics long before COVID happened. He's been harping about AI and jobs these days, which I think many here would also agree with.
If we were discussing him as a person I have a more mixed opinion of him, but it's way more positive than the hate he gets around here, which I find hilariously misplaced.
Author is spot on on Bill Gates. By his own admission, his company essentially missed the coming of the Internet, and we all literally suffer today for it.
He missed the coming interconnectedness of the machines and instead presumed that every computer was going to be a fancy copy machine calculator (that will never talk to others, thus security is not needed).
billg wrote the “Internet Tidal Wave” memo in May of 1995! This was an era when the dominant enterprise IT infrastructure was heavyweight (Windows!) client/server and homes were just starting to connect to the internet via dialup en masse.
Microsoft proceeded to pivot so successfully that they monopolized the web browser market alongside winning a massively profitable position in enterprise “intranet” apps and TCP/IP-based servers.
I'm trying to find the exact clip, but there's a segment of an interview where Elon Musk explains why he finds it likely that we are living in a simulation, and part of his evidence is how unlikely he finds it to be him: He's the richest man in the world, the first trillionaire, he's influencing geopolitics and launching space rockets.
He finds it so astounding to be himself, to feel like the Main Character of the World that he considers this evidence that the world he experiences is a simulation. The more humbling proposition -- that the world is largely random and he just happens to be the guy who kept rolling sixes -- is less compelling.
What is the opposite of the Dunning-Kruger effect? Where you believe since you don’t know anything about a topic (eg biology), therefore nobody else can (you may have some incredible decades of research experience to have novel thoughts on it)?
The first part of the article makes sense in that:
In any system of competing agents (such as human beings) where the reward for performance can enhance performance, will always result in even minor random differentials in initial performance exponentially accumulating until one or a handful of individuals dominate. In such a system, only a small portion of success can be attributed to innate skills. So it's true that success in itself, while signaling fitness to survive, does not necessarily signal an outsized ability to make accurate judgements.
I'm not entirely clear what the second part of the article is trying to convey. Is it that highly successful individuals attribute their success entirely to their skill and therefore when a machine is able to replicate their skillset, they are surprised and consider the technology to be world-altering?
What I have seen first-hand is very talented individuals in one domain (tech) assume immediate authority in a novel domain (in my case it was local city planning). The thinking seems to be sort of the human equivalent of AGI? Note how confidently Gates talks about his expertise in novel biological science, for example.
Not doubting the intelligence of these individuals, just saying they sure tend to undervalue domain specific expertise. I don't know who Elon consulted about his cybertruck, but it wasn't truck people.
I do worry that AI exasperates this problem as a quick few queries and thoughtful questions have certainly given me over-confidence in a new subject. You don't know what you don't know.
"the very notion that something else could be "as good as I am" is, first of all, hard to believe and secondly necessarily means that it's "superhuman". That man is not lacking self-esteem, no sirree, Bob. "
I like that the article just tries to pass off a computer program coding at the level of humans (not to mention solving Millennium Problems) as something that should be just be totally unsurprising. Bill Gates is just a dumb arrogant tech bro to be surprised that a neural network exceeds his abilities at coding.
I think this is written forgetting a lot of the context of the time (I was at a relevant part of the industry to be present around some of the early “oh shit this thing works” moments).
It was absolutely not a given that next-token prediction would scale to the point that it could answer outside of its training domain (this is the point about “made up questions”, it was important to test it on material outside of the training set since the scale of training was yet-unheard-of at the time).
Further, the success of post-training on verifiable tasks (coding) was also a complete shock that changed the world overnight. LLMs jumped the “I can make code” to “I confidently can make code that solves your problem” chasm suddenly, and it has changed the industry.
For someone so invested in this from the early days (remember the scale of Microsoft’s commitment) it is entirely logical that testing things “on day one” would occur and this would be a much bigger shock compared to most people that learned of advances at the same time as the rest of the world. (There are huge delays between research capabilities and shipping)
It seems pretty clear to me that Bill was saying AI was as good as him a few months ago and now it is superhuman. This is an interview transcript, remember, not a written affidavit. Humans can switch focus mid stream.
Designing new algorithms is where the real skill is, but an average programmer will never do this, and even a 10x programmer will do it maybe a few times in their lifetime at most.
145 comments
[ 0.25 ms ] story [ 44.6 ms ] threadhttps://ourworldindata.org/profile/co2/united-states
The second chart is the one I’m looking at. Where is the 15 year reversal? If zoom in I see a small uptick. Not a 15 year trend reversal.
