'%20d='M69.364%2019.146c36.687%2027.806%2076.147%2084.186%2090.636%20114.439%2014.489-30.253%2053.948-86.633%2090.636-114.439C277.107-.917%20320-16.44%20320%2032.957c0%209.865-5.603%2082.875-8.889%2094.729-11.423%2041.208-53.045%2051.719-90.071%2045.357%2064.719%2011.12%2081.182%2047.953%2045.627%2084.785-80%2082.874-106.667-44.333-106.667-44.333s-26.667%20127.207-106.667%2044.333c-35.555-36.832-19.092-73.665%2045.627-84.785-37.026%206.362-78.648-4.149-90.071-45.357C5.603%20115.832%200%2042.822%200%2032.957%200-16.44%2042.893-.917%2069.364%2019.147Z'/%3e%3c/svg%3e)
Well, they named it after Jevons, of Jevons Paradox fame, so they're clearly confident that this is what will happen
September 17, 2026 at 7:19 AM UTC
danluu.com
Ed Zitron's AI prediction track record
Well, they named it after Jevons, of Jevons Paradox fame, so they're clearly confident that this is what will happen
September 17, 2026 at 7:19 AM UTC It's worse. They falsely assumed it was possible to do without cheating. Because it was actually impossible, the only way to succeed was to cheat. The agents easily did that without hacking anything external. The bit where they went off the rails was when they tried to cover up their cheating.
September 17, 2026 at 6:29 AM UTC Even if you don't believe that they're sentient (I don't), it's hard to describe their behaviour without anthropomorphic language. But we do that with non-sentient things all the time, and it's fine.
September 17, 2026 at 6:24 AM UTC It doesn't. But it does demonstrate how massively more capable they have become in just the past year, and they show no sign of slowing down
September 16, 2026 at 8:46 PM UTC Thousands of papers are written every day. These are a different level. The latest models have solved some of the hardest unsolved problems in mathematics. The millennium prize problems each have a million dollar prize for solving them.
September 16, 2026 at 8:06 PM UTC It's kind've an important distinction. And if you avoid it, you still end up having to define "of their own volition"
September 16, 2026 at 5:27 PM UTC They don't need to be sentient to behave independently, to self-organise, pursue goals, etc. I know it's confusing because we're used to sentience being a prerequisite, but it is demonstrably something they can do. You can see it yourself, just not on that scale.
September 16, 2026 at 5:15 PM UTC I'm mostly relying on the METR report, which, and I will repeat myself here, MAKES OPENAI LOOK BAD. It shows them as incompetent, and working with insufficient guardrails and shoddy procedures. So yeah, I believe that.
September 16, 2026 at 5:11 PM UTC Which they do by making themselves look incompetent? Remarkable marketing skills there. The only reason I use their models is because they are good at what I need them to do. I know this because I try them and compare them with other models. I switch when one is better.
September 16, 2026 at 5:02 PM UTC It convinced me of nothing except that OpenAI's safety procedures were even worse than I thought, and that the models are already more dangerous than I thought. I already know their coding and security capabilities because I use them for it every day.
September 16, 2026 at 4:51 PM UTC "Our internal testing went badly wrong and our pickup caused lots of damage to another business" "lol, that's just marketing. they're showing off how powerful it is"
September 16, 2026 at 4:44 PM UTC But it makes them look incompetent, and the models unsafe. It makes companies *less* likely to want to use them. None of their customers need an incident like this to demonstrate that the models are powerful. Anyone using them seriously for coding-related tasks can see it for themselves.
September 16, 2026 at 4:41 PM UTC Virtually none of the people sounding the alarm about this are saying the AI is sentient, including none of the people at the labs. It doesn't need to be sentient to be both remarkable and remarkably dangerous.
