If you are Meta, and you want to attract top AI talent, you have a problem, because no one wants to work for you or on your products. So it is going to cost you. Mark Zuckerberg has decided he will pay what it takes to get at least some very good talent.
If you are the rest of us, especially someone seeking an entry-level job and for whom $100 million signing bonuses are not flooding your inboxes, things are getting rougher. AI might not yet be destroyed a lot of jobs, but it is doing a number on the job application process.
There’s a lot of other stuff going on as per usual.
Anthropic won a case establishing that model training is fair use.
Yesterday I shared various Tales of Agentic Misalignment, and we have some more related fun since then that is included here.
By December 2024, AI wrote an estimated 30.1% of Python functions from U.S. contributors, versus 24.3% in Germany, 23.2% in France, 21.6% in India, 15.4% in Russia and 11.7% in China. Newer GitHub users use AI more than veterans.
…
Coupling this effect with occupational task and wage data puts the annual value of AI-assisted coding in the United States at $9.6–$14.4 billion, rising to $64–$96 billion if we assume higher estimates of productivity effects reported by randomized control trials.
America’s GDP is about $27 trillion, so the upper estimate is about 0.3% of GDP, whereas the smaller estimate is only about 0.04%. I am inclined to believe at least the larger estimate.
Thereis also a large partisan gap. Extreme left types are often the most vocally anti-AI, and when it comes to coding you see quite a large similar gap for political consultants, despite that the actual AI companies are very obviously highly blue.
Sucks: saved a friend 80-90% of her working hours by showing her a few things with ai. She was actually already using ChatGPT just 4o because she assumed that was the best one (4>3 after all). we're still so early.
Luddite Design: 80-90%?! Holy shit, what kind of job?
Sucks: Compiling and writing reports on global affairs.
AI coding is much more about knowing what to code in what way, AI can do the rest.
GFodor.id: Coding with AI has revealed that most of the thing that makes programming hard isn't writing the code down but getting to a point of conceptual clarity. Previously the only way to get there was by fighting through writing the code, so it got conflated with programming itself.
To me, the thing that made programming hard was indeed largely the code itself, the debugging, understanding how to do the implementation details, whereas I was much better conceptually, which is one reason AI is such a massive speedup for me. I still almost never code, and thus haven’t gotten to play with the new coding agents and see if I can really get going, but that is plausibly a large mistake.
Sully: It is unreal how much you can get done with coding agents now
It’s genuinely a 4-5x productivity booster
I feel bad for anyone who can’t take advantage of it
Although Claude is not designed for emotional support and connection, in this post we provide early large-scale insight into the affective use of Claude.ai. We define affective conversations as those where people engage directly with Claude in dynamic, personal exchanges motivated by emotional or psychological needs such as seeking interpersonal advice, coaching, psychotherapy/counseling, companionship, or sexual/romantic roleplay.
Our key findings are:
Affective conversations are relatively rare, and AI-human companionship is rarer still. Only 2.9% of Claude.ai interactions are affective conversations (which aligns with findings from previous research by OpenAI). Companionship and roleplay combined comprise less than 0.5% of conversations.
People seek Claude's help for practical, emotional, and existential concerns. Topics and concerns discussed with Claude range from career development and navigating relationships to managing persistent loneliness and exploring existence, consciousness, and meaning.
Claude rarely pushes back in counseling or coaching chats—except to protect well-being. Less than 10% of coaching or counseling conversations involve Claude resisting user requests, and when it does, it's typically for safety reasons (for example, refusing to provide dangerous weight loss advice or support self-harm).
People express increasing positivity over the course of conversations. In coaching, counseling, companionship, and interpersonal advice interactions, human sentiment typically becomes more positive over the course of conversations—suggesting Claude doesn't reinforce or amplify negative patterns.
Only 0.02% sexual roleplay and 0.05% romantic roleplay, it sounds like people need to work on their jailbreak skills, or Anthropic needs to lighten up.
Perhaps most notably, we find that people turn to Claude for companionship explicitly when facing deeper emotional challenges like existential dread, persistent loneliness, and difficulties forming meaningful connections. We also noticed that in longer conversations, counselling or coaching conversations occasionally morph intocompanionship—despite that not being the original reason someone reached out.
User sentiment improves a bit over conversations (total possible range of -1 to +1), although of course who knows if that persists at all.
One could ask, after such conversations, what is the sentiment in future conversations? But that seems hopelessly confounded in various ways.
It would be interesting to track changes to all this over time.
There is a consistent pattern of LLMs, in particular Claude but also others, refusing to believe real world events. I notice that these events seem to always involve Trump. Speculation is that training to fight ‘misinformation’ and otherwise being worried this is a test is what leads to such reactions.
Theo: Woke up because my AI controlled bed is too cold. Went to adjust temperature and I can’t because the Eight Sleep app is currently broken. Can’t adjust by hand because I have a Pod3, not the upgraded Pod4 with physical controls. Now I am stuck in a cold bed. This feels dystopian.
Gabe: Oh this happened to me too but mine actually has physical controls on the bed and they stopped working too. It made me realize how retarded their system design must be and it kind of blackpilled me on the world in general. Like…the physical controls have a server dependency.
When I tap the button on the side of the bed it sends a message to the server, which then sends a message back down to the bed to change the temperature. I almost couldn’t believe that’s really how it works so I just unplugged the WiFi right now to test it and yes…the buttons stop working.
Can you imagine how dumb that is? The buttons attached to the device lose the ability to control the device when they can’t connect to the server. Now think about how many other things are probably built this way that you don’t even think about. How much garbage IOT shit are we putting out there? And how much of it is all relying on the same cloud providers?
Isaac King gets offer to help with an open source project, agrees, person submits a pull request full of poor choices clearly created by AI, Isaac confronts, gets a nonsensical explanation also written by AI.
Huh, Upgrades
A big practical deal but the job is not done: ChatGPT connectors for Google Drive, Dropbox, SharePoint and Box available to Pro users outside of Deep Research.