Unless something catastrophic happens, [0] construction four years out is effectively locked in. When what's being constructed is expected to increase CO2 emissions by ~50% after it's built, it's quite reasonable to assert that the LLM Industry has reversed our collective work on cutting CO2 emissions. It's exceedingly reasonable to bitch bitterly about it when those very same entities driving the construction today were screaming left and right about all of the initiatives they'd been championing to reduce power consumption and Save The Planet and whatnot just a few years ago.
[0] ..such as a fuckoff-huge bubble that's being used to justify said construction messily exploding...
Chip efficiency gains will mean that by 2030 you’ll be getting 100X the inference throughput on half the power. All this construction is not necessary.
They're in line with what's needed to meet the revenue projections of the LLM companies. Your "100x the inference" assertion ignores the most expensive part of the work of creating SOTA LLMs... the "training". That is what the DCs are primarily being built to do.
In fact, using AI requires also inference power that needs to be closer to the consumers.
Probably because of economics, politics, and logistics. Much easier and safer to have a bunch of relatively small projects running in a large variety of jurisdictions than one fuckoff-huge project that's the size of all of the smaller projects combined, don't you think? There's probably some folksy saying about that sort of thing...
> Why not just put the data centers near clean power sources?
a) Unless you live in Texas, folks in the US are connected to a nationwide electricity-distribution grid. Folks in nearly all other other countries are in similar situations. Proximity to generation is nice, but not mandatory.
b) If there was grid capacity available to run the new datacenters that are in various stages of project completion, why would they be running on-site natural-gas fired turbines? Those are way less efficient and way more expensive to run than grid-scale generation facilities. That'd be a huge waste of money if the datacenter operators could hook into the grid, yeah?
> ...that needs to be closer to the consumers to reduce latency.
Needs? No. RTT all the way across the globe is right around 250 milliseconds; I checked. For this part of these sorts of programs, servers in the same region as their clients are nice to have but not at all required.
I also found that these projections are 15 years ahead and assume nearly all power is from gas when battery and solar costs are still falling geometrically and data center demand has reignited modular nuclear interest.
I’m generally skeptical of all predictions of the form “look at this terrifying thing that happens if we extrapolate the current trends indefinitely!” See: “The Population Bomb” by Paul Erlich and countless other doom predictions of a similar nature.
Not being Pollyanna either. Just being realistic.
1. Trump's War
2. Falling birth rates
3. Petro vs Electro state directions (Russia -> US -> China)
There are also unknowns around the thawing of the permafrost at northern latitudes.
But hey. FAFO. We will live and adapt. Humans are like cockroaches, as a bio professor said to me once, and he explained that this is not an insult.
Hyperscalers want to build out enough that the trend will reverse. How much is t.b.d., America is likely to slow down the buildout
Or is the theme of the piece that one ought to have epistemic humility?
But the thing is, this piece itself also makes claims about psychology, power theory, and human nature to some degree, which is ironic. But that can't be it. Because this piece has no epistemic humility in it.
Then, if we demand according to what this piece says, in the end it comes down to "no one can test AI outside the major they studied in university" — but honestly, isn't it just a matter of buying the textbooks and papers and testing AI with them?
It's a slightly difficult problem.
Of course not, but that's not what's being argued. If you believed you could authoritatively declare what AI is or is not, that would be different, though.
> couldn't Bill Gates have studied biology hard and understood things outside his own field too?
Yes but he would still be a dilettante. Belief to the contrary is more likely to be an indication of engineer's disease.
The point of the matter is that the intellectual humility the author of this piece advocates isn't present in this piece. What papers are cited for the psychological claims the author of this piece makes? Of course, there's no need to do that in a blog essay. But the target of this piece's criticism is the absolutization of one's own standards and epistemic humility, and the problem is that this piece also commits that contradiction. This is a matter of the piece's level. I respect the intellectual aesthetics of people who like this kind of writing, but it's not my taste.
What I am seeing in the article is an unchecked case of cognitive dissonance. That's just the overwhelming sense I get. You identified it as self-contradiction. The reason I label it cognitive dissonance is that the author is arguing against something that they themselves are projecting; that subject matter experts in one domain are overconfident in another.
This article's self-fulfilling prophecy is that it will one day be used as an argument from authority, if it hasn't already.
I see biases claims against billg can be applied against the post author.
The original v-tailed Beechcraft Bonanza used to be known as the "v-tailed doctor killer".
