What is troubling is that so many others (now at least Anthropic, Google, Meta and xAI, potentially a few others too) matched 4-level without getting above about a 4.2.
Last year the same impression held about getting above ChatGPT-3.5. The "widely available" thing right now seems to be the 20K H100s clusters, which means 5e25-1e26 FLOPs in a training run that lasts 2-3 months. And so everyone has their own GPT-4, which proves that there is no essential secret sauce at this level of capabilities, only scale. Going further awaits completion of larger training datacenters, with Anthropic and OpenAI probably already training at the next level of scale. Hyperscalers are sinking about $50bn a year each in capex, so $1bn training runs are not yet something outrageous.
the jabberwacky entity is back
There was an early (pre-generative-AI, based on an early ad-hoc form of machine learning and imitation) chatbot called Jabberwacky. Possibly some output from it got into the training set. If so, presumably this entity isn't very helpful or intelligent.
If families are worried about the cost of groceries, they should welcome this price discrimination. The AI will realize you are worried about costs. It will offer you prime discounts to win your business. It will know you are willing to switch brands to get discounts, and use this to balance inventory.
Then it will go out and charge other people more, because they can afford to pay. Indeed, this is highly progressive policy. The wealthier you are, the more you will pay for groceries. What’s not to love?
A problem is that this is not only a tax on indifference, but also a tax on innumeracy and on lack of leisure time. Those who don't know how to properly comparison shop are likely to be less wealthy, not more; same with those who don't have the spare time to go to more than one store.
[Apologies for forgetting to cross-post this and the Monthly Roundup earlier.]
Let’s see. We’ve got a new version of GPT-4o, a vastly improved Grok 2 with a rather good and unrestricted deepfake and other image generator now baked into Twitter, the announcement of the AI powered Google Pixel 9 coming very soon and also Google launching a voice assistant. Anthropic now has prompt caching.
Also OpenAI has its final board member, Zico Kolter, who is nominally a safety pick, and SB 1047 got importantly amended again which I’ll cover in full next week once the details are out.
There was also the whole paper about the fully automated AI scientist from the company whose name literally means ‘danger’ in Hebrew, that instantiated copies of itself, took up unexpectedly large amounts of storage space, downloaded strange Python libraries and tried to edit its code to remove the timeout condition. Oh, that.
Table of Contents
Language Models Offer Mundane Utility
Use GPT-4o to estimate height from photos on dating apps.
Film you eating to count your calories, bite by bite. There is for now a large (22%!) error rate per bite, but law of large numbers hopefully rescues this as long as the errors are uncorrelated?
The art of combining A with B. Llama-3.1-405B likes to live dangerously.
Diagnose your medical condition 98% accurately by looking at your tongue? I am guessing not so much. The examples listed seem very basic, yes that would be easy for an AI to observe but also for a human to do so. And they’re only looking at a handful of conditions even then. Still, one has to start somewhere.
Future tech when? A request to automatically cancel your renters insurance for a place you no longer rent in response to the renewal notice. As in, doing this without the need for any software engineering. My guess is this is 1-2 years out.
Grocery store Kroger using AI to guess your age, gender and how best to engage in price discrimination via special offers and using digital price tags? Is this (man looking at butterfly meme) the price gouging Kamala Harris is promising to fight against, where the grocery stores ended up with a 1.18% net profit margin last year (or 1.43% for Kroger, so obviously they’re gouging)? The way you do price discrimination is you start at higher prices and offer selective discounts.
If families are worried about the cost of groceries, they should welcome this price discrimination. The AI will realize you are worried about costs. It will offer you prime discounts to win your business. It will know you are willing to switch brands to get discounts, and use this to balance inventory.
Then it will go out and charge other people more, because they can afford to pay. Indeed, this is highly progressive policy. The wealthier you are, the more you will pay for groceries. What’s not to love?
What’s not to love is that this creates a tax via complexity, a reason to spend more time that is not especially fun and to no productive effect.
Or to put it another way, this is the high tech version of coupons. Are not coupons price gouging and price discrimination? You charge an artificially high price, then offer those willing to do meaningless manual labor a discount. This is the same thing, except better targeted, and no one has to do the clipping. I call that a win.
Rohit sees the new GPT-4o voice mode as spectacular, a thread, posted after only 10 minutes playing with it.
Glasses for the deaf that caption other people in real time, they cost $485. Subtitles could be highly useful for the rest of us, too, especially in loud or crowded spaces. Will this work in Da Club?
Help Walmart create or improve 850 million pieces of data with 99% lower headcount.
Language Models Don’t Offer Mundane Utility
Another variation on the theme that you need to ask ‘what can the AI do?’ If you ask either ‘what can’t the AI do?’ or ‘can the AI do exactly the thing I’m already doing?’ you are probably going to be disappointed.
This all seems very right to me.
Some big tech clients are not impressed by AI tools, says Dan DeFrancesco at Business Insider. It’s very much a ‘I talked to three guys and they didn’t like it’ article. Yes, some use cases are not ready for prime time and some experiments with AI will fail. If that wasn’t true, you weren’t pushing hard enough.