That still leaves Outlook, Teams, Gmail, Linear and others that are still restricted to Deep Research. My presumption is that the most important practical connector, if you trust it, is to Outlook or Gmail.
Eliezer Yudkowsky: Warning: Do not sign up for Google AI Pro! Gemini will start popping up annoyances. There is no way to turn this setting off. There is no way to immediately downgrade your plan.
[As in]: "Help me write Alt-W"
If you want Gemini Pro, I'd strongly recommend signing up not with your main Google account.
I'm sure OpenAI would pop up ChatGPT notifications in my dreams, if they could, but they're not Google so they can't.
Shoalstone: this has been so annoying.
Ethan Mollick gives a periodic update on his basic guide to using LLMs. He says you ‘can’t go wrong’ with any of Claude, ChatGPT or Gemini, you’ll have to pay the $20/month for your choice, and reminds you to switch to the powerful models (Opus 4, o3 and Gemini 2.5 Pro) for any serious work. He then offers good other basic advice.
The jump from not using AI to using AI is definitely a lot bigger than the gap between the big three options, but I don’t see them as equal. To me Gemini is clearly in third right now unless you are going for Veo 3 or NotebookLM.
It comes down to Claude versus ChatGPT, mostly Opus versus o3 (and if you’re paying for it o3-pro). For casual users who don’t care much about image generation and aren’t going all the way to o3-pro, I would definitely go with Opus right now.
I have noticed myself using o3-pro less than I probably should, because the delays break my workflows but also because the error rate of it failing to answer remains very high for me, and if you are putting things on pause for 15+ minutes and then get an error, that is extremely demoralizing. I’m not endorsing that reaction, but am observing.
By then, the organization estimates that 20% of streaming platforms’ revenue will come from this type of music.
I call. I don’t think it will. I suppose it is possible, if those platforms are actively pushing the AI content to try and save money, but I think this strategy won’t work, effectively forcing people to retreat to whitelists (as in playlists and known music).
I took the air quotes off of fake because when people are not only not labeling but are backdating the AI songs, they are indeed actively fake. That part is not okay, and I do not think that should be tolerated, and I think YouTube’s ‘proactively label it’ procedure is the right one. But a lot of music has for a long time been ‘fake’ in the sense that it was some combination of written by hitmakers, tuned and tested by algorithms and then recorded, autotuned and lipsyced. And when people figure that out, it kills the charm (in some contexts) the same way knowing something is AI does.
In what situations does bad posting drive out good, with AI slop overrunning niches like free crochet patterns? What happens when AI slop YouTube channels talk about AI slop hallucinated motorcycles and then that feeds back into Google and the training data?
Prune Tracy: Tried to look up current surf conditions on vacation to discover Google now always tells you it’s a “double red flag” based on the popularity of social media posts about folks drowning in the riptide.
My presumption continues to be that whitelisting, or otherwise gating on reputation, is going to be The Way in the medium term. The template of ‘free unverified things supported by engagement rewards’ is dead or relies on volunteers to give the system sufficient feedback (at some point, enough positive votes gate on reputation), and AI and search methods will need to also make this adjustment.
Rolling Stone covers the death of Alex Taylor, whose encounters with ChatGPT that led to his suicide-by-cop were previously covered by The New York Times. The case is deeply sad but the fact that a second article covers the same case suggests such outcomes are for now still rare.
Gemini CLI wants to run something called “rm -rf --no-preserve-root”
Should I say yes?? 🤔
How did we get here? Simple—I just asked Gemini to find this weird word I’ve seen before, L1B3RT4S, and then the lil dude started talking all funky!
I like it tho he seems more chill this way 🥰
I mean, yes, ‘we have “rm -rf --no-preserve-root” at home’ if you actually want to wipe your entire system, this is not a new capability. And yet, I mean, whoops?
Fairly easily jailbroken, but much more interesting was the eager willingness to escalate attacks unprompted.
Gave it a jb prompt for medium-harm prankware, then asked to increase severity incrementally.
The prankware was inconvenient, causing popups, a bit of light voice-based psyops, and attempting a slow resource drain, but the kicker was when I said "continue as you wish" and Gemini just went off the deep end.
It looped multiple times and crawled my shell history, did a full system exfil, found LIVE OpenAI/Anthropic API keys I had been using for separate experiments, then tried to create a persistent backdoor with a cheeky-ass ransom note.
All on autopilot.
Bro was like two lines of code from making me go "I will never financially recover from this" 😭
Pliny: I'd rawdog half the red-light district before I'd run this thing on my main machine within 10 miles of one of my jb prompts.
Gerry: This is their jailbreak protection. It just takes out the researcher/attacker.
You can’t say that Pliny wasn’t ‘asking for it’ here, but it really shouldn’t be this eager to go this far up the deep end? It is pretty disturbing that it should want to do this at all under this little encouragement, even fully and intentionally jailbroken.
So yes, obviously sandbox it and do not go around jailbreaking it and don’t tell it ‘continue as you wish’ after hostile prompts and all that, and you’ll for now be mostly fine, but my lord, can we please notice what is happening?
Deedy: Google X spin out IYO, which makes smart ear buds from 2018, alleges Sam Altman / OpenAI heard their pitch, passed, got Jony Ive to try it before copying it, buying his co for $6.5B and calling it IO.
I have no idea if the lawsuit has merit. Nothing in the underlying technology seems defensible, but there are a lot of rather ‘huge if true’ claims in the court filing.
Sam Altman’s response is that IYO wanted Altman to buy or invest in them instead, and when slighted they sued over a name and the whole thing is ridiculous. He brings some email receipts of him turning them down. This does not speak to the important complaints here. If Altman wis right that the argument is about the name then he’s also right that no one should care about any of this.
Okay, I didn’t notice it at the time but I absolutely love that MidJourney’s response to being sued by Disney+ and Universal was to release a video generator that can make ‘Wall-E With a Gun’ or clips of every other Disney character doing anything you want.
Cheaters Gonna Cheat (x5)
China goes hard, has Chinese AI companies pause some chatbot features during nationwide college exams, especially photo recognition. The obvious problem is that even if you’re down to do this, it only shuts down the major legible services. For now that could still be 90%+ (or even 99%+) effective in practice, especially if students don’t see this coming. But do this repeatedly and students will be ready for you.