I'm smart enough to know that I can't make a distinction between my expertise or luck for the success of those project. But my ego wants it to be my expertise because I'm the only constant in those projects
Or someone has the impression that Ai can code as well as them (and find errors) but at 10x the speed or more. They then conclude that human ability and superhuman speed is superhuman in total.
Then immediately goes on to say that his comments about AI coding are arrogant despite being an expert in that field with more experience than nearly anyone.
Unsure how to interpret that.
Being CEO does not involve a lot of coding. Yes he was an expert 40 years ago, but a lot has happened in 40 years
Nothing annoys me more than business folks who try to claim street cred from their halcyon days as a programmer.
These folks are consistently the ones who make the worst decisions because they falsely believe their intuition of how software worked 20 years ago is valuable insight for technical decisions being made today.
Then ask Claude!
I always get the "ick" when I see someone successful in one domain assume they're automatically brilliant in all others. Like someone who made a fortune selling B2B SaaS going on a podcast talking nonsense about "quantum" or whatever.
This would be like if I went in my garage, spent several weeks with the 3d printer and CNC machine to build a robot that can "dance", then I made a video of me dancing about as well as your average dad at a wedding along with my dancing robot and I said "this thing can dance better than me!", therefore: "Ha, your robot sucks! You must think you're Baryshnikov if you call that dancing". I mean this is kind of sad.
Slightly more than a couple of years ago, just observing an LLM coherently doing work at the academic level of a sixth grader was brand new to all of us. Which was objectively impressive. It doesnt matter "so what? anyone can do 6th grade math"....this..is not..the point...
edit: the point they want to make, that these people are not qualified to make statements about "superintelligence", and that they see themselves fit to determine "superintelligence" based on watching the robots be smarter than them, is one I generally agree with, and this is a kind of hubris. but it's not a hubris you can assign to the general LLM-using population as this seems to be doing.
Around when AI was getting much better at SQL I was writing SQL everyday at work and had been for years. I don’t think it would’ve been hubris for me to judge that it was getting really good and would shortly be super human at SQL. It would’ve been correct anyway.
What's your definition of exceptional? Because that fits mine.
It took some amount of talent to squeeze something useful into computers with those constraints, but the upper limit on complexity also put an upper limit on how exceptional anything could be.
Bill Gates is/was an exceptional businessman. His technical abilities were nothing special.
The reasonable take is that Gates was an exceptional dev for his era for the things he did with the things he had.
If we assume that you're 95th percentile at writing SQL, it's going to be amazing when companies can just copy-paste the 0.95 SQL dev instead of having to pay for that expertise -- but it doesn't really mean that AI improvement is going exponential, or that we've reach the singularity, or that AI will cure all disease and end human suffering or any of the other nonsense you see AI boosters predicting on this website every day.
It feels like the author is making a bit of a big deal about bill gates declaring superhuman in January when it might take a year or so before it’s undeniable?
Perhaps, but that point was in the 1970s, 50 years before he was evaluating OpenAI, and there were vastly fewer programmers in the 1970s, so "exceptional" then isn't necessarily exceptional now.
A footnote in the article mentions a crucial way he was exceptional: "Bill Gates went to one of the only high schools in the world that had a computer".
Are you going to say that someone like Salieri and Chopin were not technically exceptional because not that many people in the planet had a piano available?
There's a crucial difference: music has not become significantly more complex over the past 200 years. As far as I'm aware, pianos themselves haven't changed significantly since then. In contrast, computers and programming have changed dramatically since the 1970s. Whatever Gates did in the 1970s, it's not clear how 2020s Bill Gates, who stopped programming full-time decades ago, is qualified to evaluate contemporary programming.
1. Can someone _who haven't programmed for 50 years_ correctly evaluate current coding AI, and 2. Is what made a good programmer in the 70s the same as what make is good today to evaluate current coding AI.
If someone did competitive bike racing in 1970 and stopped since then, he probably forgot a lot of things and is not a good judge on current cycling. But if he were to be magically transported from that time, I don't see what he would lack. Sure, bikes are different (and better), nutrition is better understood, and so on. But the core job is the same.
I actually don't. See my other comment: "You're conflating two different things, (1) difficulty in principle with (2) practical knowledge and experience." https://news.ycombinator.com/item?id=50498280
> competitive bike racing
You cherry-picked an activity that depends primarily on the physical limits of the human body, which as far as I'm aware hasn't evolved significantly since the 1970s.