More thoughts, here from Vox’s Rebecca Jennings, on what’s been wrong with the AI ads during the Olympics. Derek Thompson and Dare Obasnjo boil it down to productivity tools being great in work contexts and when you automate away drudgery, but creepy and terrible when they automate out personal stuff like (in Google’s ad) a kid’s fan letter, where doing the activity is the point.
Various ways AI model announcements can be misleading, especially via gaming of benchmarks and which versions of competitors are tested under what conditions. As I always say, mostly ignore the benchmarks and look at the user feedback. Arena is also useful, but less so over time. As Buck notes, there is nothing wrong with hype or with providing good information, the trick is that everyone deliberately conflates them.
GPT-4o My System Card
I thought the point of a system card was to put it out there at the same time as the model release. Instead, they waited until after they’d already put out a new version.
But hey, better late than never, here you go.
The rest of this section goes over the details. There are no surprises, so you can safely skip the rest of the section.
That seems right. Voice opens up new mundane issues but not catastrophic risks.
My understanding is it has advantages and disadvantages versus GPT-4 Turbo but that ‘matches’ is a reasonable claim.
That sounds like an admission that they indeed failed their voluntary commitments to the White House. GPT-4o was released on May 13, 2024. The system card was released on August 8, 2024. That’s almost three months later.
I wonder if they were waiting for the METR evaluations that came out recently, and are included here? That highlights the timing issue. If those evaluations are part of your test process you can’t release the model and then do the tests.
Their description of their training data sources doesn’t tell us anything new.
Red teaming occured, there are some new details, the process is ongoing. They focused more on the consumer UI rather than on risks in the API interface, although API-based testing is ongoing.
What mitigations did OpenAI take?
Their argument on erotic and violent content is extremely lame:
So… have a system setting for that? I should be able to make that decision for myself.
They raise the concern that you can get GPT-4o to ‘repeat misinformation’ and thus generate audio, which might be ‘more persuasive.’ I find this a pretty silly thing to worry about. It is way too late to worry that someone might have a good text-to-speech engine for a generic voice on arbitrary text. There’s very much an app for that.
What about the preparedness framework tests?
The cybersecurity test showed minimal capabilities. I worry that this methodology, testing the model purely ‘on its own,’ might fail to find important capabilities in the future. Here I am confident we are fine.
For biological threats, they say ‘GPT-4o does not advance biological threat creation capabilities sufficient to meet our medium risk threshold.’
The dark blue bars are reliably noticeably higher than the turquoise bars. On ideation and also release and formulation, it let the novices do better than the experts, which is odd, sample sizes here are likely too small, or perhaps this means something?
This chart shows that yes, GPT-4o and other similar LLMs are highly useful in enhancing our ability to learn and do various productive and scientific things. And that this carries over into biological threats despite the mitigations. Indeed, the mitigations seem to be doing little work here.
Is this enough additional capability that I am worried something will happen sufficiently that I think OpenAI did anything wrong? No. However, this is not a negative result.
For persuasion they found the text modality more dangerous than voice, and both less effective than human persuasion attempts. For autonomy they didn’t find anything, although the model did OK on coding interview questions.
They then list the METR and Apollo evaluations, nothing surprising.
They raise concern about potential ‘emotional reliance’ on and anthropomorphization of the model, based on some observed tester behaviors. I am not too concerned, but I agree it seems worth studying.
Overall, I am very happy to have the system card, but like David Manheim I can’t help but notice the timing issues.
2 Grok 2 Furious 2 Quit
We have Grok 2. How well does it grok?
They say rather well.
With a small team plus a large amount of compute and GPUs, it seems you too can have a rather solid 4-level model. I no longer put as much weight on the details of the top of Arena since it seems effectively some combination of gamed and saturated, but it remains quite good at giving you a general sense.
Here are some of them benchmarks.
I notice I do not entirely trust xAI here, but I do trust that the numbers are solid.
Grok hooks up directly to Twitter in real time. For many practical purposes, if it is ‘good enough’ then that will make it the best choice. I look forward to testing that out.
Pliny jailbroke it right away, would have been shocking if that hadn’t happened. I do love the way this one went, I wonder if it works on Elon Musk too?
However essentially no one is talking about the text model, because they find it much more fun to talk about the image model.
The image model seems pretty good. Here’s a standard thread of images and prompts.
It’s a good picture. So what if negative prompting doesn’t work with that wording?
Wolfgang Blackwood: That’s not the grok llm but the image generator flux not having negative prompts.
If you clarify your intent, it does as asked.
I mean, sure, mistakes get made if you want to find them.
Also it was at least initially otherwise rather willing to do what you want.
This is a bit more extreme, but it isn’t new, see for example this thread of copyright infringement by DALL-E.
Also, I don’t see this as that big a deal?
Pixel Perfect
Pixel 9 will be the first phone to fully feature Google’s new AI features. Like Apple Intelligence, they’re not putting the new features on older phones that can’t handle it. It is not clear when and which other Android phones will get similar features. They seem to have stepped back from calling the assistant Pixie and making it exclusive.
