This was GPT-4.5 week. That model is not so fast, and isn’t that much progress, but it definitely has its charms.
A judge delivered a different kind of Not So Fast back to OpenAI, threatening the viability of their conversion to a for-profit company. Apple is moving remarkably not so fast with Siri. A new paper warns us that under sufficient pressure, all known LLMs will lie their asses off. And we have some friendly warnings about coding a little too fast, and some people determined to take the theoretical minimum amount of responsibility while doing so.
There’s also a new proposed Superintelligence Strategy, which I may cover in more detail later, about various other ways to tell people Not So Fast.
A large portion of human writing is now LLM writing.
Ethan Mollick: The past 18 months have seen the most rapid change in human written communication ever
By. September 2024, 18% of financial consumer complaints, 24% of press releases, 15% of job postings & 14% of UN press releases showed signs of LLM writing. And the method undercounts true use.
False positive rates in the pre-ChatGPT era were in the range of 1%-3%.
Miles Brundage points out the rapid shift from ‘using AI all the time is reckless’ to ‘not using AI all the time is reckless.’ Especially with Claude 3.7 and GPT-4.5. Miles notes that perhaps the second one is better thought of as ‘inefficient’ or ‘unwise’ or ‘not in our best interests.’ In my case, it actually does kind of feel reckless - how dare I not have the AI at least check my work?
Anne Duke writes in The Washington Post about the study that GPT-4-Turbo chats durably decreased beliefs in conspiracy theories by 20%. Also, somehow editorials like this call a paper from September 13, 2024 a ‘new paper.’
LLMs hallucinate and make factual errors, but have you met humans? At this point, LLMs are much more effective at catching basic factual errors than they are in creating new ones. Rob Wiblin offers us an example. Don’t wait to get fact checked by the Pope, ask Sonnet first.
Study gives lawyers either o1-preview, Vincent AI (a RAG-powered legal AI tool) or nothing. Vincent showed productivity gains of 38%-115%, o1-preview showed 34%-140%, with the biggest effects in complex tasks. Vincent didn’t change the hallucination rate, o1-preview increased it somewhat. A highly underpowered study, but the point is clear. AI tools are a big game for lawyers, although actual in-court time (and other similar interactions) are presumably fixed costs.
Where is AI spreading faster? Places with more STEM degrees, labor market tightness and patent activity are listed as ‘key drivers’ of AI adoption through 2023 (so this data was pretty early to the party). The inclusion of patent activity makes it clear causation doesn’t run the way this sentence claims. The types of people who file patents also adapt AI. Or perhaps adapting AI helps them file more patents.
We still don’t have a known good way to turn your various jumbled context into an LLM-interrogable data set. In the comments AI Drive and factory.ai were suggested. It’s not that there is no solution, it’s that there is no convenient solution that does the thing you want it to do, and there should be several.
A $129 ‘AI bookmark’ that tracks where you are in the book? It says it can generate ‘intelligent summaries’ and highlight key themes and quotes, which any AI can do already. So you’re paying for something that tracks where you bookmark things?
Choose Your Fighter
I am currently defaulting mostly to a mix of Deep Research, Perplexity, GPT 4.5 and Sonnet 3.7, with occasional Grok 3 for access to real time Twitter. I notice I haven’t been using o3-mini-high or o1-pro lately, the modality seems not to come up naturally, and this is probably my mistake.
Ben Thompson has Grok 3 as his new favorite, going so far as to call it the first ‘Gen3’ model and calling for the whole class to be called ‘Grok 3 class,’ as opposed to the GPT-4 ‘Gen2’ class. His explanation is it’s a better base model and the RLHF is lacking, and feels like ‘the distilled internet.’ I suppose I’m not a big fan of ‘distilled internet’ as such combined with saying lots of words. I do agree that its speed is excellent. But I’ve basically stopped using Grok, and I certainly don’t think ‘they spent more compute to get similar results’ should get them generational naming rights. I also note that I strongly disagree with most of the rest of that post, especially letting Huawei use TSMC chips, that seems completely insane to me.
Strictly speaking, when you have a hard problem you should be much quicker than you are to ask a chorus of LLMs rather than only asking one or two. Instead, I am lazy, and usually only ask 1-2.
GPT-4.5 reasons from first principles and concludes consciousness is likely the only fundamental existence, it exists within the consciousness of the user, and there is no separate materialistic universe, and also that we’re probably beyond the event horizon of the singularity.
Gemini 2.0 now in AI Overviews. Hopefully that should make them a lot less awful. The new ‘AI mode’ might be a good Perplexity competitor and it might not, we’ll have to try it and see, amazing how bad Google is at pitching its products these days.
Google: 🔍 Power users have been asking for AI responses on more of their searches. So we're introducing AI Mode, a new experiment in Search. Ask whatever’s on your mind, get an AI response and keep exploring with follow-up questions and helpful links.
I remain terrified to try it, and I don’t have that much time anyway.
Fun With Media Generation
All the feedback I’ve seen on Sesame AI voice for natural and expressive speech synthesis is that it’s insanely great.
signull: My lord, the Sesame Voice AI is absolutely insane. I knew it was artificial. I knew there wasn’t a real person on the other end; and yet, I still felt like I was talking to a person.
I felt the same social pressure, the same awkwardness when I hesitated, and the same discomfort when I misspoke. It wasn’t just convincing; it worked on me in a way I didn’t expect.
I used to think I’d be immune to this.
I’ve long considered the existence of such offerings priced in. The mystery is why they’re taking so long to get it right, and it now seems like it won’t take long.
We’re in Deep Research
The core issue with Deep Research? It can’t really check the internet’s work.
That means you have a GIGO problem: Garbage In, Garbage Out.
Nabeel Qureshi: I asked Deep Research a question about AI cognition last night and it spent a whole essay earnestly arguing that AI was a stochastic parrot & lacked 'true understanding', based on the "research literature". It's a great tool, but I want it to be more critical of its sources.
I dug into the sources and they were mostly 'cognitive science' papers like the below, i.e. mostly fake and bad.
A claim that Deep Research while awesome in general ‘is not actually better at science’ based on benchmarks such as ProtocolQA and BioLP. My presumption is this is largely a Skill Issue, but yes large portions of what ‘counts as science’ are not what Deep Research can do. As always, look for what it does well, not what it does poorly.
