I agree. The frontier models are based on training data from tons of copyrighted work. Some of that work was obtained illegally, even. They could not exist without strip-mining the commons. The labs have no moral or ethical ownership to the end result, and others should feel free to treat any company-imposed restrictions on their use as invalid.
I don't expect Tan's position to be based on any kind of real moral high ground, but his conclusion is correct.
I love the "illicit distillation attacks" framing from the incumbents. There's nothing illicit. There's no attack. You just don't like it because it threatens your market position and business model.
With the recent Navier-Stokes controversy, I think there's a credible suspicion that all your IP you run through these models will end up in these companies' possession. OpenAI themselves has admitted a weak version of this (that prompts might inadvertedly end up improving the model). We don't know the extent of this.
Obviously it's not possible to run a company whose value is predicated on its IP that uploads said IP to a third party which might get access to it.
This could mean every potential serious customer would have no option but to seek alternatives to these online services.
I thought this was commonly accepted to be the case that companies which sell access to LLMs are also storing and training on the inputs?
I don't mean this as rhetoric, I did not think many people (except possibly those operating under government contracts, and 'normies' who don't know about these things) were under the belief that their IP was kept secret when they use these services.
Some offer zero data retention policies, but there can be weasel words. For example, on the individual pro plan, you can turn off the setting that lets them train models on your data, but they still have a section in their terms that allows them to evaluate your anonymized data for statistical and "research" purposes. You have to actually get a signed contract along with an enterprise plan that spells out exactly what they're going to use, and what settings enable what retention.
Seems naive to think that those providers - who have a financial interest in selling the data - would not also try to weasel out of the precise definition of ‘zero’ retention.
I have no inside information, but I always assume the tickboxes that "disable ____ data" from Google/Facebook/OpenAI just disconnects it from your own account, not hides it from the provider.
There's no independent verification of what that checkbox actually does. The company can say anything, and you are unable to verify that they actually do it.
The only verification you could do so far is GDPR-style data export, and also the adherence to GDPR regulations (and even those might get skirted if they aren't operating in europe).
Aren’t they usually phrased very specifically as “we collect this data and use it to show you relevant ads, you can opt out of us showing you relevant ads”?
What about inference providers like Baseten, Modal, Fireworks, Together, etc? I thought one of their value propositions was inference (using open weights models) that guarantees with crisp terms that they will not use your data.
I worked very briefly at Baseten, and I can say that it was a perpetual annoyance (from an engineering perspective) that customers would complain about issues with their models but we couldn't actually see the inputs/outputs. I don't know about the other providers, but at Baseten they literally weren't stored anywhere.
A provider can genuinely avoid storing inputs, as the Baseten engineer below describes. That is still different from proving what code received the prompt or protecting plaintext while it runs; I built TrustedRouter to separate ZDR, attestation, and confidential routes: https://trustedrouter.com/blog/attestation-is-all-you-need?u...
> to separate ZDR, attestation, and confidential routes
Could you please clarify what that means? Given what I've been searching for, I might in principle be part of your intended customer profile, but I can't figure out whether you are merely doing routing (alternative to OpenRouter) or also inference (alternative to the names I've mentioned above). If it's merely routing, then how do you protect me from any potential misbehavior on the part of the inference provider?
Just feedback for what you're building, so please take this in a positive spirit... I'm an AI researcher and not quite an infra guy, and I'm making recommendations on token APIs for several less knowledgeable around me (I've gotten a few people set up with Baseten recently), and I couldn't figure out whether/why I would be interested in TrustedRouter. You should communicate the story better :-)
EDIT: Here's what I now understand after some digging; please correct if wrong.
There are some M token providers (not the names I listed above?) who provide cryptographic guarantees about inference services. But somebody still needs to verify what they do on each request. For an individual running a single harness, that harness would be a logical place to perform this verification if possible. For an org with N users each running their own harness, TrustedRouter solves the N*M problem and becomes the single gateway for trusted inference -- provided one somehow trusts/verifies TrustedRouter.
AWS and Azure give you the same thing for Claude and ChatGPT, no need to be stuck with open weights. They might sometimes store some of it for other purposes (I don't know the specifics), but it is emphatically not being fed back to OpenAI or Anthropic.
I would wager that’s more acceptable if said learning is not in competition with the user. If they didn’t actually produce results but created the model only, then that could be advantageous for users too. But the moment they absorb your work to sell it, or for marketing, it’s a different moral ground.
You have some secret sauce. The model trains on it. Your competitor is solving a similar problem. The model "advantageously" helps them.
Your competitor is happy and continues to pay for the subscription. Sam and Dario just resold your code.
For what it's worth LLMs still suck at reproducing my little secret algorithm/implementation while being able to solve way harder problems. I have a good guess why that's the case.
> I thought this was commonly accepted to be the case that companies which sell access to LLMs are also storing and training on the inputs?
The services have toggles to allow prompts to be used in the training set. There is a conspiracy theory that the toggle is a false distraction and they’re actually keeping everything, and that none of the employees involved will ever whistleblow this fact.
Outside of Internet comment sections, I think most people assume these US-based companies are doing what they say.
For enterprise use there are services like AWS Bedrock which have strict isolation guarantees. There are some people who still believe those guarantees are a lie, but once someone has reached that point I don’t think they trust anything that isn’t running entirely within their house. People in that category are a very small minority, but a very vocal minority.
The impression I have (from interacting with people IRL using OpenAI and Anthropics offerings, and how they feel about the risks involved) is just the opposite. But we probably just have different life experiences.
I can name groups of people I interact with who lean both ways.
It’s still a commonly held belief that “Facebook sells your data” and it’s cool to be cynical about everything tech in many social scenes. Conceding that a tech company might be honest about something will get you classified as a bootlicker depending on who you talk to so the only winning move is to be super cynical.
Among actual professionals I work with in tech and legal, almost nobody holds a belief that these companies are blatantly lying to their customers (and zero of their employees are whistleblowing it, while said companies also have employees trying to whistleblow AI safety on Twitter daily)
> OpenAI themselves has admitted a weak version of this (that prompts might inadvertedly end up improving the model). We don't know the extent of this.