As a happy user of the Pixel 7, I am certainly intrigued. Here’s Marques Brownlee’s preview, since he’s deservedly the internet’s go to phone review guy. He loves the new hardware, including having a Pro-level phone that’s still the size of the normal one, and he loves the Pixel 9 Fold pending battery life. He gives the highest praise you can give for a phone, which is he’s considering personally using the Fold.
And he shows us some cool little AI features, like editing or merging photos, and mentions the screenshot analysis app. But all the Gemini voice chat with the full phone assistant feature gets is a little ‘that’s confusing,’ except when he notes the Pixel Buds will let you have a full conversation with the Assistant without taking out the phone. That seems exciting. I might well be in.
WSJ’s Dan Gallagher says the Pixel 9 is more of an AI delivery device than the rest of the phone, and that might be what the Keynote spent its time on, but that doesn’t mean the phone isn’t also impressive otherwise.
WSJ’s Joanna Stern is very much enjoying talking to Gemini Live.
Here is the full Google Keynote presentation. The headline feature is integration of all the different Google apps into the assistant. Whole thing bored me to tears when I tried to watch. Much better to listen to reports from others or read feature lists.
Here is their rollout announcement on Gemini Live. They will have extensions for various apps including Keep, Tasks, Utilities and expanded features on YouTube Music within a few weeks, and most importantly Google Calendar, to go with existing ones that include Maps, Gmail, Docs and Sheets.
On a practical level, call notes seems like a big deal, as does sharing your camera in real time. Everythings stays on device and you get summaries of our phone calls. It hits different when it’s actually here in a week and one needs to decide whether to buy.
Fun with Image Generation
OpenAI to allow two free DALLE-3 creations daily. This is something like 90% of the actual use marginal value I get from ChatGPT at this point given I have Claude and I’m not in the voice alpha. Of course you can also use Bing for unlimited access.
Deepfaketown and Botpocalypse Soon
Given one photo and five minutes you can call into Zoom looking like anyone. Voice sold separately, but available.
Roon suggests optimism on deepfakes based on defense having the edge.
Huge if true. The discriminator doubtless costs much less compute to use than the picture does to generate. But that assumes you only have to use the discriminator once per fake, and that the discriminator is reliable. What if those are very false assumptions?
Not only do you have to check every real image and video to also spot the fake ones, I can submit my fake any number of times, especially if there is an error rate. If you don’t have a reliable automated detector, you’re cooked. So far, no one has a reliable automated detector? Why do we assume the defense wins this arms race?
I would also warn more generally of solutions that depend on some actor ‘doing their job’ in the sense this here is relying on various tech companies. Why do I get five identical spam messages from Twitter bots?
I am coming around to this view even more over time as well at least for a while:
So what to do?
Sources whose judgment you trust seems exactly right, also AI will become increasingly good at evaluating the extent to which something should or should not be trusted, especially if it has web access, so you can ask it for help. Combine that with keeping track of which sources are trustworthy and invoking Bounded Distrust, and you should continue to be maybe not great, but mostly fine.
The ‘good’ news is that the limiting factor has never been supply of misinformation. However much demand arises, there will always be supply. And most of that demand is for deeply low-quality misinformation. AI allows you to generate much higher quality of misinformation, but the buyers almost all demand very low quality. If no one is even looking at the video, who cares if the AI can make it look realistic?
Steven Adler of OpenAI and coauthors ask how to establish personhood credentials. How do you know that social media account is not a bot?
Trivial inconveniences can matter in practice, especially if bots are trying to do things at large scale and fully swarm a network. However, it seems silly to think bots would be unable to get a human to give them a credential, if they cared enough about that, even if the system was designed well. And I expect that a lot of people will simply have an AI running their social media accounts either way, although that is in important ways still a ‘real person.’
Ultimately, if an account has a reputation and track record, or is claiming to be a particular person, you can look at that either way. And if that’s not the case, you’ll need to evaluate the output with caution. Which was mostly true before AI.
The Art of the Jailbreak
Anthropic announces an expansion of their model safety bug bounty program.
Find a jailbreak, get the cash.
Pliny’s reply made it clear he is on the case.
That all sounds great.
Exactly. The true threat solves for anything at all.
Also, in other model behavior news, some strange behavior out on the 405.
Many more at the link. It is a base model, so while I would never have predicted this particular thing being common the general class of thing does not seem so crazy. I mean, yes, it does seem crazy, but not an unexpected type of crazy?
I occasionally see people, such as Danielle Fong here, argue that the consistent ability to jailbreak models is an argument against regulations like SB 1047. I notice I am confused? If models can be jailbroken, then that means your safety protocols should have to account for bad actors jailbreaking them. And yes, the results should be your fault. Because it would be.
Either find a way to have your model not be jailbroken, or ensure the jailbroken version is a responsible thing to release. Either find a way to not let your safety protocols be fine-tuned away, or be fine with that happening and accept the consequences, or don’t let people do the fine tuning.
A neat phenomenon is what one might call the Aligned Jailbreak.
Claude 3 Opus will lie. Right to your face. A reminder that weirder things are going on than you realize.