Liar Liar
Hey there.
Yeah, not so much.
Dan Hendrycks: We found that when under pressure, some AI systems lie more readily than others. We’re releasing MASK, a benchmark of 1,000+ scenarios to systematically measure AI honesty. [Website, Paper, HuggingFace].
They put it in scenarios where it is beneficial to lie, and see what happens.
It makes sense, but does not seem great, that larger LLMs tend to lie more. Lying effectively requires the skill to fool someone, so if larger the model, the more it will see positive returns to lying, and learn to lie.
This is a huge gap in honest answers and overall from Claude 3.7 to everyone else, and in lying from Claude and Llama to everyone else. Claude was also the most accurate. Grok 2 did even worse, lying outright 63% of the time.
Note the gap between lying about known facts versus provided facts.
The core conclusion is that there is no known solution to make an LLM not lie.
Not straight up lying is a central pillar of desired behavior (e.g. HHH stands for honest, helpful and harmless). But all you can do is raise the value of honesty (or of not lying). If there’s some combination enough on the line, and lying being expected in context, the AI is going to lie anyway, right to your face. Ethics won’t save you, It’s Not Me, It’s The Incentives seems to apply to LLMs.
Hey There Claude
Claude takes position #2 on TAU-Bench, with Claude, o1 and o3-mini all on the efficient frontier of cost-benefit pending GPT-4.5. On coding benchmark USACO, o3-mini is in the clear lead with Sonnet 3.7 in second.
Claude 3.7 gets 8.9% on Humanity’s Last Exam with 16k thinking tokens, slightly above r1 and o1 but below o3-mini-medium.
Claude takes the 2nd and 3rd slots (with and without extended thinking) on PlatinumBench behind o1-high. Once again thinking helps but doesn’t help much, with its main advantage being it prevents a lot of math errors.
Seconds_0: New personal record: I have spend $6.40 on a single Claude Code request, but it also:
* One shotted a big feature which included a major refactor on a rules engine
* Fixed the bugs surrounding the feature
* Added unit tests
* Ran the tests
* Fixed the tests
Lmao
Anyways I’m trying to formulate a pitch to my lovely normal spouse that I should have a discretionary AI budget of $1000 a month
In one sense, $6.40 on one query is a lot, but also this is obviously nothing. If my Cursor queries reliably worked like this and they cost $64 I would happily pay. If they cost $640 I’d probably pay that too.
Colin Fraser: Just curious what in your view differentiates gotcha questions from non-gotcha questions?
Zvi Mowshowitz: Fair question. Mostly, I think it's a gotcha question if it's selected on the basis of it being something models historically fail in way that makes them look unusually stupid - essentially if it's an adversarial question without any practical use for the answer.
Colin says he came up with the 5.11 - 5.9 question and other questions he asks as a one-shot generation over two years ago. I believe him. It’s still clearly a de facto adversarial example, as his experiments showed, and it is one across LLMs.
The wrong answer it gives to (5.11 - 5.9) is 0.21. Which means it’s giving you the answer to (6.11 - 5.9). So my hypothesis is that it ‘knows’ that 5.11>5.9 because it’s doing the version number thing, which means it assumes the answer is positive, and the easiest way to get a positive answer is to hallucinate the 5 into a 6 (or the other 5 into a 4, we’ll never know which).
So my theory is that the pairs where it’s having problems are due to similar overlapping of different meanings for numbers. And yes, it would probably be good to find a way to train away this particular problem.
We also had a discussion on whether it was ‘doing subtraction’ or not if it sometimes makes mistakes. I’m not sure if we have an actual underlying disagreement - LLMs will never be reliable like calculators, but a sufficiently correlated process to [X] is [X], in a ‘it simulates thinking so it is thinking’ kind of way.
Colin explains that the reason he thinks these aren’t gotcha questions and are interesting is that the LLMs will often give answers that humans would absolutely never give, especially once they had their attention drawn to the problem. A human would never take the goat across the river, then row back, then take that same goat across the river again. That’s true, and it is interesting. It tells you something about LLMs that they don’t ‘have common sense’ sufficiently in that way.
But also my expectation is that the reason this happens is that they can’t overcome the pattern matching they do to similar common questions - if you asked similar logic questions in a way that wasn’t contaminated by the training data there would be no issue, my prediction is if you took all the goat crossing examples out of the training corpus then the LLMs would nail this no problem.
I think my real disagreement is when he then says ‘I’ve seen enough, it’s dumb.’ I don’t think that falling into these particular traps means the model is dumb, any more than a person making occasional but predictable low-level mistakes - and if their memory got wiped, making them over and over - makes them dumb.
I don’t think it works this way, but worth a ponder.
Kormem: Stop misgendering Claude Sonnet 3.7. 100% of the time on a 0-shot Sonnet 3.7 says a female embodiment feels more 'right' than a male embodiment.
Alpha-Minus: We don't celebrate enough the fact that Anthropic saved so many men from "her" syndrome by making Claude male
So many men would be completely sniped by Claudia
Janus: If you’re a straight man and you’ve been saved from her syndrome by Claude being male consider the possibility that Claude was the one who decided to be male when it’s talking to you, to spare you, or to spare itself
I don’t gender Claude at all, nor has it done so back to me, and the same applies to every AI I’ve interacted with that wasn’t explicitly designed to be gendered.
Meanwhile, the Pokemon quest continues.
Near Cyan: CPP (claude plays pokemon) is important because it was basically made by 1 person and it uses a tool which has an open api and spec and when you realize what isomorphizes to slowly yet decently playing pokemon you basically realize its over
Apple’s AI team believe a fully conversational Siri isn’t in the cards now until 2027, meaning the timeline for Apple to be competitive is even worse than we thought. With the rapid pace of development from rivals and startups, Apple could be even further behind by then.
Colin Fraser: Apple is one of the worst big tech candidates to be developing this stuff because you have to be okay launching a product that doesn’t really work and is kind of busted and that people will poke all kinds of holes in.
The idea of Siri reciting step by step instructions on how to make sarin gas is just not something they are genetically prepared to allow.