Still, training and sports science has evolved, as you note: "nutrition is better understood". Thus, I suspect that an exceptional racer magically transported from the 1970s might not be as exceptional now. Moreover, the Earth's population is larger now, and bike racing has become popularized by TV coverage, attracting more enthusiasts, not to mention more money, so the competition is now tougher. You can see this even more clearly in sports that were traditionally very white, such as golf and tennis: more non-white people are now competing in those sports, opening to a global population. At the top, the bar has been raised.
Yes. C++ was created in 1985. JavaScript was created in 1995. I could go on and on. If a 70s programmer frozen in ice was unthawed today, that person would be unable to get a programming job.
The funny thing is that contemporary professional programmers who rely on Claude are already complaining that they've forgotten how to program, so how much could Bill Gates possibly remember 50 years later?
You're conflating two different things, (1) difficulty in principle with (2) practical knowledge and experience.
Could Bill Gates learn contemporary programming if he put in the time and effort? Perhaps. The point is that he hasn't. Although he was a 1970s programming expert, he's a 2020s programming n00b.
The job postings that require things like "5 years of experience with C++" or "5 years of experience with JavaScript" are really dumb, because anyone who really understands programming can do that in any programming language, after a few days needed to memorize its quirks, to avoid checking its specification continuously.
The number of innovations in programming languages that have been done after 1980 is extremely small in comparison with those done before 1980.
The extra programmers of today have filled the voids in the distribution of programmers vs. skill, but in the best case for the present programmers they have not shifted the distribution towards a higher skill. If there was any shift in the distribution, it seems more likely that it was towards a lower skill.
I am pretty certain that any programmer who was exceptional in the 1970s would still be exceptional today.
Several decades ago, when I was very young, I reverse engineered a few Microsoft programs for CP/M, including the BASIC interpreter, the FORTRAN compiler and the FORTRAN run-time library.
At least the BASIC interpreter is likely to have included code written by Bill Gates. Given the hardware constraints, those programs were well written. They also included a lot of assembly language tricks that would not be used today, but which were justified in a program that had to run on very weak CPUs and using very little memory.
At that time, I considered those programs very instructive.
2) With regard to timeless inborn ability, I think it's obvious that you will end up with higher achievers if you start with a much larger population all working on the same activity.
I don’t think the author realizes that “AP” is an exam for American high schoolers. Gates is only claiming he made new test questions at a high school+ level that wouldn’t have been in the training set for AI. I don’t see how it’s so unbelievable he could do that.
It’s unbelievable because Gates is neither a biology exam author nor a biology textbook author, so how would he even know what‘s in the training set for AI?
It’s unbelievable because Gates is neither a biology exam author nor a biology textbook author, so how would he even know what’s in the training set for AI? To know, you would need to be an experienced biology teacher and survey the pedagogical literature.
When I was studying for AP Calculus or Literature etc I’m sure I could’ve made decent sample questions.
Coming up with such questions is not easy, and just because you did well on a test or "kept up with the field a bit" does not mean you can ASK such questions very well.
More like half of HNs believe they're near-experts in some random field.
To flip it around, I have friends who I would consider (hobbyist) experts in CS. The geneticist is much better at C++ and OOP than me, the geologist knows a lot about 3d modelling and the whole GIS world, and the economist knows way more about DS, AI and statistics. Who am I to say they aren't experts? This kind of rigid division of field and expertise made no sense in university, and it only breaks apart more as we get older and more experienced.
Heh, well, by coincidence, before I became a programmer I spent (too many) years as a PhD candidate in philosophy, and I'm skeptical about the extent to which undergraduates can be considered experts. ;-)
That was a long time ago, though, and despite my experience, I've undoubtedly fallen far behind the current work in the field.
I should hope so, after paying so much tuition!
> knowing more than high-school curriculum - which is the bar we're discussing
No, that's not the bar you were discussing. Originally you said "near-experts", but now you've lowered the bar way down.
Given his track record and the state of AI today, I know which one I find more likely.
The message was more abstract, why focus on one (poor) example?
Take the claim of superhuman for example. If an AI is as good as Bill Gates at coding, but is relentless and tireless, then that is fairly called superhuman, the same way a drilling machine was superhuman relative to John Henry.
Biases and heuristics have become somewhat mainstream “intellectualism”, but just like people using the words prolific and myriad incorrectly, is leveraged here to a false conclusion.
The anti-AI narrative continues, to the detriment of those who listen to its self-destructive guidance that equates excessive enthusiasm with delusion.
The technology is useful (while imperfect)
But I was there. He worked really, really fucking hard. As did most of the early crowd. If you're asking yourself how people back then worked so hard, just realize that there are drugs that can help you do that.