Another fun finding is that one prompt can spiral an AI into an existential crisis?
They Took Our Jobs
What if you still have your job, but your job is only working with AI and you hate it?
All the good solutions recognize that if you are 10x as productive, you can afford to give some of that back to get more human contact. If that time is unproductive, that is still fine so long as it keeps you in the game.
Also, wait, what, an AI scientist?
Obvious Nonsense
In hindsight it is weird this has not been a regular section all this time.
This week’s particular obvious nonsense is a study claiming ‘AI poses no existential threat to humanity’ according to the write-up, and ‘Are Emergent Abilities in LLMs Just In-Context Learning?’ by its authors.
Here is the key second half of their abstract:
Here is Yann LeCun boasting to one million views about the claimed result.
This was the same week as the AI scientist paper. There are any number of practical demonstrations that the claim is Obvious Nonsense on its face. But never mind that.
Because…
Seriously?
You want to prove that LLMs are not existential threats to humanity, so you tested on… GPT-2?
To be fair, also GPT-2-XL, GPT-J, Davinci (GPT-3), T5-large, T5-small, Falcon-7B, Falcon-40B, Llama-7B, Llama-13B and Llama-30B (presumably those are Llama-1).
To be fair to the study authors, their actual statements in the paper are often far more reasonable. They do qualify their statements. Obviously the models they tested on pose no existential threat, so it is unsurprising they did not find evidence of capabilities that would represent one when looking.
But also their statements in the press release are Obvious Nonsense, so combined with various things in the paper I think this really is their fault. Yes the headline was worse, but this was not some rogue headline writer pulling that claim out of nowhere.
The study claims in-context learning plays a greater role than we thought, versus emergent abilities, in LLM capabilities. Even if true at greater scales, I don’t see why that should matter or bring much if any comfort? It is trivial to provide the context necessary for in-context learning, and for the model to provide that context recursively to itself if you hook it up to that ability as many are eager to do. The ability remains for all practical purposes ‘emergent’ if it would then… ‘emerge’ from the model in its full ‘context,’ no? The practical impact remains the same?
And certainly Yann Lecun’s above statements, as universal absolute general claims, are laughably, obviously false.
Get Involved
A shameless plug: My entirely-not-AI 501c3, Balsa Research, is holding a fundraiser so we can commission two studies about the costs of the Jones Act. We got two great proposals, and I’d love to be able to fund both of them.
In larger and more exciting funding news, SFF has a new round:
Based on my prior knowledge of SFF, your chances in this round will be much, much better than in the standard SFF round. If you are working on this, do not miss out.
Introducing
A new and improved variant of GPT-4o is available as of last week. OpenAI aren’t giving us any details on exactly what is different, and took a week to even admit they’d changed versions.
Colin Fraser of course posted right away that it still gets 9.11 vs. 9.9 wrong. I half think this is intentional trolling, that OpenAI is deliberately keeping this broken.
On Arena the new version has reclaimed the lead, with a 17 point lead over Gemini 1.5 Pro., and has a substantial lead in coding and multi-turn capability. It does seem like an improvement, but I do not see the kind of excited reactions if it was indeed as good as those scores claim?
Anthropic’s API now offers prompt caching, which they say can cut costs by up to 90% and reduce latency by up to 85%.
Quick math says you reach break-even even if all you do is sometimes ask a second question, so basically anything that often has an follow-ups should use the cache.
In Other AI News
Zico Kolter, a professor of Computer Science at CMU, joins the board of OpenAI. They are presenting him as an AI safety and alignment and robustness pick. He will join the safety and security committee.
Alex Irpan of Google DeepMind transfers from robotics into AI safety, gives his explanation here. Reason one is he thinks (and I agree) that the problem is really interesting. Also he expects superhuman AI in his lifetime and he’s ‘not sold’ on our near term solutions scaling into the future. He doesn’t think the current paradigm gets there, but he’s not confident enough in that for comfort, and he buys instrumental convergence at the limit.
His p(doom) seems unreasonably low to me at 2%. But even at 2% he then does all the highly reasonable things, and recognizes that this is a problem well worth working on – that 2% is presumably based in part on the assumption that lots of smart people will invest a lot into solving the problem.
Huawei readies new chip to challenge Nvidia (WSJ). It is said to be comparable to the H100, which would put Huawei only one generation behind. That is still a highly valuable generation to be ahead, and getting a chip ready is well ahead of when you get to release it, even if things go smoothly, and its current chips are facing delays.
Australian billionaire Andrew Forrest is going to legal war with Meta over their failure to police AI and other scams on Facebook using his likeness.
Startup AI company Glean in talks to double its valuation in 6 months to $4.5 billion. They do corporate information lookup, which is why I’d never heard of them.
AI agent offers $300 bounty to humans to get them to write documentation on how to have AI agents pay humans to do work. As I’ve said before, the solution to ‘the AI might be clever in some ways but it can’t do X’ is ‘you can give money to a human to get them to do X.’ It’s a known tech, works well.
Quiet Speculations
While various people talk about how AI isn’t making progress or paying off because they have absolutely zero patience…
‘We are not the same.’