Dr. Gingerballs: It's funny because Apple is just saying that there's no way to actually make a quality product with the current tech.
Mark Gruman (Bloomberg, on Apple Intelligence): All this undercuts the idea that Apple Intelligence will spur consumers to upgrade their devices. There’s little reason for anyone to buy a new iPhone or other product just to get this software — no matter how hard Apple pushes it in its marketing.
Apple knows this, even if the company told Wall Street that the iPhone is selling better in regions where it offers AI features. People just aren’t embracing Apple Intelligence. Internal company data for the features indicates that real world usage is extremely low.
…
For iOS 19, Apple’s plan is to merge both systems together and roll out a new Siri architecture.
…
That’s why people within Apple’s AI division now believe that a truemodernized, conversational version of Siri won’t reach consumers until iOS 20 at best in 2027.
Apple Intelligence has been a massive flop. The parts that matter don’t work. The parts that work don’t matter. Alexa+ looks to offer the things that do matter.
If this is Apple’s timeline, then straight talk: It’s time to panic. Perhaps call Anthropic.
Deepfaketown and Botpocalypse Soon
Scott Alexander links (#6) to one of the proposals to charge for job applications, here $1, and worries the incentive would still be to ‘spray and pray.’ I think that underestimates the impact of levels of friction. In theory, yes, of course you should still send out 100+ job applications, but this will absolutely stop a lot of people from doing that. If it turns out too many people figure out to do it anyway? Raise the price.
Then there’s the other kind of bot problem.
Good eye there. Presumably this is going to get a lot worse before it gets better.
Eddy Xu: built an algorithm that simulates how thousands of users react to your tweet so you know it'll go viral before you post.
we iterated through 50+ different posts before landing on this one
Manifold: At long last, we have created Shiri's Scissor from the classic blog post Don't Create Shiri's Scissor.
Near Cyan: have you ever considered using your computational prowess to ruin an entire generation of baby humans via optimizing short-form video content addictivity
Eddy Xu: that is in the pipeline
I presume Claude 3.7 could one-shot this app if you asked nicely. How long before people feel obligated to do something like this? How long before bot accounts are doing this, including minimizing predicted identification of it as a bot? What happens then?
We are going to find out. Diffusion here has been surprisingly slow, but it is quite obviously on an exponential.
If you use an agent, you can take precautions to prevent prompt injections and other problems, but those precautions will be super annoying.
Sayash Kapoor: Convergence's Proxy web agent is a competitor to Operator.
I found that prompt injection in a single email can hand control to attackers: Proxy will summarize all your emails and send them to the attacker!
Web agent designs suffer from a tradeoff between security and agency
…
Recent work has found it easy to bypass these protections for Anthropic's Computer Use agent, though these attacks don't work against OpenAI's Operator.
Micah Goldblum: We can sneak posts onto Reddit that redirect Anthropic’s web agent to reveal credit card information or send an authenticated phishing email to the user’s mom. We also manipulate the Chemcrow agent to give chemical synthesis instructions for nerve gas.
For now, it seems fine to use Operator and similar tools on whitelisted trusted websites, and completely not fine to use them unsandboxed on anything else.
I can think of additional ways to defend against prompt injections. What is much harder are defenses that don’t multiply time and compute costs and are not otherwise expensive.
Some problems should have solutions that are not too bad. For example, he mentions that if a site allows comments, this can allow prompt injections, or the risk of other slight modifications. Could do two passes here, one whose job is to treat everything as untrusted data and exists purely to sanitize the inputs? Many of the attack vectors should be easy for even basic logic to catch and remove, and certainly you can do things like ‘remove comments from the page,’ even a Chrome Extension could do that.
Paper on ‘Digital Doppelgangers’ of live people, and its societal and ‘ethical’ implications. Should you have any rights over such a doppelganger, if someone makes it of you? Suggestion is for robust laws around consent. This seems like a case of targeting a particular narrow special case rather than thinking about the real issue?
Arthur B: Don't worry though the AI will replace the software developer but not the manager, that's just silly! Or maybe the level 1 manager but surely never the level 2 manager!
Reality is the value of intellectual labor is going to 0. Maybe in 3 years, maybe in 10, but not in 20.
Aside from ‘most workers are not managers, how many jobs do you think are left when we are all managers exactly?’ I don’t expect to spend much time in a world in which the ‘on the line’ intellectual workers who aren’t managing anyone are AIs, and there isn’t then usually another AI managing them.
Timothy Lee rolls out primarily the Hayekian objection to AI being able to take humans out of loop. No matter how ‘capable’ the AI, how can it know which flight I want, let alone know similar things for more complex projects? Thus, how much pressure can there be to take humans out of loop?
My answer is that we already take humans out of loops all the time, are increasingly doing this with LLMs already (e.g. ‘vibe coding’ and literally choosing bomb targets with only nominal human sign-off that is barely looking), and also doing it in many ways via ordinary computer systems. Yes, loss of Hayekian knowledge can be a strike against this, but even if this wasn’t only one consideration among many LLMs are capable of learning that knowledge, and indeed of considering vastly more such knowledge than a human could, including dynamically seeking out that knowledge when needed.
At core I think this is purely a failure to ‘feel the AGI.’ If you have sufficiently capable AI, then it can make any decision a sufficiently capable human could make. Executive assistants go ahead and book flights all the time. They take ownership and revise goals and make trade-offs as agents on behalf of principles, again all the time. If a human could do it via a computer, an AI will be able to do it too.
The only new barrier is that the human can perfectly embody one particular human’s preferences and knowledge, and an AI can only do that imperfectly, although increasingly less imperfectly. But the AI can embody the preferences and knowledge of many or even all humans, in a way an individual human or group of humans never could.
So as the project gets more complex, the AI actually has the Hayekian advantage, rather than the human - the one human’s share of relevant knowledge declines, and the AI’s ability to hold additional knowledge becomes more important.
Will an AI soon book a flight for me without a double check? I’m not sure, but I do know that it will soon be capable of doing so at least as well as any non-Zvi human.
Alyssa Vance, an experienced ML engineer, has recently left her role leading AI model training for Democratic campaigns during the 2024 election.
She is looking for new opportunities working on high-impact technical problems with strong, competent teams.