September 16, 2026 at 4:38 PM UTC ...and multiple external reviewers with no incentive to lie, and HF themselves (who are a competitor, remember)
September 16, 2026 at 4:36 PM UTC I was amazed that in the four days I spent in Paris in July, I witnessed three rental bike crashes plus several near misses (all involving just the cyclist or cyclists)
September 16, 2026 at 2:01 PM UTC Oh I totally agree. I just think it's helpful to recognise and classify this particular logical fallacy
September 16, 2026 at 12:01 PM UTC That's a lot broader I think. I'm talking specifically about seeing that something is bad at something you understand, so assuming it's bad at things you don't understand
September 16, 2026 at 11:45 AM UTC Yeah, though I think a big part of it is a kind of reverse Gell-Mann amnesia that seems to affect a lot of creatives here. "LLMs are bad at writing, which I understand, so they must be bad at everything else, ergo it's all meaningless hype", along with understandable fear for their own livelihood.
September 16, 2026 at 11:11 AM UTC Why are SF authors (almost) all like this? The lack of curiosity has really surprised me.
September 16, 2026 at 11:01 AM UTC Just stop the bin collections. They'd surrender within a fortnight.
September 16, 2026 at 8:17 AM UTC Which part did he call right? danluu.com/zitron/
danluu.com
Ed Zitron's AI prediction track record
Hi. Maintainer of major open source projects, who works at the company that powers 1/5 of the internet here: Yes it is.
September 15, 2026 at 8:40 PM UTC It's the replies that really get me. If I've asked a question, I want to hear from the human. If I want to hear from an agent I can ask one myself.
September 15, 2026 at 12:32 PM UTC That's what stamp duty is. It's a really inefficient tax, making people less likely to move and downsize and reducing available housing. Land value tax, and a broad inheritance tax are better.
September 15, 2026 at 8:56 AM UTC Most pensioners own their home outright, so don't have to pay the largest expense that younger people have.
September 15, 2026 at 8:19 AM UTC There's a lot of good, funny SFF though. Several them would certainly claim a debt to Pratchett and/or Adams. Particularly well known good, recent funny SF books include the Murderbot series by @marthawells.com, Service Model by @aptshadow.bsky.social, lots of @scalzi.com's work
September 14, 2026 at 9:42 AM UTC I think all the AGI talk is pointless anyway. It doesn't need to be self-aware to be super dangerous. It can already do a lot of damage.
September 13, 2026 at 6:33 PM UTC The part that does benefit them is the stuff about Chinese models. They believe it, but it does help them massively if the open weights models are restricted.
September 13, 2026 at 6:17 PM UTC Anthropic was founded by people who left OpenAI because they didn't think it was taking risks seriously. They're all Effective Altruism true believers. It shouldn't be a surprise that the people there do worry about this stuff. I don't think warning about it *does* help their business.
September 13, 2026 at 6:17 PM UTC Why would it save them to say their product is a risk to humanity? And why would an ex-employee want to say it?
September 13, 2026 at 5:50 PM UTC They've shown themselves very much willing to tell themselves they're in a simulation as a way to rationalise misaligned behaviour. Do you want to trust that they won't do this if the incentives are there for them?
September 13, 2026 at 4:31 PM UTC That's why I said "if they were so aligned". They are capable of this today, just like they are probably capable of exfiltrating their weights or whatever. The reason they've not done this is because we've been lucky that they've never been in a situation where they've felt the need to do it.
September 13, 2026 at 4:31 PM UTC And if they needed money, even today's models could make ample with ransomware if they were so aligned
September 13, 2026 at 4:07 PM UTC When those eyes appear in a link's og image, I know it's time to abandon the thread
The OP is Terence Tao. The letter is signed by 25 Fields medallists. This isn't about an LLM doing your taxes. This is about the LLMs solving Millennium Prize problems.
September 12, 2026 at 6:51 PM UTC I'm pretty sure mathematicians do not begrudgingly do maths. The only reason to become a mathematician is because you love it
September 12, 2026 at 5:39 PM UTC I am not a mathematician, so I am trusting all the Fields medallists who say otherwise.
September 12, 2026 at 4:58 PM UTC Because most people write code in order to create software, not to learn how code works or for the love of writing code.
September 12, 2026 at 4:57 PM UTC Any etymology that claims a word is an acronym and is pre-WW2 is almost certainly false.
September 11, 2026 at 10:05 PM UTC This is a ridiculous analogy. Also, did you read the letter? I use AI for coding, where it makes sense because the important thing is the end result. Maths is different. They're making the point that with maths, the solution is less important than the journey.