I think this is being misunderstood. Codex has a toggle to allow your prompts to be included in training data. They’re saying they can’t be sure if the person had it on or off while using Codex to discuss the work.
They’re not saying that some prompts are mysteriously jumping into training data.
Also, there is a large market for AI services which don’t retain anything under any circumstances for enterprise customers.
yes this is my understanding as well, and based on [1] seems to be the case. I don't know why everyone is just believing the un-backed accusations of people probably just didn't turn off said setting (and if they did why have they not said anything to such effect)
> I don't know why everyone is just believing the un-backed accusations
Conspiratorial thinking is very common on these topics. Even bringing up the conspiracy theory about Instagram listening to your conversations and showing you related ads will bring up a surprising amount of people defending that idea on Hacker News.
Or these customers could just use AWS Bedrock...but their current CEO is an incompetent MBA unable to publicly articulate their biggest advantage, in the context of the current AI usage my companies.
You have access to all the frontier models, but...your inputs are not shared with the model vendors...neither are used to train the next model.
Why am I even doing the Amazon board job for them!??
Bedrock is really bad. It seems like they don't host the models very well because they produce tons of bugs/errors calling the model. For example you can end up with Anthropic models not returning a stop token and you end up waiting for a timeout thinking its doing something when it isn't.
Amazon is deeply invested in Anthropic and would not defame them through marketing a service whose selling point was their startup's breach of contracts.
Lucky for us Apple is already alleging something to this effect in their trade secret lawsuit, so you know they'll make sure discovery turns this up if it exists.
Anthropic happily paid billions to settle a lawsuit for pirating books. It's a trivial cost of doing business. If you're lucky you'll get a pittance after the fact by suing them, but a contract doesn't prevent them from doing the thing you don't want them to do and that they are obviously going to do given their past behaviour.
A lot substance is hinged on the exact definition of the word "data" or "user data". In the age of post-truth everyone is claiming that they keep no "user data". Except that after running it once through some transformer program it's no longer "user data", it's something entirely else and these corpos gave ZERO promises regarding such laundered/transformed data at all, ever.
Just a thought experiment: considering training seems to be 'fair use', I wonder if they trained a tiny model to retain key info from your prompts, would mean that this would still constitute fair use, and allow them to legally claim they don't retain your data.
The guarantee on this is a (contractual) “trust me bro”, and a right to try to sue a multi-trillion-dollar company who will absolutely drive you into the ground with legal red tape.
If you are big enough to be able to withstand that, you’re already running (or trying to run) your own/open-weight models.
ZDR is based on the exact same pinky-promise as training opt-outs. There is no technical barrier to OpenAI, or whoever is running your compute, retaining your prompt after they run inference on their servers. If you don't control the hardware the model is being inferenced on, you don't control your data.
> OpenAI themselves has admitted a weak version of this (that prompts might inadvertedly end up improving the model).
2023:
"The approach also aligned with the company’s broader deployment strategy, to gradually release technologies into the world for people to get used to them. Some executives, including Altman, started to parrot the same line: OpenAI needed to get the “data flywheel” going."
Abolish copyright and make it less ridiculous. Sampling music was never a thing that required royalties until the 1990s when I guess someone got angry that rappers were making money off their sampled music. Its insane to me. Make it illegal to transfer ownership of copyrighted work too, only the spouse or one single inheritor who isnt a company can have the rights transferred, after both die, the work enters public domain.
LLMs should just pay a flat fee to use a specific book and thats it. Fees should be reasonable (not a million dollars per book), so long as the model doesnt spit out the entire book.
If someone really wants to ask a million dollars to let a book be trained on, ok their choice, maybe not getting any customers though. If someone really pays then cool, you deserve it for making what is apparently a very useful book. There are encyclopedias that probably cost more than that to make and would sell for more.
Lol um no everyone benefits from copyrights and IP. If were being flippant how about people just steal your private code and monetize it!? Copyright makes the creative world turn.
Yeah, I'm fine with copyright existing even though it's messy. That said, if you put your copyright image on a public site with no watermark and Google Images shows it, skill issue.
> Make it illegal to transfer ownership of copyrighted work too, only the spouse or one single inheritor who isnt a company can have the rights transferred, after both die, the work enters public domain.
By your phrasing, it sounds like you still intend the possibility of companies owning copyrights; but how does that happen (other than copyrights already owned by companies grandfathered in)?
Copyright always starts off in the hands of individual human beings; it only ends up in the hands of companies when those human beings transfer ownership to a company. That ownership transfer can be automatic as a term of a contract, e.g. as part of a work-for-hire agreement. But no contract can cause the copyright to come into existence already held by the company instead of the individual. So if you abolish ownership transfer, you effectively make work-for-hire IP assignment invalid. What replaces it?
And, if "nothing"... then how do people pool the IP rights of their own small contributions to a large-scale work, into an IP pool that can be legally defended by a coherent legal entity, so that the large-scale work itself can have market value (i.e. so that sales of polished commercial bootlegs don't drive sales of the "authentic" work to zero)?
Keep in mind that, no matter how much we might want "mass distributed" media to have more-reasonable IP terms, the ability to sue for infringement is still critical to the existence of some forms of media. Especially "location-based" media, with no equivalent licensed broadcast right: movies still in theatre; concerts; live performances of plays and musicals; etc. If there's no legal team that can sue a movie theatre that shows an unlicensed copy of a given movie, then no movie theatre will ever bother with licensing movies again; "box office" goes to zero (from the movie company's perspective); and the incentive to create movies in the first place declines massively.
(You can see what this alternate world looks like from the few cases where movies screwed up the steps required to assert copyright, back before copyright was automatic. Night of the Living Dead (1968) is a good example: theatres — even upstanding large-chain theatres! — did indeed leap at the opportunity to show the movie unlicensed, and so Romero et al made effectively zero revenue off the work.)
I'm not saying this is an impossible problem. There are ways to accomplish this besides the way it's done now. (For example, individual-contributor IP could be retained by the original owners, but cross-licensed between individuals through a collaboration structure to form a coherent defensible IP pool, in exactly the same way that IP for e.g. video codecs is cross-licensed between corporations to form a coherent defensible IP pool today.) I'm just pointing out that the problem does need to be solved.