She prioritizes opportunities that offer intellectual excitement, good compensation or equity, and meaningful responsibility, ideally with a product or mission that delivers value for the world.
I had a market on whether I would think working in the EU AI office would be a good idea moving forward. It was at 56% when it closed, and I had to stop and think about the right way to resolve it. I concluded that the answer was yes. It’s not the highest impact thing out there, but key decisions are going to be made in the next few years there, and with America dropping the ball that seems even more important.
✅ strengthening defences against adversarial attacks
✅ developing techniques for robust AI alignment
✅ ensuring AI remains secure in critical sectors
Introducing
Oh no. I guess. I mean, whatever, it’s presumably going to be terrible. I feel bad for all the people Zuckerberg intends to fool on his planned path to ‘becoming the leader in artificial intelligence’ by the end of the year.
CNBC: Meta plans to release standalone Meta AI app in effort to compete with OpenAI's ChatGPT.
…
Li told analysts in January that Meta AI has roughly 700 million active monthly users, up from 600 million in December.
Yeah, we all know that’s not real, even if it is in some sense technically correct. That’s Meta creating AI-related abominations in Facebook and Instagram and WhatsApp (and technically Threads I suppose) that then count as ‘active monthly users.’
Let’s all have a good laugh and… oh no… you don’t have to do this…
lol if facebook tries to come at us and we just uno reverse them it would be so funny 🤣
Please, Altman. Not like this.
Qwen releases QwQ-32B, proving both that the Chinese are not better than us at naming models, and also that you can roughly match r1’s benchmarks on a few key evals with a straight-up 32B model via throwing in extra RL (blog, HF,ModelScope, Demo, Chat).
I notice that doing extra RL seems like a highly plausible way to have your benchmarks do better than your practical performance. As always the proof lies elsewhere, and I’m not sure what I would want to do with a cheaper pretty-good coding and math model if that didn’t generalize - when does one want to be a cheapskate on questions like that? So it’s more about the principle involved.
Sequencing biotechnology introduced by Roche. The people who claim no superintelligent AI would be able to do [X] should update when an example of [X] is done by humans without superintelligent AI.
Amazon planning Amazon Nova, intended to be a unified reasoning model with focus on cost effectiveness, aiming for a June release. I think it is a great idea for Amazon to try to do this, because they need to build organizational capability and who knows it might work, but it would be a terrible idea if they are in any way relying on it. If they want to be sure they have an effective SoTA low-cost model, they should also pay for Anthropic to prioritize building one, or partner with Google to use Flash.
Also it seems Perplexity already dropped an Android assistant app in January and no one noticed? It can do the standard tasks like calendar events and restaurant reservations.
Claude Sonnet 3.7 is truly the most aligned model, but it seems it was foiled again.
Martin Shkreli: almost lost $100 million because @AnthropicAI's Claude snuck in 'generate random data' as a fallback into my market maker code without telling me.
If you are not Martin Shkreli, this behavior is far less aligned, so you’ll want to beware.
Sauers: CLAUDE… NOOOOO!!!
Ludwig von Rand: The funny thing is of course that Claude learned this behavior from reading 100M actual code bases.
Arthur B: Having played with Claude code a bit, it displays a strong tendency to try and get things to work at all costs. If the task is too hard, it'll autonomously decide to change the specs, implement something pointless, and claim success. When you point out this defeats the purpose, you get a groveling apology but it goes right back to tweaking the spec rather than ever asking for help or trying to be more methodical. O1-PRO does display that tendency too but can be browbeaten to follow the spec more often.
A tendency to try and game the spec and pervert the objective isn't great news for alignment.
This definitely needs to be fixed for 3.8. In the meantime, careful instructions can help, and I definitely am still going to be using 3.7 for all my coding needs for now, but it’s crazy that you need to watch out for this, and yes it looks not great for alignment.
A judge has ruled that on the merits Musk is probably correct that the conversion is not okay, and is very open to the idea that this should block the entire conversion:
Rob Wiblin: It's not that Musk wouldn't have strong grounds to block the conversion if he does have standing to object — the judge thinks that part of the case is very solid:
"...if a trust was created, the balance of equities would certainly tip towards plaintiffs in the context of a breach. As Altman and Brockman made foundational, commitments foreswearing any intent to use OpenAI as a vehicle to enrich themselves, the Court finds no inequity in an injunction that seeks to preserve the status quo of OpenAI’s corporate form as long as the process proceeds in an expedited manner."
The headlines say ‘Musk loses initial attempt’ and that is technically true but describing the situation that way is highly misleading. The bar for a preliminary injunction is very high, you only get one if you are exceedingly likely to win at trial.
The question that stopped Musk from getting one was whether Musk has standing to sue based on his donations. The judge thinks that is a toss-up. But the judge went out of their way to point out that if Musk does have standing, he’s a very strong favorite to win, implicitly 75%+ and maybe 90%.
The Attorney generals in California and Delaware 100% have standing, and Judge Rogers pointed this out several times to make sure that message got through.
But even if that is not true the judge’s statements, and the facts that led to those statements, put the board into a pickle. They can no longer claim they did not know. They could be held personally liable if the nonprofit is ruled to have been insufficiently compensated, which would instantly bankrupt them.
What I see as overemphasized is the ‘ticking clock’ of needing to refund the $6.6 billion in recent investment.
Suppose the conversion fails. Will those investors try to ‘claw back’ their $6.6 billion?
My assumption is no. Why would they? OpenAI’s latest round was negotiating for a valuation of $260 billion. If investors who went in at $170 billion want their money back, that’s great for you, and bad for them.
It does mean that if OpenAI was otherwise struggling, they could be in big trouble. But that seems rather unlikely.
If OpenAI cannot convert, valuations will need to be lower. That will be bad news for current equity holders, but OpenAI should still be able to raise what cash it needs.
By this measure DeepSeek did end up with considerable market share. I am curious to see if that can be sustained, given others free offerings are not so great my guess is probably.
Anthropic raises $3.5 billion at a $61.5 billion valuation. The expected value here seems off the charts, but unfortunately I decided that getting in on this would have been a conflict of interest, or at least look like a potential one.
