TIL: xAI burn rate

[timelets]

Elon Musk’s artificial intelligence startup xAI is burning through $1 billion a month as the cost of building its advanced AI models

https://www.bloomberg.com/news/articles/2025-06-17/musk-s-xai-burning-through-1-billion-a-month-as-costs-pile-up

Comments

[dmm]

> He is rich because people believe he is rich.

He owns good chunks of expensive companies.

> Did you see his talk at the recent Y-combinator event? Clearly he was on drugs (in a bad way), he looked spent and out of ideas.

I don't remember if I've seen that. He is known to have rather serious drug problems, yes...

I did see the recording of yesterday's Grok 4 release livestream, and he was doing very well during that event (to my eyes, at least). Also their numbers on ARC-AGI-2 eval are rather mind-blowing for anyone keeping track of the dynamics of this year Arcprize contest (making progress on ARC-AGI-2 seems to be really difficult).

> Altman, Zuckerberg or Musk

I would not bet on Zuck at the moment. Llama 4 release failure was rather spectacular. Meta's attempt to buy itself a different AI team for tons of money is also telling. (They do have plenty of in-house talent, but they don't know how to use that talent.)

Perhaps they might recover, we'll see. But this year is less good for Meta AI. Llama 3 was great, even if a bit too brute-force. But this year their internal organizational pathologies are transpiring...

Generally speaking, the list of contenders is longer (and as time goes by, it gets larger). I think the importance of having extreme amounts of compute is exaggerated. Algorithmic progress is contributing more than hardware scaling in recent years, and as time goes by, the ability to use AI for AI research in a more productive fashion is the real thing they all are competing for (if an explosion happens, it will be due to AI algorithmic progress accelerating AI algorithmic progress).

A number of large and small American (or partially American) labs and a number of Chinese labs are viable contenders, at least one Japanese lab is a contender, we do see some very interesting Saudi-Swiss activity, and with all this rather extreme Saudi and UAE hardware build I would not be surprised if we see more in that region, perhaps France is not completely out yet, perhaps Singapore and South Korea will build further on their research momentum. It's fairly open-ended (with some open-weight models being rather high on the leaderboards, and thus everyone having a chance to build innovations on top of them in various ways).

[timelets]

The value of companies he owns depends on people's perception of his promises and to keep it up, he'd have to deliver an awful lot of business within the next couple of years. We'll see. Usually, drug problems don't go away. Rather, they get worse over time.

Of course, I totally agree that there will be many good models out there. Further, I also see good models that work outside of the usual LLM spaces; they already make a big difference in the field today. But it takes way more than a good core technology to create a highly profitable business that scales to hundreds of millions and even billions of users. Musk has never done that because the "construction" process requires many partners and he's not good at partnering.

In any case, in the next five or so years technology developments are going to be very exciting. There are serious issues with AI, including public trust, that are way bigger than Musk and he's more of a distraction here, rather than the mover and the shaker he used to be with SpaceX and Tesla. The good news is that the emperor still has some clothes on.

[timelets]

BTW, users don't rank Grok 4 above the current version of ChatGPT — "Grok 4 (incl. Heavy) is not a great AI model."
Here's more detail: https://x.com/Sherveen/status/1943306220170809827

[dmm]

oh, users are very different; people are still evaluating; we'll have a better picture in a few days

preliminary indications are that it's great in math and reasoning, but not so much in coding (surprisingly) or in writing

also it knows how to score super on tests in physics, but not necessarily on a more deep understanding of physics (this one is something none of the models seem to do well yet)

anyway, if no one summarizes the results earlier, Zvi will surely get a good summary to us next Thursday in his next weekly AI review (following this one: https://www.greaterwrong.com/posts/FczrW2kQ7WxGW39Yv/ai-124-grokless-interlude) at the latest

[timelets]

The latest version of Elon Musk’s artificial intelligence chatbot Grok is echoing the views of its billionaire creator, so much so that it will sometimes search online for Musk’s stance on an issue before offering up an opinion.

The unusual behavior of Grok 4, the AI model that Musk’s company xAI released late Wednesday, has surprised some experts.
...
“It’s extraordinary,” said Simon Willison, an independent AI researcher who’s been testing the tool. “You can ask it a sort of pointed question that is around controversial topics. And then you can watch it literally do a search on X for what Elon Musk said about this, as part of its research into how it should reply.”

[dmm]

I know. Now people are going to investigate if it was explicitly instructed to do so (which would be pretty bad), or whether it is a side effect of their training process (which would be interesting):

https://x.com/TheZvi/status/1943439850608431388

[timelets]

He's going to shove his opinions down everybody's throat for as long as he's in charge. He's done that when he bought twitter and he's doing it now.

[dmm]

On Twitter, not really. He tried a bit, mostly caused people to laugh at him (while being quite mad at him, and rightly so), and more or less stopped.

I am using Twitter all the time, and I don't see much traces of Musk (although a bit too much of his crowd, not in my feeds, but in those stupid "trending things" (I still have not figured out how to block politics and sport from "trending" words shown to me)).

[timelets]

He is a repeat offender. When he bought twitter he made engineers insert his tweets into millions of feeds. I, personally, had to block him, as did many other people. After seeing that users are blocking him, they tweaked the algorithms to promote feeds that engage with his tweets and demote those that don't. It's disgusting.

[dmm]

Elon's behavior is quite disgusting along multiple dimensions...

But this effect you are talking about was not pronounced enough for me to notice; I only knew about it because people were discussing that. My feeds were always free of all those people.

I am way more concerned about technical degradation of Twitter (and about the even more pronounced technical degradation of Facebook, and general enshittification of more-or-less everything, so that's not really all that Elon-specific)...