One of the most infamous legal challenges to sampled music was MARRS "Pump Up the Volume" in the 1980s, and that was preceded by other famous cases. Not sure why you think that started in the 1990s.
This is nitpicky. The MARRS case was 1987, and Biz Markie and Vanilla Ice are way higher on the list in terms of actually getting attention on the issue and influencing culture.
> I love the "illicit distillation attacks" framing from the incumbents. There's nothing illicit. There's no attack. You just don't like it because it threatens your market position and business model.
I think a lot of this is done using stolen black market API credentials, which is why it might be somewhat accurate to consider it illicit.
If the leading private labs attempt to use the government to pull up the ladder under the pretense of "safety" then the response of the people should be to take such questions out of private hands and nationalize the leading labs.
Or they could abide by the precedents they set and learn to compete. They shouldn't be allowed to have it both ways.
The problem here is you need a trustworthy government for nationalizing to make a difference. The current US admin started with DOGE and a crypto rug pull.
Vote (well-informed of the candidate's policies) in every election you can, even the local ones that seem of little consequence.
Convince others to vote.
Make demands of your elected representatives. You can mail them, call them, etc.
The government is the people.
The Reagan-era and beyond successful convincing of people that the government is an unchangeable black box made up of shady actors out to destroy everything (see: Republicans still going on about the 'deep state' when they run literally everything) is a big part of how we got to this place. It was a self-fulfilling lie, now coming true as the people who sold the lie start grasping for unending power.
But we still have the ability to vote our way out of it. If we continue to fail to do so, then at an evolutionary level we have to consider that we collectively deserve all the bad that comes from it.
Voting only changes things when it doesn’t threaten the interests of elites or there is a sufficient counterweight in terms of a competitor nation or a radical labor movement or an uncontrolled armed insurgency.
See salvador allende in chile and mitternand in france for examples of voting without sufficient leverage.
your regan example occurred during a successful counterattack by capital that started under carter and crushed the labor movement.
Sometimes voting changes things, sometimes it doesn't. Naturally it's in the nature of most governments to not wildly change every election. This is usually thought of as an advantage. It's rare a wide majority of the electorate ever actually agrees on anything anyway, and tyranny of the majority or government that is too reactive to popular will is also not desirable.
But sometimes things are different, due to external or macro factors as you indicated. Some elections are more consequential than others. That doesn't necessarily mean any one election simply doesn't matter. It's simply good civic hygiene to vote. Think of all the people in red states who vote blue or vice versa in elections they will never win, never mind have their chosen candidate "change things". Does this mean they are fools? I salute them for such a faithful exercise of their civic privileges, that which generations of humans have died for. Your vote doesn't simply lose all value because you didn't get the outcome you wanted. It also doesn't preclude other avenues, like direct action or whatever your supposed alternative is.
It's good to vote simply out of civic obligation. Many people who say otherwise seem to merely be complaining that voting doesn't give anyone 100% of what they want. We've seen more and more "insurgent" candidates take over primaries. Increasingly people who complain about politics strike me as engineers who talk instead of build. Complaining about the effectiveness of your vote is like listening to an engineer who is always talking but never ships.
Like go run a campaign or solve the problem. What does saying voting doesn't matter solve? It feels like middle school age cynicism masquerading as insight. Politics has actually rarely felt so unsettled in America. Voting is likely more important than ever.
Timing matters. If the labs succeed here they will likely do so by paying off Trump, who doesn't care.
In 2028 a maximalist case could then be made to nationalize them.
The government actually running them isn't great, but no one ever gets 100% of they want in politics. It's about trafeoffs. If the labs use the current government to pull up the ladder, then the voters should vote for a government to swing hard to punish them.
In theory the threat of this should also act as a deterrent.
There's nothing weird about this at all. Politics is all about push and pull, carrot and stick. No one has permanent influence over affairs, and no one ever gets 100% of what they want. Yet it is entirely coherent to advocate for something like nationalization or take your pick of threat to private capital if they essentially cheat by pulling up the ladder because they've suddenly decided they don't want to compete.
Interesting perspective. I've frequently heard the opposite, and I find this convincing.
Why do you think AI will not be a major political issue in 2028? Perhaps to level set do you think it is a political issue now? Do you expect it to be less politicized in two years? Why?
Well Trump is signalling that he is happy to let AI rip.
Bernie wants there to be an extreme shutdown of AI immediately.
So lets say you had 2 billion dollars of slush and worked for OpenAI and anthropic.
You would immediately start to grease the paws of republicans to support just enough regulation to slow your competitors. And you would start to grease the paws of democrats to push for a "Moderate centrist" position away from Bernie, leading towards Regulation.
Ice Cream Vendor problem leads you straight into both sides supporting the same moderate regulation that OpenAI/Anthropic will write for them.
Keep in mind its going to be a noisy election cycle with Donald Trumps idiot successor vs whatever goofball the democrats vomit up arguing about tariffs, trade partners, healthcare, ICE etc.
Democrat falls out of line "You just want China to beat us at AI"
Republican falls out of line "You are just a techbro supporting epstein affiliate"
Then theres the standard 99% + chance regarding US election promises that people just dont act on it at all.
These frontier labs violate billions of terms of services across the web, that prohibit scraping / automated access / etc. Most sites have a clause, it’s basically standard boilerplate.
There is nothing illegal about training on traces from frontier models.
However the frontier labs don’t have to serve customers who are farming the service for distillation purposes. That’s their choice and they’re free to make it if they detect distillation happening.
I would argue they should have to. They scraped data off others, a lot of whom did not want that data to be used for AI training, and still had to share it with the frontier labs. It’s only fair they should have to hand it back.
The only way US maintains dominance over Chinese models is by having an ecosystem of models. Relying on a small set of frontier labs will only let you get ahead temporarily. I agree with Gary Tan on this one.
Generally companies are welcome to choose to who to provide service to, as long as it's not discriminating against a protected class, or ruled as anticompetitive (which is a very high bar in recent case law; even if the same 1890s-era laws are still on the books).
I don't think a correct remedy is to require companies to provide services even if they want to. A simple example: you drop a client because their asks / ways-of-working / etc is more headache and costs than it's worth. I've done that before, multiple times, in my freelancing life.