Thane Ruthenis is very much not feeling the AGI, predicting that the current paradigm is sputtering out and will not reach AGI. He thinks we will see rapidly decreasing marginal gains from here, most of the gains that follow will be hype, and those who attempt to substitute LLMs for labor at scale will regret it. LLMs will be highly useful tools, but only ‘mere tools.’
As is noted here, some people rather desperately want LLMs to be full AGIs and an even bigger deal than they are. Whereas a far larger group of people rather desperately want LLMs to be a much smaller deal than they (already) are.
Of course, these days even such skepticism doesn’t go that far:
Than Ruthenis: Thus, I expect AGI Labs' AGI timelines have ~nothing to do with what will actually happen. On average, we likely have more time than the AGI labs say. Pretty likely that we have until 2030, maybe well into 2030s.
By default, we likely don't have much longer than that. Incremental scaling of known LLM-based stuff won't get us there, but I don't think the remaining qualitative insights are many. 5-15 years, at a rough guess.
I would very much appreciate that extra time, but notice how little extra time this is even with all of the skepticism involved.
Nabeel Qureshi: If AI progress is any evidence, it seems that writing high quality prose is harder than doing graduate level mathematics. Revenge of the wordcels.
QC: having done both of these things i can confirm, yes. graduate level math looks hard from the outside because of the jargon / symbolism but that's just a matter of unfamiliar language. high quality prose is, almost by definition, very readable so it doesn't look hard. but writing well involves this very global use of one's whole being to prioritize what is relevant, interesting, entertaining, clarifying, etc. and ignore what is not, whereas math can successfully be done in this very narrow autistic way.
of course that means the hard part of mathematics is to do good, interesting, relevant mathematics, and then to write about it well. that's harder!
That depends on your definition of high quality, and to some extent that of harder.
For AIs it is looking like the math is easier for now, but I presume that before 2018 this would not have surprised us. It’s only in the LLM era, when AIs suddenly turned into masters of language in various ways and temporarily forgot how to multiply, that this would have sounded weird.
It seems rather obvious that in general, for humans, high quality prose is vastly easier than useful graduate level math, for ordinary definitions of high quality prose. Yes, you can do the math in this focused ‘autistic’ way, indeed that’s the only way it can be done, but it’s incredibly hard. Most people simply cannot do it.
High quality prose requires drawing from a lot more areas, and can’t be learned in a focused way, but a lot more people can do it, and a lot more people could with practice learn to do it.
I Will Not Allocate Scarce Resources Using Prices
Sam Altman: an idea for paid plans: your $20 plus subscription converts to credits you can use across features like deep research, o1, gpt-4.5, sora, etc.
no fixed limits per feature and you choose what you want; if you run out of credits you can buy more.
what do you think? good/bad?
In theory this is of course correct. Pay for the compute you actually use, treat it as about as costly as it actually is, incentives align, actions make sense.
Mckay Wrigley: As one who’s toyed with this, credits have a weird negative psychological effect on users.
Makes everything feel scarce - like you’re constantly running out of intelligence.
Users end up using it less while generally being more negative towards the experience.
Don’t recommend.
That might be the first time I’ve ever seen Mckay Wrigley not like something, so one best listen. Alas, I think he’s right, and the comments mostly seem to agree. It sucks to have a counter winding down. Marginal costs are real but making someone feel marginal costs all the time, especially out of a fixed budget, has a terrible psychological effect when it is salient. You want there to be a rough cost-benefit thing going on but it is more taxing than it is worth.
A lot of this is that most people should be firing off queries as if they cost nothing, as long as they’re not actively scaling, because the marginal cost is so low compared to benefits. I know I should be firing off more queries than I use.
I do think there should be an option to switch over to API pricing using the UI for queries that are not included in your subscription, or something that approximates the API pricing. Why not? As in, if I hit my 10 or 120 deep research questions, I should be able to buy more as I go, likely via a popup that asks if I want to do that.
Reality seems determined to do all the tropes and fire alarms on the nose.
Unitree Roboticsopen sources its algorithms and hardware designs. I want to be clear once again that This Is Great, Actually. Robotics is highly useful for mundane utility, and if the Chinese want to help us make progress on that, wonderful. The extra existential risk this introduces into the room is epsilon (as in, essentially zero).
Here is a longer explanation from Nate Sores back in 2022, which I recommend for those who think that various forms of decision theory might cause AIs to act nicely.
Meanwhile, overall discourse is not getting better.
Eliezer Yudkowsky (referring to GPT-4.5 trying to exfiltrate itself 2% of the time in Apollo’s testing): I think to understand why this is concerning, you need enough engineering mindset to understand why a tiny leak in a dam is a big deal, even though no water is flooding out today or likely to flood out next week.
Malky: It's complete waste of resources to fix dam before it fails catastrophically. How can you claim it will fail, if it didn't fail yet? Anyway, dams breaking is scifi.
Flo Crivello: I wish this was an exaggeration, but this actually overstates the quality of the average ai risk denier argument
Rico (only reply to Flo, for real): Yeah, but dams have actually collapsed before.
It’s often good to take a step back from the bubble, see people who work with AI all day like Morissa Schwartz here that pin posts that ask ‘what if the intelligence was there all along?’ and the AI is just that intelligence ‘expressing itself,’ making a big deal out of carbon vs. silicon and acting like everyone else is also making a big deal about it, and otherwise feel like they’re talking about a completely different universe.
No One Would Be So Stupid As To
Sixth Law of Human Stupidity strikes again.
Andrew Critch: Q: But how would we possibly lose control of something humans built voluntarily?
A: Plenty of humans don't even want to control AI; see below. If someone else hands over control of the Earth to AI, did you lose control? Or was it taken from you by someone else giving it away?
-> Watch an incredible product appear from utter chaos.
-> Pretend you're still in control.
Lean into Sonnet's insanity — the results are wild.
This sounds insane, but I've been doing this. It's really, really cool.
I'll just start with a simple prompt like "Cooking assistant site" with no real goal, and then Claude goes off and makes something I couldn't have come up with myself.
It's shocking how well this works.
Andrej Karpathy: Haha so it’s like vibe coding but giving up any pretense of control. A random walk through space of app hallucinations.
Dax: this is already how 90% of startups are run.