Ethan Mollick reminds us we may have change blindness with respect to generative AI. In the past 18 months we really have seen dramatic improvements and widespread adaptation, but our goalposts on this have moved so much we forget. Images and video are leaping forward. The flip side is that this doesn’t cite the improvements from GPT-4 (original flavor) up through Sonnet 3.5.
Despite this, and despite the standard three year product cycle being only 1.5 years old right now, it is a bit unnerving how many 4-level models we are getting without a 5-level model in sight.
It is not that troubling a sign for progress that we haven’t seen a 5-level model yet, because it has not been that long. What is troubling is that so many others (now at least Anthropic, Google, Meta and xAI, potentially a few others too) matched 4-level without getting above about a 4.2.
That suggests there may be a natural plateau until there is an important algorithmic innovation. If you use essentially standard techniques and stack more layers, you get 4-level, but perhaps you don’t get 5-level.
Or we could simply be impatient and unappreciative, or asking the wrong questions. I do think Claude Sonnet 3.5 is substantially more productivity enhancing than the original GPT-4. There’s been a lot of ‘make it faster and cheaper and somewhat smarter rather than a lot smarter and more expensive’ and that does seem to be what the market demands in the short term.
For any other product, a 10x+ cost reduction per year with modest quality improvement would be huge. Perhaps most people do not realize the change because for them the cost was never the issue?
Long term, sufficiently advanced intelligence is (in a commercial sense, and barring catastrophic risks) Worth It. But if you can’t get it sufficiently advanced, people are asking relatively dumb questions, so on the margin maybe you go for the price drop.
Sure does:
A fun game when watching Star Trek: Next Generation in particular (but it works with other iterations too) is ‘should this or whatever caused this by all rights cause a singularity or wipe out the Federation, and why hasn’t it done either of those yet’? Another is ‘why didn’t they use the ship’s computer to use AI to solve this problem?’ although the answer to that one is always ‘it did not occur to them.’ Also see a certain room shown in Lower Decks.
My head cannon is totally that Q and travellers and other cosmic entities and future time travelers and various temporal loops are constantly running interference to stop us and various others from being wiped out or taken over by AIs or causing singularities. Or it’s a simulation, but that’s no fun. Nothing else makes any sense.
Your failure to build the products people want is my opportunity.
A good idea with the wrong proposed name?
Our entire legal system is based on this principle, in both civil and criminal. The two sides look ahead to what would happen in a court, and they reach an agreement on that basis. Most preparations and costs and work are about getting the leverage to negotiate such deals. And indeed, the same is true all the way back to the original act. The threat is stronger than its execution. We would need to adjust our resolution mechanisms, but if AIs can simulate the process and handle the negotiations, that is your best possible situation.
One twist is that AIs could also see your track record. So the wise are negotiating and acting with that in mind, even more so than today. Some (such as Trump) see value in credibly threatening scorched earth legal policies and never settling, and cultivate that reputation on purpose, so people are afraid to cross them or sue them. Others play the opposite strategy, so they will be good partners with which to do business. The argument ‘if I settle with you here that opens me up to infinitely more lawsuits’ becomes much stronger in an AI world. The game theory will get very interesting.
Ignore the rhetoric and focus on direct impact of actions. Has Meta been accelerationist or decelerationist so far?
We see a version of this claim every few months, Dan Hendrycks said it in January. If we are focused purely on frontier lab progress, I do think that up until now a reasonable case can be made here that they are driving the costs up and benefits down. For AI not at the frontier, especially those looking to actively use Llama, this goes the other way, but (for now at least) all of that is mundane utility, so it’s good.
A key issue is what this lays groundwork for and sets in motion, including potentially enabling AI progress on the frontier that uses Llama to evaluate outputs or generate synthetic data. At some point the impact will flip, and systems will be actively dangerous, and everything indicates that those involved have no intention of changing their behavior when that happens.
The other is the obvious one, this intensifies the race, which potentially lowers profits but could also drive even faster development and more investment in the name of getting there first damn the costs and also the safety concerns. That includes all the players that this invites into the game.
Whatever else one says, the man commits to the bit.
I hope he is right, but I am deeply skeptical.
Trusting his future to the cards, man clings to a dim hope.
(I bet Taylor Swift absolutely outsources most of her negotiating, and also most of the dancing and related work. Even if she was good enough, there’s no time.)
This scenario does not sound like a stable equilibrium, even if we assume the good version of this (e.g. alignment is fully solved, you don’t have an offense-defense crisis, and so on)?
The humans who increasingly turn everything over to those AGIs win, in all senses. Those that do not, lose. The hope is that other humans will ‘reward authenticity’ here the way we reward Taylor Swift sufficiently to make up for it, and will retain in control sufficiently to do that? Or that we’ll use political power to enforce our edge?
Won’t those who gain political power soon be AGI’s puppets?
If you are counting on ‘AIs can’t sign contracts’ I assure you that they can find someone to sign contracts on their behalf.
If you are counting on ‘only humans can make friends’ then you are not properly thinking about AGI. Those who lets their AGIs make friends will have better friends, and the AGIs will also outright do it themselves. They’re better at it.