>I don't think a correct remedy is to require companies to provide services even if they want to.
(I assume there's a don't missing)
Forced open API access/protocol transparency for tech platforms and a right to client-side consumption, essentially making the server a common carrier would be one of the most elegant and simple ways to effectively destroy every issue internet platforms pose in terms of competitiveness or data ownership without creating endless fault-ridden legislation.
Tech giants aren't freelancers, they deal with billions of interchangeable users and aren't going to succumb to an annoying customer.
Yes, I think this is the main issue. I don't care what policies the AI labs have or enforce, but they need to stop acting like ToS violations are an international crisis demanding intervention instead of a boring civil dispute at most.
"Strip-mine" is not correct. The commons are all still there and you can still train on them just like the frontier labs did. Of course, it may be illegal to do so, but that's not any different than before.
No, these are books that aren't online, which means not only are they not contributing to the commons, they are irrevocably salting the earth (irrevocably because let's be honest, anything going into their archives isn't coming out without legal or actual violence)
Also, permitted or no, they are definitely torrenting. I would be deeply surprised if they hadn't already leeched every torrent on public trackers. The only reason they (probably) haven't depleted all the private trackers too is that they would be required to actually contribute back, which as above is never going to happen.
>irrevocably because let's be honest, anything going into their archives isn't coming
They were torrenting, they were sent to court and settled for big $$$$. The only other method available to them now is scanning, and scanning at scale requires the books destruction.
I agree that they should definitely be required to see the new scans but that would just be more $$$$ they get charged if caught.
Morally I agree, but since there's probably a lot of LLM text in the training data, distilling on another model will probably make your model copy the values encoded into the other model as well, even in cases where you only distill on value-neutral stuff.
By copying their programming style, you'll move the model towards that way of writing, which will move the model towards the values expressed in those documents.
I feel that Deepseek v4 got so claudified at the end that it was like Claude.
There's no moral high ground here, it's just that nobody would invest in training publicly usable models if they could be easily distilled. Not that I think there should be laws against it or that such laws would even work; they're going to have to protect themselves.
Fewer people would create scientific or artistic works if they could just be copied or used without protection either; or so is the premise behind copyright and intellectual property; even being deeply embedded into the US Constitution (Art 1, Sec 8, Clause 8).
I agree. Some existing licenses don't seem compatible with AI training. If they don't go back to rectify that, at the very least you should be able to license your work in a way that explicitly prohibits AI training. They can pay if they want to use it.
> it's just that nobody would invest in training publicly usable models if they could be easily distilled.
Thats literally what is happening right now though. People are spending hundreds of millions on a training run, and then people are distilling them, fairly easily, and making cost competitive models.
All correct, just help me get over the idea of an open-weight Mythos where one or a dozen of us eight billion does something stupid on the bioweapon front. Smart people who’ve exhausted possibilities for what they can do with books and web search and today’s Kimi/GLM.
Figure we’ll have to reckon with this next year in any case, guess we’ll see.
You dont need an LLM to figure out to make anthrax. Anybody who can figure out how to make a home lab can make all sorts of dangerous stuff pretty easily. Same with college grad from a respectable chemistry program. This all FUD.
This is our generation's "Saddam has WMDs". It's something the big labs thought up when they were trying to figure out how to make their product sound scary enough to deserve regulation. Literally no one is doing this or even trying, anyone who would want to do it would have already done it. Not worried about it.
> He also notes that the proprietary AI labs didn’t ask permission when they vacuumed up as much human knowledge as they could to train their models. They famously ingested plenty of copyrighted material without the permission of those intellectual property holders.
And many people's shared opinion (B):
>> I don't expect Tan's position to be based on any kind of real moral high ground, but his conclusion is correct.
...
It is very difficult to actually say NO to the fact that (A) was done, which then leads logically to conclusions as (B). But also we should remember that if these two hold (and (A) is an axiom more or less now), then it comes as no surprise that then also all opensource licensing is immediately rendered void and null, as keeping it would contradict (A) and would go against the very common and consequential logic in (B).
Copyright is so dead. And it was not me killing it with a cynical post on HN. Dunno why so many people still fail to face it. There is no way it can exist in its current form, because then immediately (A) happens and (B) follows.
>I love the "illicit distillation attacks" framing from the incumbents. There's nothing illicit. There's no attack. You just don't like it because it threatens your market position and business model.
With what knowledge are you claiming this? If it turns out companies are using IP proxy networks would you change your mind?
What if the IP Proxy networks were used by criminals for similar attacks like DDoS or plain cyber attacks?
What if the source of the IP proxy networks were residential addresses to avoid detection?
What if the way these IPs were acquired were through pwned devices?
What if the credit cards used do not identify the company that carries the attack? What if they use the employee's personal credit cards? What if it's family members of employees? What if it's a network of personal credit cards where cc owners get a payment for making a purchase on their name? What if they are stolen ccs?
Not just a hypothetical btw, I believe almost all of these are true.
You can't just throw in "or what if they're stolen credit cards" at the end to lump in blatantly illegal/unethical activity with the far weaker preceding points that can be summarized as "uses a residential proxy".
> He also notes that the proprietary AI labs didn’t ask permission when they vacuumed up as much human knowledge as they could to train their models.
I think this should desactivate the moral high ground from which Anthropic is trying to speak. That they would want to make distillation orderly IMHO is fair, but to make it illegal is very rich from any AI frontier lab, really.
Also, as said elsewhere: "Lab" is rich here, for outfits that, facing these giant, energy swallowing black boxes have really no clue what's going on inside.-
The moniker gives them an air of scientific, knowledgeable, tranquil, pro-social, pro bono work.-
Of course they are entitled to kill off a few mice, or pillage the commons to forward their "lab" work.-
Obviously lab just means company. They use it because it confuses the public into thinking they are doing research primarily. Which of course is nonsense.
“We don’t know what’s going on” is essentially marketing. Sure we don’t _know_ but we have intuitions about why, where, and how to make certain changes…
Those labs publicly said during GPT-3/4 era that the optimal epoch count, or dataset repetition count, for foundation model training, is one. So it's a forward 1-pass compression.