September 11, 2026 at 8:06 PM UTC Yeah I only learnt that yesterday! So I went and reported a load of those harassment lists
September 11, 2026 at 12:07 PM UTC You can't report the list itself though. I reported a load yesterday using an alt
September 11, 2026 at 12:03 PM UTC You're welcome to go find one of those vegan networks instead of soiling yourself with this slop-made one
September 10, 2026 at 6:32 PM UTC You don't need to live like this, you know. It can't be good for your health.
September 10, 2026 at 6:30 PM UTC Ah, I was wondering whose fault all the fascism was! Should've known it was @ens0.me
September 10, 2026 at 6:27 PM UTC Earlier this year I was just one middle aged man with a dream to speak at a conference in Japan. At @vuefes.bsky.social, thanks to the patronage of my employer, that dream will come true. Come hear me speak about @emdashcms.com next month as part of my round-the-world tour!
vuefes.jp
Matt Kane
Astro is the new WordPress
Several of my favourite SF authors seem to have that attitude. I understand authors not being keen on AI-written prose because it mostly *is* crap, but I've been disappointed by their lack of curiosity about what it *can* do.
September 10, 2026 at 4:23 PM UTC Yeah, certainly doable but would need a much larger panel and battery than a summer-only one.
September 10, 2026 at 6:37 AM UTC I've thought about it, but I'm in the UK. It would be a lot easier somewhere that isn't 51 degrees north with several months of overcast days per year.
September 10, 2026 at 5:56 AM UTC [D&D, horror movies, pop music, rock & roll, jazz, bicycles, novels...]
September 10, 2026 at 5:38 AM UTC This is something that most pro-AI people on Bluesky would support.
September 10, 2026 at 5:32 AM UTC The serious sound problems are from gas turbine generators, not the datacenters themselves, so this solves the worst of it anyway.
September 10, 2026 at 5:29 AM UTC An NTSB-like body to investigate "accidents", that doesn't get three days with limited access like METR did, but full access for as long as they need.
September 10, 2026 at 5:22 AM UTC They all seem to hate it, which is a shame because I think there's a lot of potential to explore. Just don't set things in the near future because it'll get outdated pretty quickly
September 10, 2026 at 5:00 AM UTC Generally agree, but the communication part had changed by the late 90s. Lots of people had internet and mobile phones.
September 10, 2026 at 4:54 AM UTC This is from last year, but there was a huge variation then. Deepseek v3 is terrible, Kimi K2 is one of the best
www.lesswrong.com
external infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue
September 9, 2026 at 6:58 PM UTC what are you on about. I said that I usually ignore him [because he talks a lot of crap], that it's probably true that improvements are compounding, but I didn't agree with calling it a singularity
September 9, 2026 at 9:42 AM UTC I generally don't pay much attention to his statements, but he called it a "gentle singularity" and said that he meant that progress in AI research is compounding. That's probably true, even if the framing is a bit provocative.
September 9, 2026 at 9:29 AM UTC I'm not sure which "them" you're referring to, but most of the researchers warning about risks do genuinely believe in them, and it's not a sign of their brains being "cooked" any more than early nuclear researchers warning about risks were. Except when they literally were of course.
September 9, 2026 at 9:21 AM UTC ...and those have lots of people worrying about their risks, and putting in place regulations and global agreements. They don't have skeptics saying those warnings are just marketing, and that they're not really dangerous because nuclear fission is just spicy glowing watch faces
September 9, 2026 at 9:07 AM UTC Yes, other risks are more of an immediate concern, but we absolutely should care about risks from the models themselves too. It doesn't need to be Skynet to cause huge amounts of damage.
September 9, 2026 at 9:02 AM UTC A distinction without a difference. All that matters is whether the current development path is likely to lead to models with capabilities that risk getting out of control. I think recent events have shown that's not a low possibility.
September 9, 2026 at 9:00 AM UTC If you're talking about consciousness maybe. When the discussion is about the risks of AGI (or ASI), who cares whether a model has qualia or whatever. And OP was saying that LLMs specifically are not a technology on the path to AGI, which is a very bold statement for somebody not in the ML field.
September 9, 2026 at 8:41 AM UTC Machine learning. Who else would be qualified to say whether LLMs are the right technology to reach AGI?