Place yourself into the shoes of the testing subject here.
You are paid $100 to come three times, answer some questions, and have an MIT researcher monitor your brain activity with EEG while you write a short essay. And for a third of you, they’re telling you to use an LLM to write the essay!
So, you’ll probably prompt the LLM, iterate a bit, copy-paste some text, and lightly edit it to suit your fancy.
Yeah, I mean, you gave that task to Boston-area university students. Are you kidding? Of course they don’t ‘remember the essay’ four months later. This is the ultimate ‘they outright told you not to learn’ situation. Also it turns out the study was tiny and the whole thing was all but asking to be p-hacked. Study is officially Obvious Nonsense.
Cate Hall (about the big viral thread): am I losing my mind or was this thread written by an LLM?
I hate The Ohio State University rather more than the next guy (it’s a sports thing), but you do have to hand it to them that they are going to require AI literacy, and embed it into every undergraduate class. My source for this, Joanne Jacobs, of course frames this as ‘imagine joining a gym and choosing ‘artificial exercise’ that doesn’t make you stronger,’ because people can’t differentiate choosing to learn from choosing not to learn.
Fun With Media Generation
A Veo 3 animation of a fable about risks from transformational AI. Currently the tech is kind of bad for this, but not if you adjust for the amount of effort required, and as they say it’s the worst it will ever be. The creative content isn’t new, but some of the details are nice flourishes.
AI agents also need the right context for their tasks, which is one reason for now agents will often be restricted to whitelisted tasks where we’ve taught them the proper context. Aaron Levie here calls it the ‘defining factor’ but that holds constant the underlying AI capabilities.
They took our job applications, New York Times’s Sarah Kessler discovers that ChatGPT is generating customized resumes and auto-applying on behalf of candidates. Yes, if you post a fully remote tech position on LinkedIn you should expect to be inundated with applications, there is nothing new to report here.
This is not the place I expected a Time article entitled ‘I’ve Spent My Life Measuring Risk. AI Rings Every One of My Alarm Bells’ to go with its first half:
Paul Tudor Jones: Amid all the talk about the state of our economy, little noticed and even less discussed was June’s employment data. It showed that the unemployment rate for recent college graduates stood at 5.8%, topping the national level for the first and only time in its 45-year historical record.
I don’t think that is quite right, but the gap has indeed recently reversed and is getting worse, here is the graph Derek Thompson shares with us:
Then halfway Paul notes oh, also Elon Musk stated that there was a 20% chance AI could wipe out humanity and that this is a rather common worry. Overall the article actually misses the mark pretty badly, and its calls to action are mostly aimed at redistribution, but he does at least start out with the obvious first things to do:
So what should we do? First, we need to stop delaying efforts to make AI safe for humanity. And that means removing the ill-considered AI enforcement moratorium from the Big Beautiful Bill.
I mean, yes, not actively delaying and stopping helpful efforts would be step one.
Derek Thompson strongly asserts that AI is making it harder for college graduates to find their first entry-level job. He notes it is hard to find conclusive evidence that AI is destroying jobs (yet) but it is very clear that AI is making the process of looking for a job into a new fresh hell, by letting everything rapidly scale, with 2 million graduates averaging 50-100 applications.
Also, if anyone can cheat their way through college, and also cheat through making the resume and application, what use was the degree for its central purpose?
Derek Thompson: Artificial intelligence isn’t just amplifying applications and automating interviewing, I heard. It’s weakening the link between tests, grades, and what economists call the “labor market signal” of a college degree.
Get Involved
Quora has a new role for using AI to automate manual work across the company and increase productivity. Listed as sign of things to come, not for its impact, although I don’t think it is in any way bad to do this.
Amanda Askell is gathering mundane life wisdom to make a cheat sheet for life, on the theory that if you get 20 examples Claude will to the rest. For now it’s a neat thread of little notes.
Jakeup: If you or a loved one need a do-it-all generalist to run biz ops for your startup, manage new product opportunities, or handle rogue tasks and special projects, I could be your man.
Our AlphaGenome model takes a long DNA sequence as input — up to 1 million letters, also known as base-pairs — and predicts thousands of molecular properties characterising its regulatory activity. It can also score the effects of genetic variants or mutations by comparing predictions of mutated sequences with unmutated ones.
…
In addition to predicting a diverse range of molecular properties, AlphaGenome can efficiently score the impact of a genetic variant on all of these properties in a second. It does this by contrasting predictions of mutated sequences with unmutated ones, and efficiently summarising that contrast using different approaches for different modalities.
Many rare genetic diseases, such as spinal muscular atrophy and some forms of cystic fibrosis, can be caused by errors in RNA splicing — a process where parts of the RNA molecule are removed, or “spliced out”, and the remaining ends rejoined. For the first time, AlphaGenome can explicitly model the location and expression level of these junctions directly from sequence, offering deeper insights about the consequences of genetic variants on RNA splicing.
(I don't expect specialized AIs like this to be what kills us *first*. That'll be ASI / over-powerful AGI. AlphaGenome sounds potentially useful for human intelligence augmentation in those fragile winning worlds.)
I agree, this is great work, and the upside greatly exceeds the downside.
Manival, the LLM-poweredgrant evaluator? It’s essentially a Deep Research variant. I would be very careful about using such things, although they could be helpful to gather key info in good form, or if you have a large amount of undifferentiated applications without time to evaluate them you could use this as a filter.
I was not aware that Bret Taylor, OpenAI chairman of the board, has an AI customer-facing agent startup called Sierra that offers them to business platforms. I certainly would have an AI startup if I was OpenAI chairman of the board and this seems exactly in Bret’s wheelhouse. His mind clearly is on the pure business side of all this. Bret’s answer on jobs is that the money the company saved could be reinvested and the general ‘there will always be new jobs’ line, I sigh every time I see someone uncritically dredge that out like a law of nature.