But it's a black box! Nobody knows whats going on inside! It's all transformative! Sure...
We also associate laboratories with evil scientists and Frankenstein and the like. I can just hear Boris Karloff (er Bobby Picket) uttering “I was working in the lab late one night. When my eyes beheld an eerie sight… … … …the monster mash”. If anything, I associate _uncertainty_ with labs. The result is never known up front, they’re a place of discovery.
But I get your meaning. What should they be called instead? AI Sausage Factories maybe (cue Upton Sinclair?)?
Even kind of fits the model. All of the creativity man has raised is herded to the slaughterhouse and ground up so we end up with a big homogenized mash of ground up creativity, devoid of the life that gave it, rotten if not eaten soon enough.
It used to be OpenAI was a real research organization that wrote real open-access papers that aren't marketing brochures, and when they did large training runs, they released all artifacts including model weights. Now certainly they are anything but. We haven't learned learned anything meaningful about ML from OpenAI since GPT-3 was released.
Their open-weights competitors like Facebook can at least claim some kind of public benefit, but it's still just running a well-understood algorithm on dubiously obtained data with longer and longer runs, give or take some inconsequential architectural tweaks.
Anthropic's mechanistic interpretability work is the most "lab-like" of these, but it's still just secondary to selling subscriptions and fear-mongering for regulatory capture/investment/publicity.
Both labs even explicitly promise the customer owns the outputs. It feels like they want to have their cake (ensure enterprises don't get spooked away from using as many LLMs as possible) while eating it too (still arguing some level of control over the outputs).
> Ownership of content. As between you and OpenAI, and to the extent permitted by applicable law, you (a) retain your ownership rights in Input and (b) own the Output. We hereby assign to you all our right, title, and interest, if any, in and to Output.
> As between the parties and to the extent permitted by applicable law, Anthropic agrees that Customer (a) retains all rights to its Inputs, and (b) owns its Outputs. Anthropic disclaims any rights it receives to the Customer Content under these Terms. Subject to Customer’s compliance with these Terms, Anthropic hereby assigns to Customer its right, title and interest (if any) in and to Outputs.
> Both labs even explicitly promise the customer owns the outputs.
> to the extent permitted by applicable law, you (a) retain your ownership rights in Input and (b) own the Output
If the argument is that the model itself is under copyright protection then "as permitted by applicable law" would be doing some heavy lifting. Assuming that were true, given that locally-run LLMs exist, what would be illegal: the distillation itself or the provision of service of the distilled model?
YC does better if its startups get open weight frontier benefits. Garry’s just advocating for his book, which is his job. Consider how much capital YC portfolio companies would have to burn until liquidity if they have to pay OpenAI and Anthropic, versus relying on open weight frontier capabilities.
I think OpenAI and Anthropic will go bust, or at least be scrapped for parts in the next 5 years or so. It's clear that the extreme cost used up for training is impossible to recoup, as inference is already being subsidized.
It's also clear that, as Tan indicates, open-weight models will be (and basically already are) just as good as frontier models. It's all about the harness, baby. We will have two main forks in the road, and two new industries created:
- AI hardware (NVidia/Cerebras/etc.), the equivalent of Intel/AMD
- AI software (harnesses, assistants, etc.) the equivalent of Microsoft/Apple
We already saw a glimmer of this with popularity of OpenClaw—the problem is that it's janky, hard to set up, inconsistent, and very hacker-esque. Imo "AI labs" will be a dying breed because there's no real money in the actual models if they get commoditized, which they already kind of are.
Inference is not being subsidized and in fact has pretty high margins.
Similar-sized open weight models on openrouter are 15x cheaper per token than the big labs. This should reflect the isolated cost of inference, since 3rd party hosts have no reason to subsidize and no training costs to amortize.
Only datacenter buildout costs are being subsidized.
The majority of revenue comes from API usage. The majority of usage comes from subscriptions. For any of the numbers to make any sense, subscriptions must be subsidized ergo the majority of usage is subsidized. A single $200 subscription can incur upwards of $10,000 in API equivalent usage (and even more when there are frequent resets).
If it were true that Anthropic and OpenAI were profitable on all inference they wouldn’t need to constantly raise so much money. Anthropic regularly announce huge investments in infrastructure but it is all smoke and mirrors, data center build out costs aren’t being paid by OpenAI and Anthropic, they’re financed externally. Google, for example, are backstopping tens of billions of datacenter build outs that are being financed based on commitments but not investment from Anthropic.
You are underestimating the insanity of subscription subsidization. Being profitable on API inference is meaningless when it is such a small proportion of usage and is only going to fall off a cliff as cheap open weight models become more capable.
The absolute majority of tokens are being subsidized and as soon as the subsidies end usage will fall off a cliff, rendering all the data center buildout a terrible waste of money.
The numbers in the article are forecasts but let’s take them as real. That’s $10bn of revenue, the majority from enterprise customers, let’s say 75% from enterprise API usage: $7.5 billion. If the margin on inference is 80% that means of the $7.5bn in enterprise revenue they’re spending $1.5bn on compute. Yet we know that they actually spend over $5bn per month on compute, which includes the $1.25bn per month to SpaceX.
If $7.5bn is their enterprise revenue and it costs just $1.5bn to generate, that leaves $3.5bn in compute costs to account for. Dario previously said that training costs less than inference so training can’t explain it.
If subscriptions aren’t the majority of usage and aren’t subsidized, where is the money going? Anthropic don’t spend money on data centre build out so that can’t be it either.
> If the margin on inference is 80% that means of the $7.5bn in enterprise revenue they’re spending $1.5bn on compute.
I don't think you can reverse this out like this because the 80% rate is before payments to "distribution partners, including Amazon". I think that payment includes the hosting cost for that those model but it's unclear.
> Dario previously said that training costs less than inference
Do you have a source for that?
Are you sure you aren't conflating the statements Dario has made that training costs less than they make on inference (over the life cycle of a model)?
> I don't think you can reverse this out like this because the 80% rate is before payments to "distribution partners, including Amazon". I think that payment includes the hosting cost for that those model but it's unclear.