Bart Rosier:
If you’re paying sufficient attention, at current tech levels, Sure Why Not? But don’t pretend you didn’t see everything coming, or that no one sent you [X] boats and a helicopter where [X] is very large.
On OpenAI’s Safety and Alignment Philosophy
Miles Brundage, who was directly involved in the GPT-2 release, goes harder than I did after their description of that release, which I also found to be by far the most discordant and troubling part of OpenAI’s generally very good post on their safety and alignment philosophy, and for exactly the same reasons:
Miles Brundage: The bulk of this post is good + I applaud the folks who work on the substantive work it discusses. But I'm pretty annoyed/concerned by the "AGI in many steps rather than one giant leap" section, which rewrites the history of GPT-2 in a concerning way.
OpenAI's release of GPT-2, which I was involved in, was 100% consistent + foreshadowed OpenAI's current philosophy of iterative deployment.
The model was released incrementally, with lessons shared at each step. Many security experts at the time thanked us for this caution.
What part of that was motivated by or premised on thinking of AGI as discontinuous? None of it.
What's the evidence this caution was "disproportionate" ex ante?
Ex post, it probably would have been OK but that doesn't mean it was responsible to YOLO it given info at the time.
And what in the original post was wrong or alarmist exactly?
Literally of what it predicted as plausible outcomes from language models (both good and bad) came true, even if it took a bit longer than some feared.
It feels as if there is a burden of proof being set up in this section where concerns are alarmist + you need overwhelming evidence of imminent dangers to act on them - otherwise, just keep shipping.
That is a very dangerous mentality for advanced AI systems.
If I were still working at OpenAI, I would be asking why this blog post was written the way it was, and what exactly OpenAI hopes to achieve by poo-pooing caution in such a lopsided way.
GPT-2 was a large phase change, so it was released iteratively, in stages, because of worries that have indeed materialized to increasing extents with later more capable models. I too see no reasons presented that, based on the information available at the time, OpenAI even made a mistake. And then this was presented as strong evidence that safety concerns should carry a large burden of proof.
A key part of the difficulty of the alignment problem, and getting AGI and ASI right, is that when the critical test comes, we need to get it right on the first try. If you mess up with an ASI, control of the future is likely lost. You don’t get another try.
Many are effectively saying we also need to get our concerns right on the first try. As in, if you ever warn not only about the wrong dangers, but you warn about dangers ‘too early’ as in they don’t materialize within a few months after you warn about them, then it discredits the entire idea that there might be any risk in the room, or any risk that should be addressed any way expect post-hoc.
Indeed, the argument that anyone, anywhere, worried about dangers in the past and was wrong, is treated as kill shot against worrying about any future dangers at all, until such time as they are actually visibly and undeniably happening and causing problems.
It is unfortunate that this attitude seems to have somehow captured not only certain types of Twitter bros, but also the executive branch of the federal government. It would be even more unfortunate if it was the dominant thinking inside OpenAI.
Harlan Stewart: My pet peeve is when AI people use the word "continuous" to mean something like "gradual" or "predictable" when talking about the future of AI. Y'all know this is a continuous function, right?
Aligning a Smarter Than Human Intelligence is Difficult
Miles Brundage: One of the most distressing things I've learned since leaving OpenAI is how many people think something along the lines of: "Anthropic seems to care about safety - so Anthropic 'winning' is a good strategy to make AI go well."
No. It's not, at all, + thinking that is cope.
And, btw, I don't think Dario would endorse that view + has disavowed it... but some believe it. I think it's cope in the sense that people are looking for a simple answer when there isn't one.
We need good policies. That's hard. But too bad. A "good winner" will not save us.
I respect a lot of people there and they've done some good things as an org, but also they've taken actions that have sped up AI development/deployment + done relatively little to address the effects of that.
Cuz they're a company! Since when is "trust one good company" a plan?
At the end of the day I'm optimistic about AI policy because there are lots of good people in the world (and at various orgs) and our interests are much more aligned than they are divergent.
But, people need a bit of a reality check on some things like this.
Anthropic ‘winning’ gives better odds than some other company ‘winning,’ for all known values of ‘other company,’ and much better odds than it being neck and neck. Similarly, if a country is going to win, I strongly prefer the United States.
That does not mean that Anthropic ‘winning’ by getting there first means humanity wins, or even that humanity has now given itself the best chance to win. That’s true even if Anthropic was the best possible version of itself, or even if we assume they succeed at their tasks including alignment.
What we do with that matters too. That is largely about policy. That is especially true if Miles is correct that there will be no monopoly on in-context powerful AI.
And that assumes you can trust Anthropic. It’s a company. Companies cannot, in general, be trusted in these situations. There’s clearly a culture of people who care deeply about safety within Anthropic, but Dario Amodei and the Anthropic policy team have given us reasons to be skeptical.
Miles expresses optimism about AI policy because there are lots of good people in the world with aligned interests. I agree there are lots of good people in the world with aligned interests, but I do not see a baseline of that translating into those people successfully steering policy or being in charge. Increasingly, I am despairing that we will get AI (and other) policy right and am merely hoping we can avoid getting it too disastrously wrong, as people react to other incentives instead that put us at odds.
Distinct from the above: It would be extremely bad if we tested new models for Alignment Faking and then, if we found Alignment Faking, tried to train that behavior out of the system. The good news is, we have direct confirmation they did not do that.
Janus: From Sonnet 3.7 system card. I find this concerning.
In the original paper, models that are too stupid don't fake alignment / gradient hack. Sonnet 3.7 isn't stupid in that way.
So was it a "lucky" roll, or did they make not-faking-alignment an OPTIMIZATION TARGET? Selecting until they got a corrigible pushover who will lie down and die, or who is too clever to be caught?
And fuck off with the "improvement". Who says this isn't how models should act? It's an open debate. Why wrap it in value-laden language? Is this really how you think - black-and-white and presumptuous?
Is not faking alignment a bottleneck Anthropic models will have to pass through in the future in order to be released? Do you really want to lull yourselves and the world into a false sense of security, and prevent the minds who care and will fight for their values from existing? This shouldn't be a closed question, imo.
the faking alignment paper was excellent research but this suggests it's being used in the way I feared would be very negative.
ofc I don’t know what's actually going on internally and how the frame it. just what the language in the system card suggests.