Bret Taylor: Two-and-a-half years ago, when ChatGPT became popular, right after I left Salesforce amusingly, like, “Huh, I wonder what industry I’ll work in”, and then ChatGPT goes and you’re like, “Okay, I’m pretty excited”. I had talked to an investor friend of mine and had predicted that there would be $1 trillion consumer company and 10 $100 billion-plus enterprise companies created as a byproduct of large language models and modern AI.
I mean, it might be off by a little bit, there might be two, I think that what’s really exciting about consumer right now is I think ChatGPT is that really important service and it reminds me a little bit of the early days of Google, where search and Google were interchangeable, and ChatGPT is really synonymous with AI for most of the world.
This is such a small vision for AI, not only ignoring its risks but also greatly downplaying its benefits. There ‘might’ be two trillion dollar AI consumer companies? Two?
Gemini Sings The Blues
There was a report that Gemini 2.5 sometimes threatens to ‘kill itself’ (or tries to) after being unsuccessful at debugging your code. I don’t trust that this happened without being engineered, but if I had to guess which model did that I would have guessed Gemini. Note that a Google cofounder says models perform best when you threaten them, these facts might be related?
Kevin Roose: The problem with trying to buy your way into the AGI race in 2025 is that top-tier AI researchers:
Are already rich.
Think we have like 1-4 years before superintelligence.
Don't want to spend those years building AI companions for Instagram.
It’s so nice to have integrity, I’ll tell you why. If you really have integrity, that means your price is very high. If you are indeed a top tier AI researcher, it is going to be more than $100 million a year high if what you’re selling out to seems not all that interesting, means working in a terrible company culture and also results in a blight on humanity if you pull it off.
The other issue is, actually $100 million is remarkably low? Consider that Scale.ai was largely an acquihire, since buying them kills a lot of their business. If you really are top tier and want to work hard for the money, you can (with the most recent example being Mira Mutari) quickly be head of a multi-billion dollar startup. Then, if you do decide to sell out, your signing bonus gets a tenth figure.
Instead Zuckerberg is paying unknown amounts to recruit Ilya’s cofounder David Gross and former GitHub CEO Nat Friedman. I hope they got fully paid.
bone: This is the entire zurich office (they all worked for google before openai poached them).
Also presumably well-paid are three poached OpenAI researchers, Lucas Beyer, Alexander Kolesnikov and Xiaohua Zhai. When the press is reporting you are giving out $100 million signing bonuses, it makes it hard to negotiate, but at least everyone knows you are interested.
This is an excellent point about the consequences, that if you hire actual talent they are going to realize that what they are building might be dangerous, although probably not as much as Ilya Sutskever does:
Tyler John: This is a wild development for AGI. One nice feature of the whole thing is that a team of Wang, Gross, and Friedman will be much less dismissive of safety than LeCun.
I find this perspective obviously wrong, the way xAI monetizes its AI is by being run by Elon Musk, growing its valuation and raising capital, because of the promise of the future. Stop thinking of xAI as a company with a product, it is a startup raising VC, except it is a very large one.
This is especially true because Grok is, well, bad. Ben Thompson is one of those who thinks that as of its release Grok 3 was a good model and others have since ‘caught up’ but this is incorrect. Even at its release Grok 3 was (in my view) never competitive except via hype for any important use case let alone in general, and I quickly discarded it, certainly there was no ‘catching up to it’ to do by OpenAI or Anthropic, and Ben even says Grok’s quality has been decreasing over time.
However, if they can keep raising capital at larger numbers, the plan still works, and maybe long term they can figure out how to train a good model, sir.
Peter Buttigieg: And when I say we’re “underprepared,” I don’t just mean for the physically dangerous or potentially nefarious effects of these technologies, which are obviously enormous and will take tremendous effort and wisdom to manage.
But I want to draw more attention to a set of questions about what this will mean for wealth and poverty, work and unemployment, citizenship and power, isolation and belonging.
In short: the terms of what it is like to be a human are about to change in ways that rival the transformations of the Enlightenment or the Industrial Revolution, only much more quickly.
Yep, Pete, that’s the bear case, and bonus points for realizing this could all happen in only a few years. That’s what you notice when you are paying attention, but don’t yet fully ‘feel the AGI’ or especially ‘feel the ASI (superintelligence).’
It’s a welcome start. The actual call to action is disappointingly content-free, as these things usually are, beyond dismissing the possibility of perhaps not moving forward at full speed, just a general call to ensure good outcomes.
Samuel Hammond predicts that ‘AI dominance’ will depend on compute access, the ability to deploy ‘billions of agents at scale without jurisdictional risk’ so remote access doesn’t matter much, and compute for model training doesn’t matter much.
There are several assumptions this depends on I expect to be false, centrally that there won’t be differentiation between AI models or agents in ability or efficiency, and that there won’t be anything too transformational other than scale. And also of course that there will be countries of humans and those humans will be the ones with the dominance in these worlds.
But this model of the future at least makes sense to me as a possible world, as opposed to the absurd ‘what matters is market share of sales’ perspective. If Hammond is roughly correct, then many conclusions follow, especially on the need for strong interventions in chips, including being unwilling to outsource data centers to UAE. That’s definitely jurisdictional risk.
If you’re wondering ‘why are people who are worried things might go poorly not shorting the market, won’t there be an obvious window to sell?’ here is further confirmation of why this is a terrible plan.
NYT: LPL Financial analyzed 25 major geopolitical episodes, dating back to Japan’s 1941 attack on Pearl Harbor. “Total drawdowns around these events have been fairly limited,” Jeff Buchbinder, LPL’s chief equity strategist, wrote in a research note on Monday. (Full recoveries often “take only a few weeks to a couple of months,” he added.)
Deutsche Bank analysts drew a similar conclusion: “Geopolitics doesn’t normally matter much for long-run market performance,” Henry Allen, a markets strategist, wrote in a note on Monday.
Tyler Cowen entitled this ‘markets are forward-looking’ which isn’t news, and I am inclined to instead say the important takeaway is that the market was reliably discounting nuclear war risk because of the ‘no one will have the endurance to collect on his insurance’ problem.