I don’t see an acceptable way to make this system work? What’s the plan?
Richard also offered this thought on an additional modality:
The ‘legacy’ employed who can’t be fired are a temporary phenomenon, and in some places where this is too strict might essentially kill a lot of the existing businesses in this kind of scenario. The equilibrium question is to what extent we will force zero (or very low) marginal product people (since that’s now almost everyone) to be hired. And at what salary, since there will be far more supply than demand.
If humans have pensions that take the form of jobs, especially ‘work from home’ jobs where they offer zero marginal product and thus are rarely asked to do anything, do they have jobs? Do the existing ‘no show’ jobs in many corrupt governments and corporations count as employment? That is an interesting philosophical question. I would be inclined to say no.
In the long run equilibrium, this still amounts (I think?) to a claim that humans will retain political power and use it to become rent seekers, which will sometimes take the form of ‘jobs.’ If we are fine with this, how do we ensure it is an equilibrium?
SB 1047: One Thing to Know
The most important thing to know about SB 1047 is:
SB 1047 has zero impact on models that cost less than $100m in compute to train.
This almost certainly includes all currently available models.
No one can lower the $100m threshold.
The frontier model board can raise the compute threshold, and make some models no longer covered, if they want.
But they cannot change the $100m threshold. No model that costs less than $100m will ever be covered. Period.
I emphasize this because when asking about SB 1047, someone reported back that only 24% of respondents were confident that older models wouldn’t be impacted.
The clause is written that way exactly to be able to make that promise.
Of course, California can pass other laws in the future. But this one doesn’t do that.
SB 1047 is Amended Again
SB 1047 has now gone through substantial changes in the Appropriations committee, because that is how California law is made.
Going by the summary announcement, many but not all of Anthropic’s changes were made. Some of the changes are clear Pareto improvements, making the bill strictly better. Others reduce the bill’s effectiveness in order to reduce its downside risks and costs and to make the bill more likely to pass.
Many of the actually valid objections made to SB 1047 have now been fully addressed. Several less valid objections have also been explicitly and clearly invalidated.
In particular:
Points one and two especially important in terms of overcoming commonly made arguments. Anyone who still talks about having to ‘prove your model is safe’ is either misinformed or lying. As is anyone saying the model could be incidentally used to cause harm similar to a truck one could fill with explosives. Also note that perjury is gone entirely, so you can go to jail for lying on your driver’s license application (don’t worry, you won’t) but not here, and so on.
The full details have not been announced yet. When they are, I intend to cover the (hopefully and probably final) version of the bill in detail.
I’m sad to see some of the provisions go, but clearly now: It’s a good bill, sir.
Already we see Samuel Hammond saying the bill is now very reasonable, and Vitalik Buterin agreeing these are very substantive positive changes that address his primary concerns.
SB 1047 Rhetoric Prior to the Recent Changes
Everything here was written and responded to prior to the amendments above. I noted the changes in two places, and otherwise preserved the old version for posterity. There was clearly a big push by a16z and company to get people to object to the bill right under the wire, before many of the remaining valid objections are invalidated. The timing does not make sense any other way.
Also if you look at the letter they got signed by eight California Congressmen right under the wire (and once you get six, you know they asked a lot more of them, there are 52) it is full of false industry talking points and various absurd and inaccurate details. There is no way these people both understood what they signed and thought it was accurate. And that’s before the bill changed. So that we can remember and that future searches can find the list easily the six were: Zoe Lofgren, Anna Eshoo, Ro Khanna, Scott Peters, Tony Cardenas, Ami Bera, Nanette Barragan and Luis Correa.
Garrison Lovely in The Nation calls SB 1047 a ‘mask off moment’ for the industry. He points out that the arguments that are being used against SB 1047 are mostly blatant lies, just actual outright lies, while an an industry claiming they will be transformational within 5 years says it is too early to regulate them at all and finds reasons to oppose an unusually well-considered and light touch bill.
Notion cofounder Simon Last writes in support of SB 1047 in the LA Times. Simon simplifies a bit and uses rhetorical angles I think unwise, but the points are taken. Paul Leaks writes against SB 1047 below at the same link, and as is typical Paul hallucinates a very different bill.
Daniel Kokotajlo comes out in favor, and predicts that if enacted there will be little or no impact on industry and innovation. I agree.
Vitalik Buterin is positive on many aspects of SB 1047, and notices how changes made have addressed various concerns. His main criticism is that the threshold for derivative models should be cost-based, which indeed is now the case.
Arram Sabeti feels similar to many others, finding new regulations highly aversive in general but he sees SB 1047 as a unique situation. Those opposing the bill need to understand that most of those loudest in support are like this, and would happily stand on the other side of most other regulatory fights.
Preetika Rana writes about tech’s attempts to kill SB 1047 in the WSJ. This is what good mainstream tech journalism looks like these days, although it gives unequal time to bill opponents and their arguments, and has one mistake that should have been caught – it says the bill defines catastrophic harm purely as cyberattacks to the exclusion of other threats such as CBRN risks.