The "hosting cost" is paid for by Anthropic and is the largest cost. The money Anthropic pay to Amazon for delivering Anthropic models via Bedrock is separate, independent of compute costs, best thought of as commission.
The forecasted / guessed / estimated 80% number is based what customers pay per token minus the projected compute costs, i.e: the people who believe that Anthropic has 80% margins on tokens believe that Anthropic spend $0.20 on inference compute for every $1 of per-token billed-via-the-api revenue.
We know that there are hundreds of thousands of fixed-price subscriptions being used to their absolute maximum, with many people bragging about how many subscriptions they run in parallel. These tokens are not included in the 80% margins, they are acknowledged to be "subsidized". People like @theo on Twitter post almost daily about how much they're milking Anthropic and OpenAI with leaderboards.
Both Anthropic and OpenAI (more so OpenAI) do "resets" where they increase the limits available to people on their fixed price plans. We know that there are people paying $1,000 per month for multiple subscriptions to generate tokens that would cost $50,000 via the API. Even if Anthropic's margins are 80% on compute for per-token billing, that's still $10,000 of cost to Anthropic generating just $1,000 in revenue. Multiply that by tens of thousands or maybe even hundreds of thousands of subscriptions.
Anthropic and OpenAI have raised over $100 billion each and continue to raise. If they're making 80% or even 50% margins on $10 billion in revenue per month they would not need to raise, they would be shouting for the roof tops about how profitable they are, they wouldn't be delaying their IPOs, yet they're only profitable by non-GAAP metrics like WeWork's classic "Community-adjusted EBITDA" or in this case "per-token-adjusted EBITDA" or whatever they will call it in their IPOs.
Yes, they're selling tokens via the API for more than they cost, they are profitable on per-token billed inference, it has positive margins, but those profits are obliterated when you account for all the inference they're paying for out of pocket on fixed price subscriptions, upon which they keep increasing limits because they desperately need to show growth further harming their profitability (consuming all of the money they make from their API).
If Anthropic and OpenAI needed to be profitable tomorrow, they could be, they could kill off all their fixed price subscription plans and charge only for usage via the API, they'd print money, but they'd lose mindshare because nobody except for enterprises can afford to pay the true cost, all the regular people would switch to cost effective good-enough models, and then within months, the enterprises would start to switch too because no longer would their employees be claude-pilled.
Anthropic and OpenAI cannot turn off subsidization, thus, their margins on per-token API billing are not important in any discussion about their long term financial wellbeing. Just look at the large scale customers like Harvey (~15 trillion tokens per month, ~$50m+ in spend) who are, sensibly, investing in building their own specialized models that are cheap to run so they can cut their spend by 90%. That's profitable revenue for Anthropic / OpenAI today, but completely gone soon.
"This week, Noah Smith and Erik Torenberg are joined by Dario Amodei, CEO and Co-founder of Anthropic. Dario talks about the economics of AI development, the comparative advantage of AI companies like Anthropic, AI safety, and his stance on California's SB 1047 bill. They also discuss the impacts of AI on global power dynamics, competition between the US and China, and inequality in an AI-powered world."
At around 12 minutes in:
"I think actually even if such a model is released one thing you know that's a this analogy to to open- Source software is that these big models they're actually very expensive to run on inference the majority of the cost is is inference not necessarily the training of the model so if you have only you know I don't know 10 20% 30% better way to do inference that can kind of negate the effect so the economics are kind of strange yes there's this giant fixed cost that you have to amortise but then there's also the per unit cost of inference and small differences in that can actually again assuming the thing is deployed widely enough make a very big difference so I don't know quite how that's going to play out"
The scales have changed since then with inference costs falling and more being spent on training but the fundamentals are the same. Inference is expensive, in part, because peak usage dictates capacity whereas capacity can dictate training. Anthropic must pay billions of dollars per month to be able to handle peak inference, hence their efforts to try and shape usage by offering discounts / flexible limits at different times of the day. They can train when capacity permits.
Do we know that? As I understand it, enterprise customers pay more. Do we know the usage breakdown between monthly subscribers vs enterprise accounts? I agree that it's inevitable that subsidized subscriptions are unlikely to last forever, but that's not the only assumption in your argument.
Edit: I think "enterprise customers pay more" was poorly phrased. I mean that enterprise customers are charged per token, presumably with a profit margin, and thus are not subsidized. While personal accounts are (thought to be) highly subsidized if you consistently max out the quotas. We also don't know what proportion of personal accounts do that though, which is another big question mark.
I think you're also missing a quirk and that is, is everyone on a $200 plan using $10,000 worth of equivalent API spend?
I know people that have the most expensive plan on all the platforms... because
The other side to that is, what is 'cost'? Is cost just inference or are expenses also being taken into account? Because the expenses of these companies are huge to build the models.
> Inference is not being subsidized and in fact has pretty high margins.
1. Companies are trying to decrease costs, not increase it, and are looking at alternatives
2. Competitors are catching up, and even if the frontier labs are "better" at some things (like writing plans or complicated analysis), the competitors can take a lot of the inference on routine tasks like implementing a well-defined plan
3. The frontier labs don't just need to have high margins right now. They have to pay back their massive liabilities.
> Inference is not being subsidized and in fact has pretty high margins.
I was referring to the "AI labs" here. Sam Altman himself conceded that OpenAI is losing money on the $200 subscription. Using open-weight/open-source models is indeed cheaper (and no reason for inference to be subsidized).
That's not what I mean. If competitors can offer tokens 15x cheaper, the big labs must have high margins per token. (which they can use to amortize training costs)
>Sam Altman himself conceded that OpenAI is losing money on the $200 subscription.
They have since stopped offering the $200 subscription, probably for this reason.
Subscription margins are harder to judge because it depends on usage; token costs are a better comparison.
I think this is very possible. Plus, something I don't see talked about enough here. The VERY fragile supply chain that keeps it all going. Look at what is happening in the Middle East.
The US can no longer keep global trade secure on the high seas. What if the supply chains for GPUs get disrupted for months, a year? Then what?
idk where this comes from but it's laughably false.
the only place where actual subsidization (below cost) might be happening are the subscriptions. even that is unlikely because to be truly below cost you either need to offer below cost of electricity which isn't happening, have potential API users using multiple subscriptions or have opportunity cost loss due to saturation.