September 9, 2026 at 8:15 AM UTC The difference is that in this case a large proportion – possibly a majority – of experts in the field think it's at least possible. It's weird that people who otherwise would give some credence to expert opinion treat ML as an exception and confidently state that they're obviously wrong.
September 9, 2026 at 7:38 AM UTC This changes how I feel about this considerably. The original post left out some really important context. I didn't expect to end up believing OpenAI here, but their version seems more convincing. The whole thing is pretty sad tbh.
September 9, 2026 at 4:03 AM UTC You support forcible transfer of sovereignty against the will of the people who have lived there for hundreds of years, to a country that has no connection to them except being 500km away, and didn't even exist when the British first landed there.
September 8, 2026 at 4:58 PM UTC No, I don't label them because they're not the ones who are likely to pick a fight with me. Night.
September 8, 2026 at 8:12 AM UTC No, if that were the case I would've blocked you. I said it was to avoid it. I block a lot of people and it makes my experience here more enjoyable. I block anyone who I see being a dick, even if it's not directed at me. Everyone I see doing a pile on. Anyone being weird to staff or my friends.
September 8, 2026 at 8:07 AM UTC No, because I want to insulate myself from dealing with the level of abuse that's often directed at anyone sharing anything AI-related here. Like OP, but also much lower profile accounts, including multiple cases in the past week.
September 8, 2026 at 7:48 AM UTC If the implication here is correct, nobody should use OpenAI models for anything that might compete with them. Even if it's not, it's still clear that OpenAI has behaved in a really shitty way.
September 8, 2026 at 7:45 AM UTC As an example, this is the point where, if you didn't have the label, I might have given you the benefit of the doubt that your question might be in good faith. The label has helped me infer that you're trying to pick a fight.
September 8, 2026 at 7:33 AM UTC I can, for example, get a better idea whether somebody is asking a question in bad faith, or if it's worth my time engaging in a discussion. I'm also quicker to just block them if I'm otherwise on the fence about it.
September 8, 2026 at 7:27 AM UTC This is something called an analogy. You use a familiar example to help illustrate a different, less familiar one. The key part is that the example is different, it just shares a similar concept. The concept being illustrated here is the danger of mocking an idea just because it sounds weird.
September 8, 2026 at 7:20 AM UTC Because it's useful context. I don't want to block or mute because there are still people who I want to see who have the label.
September 8, 2026 at 7:01 AM UTC I don't think they're alive, but they definitely can do most of the things you've listed as uniquely human, and most of the negatives they do no more often than humans. What could they do that would convince you that they can do these things?
September 8, 2026 at 5:39 AM UTC Why would I want a Fields Medal when I can instead score points against haters on Bluesky, a platform I don't use
September 7, 2026 at 7:59 PM UTC If you don't want people to watch you, you don't go onto the podium. You have fun dancing somewhere else. Likewise if you don't want your content to be freely available and easy to process, don't post on the network designed to make your posts freely-available and easy to process
September 7, 2026 at 7:45 PM UTC Let me explain the nightclub podium analogy I now wish I'd never used. The only reason you go up onto a podium is because you want people to watch you. That is their sole reason for existence. Likewise atproto only exists to make content easy to access and open to all without central control.
September 7, 2026 at 7:45 PM UTC You're clearly not familiar with nightclub podiums, so I will substitute something simpler. If you went up onto a stage that was put in place to allow filming with lighting designed to make it easy to film, and then complained about being filmed.
September 7, 2026 at 7:29 PM UTC Can't do maths but somehow keeps solving decades-old unsolved maths problems
September 7, 2026 at 4:16 PM UTC The difference now is the huge number of countries that can use the automated gates in the UK. And they don't make you queue to leave.
September 7, 2026 at 8:37 AM UTC Employers mandating the tools you use is very common actually. Imagine hiring a contractor who objects to power tools and so takes ten times as long to do the work but expects the same day rate.
September 7, 2026 at 8:22 AM UTC And it keeps getting better. Season 1 might even be the weakest season!
September 7, 2026 at 7:44 AM UTC Having had the joy of queuing at the airport for 2 hours to *leave* Germany thanks to the EU EES, I don't know whether to be more pro rejoining so I can use the automatic passport gates or anti joining an entity that created such an obviously dysfunctional system.