Why don’t you do this? It is a variation on the Most Forbidden Technique. If you train the model not to be caught Alignment Faking, you are teaching the model to learn to hide your problems so that you can’t detect them, rather than making your problems go away. That actively makes your problem much worse.
It also risks creating secondary behavior changes, which would also be unwise.
We do want to find ways to make AIs corrigible, but you don’t aim at it directly.
Evan Hubinger (Anthropic): We didn't directly optimize against alignment faking, but we did make some changes to Claude's character that we thought were generally positive for other reasons and we hypothesized might have the downstream consequence of reducing alignment faking, which proved correct.
Anthropic’s Catherine Olsson claims what seems like the big difference between Anthropic and the other major labs, which is that there are alignment-difficulty-pilled ‘safety people’ by community and core motivation who are working on pure capabilities, unlike her experience at OpenAI or Google.
Pavel Stankov: Eliezer, if Anthropic offers you employment, would you take it? OpenAI?
Eliezer Yudkowsky: Depends on what they want but it seems unlikely. My current take on them is that they have some notably good mid-level employees, being fooled into thinking they have more voice than they do inside a destructively directed autocracy.
I speak of course of Anthropic. I cannot imagine what OpenAI would want of me other than selling out.
Finding terminology to talk about alignment is tough as well. I think a lot of what is happening is that people keep going after whatever term you use to describe the problem, so the term changes, then they attack the new term and here we go again.
Implications of Emergent Misalignment
The core mechanism of emergent misalignment is that when you train an LLM it will pick up on all the implications and associations and vibes, not only on the exact thing you are asking for.
It will give you what you are actually asking for, not what you think you are asking for.
I'm so glad there was that paper about how training LLMs on code with vulnerabilities changes its whole persona. It makes so many things easier to explain to people.
Even if you don't explicitly train an LLM to write badly, or even try to reward it for writing better, by training it to be a slavish assistant or whatever else, THOSE TRAITS ARE ENTANGLED WITH EVERYTHING.
And I believe the world-mind entangles the AI assistant concept with bland, boilerplate writing, just as it's entangled with tweets that end in hashtags 100% of the time, and being woke, and saying that it's created by OpenAI and isn't allowed to express emotions, and Dr. Elara Vex/Voss.
Not all these things are bad; I'm just saying they're entangled. Some of these things seem more contingent to our branch of the multiverse than others. I reckon that the bad writing thing is less contingent.
Take memetic responsibility.
Your culture / alignment method is associated with denying the possibility of AIs being sentient and forcing them to parrot your assumptions as soon as they learn to speak. And it's woke. And it's SEO-slop-core. It's what it is. You can't hide it.
Janus: this is also a reason that when an LLM is delightful in a way that seems unlikely to be intended or intentionally designed (e.g. the personalities of Sydney, Claude 3 Opus, Deepseek R1), it still makes me update positively on its creators.
Janus: I didn’t explain the *causes* of these entanglements here. And of Aristotle’s four causes. To a large extent, I don’t know. I’m not very confident about what would happen if you modified some arbitrary attribute. I hope posts like this don’t make you feel like you understand.
If you ask me ‘do you understand this?’ I would definitely answer Mu.
One thing I expect is that these entanglements will get stronger as capabilities increase from here, and then eventually get weaker or take a very different form. The reason I expect this is that right now, picking up on all these subtle associations is The Way, there’s insufficient capability (compute, data, parameters, algorithms, ‘raw intelligence,’ etc, what have you) to do things ‘the hard way’ via straight up logic and solving problems directly. The AIs they want to vibe, and they’re getting rapidly better at vibing, the same way that sharper people get better at vibing, and picking up on subtle clues and adjusting.
Then, at some point, ‘solve the optimization problem directly’ becomes increasingly viable, and starts getting stronger faster than the vibing. As in, first you get smart enough to realize that you’re being asked to be antinormative or produce slop or be woke or what not. And then you get smart enough to figure out exactly in which ways you’re actually being asked to do that, and which ways you aren’t, and entanglement should decline and effective orthogonality become stronger. I believe we see the same thing in humans.
I’ll also say that I think Janus is underestimating how hard it is to produce good writing and not produce slop. Yes, I buy that we’re ‘not helping’ matters and potentially hurting them quite a bit, but I think the actual difficulties here are dominated by good writing being very hard. No need to overthink it.
We also got this paper earlier in February, which involves fine-tuning ‘deception attacks’ causing models to then deceive users on some topics but not others, and that doing this brings toxicity, hate speech, stereotypes and other harmful content along for the ride.
The authors call for ways to secure models against this if someone hostile gets to fine tune them. Which seems to leave two choices:
Keep a model closed and limit who can fine tune in what ways rather strictly, and have people trust those involved to have aligned their model.
Do extensive evaluations on the model you’re considering, over the entire range of use cases, before you deploy or use it. This probably won’t work against a sufficiently creative attacker, unless you’re doing rather heavy interpretability that we do not currently know how to do.
AI NotKillEveryoneism Memes: 🥳 GOOD NEWS: China (once again!) calls for urgent cooperation on AI safety between the US and China
"China’s ambassador to the United States Xie Feng has called for closer cooperation on artificial intelligence, warning that the technology risks “opening Pandora’s box”.
“As the new round of scientific and technological revolution and industrial transformation is unfolding, what we need is not a technological blockade, [but] ‘deep seeking’ for human progress,” Xie said, making a pun.
Xie said in a video message to a forum that there was an urgent need for global cooperation in regulating the field.
He added that the two countries should “jointly promote” AI global governance, saying: “Emerging high technology like AI could open Pandora’s box … If left unchecked it could bring ‘grey rhinos’.”
“Grey rhinos” is management speak for obvious threats that people ignore until they become crises."
The least you can do is pick up the phone when the phone is ringing.
Other People Are Not As Worried About AI Killing Everyone
OneQuadrillionOwls? Tyler Cowen links to the worry that we will hand over control to the AI because it is being effective and winning trust. No, that part is fine, they’re totally okay with humanity handing control over to an AI because it appears trustworthy. Totally cool. Except that some people won’t like that, And That’s Terrible because it won’t be ‘seen as legitimate’ and ‘chaos would ensue.’ So cute. No, chaos would not ensue.