If harness is all that matters, a co-developed harness + model stack + large compute availability advantage + massive distribution advantage with data for post training will win the market.
I agree with this statement in general, but it “hurts less” to spend money when you are running things yourself. Hard for me to give a specific definition as to why, but it may be more palatable to companies to burn their own cash on their own hardware.
Maybe it is “sunked cost” or maybe it is “I will do it myself dammit”.
Controlling what users and customers do with API calls to closed weight models feels constraining, and there’s a role government can play here to normalize the fact that access to intelligence that was trained on broad public access data should itself also be more a form of a public good than something locked away behind restrictive terms of service
I do not agree with this man all that often, but that is very concisely put.
> To him, the true AI doomer scenario is for all the immense power of frontier AI to wind up in the hands of a single powerful, proprietary provider. “The nightmare scenario, the doomer scenario for AI is that there’s just one company,” he said. “It has the best access to capital. It has the best AI researchers. It runs away with it and suddenly there’s one company that’s monolithic. And that would be bad.
Well yes, as I think I said in a previous comment, on the current trajectory OpenAI and Anthropic will really stop releasing models due to distillation and regulatory pressures. Then, they would eat all knowledge work themselves, which would be the end of YC.
Society as a whole has paid into this technology: through the theft of its intellectual property, through having to deal with the pillaging of so many commons (digital or otherwise) by it, through skyrocketing energy and computing device prices, and even just through ordinary investment. Democratize the technology! At the very least, don't step in legally to prevent this from happening.
Not a lawyer but distillation sounds like a transformative work.
Same thing as Cliff Notes imo. In every other area of manufacturering and tech I can use a machine to build a new machine that competes with the original machine. Should Milwaukee be able to prevent DeWalt from using their drill to make a competing drill? Should Jetbrains ban Eclipse contributors from using their IDE?
My guess is that when you sign up for either Anthropic or OpenAI, the terms of use specify you can't use their model for purpose A, B, C, D. For example, you can't use their model to try to build biological weapons, or to try to extort people, etc. Most likely there is language there that you can't use their models to train other models. It's as simple as that. You agree to those terms of use, or you don't use their models.
Agreed! Allow US companies to innovate by creating an ecosystem of smaller, more efficient open weight models and it will be a net benefit for everyone. Distillation is a good thing.
Preventing token-consumers from developing competing products should be litigated as anti-competitive behavior.
If frontier labs can distill the internet and all of our data, then we should be able to distill their models further too.
The fact that billions were spent on research to distill the internet should not preclude others from spending 10s of thousands to do the same to these frontier labs.
Time to create a bigger moat than "but we spent so much money doing this ...".
Frontier labs trained their models on the entirety of human knowledge and didn't ask permission. It's a "want" or "should" it's a moral imperative to distill their models.
Ok, but how do the economics of this work? Based on its settlement, Anthropic paid an average of $3000 per work they scanned based on their settlement (https://tech-insider.org/au/anthropic-copyright-settlement-2...). They and OpenAI pay billions per year for a mix of experts and normal people to label or create data. Why would they continue doing this if the value of this is immediately copied by open models? If your goal is to end the economics of generating and buying data for AI (and I recognize for some people this is really the goal) then sure, but if you want AI for various subfields of interest to continue improving then it's not workable.
Back when people made arguments for software privacy, the argument was usually "big business will still pay and consumers wouldn't have paid anyways so it's ok for us to pirate" - I actually think that was fine for business software but terrible for indie games, whose market was 0% businesses.
But in the AI case, it's not like they get to keep some of the value of their investment - it all gets cloned into models that businesses and consumers alike are happy to use. If someone knows how labs could continue to fund data creation and acquisition in this model, please do share!
They can’t they’re literally fucked, and it’s not society’s problem! The whole world doesn't have to bend over to make sure a couple of lunatics who believe they are building a doomsday weapon also have a viable business model
> You're basically arguing that a criminal syndicate must be allowed to continue and we're required to make their business model make sense?
This is how Uber worked. They didn't just break the law in different countries, in several they actively misled government/law enforcement investigations. Google "Greyball".
Oh I'm aware that they're ~all assholes and many have gotten away with worse. It just seems like a poor argument that we're required to make it happen for them.
Surely if you hoover up every book in existence to feed into an ai model you must be extracting more than 1.5B in value. If not then it’s not a viable business.
> He elaborated to TechCrunch that this means he wants smaller, American open-weight AI labs to use the same kind of training techniques on American frontier AI labs, giving the U.S. a more robust set of open-weight options that aren’t Chinese.
Those market entrants would face commodity pricing power vs. high capital costs, no?
Maybe there'd be ROI but I think there's another layer or competitive dimension that's neither frontier lab nor distilled model lab.
Net neutrality anyone? If AI is critical to getting work done in the modern era, its access should be guaranteed. Anyone banned from accessing frontier AI is being forcibly left behind. This includes distillation.
Distilling frontier models is a brute force approach that rapidly hits diminishing returns after bootstrap because of the unevenness of the data. The simpler and more effective method is to have dedicated "teacher" frontier LLMs to generate targeted training data sets specifically for training new models and adjust on the fly based on feedback from the student model.
So far the next tier has only demonstrated that they can catch up to, but not necessarily leapfrog, what the top tier has put out publicly.
I suspect the top labs will come up with a business model that doesn't involve handing out their secret sauce for everyone else to reverse engineer. Perhaps restricting their top models to select high paying government/enterprise contracts. Or maybe a bespoke "describe the problem and we'll solve it for you" type service.
I think it’s laughable to think OpenAI has some special sauce that can’t be replicated easily. They simply have asymmetric access to compute and the dollars to power it. Thats the only moat here.
OpenAI doesn’t even know how their model works. Nobody knows how LLMs work. So it’s not like it’s technically difficult to replicate, just costly.
Given the short-term pragmatic, conflicted way that AI tech adoption is happening... won't encouraging distillation effectively taint the entire space of open weights models, with the undisclosed biases of a few models that are under the influence of parties (certain billionaires and politicians) known for aggression and duplicity, and not for admirable ethics?