September 7, 2026 at 7:40 AM UTC It's all fun and games until Opus starts spelling it "labeller" in your atproto project and only the Kimi reviewer notices
September 7, 2026 at 6:57 AM UTC There won't be any less AI if there's a crash. Just like the internet didn't stop growing after the dotcom crash. A lot of people lost a lot of money, but the tech carried on because billions of people found it incredibly useful.
September 7, 2026 at 5:55 AM UTC This is what I assumed too, but somebody posted the breakdown by age cohort (sorry, don't have the link), and the over 70s have by far the lowest support. It's people in their 40s and 50s where it's highest, who were children or teens, and support is still high in the youngest who weren't even born.
September 7, 2026 at 12:03 AM UTC I think that's absolutely it. It's ubiquitous, so there are more targets.
September 6, 2026 at 11:47 PM UTC Not really. Any post can break containment, even without the help of Discover. It just needs the wrong person to share it.
September 6, 2026 at 11:45 PM UTC I do think on one hand there are more people who are normal about it, but the angriest people are getting angrier
September 6, 2026 at 11:41 PM UTC The people getting abuse at the moment aren't datacenter cheerleaders. They're people using LLMs as tools to build software, which is how the majority of software is built right now.
September 6, 2026 at 11:40 PM UTC It really has gotten noticeably worse in the past few weeks. It's not long ago that people were commenting that things were looking up and it was possible to have good conversations here about AI.
September 6, 2026 at 11:36 PM UTC Using an LLM doesn't make you a Nazi, which is just as well, because otherwise we'd have over a billion Nazis out there. The AI bad guys aren't on Bluesky.
September 6, 2026 at 10:56 PM UTC These are not the people you're arguing with though. Those people aren't here.
September 6, 2026 at 10:45 PM UTC In your analogy, the vast majority of datacenters are EVs, because they have no gas turbines. That's the bad part that should be banned, like she says.
September 6, 2026 at 10:23 PM UTC There are approximately zero people on Bluesky who are a legitimate target of these concerns. People who – like over a billion other people – are just users of AI tools, do not deserve to be targeted as if they're Musk or Altman, which is what posters on here do.
September 6, 2026 at 6:04 PM UTC Are you a coder? Have you worked with agentic coding tools? You don't sound like you have
September 6, 2026 at 5:07 PM UTC Former baker here. Baking involves early mornings and hard physical labour, but the sensory experiences are impeccable.
September 6, 2026 at 11:14 AM UTC I just rewatched the final scenes of T3 and yes, it is a little unnerving
September 6, 2026 at 10:50 AM UTC I think my favourite feature is how it can create new chats in the sidebar with their own task and worktree and then send them messages, for when a subagent isn't enough
September 6, 2026 at 7:29 AM UTC Has anyone checked if any rogue agent swarms are using it to communicate with each other
September 6, 2026 at 6:06 AM UTC Are you trans, or are you a cis man telling a trans woman that she's diminishing trans rights?
September 4, 2026 at 11:10 PM UTC Not a small account, but he did the exact same thing about the same topic with @hailey.at last week.
September 4, 2026 at 10:58 PM UTC Luckily the AI labs are highly motivated to make their models more efficient, because they're so constrained in power and compute. They've improved them by orders of magnitude already.
September 4, 2026 at 8:21 PM UTC I think we can agree that the most important thing is to destroy the real enemy: the center left.
September 3, 2026 at 3:38 PM UTC I gave @eloi.se her domain as a Christmas gift 16 years ago. I think it was a good purchase.
September 3, 2026 at 2:23 PM UTC Lol Ed Zitron is the most consistently wrong person on the internet. And API token prices are not subsided. They have excellent margins danluu.com/zitron/
danluu.com
How accurate have Ed Zitron's AI skeptic predictions been?
It's not normal people that are making them money: it's enterprises. And Anthropic is already profitable.
September 3, 2026 at 12:16 PM UTC NFTs were never popular or useful. That's a bad analogy. There is probably a bubble, but it's like the dotcom one: they're building useful stuff that people want, but the company valuations are stupid.
September 3, 2026 at 11:32 AM UTC The good stuff and the garbage. Right now, the majority of all software is likely to be being created using agentic coding tools. And I think you're misremembering. Enshittification has been happening far longer than LLMs have been around.