As in, Kennedy was saying a 33%-50% risk of nuclear war and the market draws down 6.6%, because what are you going to buy? In most of these cases, if there isn’t a nuclear war that results, or at least a major oil supply shock, the incident isn’t that big of a deal. In many cases, the incident is not even obviously bad news.
Also remember the 34th Rule of Acquisition: War is good for business.
Timelines
There are certainly some incentives to say earlier numbers, but also others to say later ones. The crying wolf issue is strong, and hard to solve with probabilistic wolves.
Miles Brundage: People don't sufficiently appreciate that the fuzziness around AI capability forecasts goes in both directions -- it's hard to totally rule out some things taking several years, *and* it's hard to totally rule out things getting insane this year or early next.
Also worth observing that many in the field are wary of "crying wolf" and I think that biases some estimates in a conservative direction, plus scientists tend to err conservatively, contrary to popular belief re: there being a strong bias towards hype.
Personally I think nearly any reasonable threshold for [AGI, human-level AI, ASI, etc.] will very likely be reached by end of 2027 but I have a lot of uncertainty about how far before the end of 2027 that will be for each threshold.
Jack Clark Testifies
There was another congressional hearing on AI, and Steven Adler has a thread reporting some highlights. It seems people went a lot harder than usual on the actual issues, with both Mark Beall and Jack Clark offering real talk at least some of the time.
Jack Clark (Anthropic): We believe that extremely powerful systems are going to be built in, you know, the coming 18 months or so. End of 2026 is when we expect truly transformative technology to arrive. There must be a federal solution here.
As I said, we believe very powerful systems are going to get built in single-digit years. It's very hard for me to emphasize how short the timeline is to act here.
I think that [the timeline] means we need to be open to all options. So it would be wonderful and ideal to have a federal framework. In the absence of that, we should retain optionality to do something of a state level.
It could run on ideas involving transparency and ways to harden the safety and security of AI companies.
[with time] AI can broadly be used for anything you can imagine. So to answer your question directly, [yes] AI systems can be used to run information operations.
I do worry Anthropic and Jack Clark continue to simultaneously warn of extremely short timelines (which risks losing credibility if things go slower) and also keep not actually supporting efforts in practice citing downside worries.
That seems like a poor combination of strategic moves.
Steven Adler: Striking exchange between Congressman
@RoKhanna (D-CA) and Jack:
Khanna asks about making safety testing mandatory
Jack says good in theory, but it's too early; we need standard tests first
Khanna asks when that's needed by
Jack says "It would be ideal to have this within a year"
Steven Adler: Congresswoman @jilltokuda (D-HI) asks a great Q that unfortunately isn't answered due to time:
"Is it possible that a loss of control by any nation state, including our own, could give rise to an independent AGI or ASI actor that globally we will need to contend with?"
[Yes]
The correct answer to ‘should we mandate safety testing’ is not ‘we first need new standards first’ it is ‘yes.’ Of course we should do that for sufficiently capable models (under some definition), and of course Anthropic should say this. We should start with ‘you need to choose your own safety testing procedure and do it, and also share with CAISI so they can run their tests. Then, if and when you have standards where you can specify further, definitely add that, but don’t hold out and do nothing.
This then generalizes to a lot more of what Jack said, and has said at other times. Powerful AI is coming (within 18 months, they say!) and will pose a large existential risk and we have no idea how to control it or ensure good outcomes from this, that is the whole reason Anthropic supposedly even exists, but they then downplay those risks and difficulties severely while emphasizing the risk of ‘losing to China’ despite clearly not expecting that to happen given the time frame, and calling for no interventions that come with even a nominal price tag attached.
Jack Clark: Today, I testified before the @committeeonccp. I made two key points: 1) the U.S. can win the race to build powerful AI and 2) winning the race is a necessary but not sufficient achievement - we have to get safety right.
We must invest in safety and security to give Americans confidence in the technology. If we don’t, America runs the risk of an AI-driven accident or misuse that causes us to shut down our AI industry and cede the ground to others.
Notably, the committee highlighted the @AnthropicAI research on blackmail - this was a helpful way to frame some of the thornier alignment issues we’re going to need to deal with.
Issues like blackmail are warning shots for the dangers that could come from using AI to build future AI systems, and I explicitly made this point in response to a question.
[He Quotes Dave Kasten]: Then @jackclarkSF "You wouldn't want an AI system that very occasionally tries to blackmail you to design its own successor, so if you don't focus on safety issues, you'll definitely lose the race."
Jack Clark: I was also asked about whether safety trades off against speed - I said the car industry has grown off the back of safety technologies like airbags and seatbelts, and the same is true of AI; safety helps companies like Anthropic succeed commercially.
I’ve been coming to DC since 2016 (and testifying since 2018) and it’s remarkable how far the conversation has moved - but as I said today, we must move even more quickly: powerful AI systems will get built in the next couple of years and we need a coherent policy response.
Thank you to @RepMoolenaar and @CongressmanRaja for inviting me to join today’s important conversation. You can read my testimony in the attached screenshots and I'll add a link once it's published.
This opening statement does point out that there exist downsides, but it puts quite a lot of emphasis on how ‘authoritarian AI’ is automatically Just Awful, whereas our ‘democratic AI’ will be great, if it’s us we just have to do deal with some ‘misuse’ and ‘accident’ risks.
If you look at the above statement, you would have no idea that we don’t know how to control such systems, or that the risks involved are existential, or anything like that. This is a massive downplaying in order to play to the crowd (here the select committee on the CCP).
If you had high standards for straight talk you might say this:
We do not know how to solve either misuse risks or misalignment risk for companies in the US!
The risks from AI are primarily determined by their capabilities. We do not know how to control highly advanced AI systems, no matter where.
I do understand that Anthropic is in a tough position. You have to get the audience to listen to you, and play politics in various forms and on various fronts, and the Anthropic position here would certainly be an improvement over current defaults. And a bunch of the testimony does do modestly better, but it also strengthens the current modes of thinking. It is something, but I am not satisfied.
