Having to poke through an about: page and find the thing to toggle to turn off this specific AI feature but not that one, and knowing they'd change it on you a couple of releases later, never filled me with confidence.
One switch in a readily discoverable location in the settings menu and it's all gone? Just the job, killed it now on the laptop and the mobile both.
With all this rushing to improve AI engines and whatnot? It's the uptime/reliability that's concerning me as much as anything else.
I've seen ChatGPT Work have multiple outages just in the short time our office started using it. And even when they claim systems are running normally? You can easily get timeouts trying to complete relatively simple requests in systems like Co-Pilot, at least using the free edition.
If I ask my co-worker to complete a task, I have a pretty reasonable assumption he's going to do that. If he fails to complete requests several times in a short time-frame? Then it's likely he'll get in trouble for that, even potentially resulting in his termination. AI should be held to an even higher standard since it's a machine.
What happens when everyone is counting on AI to do necessary parts of daily workflows and then it has an extended outage? That's like your whole workforce going on strike.
If you have people working on a problem for decades, and there are rumors that they have made progress, what is the motivation to independently launch your millions of dollars of AI inference at that same problem?
In a normal academic context, none and moreover this sort of deliberate scooping would be considered incredibly bad behavior. When you have two giant corporations fighting to be able to have results they can showcase to the public and their investors? Then it makes a lot more sense, even as it is incredibly damaging to the academic math culture. This is really toxic behavior and deserves pushback. But it is a mistake to go from there to think that therefore any copying or plagiarism happened.
...and I know from years of experience in I.T. that any time you give people the ability to do a database search, you'll get people putting nonsense in fields for testing purposes. So at least SOME of the "button mashing" and entries like "dickhead" sound to me like officers trying out the system for the first time, or experimenting because they weren't trying to conduct a REAL license plate search.
That doesn't disturb me as much as the overall problem of allowing these searches without a warrant or at least a judicial review. Even if it's decided that a license plate search doesn't rise to the level of requiring a warrant to conduct it? There should absolutely be someone reviewing the searches being done. (EG. The officer who is too lazy to properly document why he/she is doing them and consistently leaving fields empty needs to be punished.)
I would like to clarify a few things:
1) The screenshot is my reaching out to Levent to coordinate our releases. I hope it’s clear from the message that we came in with the best possible intentions.
2) I never ever asked for Levent to be removed from authorship of his own work (as indicated by my text). I was surprised to learn during the call with Tristan that they had only solved Euler and not Navier-Stokes; after learning this we brainstormed possible paths forward. One option we discussed was that Tristan could be the lead author on a rewrite of OpenAI’s Navier-Stokes proof. It is in that context that I said “it would be simpler if Levent was not an Anthropic employee” because I felt it would be inappropriate for an Anthropic employee to author OpenAI’s work. Importantly it was admitted that internal Anthropic models had been used in their proof of Euler blowup; I therefore felt I could not consider Levent to be an independent academic. Another option I wanted to propose (but got cut short) is to offer access to our internal model so that they could try to finish their proof and bridge the gap between Euler and NS. Again I did not know how to navigate giving access to internal OpenAI IP to an Anthropic employee.
3) To reiterate it plainly: as my text clearly indicates, and as I said during our call, OpenAI's intention was to do everything possible to celebrate their mathematical achievements and the heroic efforts that they made on Euler. In the call I was immediately met with a litany of slander, including direct threats that if we were to announce Navier-Stokes he would immediately go to the press with a barrage of unfounded accusations. I refuted all these accusations but he replied “there is nothing you can do, I simply do not trust you”. I was confused why one would turn an incredible source for celebration (of their achievements!) into such bickering, which is when I said that I did not understand why one would risk their career [over unfounded accusations]. Genuinely, at that moment, I was trying to care for him and do a last ditch attempt to get a chance to give them all the credits that they deserve. I deeply apologize for this extremely poor choice of words, it is the opposite of what I was trying to convey. (I should say that I retracted them on the spot by the way.)
4) Overall, on a personal level, it was incredibly difficult to have these conversations. Levent refused to attend any of the meetings despite my repeated asking. As Sholto Douglas said, there will need to be coordination between Anthropic and OpenAI in the future; I felt I was doing a proxy negotiation with Anthropic while the Anthropic employee refused to directly participate.
As I said in that thread, I don't know whose version of events is accurate here, but some of the details due support Bubeck's version. I suspect that a breakdown of communication occurred where both then misinterpreted what the other saying in a more hostile way than it was intended until the conversation then became genuinely hostile. In terms of the math, although I'm a mathematician, differential equations is pretty far from my expertise, so I cannot deeply evaluate how close the methods were. However, it is the case that the both were building on the methods of Cordoba-Martinez-Zoroa. In this context, my being far from this sort of work is relevant, because this was a well known enough approach that even though I'm in a pretty different subfield, I had heard of CMZ's work as an approach and that this was considered a promising approach to Navier-Stokes. Given that, my inclination is that the case that any theft occurred here is very weak.
And since that thread, I've become more convinced, as experts reading both papers point to substantial differences in the details of their approaches. One of the ironies here is that a lot of the anti-OpenAI views are coming very loudly in part due to a general anti-AI attitude. But Bumaster and Alpoge, the people who are claiming to have had work stolen (well, primarily Bumaster, Alpoge is mostly staying out of the fray) were heavily using both Claude and Codex in their work.
The compute cost for DeepSeek V3 was $5.6M. Training costs for traditional US frontier labs model is about $70 to $200M+, with estimates of up to $1B+ to train the next generation models. It's expensive, but it's not even in the ballpark of "There is no viable economic model to do so." Google revenue in Q2 was $1.32 billion *per day*. $1B is trivially less than a day; $5.6M is just 6 minutes of revenue.
The cost problem is in the compute buildout and the subsidized compute to capture users.
> It remains to be seen if even using their models can be profitable,
I presume we can stipulate there exists some tasks that are cheaper to do using an LLM than doing it "by hand". As a such, using an LLM model can be profitable. To me, it is also obvious that the business of serving an LLM model can be profitable. There's large benefits to being able to get tokens quickly, and you can expect people to pay for that, even if we presume they could run the model locally with no maintenance overhead. There's even cases where using the cloud is cheaper than you can serve locally at all - I saw one guy that measured his added cost of electricity when his computer ran a specific LLM, and it was more expensive to run it locally than to buy tokens from the cheapest inference provider for the same model, presumably because they had better hardware and cheaper electricity.
Once again: training a model with the reward being "does it solve the task?" without looking at how it solves the task is very, very dangerous.
So, people are able to recognize this, and yet in the other threads here about Anthropic calling for a slow down of AI research, everyone seems convinced that this isn't about the risks really at all.
You have never even met a practicing theoretical mathematician let alone participated in a conversation with one, have you? How would you know what constitutes a faux pas in their community, or how they collaborate?
Did you see in the comment where you are replying to where I said I'm a mathematician? My own primary areas of research are number theory and graph theory. It is pretty easy to find who I actually am and verify that yes, I am a mathematician and am in a pure field. So yes, maybe I do know something about the discipline.
A budget is just a method of worrying before you spend money, as well as afterward.