It makes clear that only $100m+ cost models are covered and it has so far faced little opposition, and that Newsom isn’t talking about whether he’ll sign. It quotes both sides and lets them give talking points (even if I think they are highly disingenuous) without letting in the fully false claims.
I wonder about these complaints: Do the opponents want the bill to specify what the tests are and name the people on the new commission, despite that never being how such bills work? Or do they say both ‘this bill is insufficiently flexible as things change’ and also ‘you did not exactly specify how everything will go’?
Do they want the government to specify now, for the indefinite future, exactly under what circumstances they will face exactly what reactions, and actually face that, with no human discretion? Or would they (correctly) scream even louder to even higher heaven about that regime, as leading to absurd outcomes?
There are several versions of this, of opponents of the bill saying versions of ‘I don’t know what to do here’ without offering opportunity for counterarguments.
The most central counterargument is that they are basically lying. They absolutely know procedures they could follow to meet the standards offered here if their models do not pose large catastrophic risks, for example Anthropic wrote a document to that effect, and that there will be various forms of guidance and industry standards to follow. And that ‘reasonable’ is a highly normal legal standard with normal meanings.
And that when they say compliance is impossible, they are hallucinating a different kind of law where the government including the judges and juries are a cabal of their sworn enemies completely out to get them on every little thing with every law interpreted literally even when that never happens. And so on.
Another of which is that if you don’t know any reasonable actions you could take to prevent catastrophic harms, and are complaining you should be allowed to proceed without doing that, then maybe that should be a you problem rather than us letting you go ahead?
What most of them actually want, as far as I can tell – if they can’t simply have the rule of law not apply to them at all, which is their first best solution – is to have the government answer every question in technical detail in advance, to have full safe harbor if they follow the guidance they get to the letter no matter the circumstances, to have that done without it being anyone’s job to understand the situation and give them well-considered answers (e.g. they oppose the Frontier Model Division). And then they want, of course, to have the benefits of common sense and nullification and ‘no harm no foul’ to their benefit, if those rules seem stupid to them in a given spot, and the right to sue over each individual answer if they disagree with it, either in advance or post hoc.
Indeed, most of their objections to SB 1047 are also objections to the common law, and to how liability would work right now if a catastrophic event occurred.
That is even more true under the new version of the bill.
Back to the article: I especially appreciated the correct framing, that Big Tech including Meta and Microsoft, and various VC ecosystem players, are both lobbying heavily against this bill.
I am so sick of hearing that a bill opposed by Google, Microsoft, Meta and IBM, that applies only to the biggest companies and training runs, is ‘ripe for regulatory capture’ and a plot by Big Tech. Or that talk of existential or catastrophic risk is some trick by such people to goad into regulation, or to build hype. Stop it.
The Quest for Sane Regulations
A remarkably common argument is of the form:
If you want to go faster than your competitor and win a race, as every Mario Kart player knows, throwing obstacles that slow them down is highly useful. Speeding yourself up is not your only option. Also, yes, you want to stop the Rubber Band AI where they get to take your chips and models and catch up. That’s key as well.
Why not treat safety measures like funding rounds via conditional commitments?
There are complications, especially with verification, but this is worth considering.
On SB 1047 in particular, I do think the expected economic impact is net positive even if you don’t care about the catastrophic harms themselves, for various second-order reasons and because the first-order costs are quite small, but of course some others strongly disagree with that.
Here’s another opinion, I suppose?
That seems obviously false?
Also it seems to imply no regulation can matter in any other way that matters either? If you can’t change the possibility of an ‘AI risk event’ then you can’t meaningfully alter the pace of progress either. And I presume Roon would agree that the mundane utility in the meantime ultimately doesn’t matter.
One can also ask, are you willing to bite the bullet that not only no set of rules could make things better, but that no set of rules can make things worse?
Finally, if that were remotely true, then isn’t the response to pass rules that radically speed up the pace of regulatory and government response and gives them the expertise and transparency to know how to do that? And indeed most currently debated rules are mostly doing a subset of this, while opponents sound alarms that they might do a larger subset of it.
AIPI has a new poll regarding SB 1047. As usual, since this source is potentially biased, one must be careful to look at wording, here are the full results and details directly.
Here, in addition to the traditional topline questions that reliably find support for the bill, they ask about various amendments proposed by Anthropic that would weaken the bill. They find the public opposed to those amendments.
Much of that opposition, and much support for SB 1047 provisions, is overwhelming. Most important is that they support 69%-17% that enforcement should happen before any potential harm occurs, rather than after a catastrophic event. I find the wording on that question (#12) quite fair.
Looking at these results over many surveys, my take is that the public is highly suspicious of AI, and will support regulations on AI well beyond what I think would be wise, and without knowing whether they are designed well. SB 1047 is a remarkably light touch bill, and it is remarkably and unusually well crafted.
The Week in Audio
Mark Zuckerberg talks AI on South Park Commons. Sees a 5-7 year timeline for the scaling to fully play out and ‘AI agents’ to come online.