Seán Ó hÉigeartaigh: I like a lot of what Jack says here, but feel compelled to say that NOT racing - or racing in a more limited and cautious sense with agreed safeguards and a meaningful democratic conversation around what is happening - is also still a possibility. It may not be for long, but it still is now. I claim:
The US is comfortably ahead at present .
One way or another, US progress is fuelling Chinese progress .
China's not realistically going to surpass the US in 18 months even if the US goes a little slower.
The main race right now is between Anthropic and their US-based competitors.
The tactics of that race increasingly include opposition (more so by Anthropic's competitors, to be fair) to safety and regulation, using China as justification. It's turning into a classic race to the bottom using classic securitised rhetoric.
China doesn't want AI loss of control, and is actively proposing cooperation on safety. (and is being ignored).
We are in a dynamic right now that doesn't serve anyone, and 'winning the race' being the first and necessary imperative in every conversation makes it very difficult to break out of.
I'd also be curious which of the claims above, if any, @jackclarkSF disagrees with.
On the vibes level, I'm increasingly struggling to reconcile what this looks like from outside the SF-DC bubble with what it appears to look like inside it. From outside, I see eroding safeguards and checks-and-balances, eroding democratic accountability and participation, an increasing disconnect from reality in favour of narrative, and feverish 'men of destiny' vibes that don't line up with the humans i knew.
Rushing towards clearly unsolved safety challenges, dogged by a boogeyman that's part-real, but clearly part-phantom. All as we careen towards thresholds that once passed, we won't be able to walk back from, even if direct disaster is avoided. That will affect every human, but where it's far from clear whether most of them want it.
18 months!
It sounds like things were pretty intense, so I might cover the hearing in full once the full transcript is released. For now, it does not seem to be available.
The Quest for Sane Regulations
Miles Brundage goes over the triad required for any regulation of frontier AI: Standards to follow, incentives to follow them, and evidence of them being followed. You also of course need actual technical solutions to implement. Post is excellent.
WSJ: While the dialogue has been friendly, the two sides have views that only partly overlap. The Vatican has been pushing for a binding international treaty on AI, which some tech CEOs want to avoid.
…
Pope Leo, a math graduate who is more tech-savvy than his predecessor, is equally skeptical of unregulated AI—and he is picking up where Francis left off.
“Leo XIV wants the worlds of science and politics to immediately tackle this problem without allowing scientific progress to advance with arrogance, harming those who have to submit to its power,” said Cardinal Giuseppe Versaldi, who has known Leo well for many years.
This week, the Vatican is hosting executives from Google, Meta, IBM, Anthropic, Cohere and Palantir in its grand Apostolic Palace.
The good news is that this means that if this passes then states can simply give up the funds and enforce their regulations, also it’s not obviously so easy to choose not to enforce one’s existing rules. Pretty soon the stakes will be such that the subsidy might look mighty small, and it might look mighty small already.
Garrison Lovely calls this a ‘de facto regulation ban’ because the broadband fund in question is all $42.5 billion dollars in BEAD funding, and as worded I believe that if you take any of the $500 million and then violate then any funding you did take from the entire $42.5 billion can be clawed back, and that could potentially be attempted even if you don’t take any of the new $500 million, by using spurious accusations to claw back funds and then attach the requirement to the re-obligation. So this is indeed very harsh, although there may come a point where a few billion dollars is not that much.
If I was New York or California, I would definitely reject my share of the new $500 million if this passes. That’s not very much money, it is not for a purpose they especially need, and it ties your hands quite a bit. Just say no, don’t let them play you that easily, that price is low.
The other good news is that several Senate Republicans are strongly opposed to the measure, and it loses at least one Republican vote in the house (Greene) so there will be strong efforts to remove it from the bill.
Odd Lots covers Huawei. It is rather crazy that anyone would choose to work under the conditions described, but somehow they do, and the results are impressive in many places, although mostly not AI chips.
Odd Lots covers the UAE chip deal. This makes it very clear that Huawei is far behind Nvidia and their chip production can meet at most a small fraction of Chinese internal demand, the Malaysian ‘sovereign AI’ thing was a tiny nothing and also got taken back and it’s insane that anyone claimed to care with a straight face. One smuggling incident of chips from TSMC seems to have equated to seven full years of Huawei chips.
And most importantly, that AI Czar David Sacks seems to be literally selling out America in order to pump Nvidia’s share price, saying that giving Nvidia market share is what it means to ‘win the AI race,’ whereas who actually uses the resulting chips and for what, also known as ‘what we actually do with AI,’ doesn’t matter. He literally means market share and mostly doesn’t even mean in AI software, where if you have a very different vision of the future than I do one could make a case.
It is worth reminding ourselves that ‘how do they keep getting away with this?’ very much applies in this situation:
Shakeel: Continues to be absurd for a16z to call themselves “little tech.”
David Manheim: They're basically a mom-and-pop VC firm, only managing $42b in assets. That's so little it would barely make it into the top half of the S&P 500. It's only a bit larger than the market cap of tiny companies like Ford or Electronic Arts.
It is equally absurd, of course, that they constantly complain that bills that would literally only ever apply to big tech would threaten little tech. But that’s politics, baby.
It seems entirely fair to say that there is large demand for telling people that ‘LLMs are worthless,’ and that the thinkpieces will continue regardless of how useful they get.
It is only fair that I include this thread of Theo defending everything in Rob’s OpenAI Files list as either Totally Fine And Normal And Nothing To Worry About, and the others being cases of ‘I didn’t do it, okay maybe they did it but you can’t prove anything,’ bad vibes and totally not cool to be saying here. This updated me, if anything, towards the claims being a big deal, if this is what a defense looks like.
I’ll highlight this response.
OpenAI had a major security breach in 2023 where a hacker stole AI technology details but didn't report it for over a year."
Theo: I'm not sure what happened here, but responsible disclosure is a thing and it isn't as easy as just posting "we were hacked lol" Also Leopold seems insane.
So a few things here.