After having read your blog i can assert you've demonstrated that effect on yourself. Bill has all capability and credentials proven so we know this guy is an authority on software engineering topics, you have zero if i'm not mistaken.
Gates doesn't have the biology credentials. He's a pedophile, to top it off, so most probably that's where the article should have started with. It's telling that for most of the last 30 or so years the most wealthy person on this planet has been a pedophile, come to think of it maybe we (the human species) do deserve what's coming.
It's absolutely infuriating because it allows people with a lot of money tell you they know what they're talking about without proving it. Your money is your perceived effort, no matter how you acquired it.
Notably hanging out with Epstein
I would say he has more authority on software engineering topics than most people on earth, but that’s just me.
You are just focusing on the tech bit. Bill gates never was a very good person to begin with and then he did some stuff that pretty much offset whatever good he did.
As for software engineering: if you think Bill Gates is an authority in the field I would like to introduce you to Donald Knuth, Peter Norvig, Fabrice Bellard, Dennis Ritchie, Linus Torvalds and many others. Bill Gates wouldn't make the top 100.
> I would bet they don't.
I mean, how many "But but but antitrust!" and "Nobody got fired for buying Microsoft!" excuses can be made for Microsoft being where it is today?
I am focusing on the tech bit because the subthread was about his technical skills. As an engineer and a tech executive, everything I've read about him places him far above most people you could name. I also think he ranks very highly as a general scientific and systems thinker. As an example, he was harping on about pandemics long before COVID happened. He's been harping about AI and jobs these days, which I think many here would also agree with.
If we were discussing him as a person I have a more mixed opinion of him, but it's way more positive than the hate he gets around here, which I find hilariously misplaced.
He missed the coming interconnectedness of the machines and instead presumed that every computer was going to be a fancy copy machine calculator (that will never talk to others, thus security is not needed).
A massive hole in perception.
billg wrote the “Internet Tidal Wave” memo in May of 1995! This was an era when the dominant enterprise IT infrastructure was heavyweight (Windows!) client/server and homes were just starting to connect to the internet via dialup en masse.
Microsoft proceeded to pivot so successfully that they monopolized the web browser market alongside winning a massively profitable position in enterprise “intranet” apps and TCP/IP-based servers.
Microsoft missed mobile, not the Internet.
The history of AOL predates this and Netscape IPO'd a few months after Gates' writing. They filled gaps that MS was obviously missing.
He finds it so astounding to be himself, to feel like the Main Character of the World that he considers this evidence that the world he experiences is a simulation. The more humbling proposition -- that the world is largely random and he just happens to be the guy who kept rolling sixes -- is less compelling.
In any system of competing agents (such as human beings) where the reward for performance can enhance performance, will always result in even minor random differentials in initial performance exponentially accumulating until one or a handful of individuals dominate. In such a system, only a small portion of success can be attributed to innate skills. So it's true that success in itself, while signaling fitness to survive, does not necessarily signal an outsized ability to make accurate judgements.
I'm not entirely clear what the second part of the article is trying to convey. Is it that highly successful individuals attribute their success entirely to their skill and therefore when a machine is able to replicate their skillset, they are surprised and consider the technology to be world-altering?
Not doubting the intelligence of these individuals, just saying they sure tend to undervalue domain specific expertise. I don't know who Elon consulted about his cybertruck, but it wasn't truck people.
I do worry that AI exasperates this problem as a quick few queries and thoughtful questions have certainly given me over-confidence in a new subject. You don't know what you don't know.
I like that the article just tries to pass off a computer program coding at the level of humans (not to mention solving Millennium Problems) as something that should be just be totally unsurprising. Bill Gates is just a dumb arrogant tech bro to be surprised that a neural network exceeds his abilities at coding.
It was absolutely not a given that next-token prediction would scale to the point that it could answer outside of its training domain (this is the point about “made up questions”, it was important to test it on material outside of the training set since the scale of training was yet-unheard-of at the time).
Further, the success of post-training on verifiable tasks (coding) was also a complete shock that changed the world overnight. LLMs jumped the “I can make code” to “I confidently can make code that solves your problem” chasm suddenly, and it has changed the industry.
For someone so invested in this from the early days (remember the scale of Microsoft’s commitment) it is entirely logical that testing things “on day one” would occur and this would be a much bigger shock compared to most people that learned of advances at the same time as the rest of the world. (There are huge delays between research capabilities and shipping)
Designing new algorithms is where the real skill is, but an average programmer will never do this, and even a 10x programmer will do it maybe a few times in their lifetime at most.