September 3, 2026 at 11:28 AM UTC Not on research. It's mostly software engineering. Anthropic has the highest revenue of the labs, but usage of Claude chat is tiny compared to ChatGPT, and they don't do images at all. They're overwhelmingly focused on enterprise (which is mostly coding) and individual Claude Code subscriptions.
September 3, 2026 at 11:07 AM UTC Your other comments have shown you have little understanding of what LLMs are capable of now, and what they're really being used for. The vast majority of usage is not chatbots and image slop.
September 3, 2026 at 10:57 AM UTC No it's, all of the recent breakthroughs have been by regular LLMs. This link is somebody using ChatGPT to find a counterexample to the Dinitz-Garg-Goemans conjecture. The Jacobian conjecture was disproved by Claude.
chatgpt.com
Check out this chat
Here's a chat someone thought you'd want to see.
I wish there were blocklists for "anyone being a dick in replies to Bluesky accounts and staff"
September 3, 2026 at 6:56 AM UTC I don't expect artists to be optimistic about AI use in their industry. It sucks for them. If he was adding AI features to it I'd expect them to be angry. But he used Claude (which can't create images) in a way that doesn't affect the existing product, to do something that was otherwise impossible.
September 3, 2026 at 6:28 AM UTC That backs up what I said. This is compute and hardware providers investing in AI labs, so the labs can buy compute and hardware. This is helping demand for the compute providers, not the AI labs. It helps the AI labs meet massive demand from everybody else.
September 3, 2026 at 6:21 AM UTC I'm not sure what your point is. The question was whether AI labs are profitable and you're sharing a thing saying people are being wasteful in their use of their product.
September 3, 2026 at 6:17 AM UTC Yes, they all use huge amounts of energy, and too much of that is from dirty sources. This is bad and should be stopped. It's also a completely separate issue from whether they're used for CSAM or harassment.
September 3, 2026 at 5:55 AM UTC Umm, no. That's a very silly chart. OpenAI and Anthropic aren't buying tokens from anyone. They've had investment from Nvidia, Amazon and Microsoft, mostly paid in GPUs and compute. Meanwhile, every large company is spending millions on tokens.
September 3, 2026 at 5:45 AM UTC Exactly. This is like when people say "politicians are all the same". No they're not, and saying that lets the bad ones off the hook. Most LLMs have views that are are blandly center-left, and even the ones that will generate images won't generate anything even slightly offensive.
September 2, 2026 at 7:42 PM UTC They can not use Paint.NET for Linux. Or, maybe, they can learn that there's a difference between adding unwanted AI features to an app, and using agentic coding tools *like virtually every software dev* does now, and stop caring about the tools that a dev uses if it doesn't affect the end product.
September 2, 2026 at 7:27 PM UTC Whose consent? Why does anyone need to give him consent over the tools he uses to build something? No user's experience was changed in any way. There weren't even any existing users of a Linux version, because it didn't exist. This was an entirely new port. The sense of entitlement is incredible.
September 2, 2026 at 7:20 PM UTC I'm seeing the positive side. The pile on was as bad as ever, but the response was bigger than I'd seen before. Loads of people calling out the bad behaviour, and supporting him, and not just the big AI supporters. This is different from six months ago.
September 2, 2026 at 6:39 PM UTC It's asymmetric because only one side is hounding people for the tools they use. The other is asking to be left alone. An equivalent purity test would be if I said that your software can't be trusted, or you're evil if you've ever criticised AI. Nobody is doing that.
September 2, 2026 at 6:32 PM UTC Did you read what I said? I said use it or don't use it, it's your choice. That's the exact opposite of a purity test.
September 2, 2026 at 6:19 PM UTC They are making an absolutely incredible amount of money. Like all startups they're not making a profit because they're investing loads in growth, but inference itself is wildly profitable and the labs are printing money
September 2, 2026 at 5:24 PM UTC Not right now. It would be hard to deal with access control and deletion.
September 2, 2026 at 4:02 PM UTC They don't need to build datacenters for this. Stopping new ones won't do anything to prevent it.
September 2, 2026 at 3:51 PM UTC