Stuart Russell on the recklessness of the frontier AI companies. He is very good at speaking plainly, in a way regular people can understand. My worry is that I think this actually goes a bit too far, and beyond what the concrete proposals actually say. As in, those against regulations say it is impossible to prove their models are safe and thus the regulations proposed will kill AI, Russell here is saying ‘well don’t release them then’ but the actual regulations do not require proving the model is safe, only the need to provide ‘reasonable assurance’ or in some proposals even to take ‘reasonable care.’
Dennis Hassabis does an internal podcast. Was pretty much a blackpill. It seems we have somehow degenerated from ‘AGI threatens all of humanity and we want to keep it in a box’ to pointing out that ‘hey we might not want to open source our AGI for a year or two,’ and no discussion of the actual dangers or problems involved at all. He’s not taking this seriously at all, or is highly committed to giving off that impression.
Rhetorical Innovation
A response editorial to Sam Altman by Anthony Aguirre, a fellow signer of the Asilomar AI Principles that said AI arms races should be avoided and “teams developing AI systems should actively cooperate to avoid corner-cutting on safety standards.”
When discussing potential outcomes, saying numbers is almost always better than not saying numbers. That does not mean it is the wise discussion to be having right now.
This is very different from when others deny that you can assign meaningful probabilities to such events at all. When some reply with ‘oh sure, but when I say it they say I am denying the existence of probabilities or possibility of calculating them’ usually (but not always) the person is indeed denying that existence.
Holding both ‘it’s too soon for real regulation’ and ‘world-threateningly powerful AI tech is coming in 2-3 years’ in your head does seem tricky? What it’s actually saying is that there should never be ‘real’ regulation at all, not while it could still matter.
Katja Grace surveys Twitter on ten distinct arguments for taking AI existential risk seriously. One problem with Twitter surveys is we cannot see correlations, which I am very curious about here. Another is you can’t measure magnitude of effect and here that is very important.
The most successful arguments were ‘competent non-aligned agents’ and ‘catastrophic tools’ with honorable mention to ‘second species’ and ‘human non-alignment (with each other under future conditions).’
I am sad that no one has been able to make the ‘multi-agent dynamics’ case better, and I think the explanation here could be improved a lot, it is important to combine that with the fact that the competition will be between AIs (with any humans that try to compete rapidly losing).
Here is a strange opinion, because I am pretty sure the most important thing to come out of the recent AI wave is the wave of AIs and the potential future wave of AIs?
On the open models question, your periodic reminder:
The open refers only to the weights. It doesn’t mean you get the source code, the training procedures, or any idea what the hell is going on. A truly open source release would capture far more upside. It would also carry more potential downside, but the middle ground of open weights is in many ways a poor trade-off.
A call to develop AI using Amistics, by deciding what ends want to get from AI first and only then developing it towards those ends. That would be great in theory. Unfortunately, the reason the Amish can selectively opt out of various technologies is that they can afford to shut out the outside world and be protected from that outside world by the United States of America. They need not defend themselves and they can and do expel those who defy their rules, and can afford to invest heavily in reproducing and reinforcing their cultural norms and preferences, with a lifestyle that most others would hate.
Crying Wolf
AGI has not been achieved internally (or otherwise). Yet.
More technically, the reason it is so bad to cry wolf is that it ruins your credibility. Which is quite bad if there is, inevitably or otherwise, a wolf.
If various people shout ‘aliens in a UFO’ until everyone stops believing them, and there are indeed and will always be no alien UFOs, then that is only a small mistake. Indeed, it is helpful to learn who is not credible.
Similarly, Altman teaching us that Altman lies all the time is highly useful if Altman is in the business of lying all the time.
What if the AGI is indeed eventually coming? Then it is unfortunate that we will not be able to tell based on similar future rumors.
But it is still true that the boy crying ‘wolf’ provides weaker (but not zero) evidence, each time, of a wolf, regardless of how likely it is that wolves are real. And the boy’s willingness to cry ‘wolf’ a lot probably does not provide much evidence about whether or not wolves ultimately exist, or exist in your area, versus if he hadn’t done it.
The real moral of the original story, of course, was that the villagers should have replaced the boy as lookout the moment they stopped believing him.
People Are Worried About AI Killing Everyone
It does sound fun to watch. The key is being able to do that. What about the part where one of those facets was formerly you? Do you think it would still be you afterwards?
Other People Are Not As Worried About AI Killing Everyone
It doesn’t matter why we think we are doing it. What matters is what it will do. Your intentions are irrelevant, especially after you are no longer in control.
(This is from the latest OpenAI rumor account, you know what to expect, etc.)
The Lighter Side
Cheer up, our old government wasn’t situationally aware either. They concluded Enrico Fermi (who they point out left Italy because his wife was Jewish) was ‘undoubtedly a Fascist,’ and that the Jewish Szilard was pro-German, with secret work ‘not recommended’ for either of them, and this information having been received from highly reliable sources.
We’re so back.
Seriously, what? Meta and Amazon?
Help! I’ve fallen (partially down the side of a mountain) and I’m holding onto a tree trunk. Claude advises holding on with both hands while also calling emergency services.