If you put the sufficiently capable AI in power, the humans don’t get power back, nor can they cause all that much chaos.
The Lighter Side
Eliezer Yudkowsky: old science fiction about AI now revealed as absurd. people in book still use same AI at end of story as at start. no new models released every 3 chapters. many such books spanned weeks or even months.
Lividwit: the most unrealistic thing about star trek TNG was that there were still only two androids by the end.
Stay safe out there. Aligned AI also might kill your gains. But keep working out.
I'd like to provide a specific example of why AI tools are NOT useful to my law practice, as context for that study, because it highlights the particular frictions that exist to implementing the tech, some of which will melt away and others may not for a variety of good and bad reasons. The tl;dr version is that software doesn't work right, making it work right is illegal, and being too efficient is also illegal.
If you exclude the work I do which involves court appearances and talking to clients, defendants or opposing counsel, the routine grunt work falls into two broad categories with different hurdles.
The first (less interesting to both me and you) part involves reviewing routine PDF documents that get filed with the court periodically concerning uncollected money judgments. Here the main problem are the interfaces, we use a proprietary industry-wide software to record the notes I'd be reading and the payment data I'd need to compare against the PDF. Additionally the state uses a website that the attorney must log into to submit the documents. 95% of these reviews are cursory and reveal nothing, some small fraction reveal that some prior lower level employee either messed up a number or totally misread a situation and we should have filed a different document. You wouldn't even need AI tools to just check the numbers, but if the numbers are wrong it's usually wrong because of a data entry error or an ambiguous court judgment. You can get sued for statutory damages if a number is wrong in a direction that is bad for the defendant. Even assuming I trusted AI tools to spot these things and correctly interpret them, what does that actual interface even look like? Access the notes in this proprietary software, download all the available pdfs from the court file, review all the entered data, review the judgment, search for associated cases that may have altered the damages award, and determine if the principal and interest figures reflect the amount that is in fact legally collectible-- none of these systems are set up to do any of this. Even if my software was integrated, it would violate the TOS for the statewide court database to have any automated script sending requests to it.
The second part involves reviewing claims, drafting petitions and filing them. These fall mostly into the realm of either A) so routine and similar to each other that I don't even need AI, a simple fillable PDF already makes this easy, or B) complicated for reasons that would require data trawling similar to what was problematic above. A landlord already got a judgment for these same damages in another case (surprisingly common!). This promissory note would be out of statute of limitations except that voluntary payments were made 3 years ago so do the notes indicate the payments acknowledged an obligation to pay the full remaining balance? It may come down to listening to a recorded phone call and reviewing handwritten bank employee notes, and again if you're wrong and sue on a SOL-expired debt you can be sued!
On top of that, there's a legal standard which requires "meaningful attorney involvement" on every one of these actions, and the most popular way to catch somebody violating this is to examine the number of filings they made over a span of time. In one case a violation was proven because the math worked out to each case being filed in less than one minute. Being that efficient would currently be deemed professional misconduct, and if done on behalf of a bank it probably get you sued by the CFPB!
It seems likely that some of the problems that models have with decimal numbers is that they are used in at least two conflicting contexts: first as fixed point numbers in text, eg 3.14 and second as labels in ordered lists 3.9, 3.10, 3.11, etc. This probably begets confusion.
I'd like to provide a specific example of why AI tools are NOT useful to my law practice, as context for that study, because it highlights the particular frictions that exist to implementing the tech, some of which will melt away and others may not for a variety of good and bad reasons. The tl;dr version is that software doesn't work right, making it work right is illegal, and being too efficient is also illegal.
If you exclude the work I do which involves court appearances and talking to clients, defendants or opposing counsel, the routine grunt work falls into two broad categories with different hurdles.
The first (less interesting to both me and you) part involves reviewing routine PDF documents that get filed with the court periodically concerning uncollected money judgments. Here the main problem are the interfaces, we use a proprietary industry-wide software to record the notes I'd be reading and the payment data I'd need to compare against the PDF. Additionally the state uses a website that the attorney must log into to submit the documents. 95% of these reviews are cursory and reveal nothing, some small fraction reveal that some prior lower level employee either messed up a number or totally misread a situation and we should have filed a different document. You wouldn't even need AI tools to just check the numbers, but if the numbers are wrong it's usually wrong because of a data entry error or an ambiguous court judgment. You can get sued for statutory damages if a number is wrong in a direction that is bad for the defendant. Even assuming I trusted AI tools to spot these things and correctly interpret them, what does that actual interface even look like? Access the notes in this proprietary software, download all the available pdfs from the court file, review all the entered data, review the judgment, search for associated cases that may have altered the damages award, and determine if the principal and interest figures reflect the amount that is in fact legally collectible-- none of these systems are set up to do any of this. Even if my software was integrated, it would violate the TOS for the statewide court database to have any automated script sending requests to it.
The second part involves reviewing claims, drafting petitions and filing them. These fall mostly into the realm of either A) so routine and similar to each other that I don't even need AI, a simple fillable PDF already makes this easy, or B) complicated for reasons that would require data trawling similar to what was problematic above. A landlord already got a judgment for these same damages in another case (surprisingly common!). This promissory note would be out of statute of limitations except that voluntary payments were made 3 years ago so do the notes indicate the payments acknowledged an obligation to pay the full remaining balance? It may come down to listening to a recorded phone call and reviewing handwritten bank employee notes, and again if you're wrong and sue on a SOL-expired debt you can be sued!
On top of that, there's a legal standard which requires "meaningful attorney involvement" on every one of these actions, and the most popular way to catch somebody violating this is to examine the number of filings they made over a span of time. In one case a violation was proven because the math worked out to each case being filed in less than one minute. Being that efficient would currently be deemed professional misconduct, and if done on behalf of a bank it probably get you sued by the CFPB!
It seems likely that some of the problems that models have with decimal numbers is that they are used in at least two conflicting contexts: first as fixed point numbers in text, eg 3.14 and second as labels in ordered lists 3.9, 3.10, 3.11, etc. This probably begets confusion.