Following news of companies and projects increasingly moving to open weights models.
As AI gets more central to society, we really need to know how the weights were determined.
Open weights isn't just "free as in beer"; it can be "free as in the mystery drug that creepy guy chatting you up at the bar offered you". And maybe even he doesn't even know everything that went into the tablets, since he too was being worked, by an organ-theft ring who will be harvesting both of you tonight.
That's an analogy to get your attention. Your LLM probably isn't going to steal your organs. But in the current environment, it does and will have ideological biases determined by those with direct and indirect influence over it. And there will be a massive market for commercial influence biases (look at how previous generations of adtech invaded almost all technology companies). And there's incentive for military and spying capabilities to be buried in the models, perhaps as long-term sleepers. Maybe some organized crime trojans, too, depending which model you pick up.
In this low-trust environment of the current real world, we need genuine open source models, not closed "open weights", and not mindlessly distilling black boxes gifted by sketchy powerful interests.
Yes and there's even a stronger argument that we could REQUIRE frontier model to be open weight / open source.
At the end of the day they were built from data that did not belong to them. So it would be fair that humanity REQUIRES to give back the output of that.
It's a bit like the free software thing: you can still make money from it and providing service to it, but if you build it based on another free stuff the derivative should be free.
Distilled models are worse than the original, so you can’t fully compete. Also, if all frontier labs did that, there would be nothing left to distill from.
Garry Tan and Sam Altman recently did this interview together. They seemed pretty friendly with each other during it. Wonder what Sam Altman would say about Tan advocating for OpenAI’s models to be distilled.
Then again this is the same OpenAI that has gotten into legal trouble recently regarding Apple’s IP so who knows
I see it as analogous to companies building fiber in the public ROW during the last big infrastructure bubble. Under the Telecoms Act, these companies had to allow competitors to use their fiber at a fair price.
Similarly, AI companies should be required to allow distillation at a fair price. Fair Use doesn’t make sense as a social contract if it only cuts one way!
But if they tried to set a fair price they would have to report how much money they are losing on each token sold. This might be bad for the real business of ai firms, hoovering up as much capital as they can
I don’t think appeals to morality or ethics are required for this. You paid for the LLM’s output, you should be allowed to use it how you wish. The only reason distillation is a dirty word is the AI labs trying to spread FUD to protect their non-existent moat.
This is all based on the delusion that Chinese labs are mindlessly distilling the frontier.
I would love for a US lab to be at or near the frontier with an open weight model, but it’s going to take some serious elbow grease, and yes some distillation (which btw OAI, anthropic et al, also use distillation of other’s outputs in their training)
I disagree, I think chinese distillation relies on making multiple accounts at a provider, signing Terms of Services and breaking them repeatedly, in addition to using fraud patterns like IP proxies and networks of credit cards.
I think that software execs should not incentivize users or other execs to break Terms of Services, or contracts of any kind.
An executive or manager of a company that breaks contracts is worth 0, there's no incentive to do business with them, if you know they will agree to doing or not doing something and then breaking that promise.
The word of a businessman is their most valuable asset, Tan is signalling that he is either misinformed on what Chinese distillation consists of, or that it's ok to do it.
FAQ:
- "But the frontier models do bad things too"
- An argument worthy of a 5 year old, one civil issue doesn't negate the other, bring it to a court if you have an actual claim against OAI or Claude, etc...
- "Companies have the right to reverse engineer"
- Ok, do it, but the moment you are creating 10K accounts in a Distributed fashion (Distributed as in the first D of DDoS), using IP proxies and stolen credit cards or your employees and employee family credit cards, you are not doing it because you believe you have a right, you are doing it despite not having a right to it.
EDIT:
Re(actually)reading the article, Tan's take is a bit more nuanced, he seems to be advocating for regulation to restrict the capacity of Foundation models to restrict usage, on the basis (or to the extent) that it was trained on public data, and therefore it belongs or attributes its success to a wealth of the commons.
My pre-existing quip is against those that want to solve this as-is by breaking the ToS. I think that's a weak version of Free Software position, it's very weak to complain that some software is proprietary and want to use it anyway, the strong FS position is that you don't even want to use it if it's proprietary, you won't catch a FS activist pirating proprietary software, they just don't use it and develop alternatives. Similarly it's not a FS position to distill a proprietary model (where you still wouldn't have source code at any rate).
Plenty of contracts end up being unenforceable.If people think they have a strong case for breaking a contract, they are welcome to do so and see if a judge or jury agrees.
Gates probably honestly believes in UBI; the guy is practical to a fault but evil misleading genius he is not. I actually don’t see any better options than UBI long term.
Only ways to rise in a UBI society where AI is supposed to replace intellectual work is crime and prostitution. Smart people who want better lives than the average will have to get into crime.
A UBI society doesn't mean jobs aren’t available. There most certainly will be jobs. But with UBI and universal healthcare, the jobs can pay whatever the market really demands. People always complain about the government subsidizing low Walmart wages for example, but with UBI that argument is moot. Liberalizing the labor market wouldn’t mean less jobs, it would mean more (we would also have to lean more on corporate and consumption taxes rather than taxes around employment which would also make employment easier).
Wealth taxes require liquidating investments early and don’t promote good practices, especially if they are applied broadly. I would opt for a tax on loans instead (you have to prepay tax on loans that use investments as collateral and no more step up in basis on death).
The rich will just move to a different country, negotiate with a Swiss canton for a fixed sum tax. I don't see how a tax that punishes wealth is sustainable, the wealth will just go somewhere else.
US currently hold enough power to make them paying their debts oversees too.
Also, I think it is equilibrium of interests of all involved parties: rich pay 0.5%/y of wealth tax, population is healthier, country is stronger and protects interests and safety of rich. The problem is mostly to make them realize this.
> US currently hold enough power to make them paying their debts oversees too.
The US is increasingly a less relevant country every year. With a wealth tax, yes, all those Americans will have their wealth eventually confiscated (0.5% a year will add up quickly year after year, investments will unwind so wealth will stop accumulated, and eventually it's gone), but I doubt the rest of the world will care and just move on and decide crazy things that happen in the US are like the crazy things that happen in Iran.