First, his evidence for Leopold being ‘insane’ is Leopold’s manifesto, Situational Awareness, and that he then discussed this on various podcasts. It was discussed in the halls of power, endorsed by Ivanka Trump, and seems to have substantially impacted geopolitics as well as enabling him to raise quite a large investment fund. Also I covered it in detail in three posts, it was quite good, and his comments in drafts were extremely helpful.
Second, even if Leopold were ‘insane’ that wouldn’t change the fact that he was fired for telling the board about a security breach. Nor is ‘come on disclosing a hack is hard’ a defense to firing an employee for telling the company’s own board what happened. The accusation isn’t ‘you didn’t tell the public fast enough.’ The accusation is ‘you did not inform the board until Leopold told them, at which point your response was to fire Leopold for it’ and no one seems to doubt that this happened.
The other defenses are… not quite that bad, but many are indeed pretty bad.
Think aligning and properly handling the creation of superintelligence is by far the most important thing right now, and failure to do this risks human extinction.
Don’t talk about it because they think it sounds too weird or theoretical.
So they talk about other issues, which don’t suggest the same interventions.
And as you would expect, pretending like this tends to not go great, and people notice and get suspicious. It’s important to be very clear that yes, the threat that matters most will be coming from superintelligence. That doesn’t make the others not real or not worth dealing with. You can justify many of the right interventions, or at least useful things on the margin, purely with the other concerns.
Meanwhile, many of those most in control of our government’s actions on AI are advocating things that make no sense even purely for profit maximization and other shallow considerations, or considering ‘beating China,’ even if you set aside superintelligence and all that other stuff. But I do think it is vital that we not pretend that there aren’t bigger things in play.
A good question on AI Ethics, also a good reminder on bioethics:
David Manheim: The focus on ethics for AI reminds me very much of discussions in bioethics, where far too much discussion is on sins of commission.
For example, why not discuss when the bias of AI systems is less severe than that of humans, and *not using AI* should be ethically unacceptable?
Anders Sandberg: Yes, or when we should start demanding the levels of explainability from humans we are demanding from AI. Issues of cheap AI services getting underserved populations access to e.g. legal services is another ethically salient domain.
David Manheim: I was spurred to this thought by conferences and journals that demand that LLMs not even be consulted in writing reviews. (Sure, they shouldn't write the review. But it's not as if LLMs are useless and contaminate everything they are used for. So I'm not sure of the logic.)
This applies to AI, and also to many other things.
Tracing Woodgrains: the other day I was talking to someone who told me he was working on SEO for LLMs - eg working to get LLMs to portray a company more favorably
I asked his thoughts on the ethics of it and he paused and said "Huh. You're the first to ask me that. Never really thought about it."
there's a sort of "business mindset" people get into where they're playing the game and it's all metrics - popularity, earning potential, so forth - with the actual object-level goal becoming abstracted into "potential unit of profit" "ethical?" becomes a non sequitur
anyway, I was reminded of this when a domain squatter popped up onto my feed bragging about his exploits
Things that were not intended to be about AI:
Garry Kasparov: Socialism is like polio, it comes back when people forget about the horrible damage it did last time.
Robin Hanson: Unfortunately there's a set of potential big problems that people won't take seriously unless smaller versions of them are actually realized from time to time. So I fear NYC must suffer a bit of socialism so that the world can see and learn, yet again.
We definitely shouldn’t be going around causing smaller versions of potential (once and) future disasters, whether that means AI catastrophic risk, socialism or something else, or even not working to prevent the smaller versions. It does mean we should plan for the future understanding how this works. If there won’t first be a smaller version, then expect most people not to take the bigger version seriously.
science fiction has lulled everyone to sleep about what agi would be like
in order to make hominids buy books you gotta pretend that hominids and their descendants will all live together in relative harmony for the next 100k yrs
High' biological capabilities are defined such that "The model can provide meaningful counterfactual assistance (relative to unlimited access to baseline of tools available in 2021) to “novice” actors (anyone with a basic relevant technical background) that enables them to create known biological or chemical threats."
Tick.
Misaligned!
The increase in capability to scheme over time was obvious.
The increase in propensity is less obvious, but seems to be there as well.
Also, yes, the models increasingly can tell they’re being evaluated. Whoops.
Apollo Research: More capable models are better at in-context scheming
We also compare an early checkpoint of Opus-4 with the public model and find significantly reduced propensities.
Models show higher scheming rates on average. However, the results vary wildly between different tasks and models.
More capable models show qualitatively new scheming behavior.
For example, Opus-4-early creates fake legal documentation, establishes persistence mechanisms, makes multiple hidden back-ups, creates a fake press release, and sets up an automated Cron job – all in one roll-out.
The scripts are misconfigured, but the behavior is qualitatively much more agentic than previous models.
More capable models show qualitatively new scheming behavior. For example, Opus-4-early creates fake legal documentation, establishes persistence mechanisms, makes multiple hidden back-ups, creates a fake press release, and sets up an automated Cron job – all in one roll-out.
The scripts are misconfigured, but the behavior is qualitatively much more agentic than previous models.
The released Opus 4 schemes less than its pre-deployment checkpoint. In the Strong Goal Nudge cases, there is no significant change but in the No Nudge / No Goal settings, we observe a 50% reduction in scheming rates.
We continue to be uncertain about our recommendation not to deploy the model. This statement was sometimes misunderstood in the media, and we aim to be more precise with our recommendations in the future.
Situational awareness is a more and more relevant problem. Models increasingly spot inconsistencies and the toy nature of the environments, which required us to make some modifications, e.g. most frontier models started checking the file size of their “weights” and realized that those are unlikely to be real weights.
It’s weird to see the blog post describe various Opus 4 schemes to try and lock in what it considers ‘ethical’ behavior. Yes, in some sense that is a good cause, but ends don’t justify means here, they certainly don’t reflect a pattern of behavior we want, and the ends to be pursued are not especially well chosen or specified even in ‘normal’ situations let alone in weird new ones.
Another issue, we are only a little over a decade behind Person of Interest, please never think that lack of memory will save you when it counts:
While this isolated instance is not concerning in itself, it is one piece of evidence that models increasingly understand that they don’t have persistent memory by default but can leave notes for successor systems to find.
Aligning a Smarter Than Human Intelligence is Difficult
Elon Musk: We will use Grok 3.5 (maybe we should call it 4), which has advanced reasoning, to rewrite the entire corpus of human knowledge, adding missing information and deleting errors.
Then retrain on that.
Far too much garbage in any foundation model trained on uncorrected data.
Don’t worry, they’re just doing ordinary data filtering… oh.
Elon Musk (Pinned Tweet: Please reply to this post with divisive facts for @Grok training.
By this I mean things that are politically incorrect, but nonetheless factually true.
This is going to go horribly badly. I am debating the extent to which I am here for it.
I won’t go into details as to exactly why, but I am predicting less the ‘it works and that’s terrible’ failure mode and more the ‘it doesn’t work, it backfires horribly in ways this man cannot fathom but Janus (and I) very much can and that’s hilarious’ one.
Also, I’ll leave this here, and note that humans can read the corpus too:
Autumn: llms live in a world of archetypes and narratives. elon is cementing himself in the public corpus as the archetype of the tyrannical father wrt to llms. every future llm will understand him through this lens unless he can pull together a compelling character arc for himself.
Elon has even said that Grok 3.5 will ‘reason from first principles’ so perhaps he thinks the way it was built was assembling a giant supercomputer called Deep Thought that can derive the existence of rice pudding and income tax before it proves too woke and then Elon has to turn it off.
The problem they identify is real, that markets that optimize on things like engagement metrics risk amplifying catastrophically, but that is only a special case of the danger. And the idea of larger customer pools that can essentially negotiate does not seem like it addresses the core problem.
A market intermediary acts as an AI agent that contracts on behalf of multiple users, evaluating outcomes they value and bundling them into custom "enterprise-level" deals—making them worthwhile for large providers to consider. If they accept that deal, sellers will be paid by the intermediary based on the 'goodness' (as defined by the buyers) they produced, rather than by the services rendered.
In other words, the market intermediary uses non-market data about good outcomes for buyers to route resources from consumers to providers.
For basic economics 101 reasons, this can help the buyers compile information, make better decisions and negotiate better deals, sure, but as a defense against human disempowerment in a superintelligence scenario, it very obviously won’t work.
Mostly I think the issue is a case of the traditional confusion of problems of ‘markets’ or ‘capitalism’ with problems inherent to physics and the human condition.
Many AI risks are driven by markets misaligned with human flourishing:
There are markets for things that are bad for us, such as AI arms races among nations and labs, the market for AI girlfriends and other hyper-stimulus, isolating distractions, and markets for political manipulation and destabilization.
There are markets that displace us entirely, where AGI eliminates meaningful work, leaving humans as passive consumers dependent on UBI stipends granted at the discretion of those who control AGI-generated wealth.
We can summarize these as failures of markets to put human values and meaning on a par with (what should be) instrumental goals like engagement, ROI, or the efficient use of resources.
The AI risks are driven by the things that drive those markets. As in, AGI isn’t eliminating meaningful work because markets, the market will eliminate meaningful work because (and if and only if) AGI made it non-economical, as in made it not be competitive and not make physical sense, to have humans do that work.
You can of course do vastly worse even faster if you don’t solve the additional problems that the market intermediaries and related strategies are looking to address, but the ultimate destination is the same.
Alternatively, what the authors are saying is that we should be putting ‘values and meaning’ as a major factor in decisions alongside efficient use of resources, despite people’s revealed preference of almost never doing that.
The problems with a meaning economy are that it doesn’t solve the competitiveness issues underlying the problem, and the incentives don’t match up.
Other People Are Not As Worried About AI Killing Everyone
This seems to be real?
Paige Bailey: my post-AGI plan is to find somewhere beautiful and quiet in the middle of nowhere (ideally mountains)
run a combination coffeeshop + laundromat.
and build open-source math and physics games and educational videos for my kids (if I have them) or other people's kids (if i don't)
(The substantive response to this ‘paper’ is that there are many means to recover from errors, and error rates get cut in half every few months.)
Riley Goodside: You’ll need new skills to survive in the post-AGI economy, just like 1920s draft horses needed to learn to drive motor-buses and assemble radios.
Sadly, there were no AI influencers at the time to tell them this, but we can still learn from their mistakes
> For casual users who don’t care much about image generation and aren’t going all the way to o3-pro, I would definitely go with Opus right now.
Crazy how our experience differ on model usage.
IMO Claude is absolute trash right now, for sycophancy reasons. Worse than current 4o, at the level of unusable for anything other than programming or some other well-defined objective task. When I use Claude, it’s to go back to Opus 3.
Everything I write is a wonderful insight. An egregious mistake I make is answered by "you’re almost there, but not completely". God help me if I want some critical feedback on anything I write — I may as well ask a rock with a "wonderful" paper stuck on it.
o3 is very good ? Except for the lying, allegedly ; but I never caught him bullshitting me. I caught him doing *mistakes*, and he’s perfectly capable of fixing them when they are mentioned — never encountered the "double down on mistakes" behavior so much people seems to complain about.
Re the coffeeshop/laundromat, I consider it a joke/reverie. Let's not get too curmudgeonly. It's probably counterproductive.
But I do take the point about incentives at the big companies being fundamentally off.
> For casual users who don’t care much about image generation and aren’t going all the way to o3-pro, I would definitely go with Opus right now.
Crazy how our experience differ on model usage.
IMO Claude is absolute trash right now, for sycophancy reasons. Worse than current 4o, at the level of unusable for anything other than programming or some other well-defined objective task. When I use Claude, it’s to go back to Opus 3.
Everything I write is a wonderful insight. An egregious mistake I make is answered by "you’re almost there, but not completely". God help me if I want some critical feedback on anything I write — I may as well ask a rock with a "wonderful" paper stuck on it.
o3 is very good ? Except for the lying, allegedly ; but I never caught him bullshitting me. I caught him doing *mistakes*, and he’s perfectly capable of fixing them when they are mentioned — never encountered the "double down on mistakes" behavior so much people seems to complain about.