The Real Story Behind Sam Altman’s Firing From OpenAI

Or someone said "If you don't switch and back me, I am going to fight every bit of your compensation. Or you can back me and leave with favorable terms."

Panic is a less likely driver.

Murati didn't demand they release the evidence, as far as I could tell. The board are describing as not wanting to throw Murati under the bus by stating the evidence came from her, which makes sense if their goal was to install her as the new CEO.

I'm surprised at how little their utility for turning unstructured data into structured data, even with some margin of error, is discussed. It doesn't even take an especially large model to accomplish it, either.

I would think entire industries could reform around having an LLM as a first pass on data, with software and/or human error checking at significant cost reduction over previous strategies.

"ChatGPT needs an array of GPUs per active user" - nit: you're exaggerating by a few orders of magnitude.

First, queries from users can be combined and fed into servers in batches so that hundreds of queries can be concurrently served by a single node. Second, people aren't on and asking ChatGPT questions every second of every day. I'd guess the median is more like ~single digit queries per day. Assuming average response length of 100 tokens and throughput of 50 tok/s at batch size 50, that's 25 QPS or 2.1M queries per day, or 420k users served per node at 5 queries per user per day.

Now, a single 8xH100 node is a lot more expensive than $5/mo, so you're directionally correct there, but I'd wager you can segment your market aggressively and serve heavily distilled/quantized models (small enough to fit onto single commodity GPUs, or even CPUs) to your free tier. Finally, this is subject to Huang's Law, which says every 2 years the cost of the same performance will more than halve.

People said similar things about Facebook. "Oh their user growth might be amazing, but they're not making any money, it's not a real business."

But it turns out that with enough funding, you can prioritize growth over profit for a very long time. And with enough growth, you can raise unlimited funds before you get to that point. And going this route is smart and effective if you want to get to a $1T valuation in under a decade.

So yeah, ChatGPT's margins might not be as high as Facebook's. But it doesn't really matter at this point, they're in growth mode. What matters is whether or not they'll be able to turn their lead and their mindshare into massive profits eventually, and while we can speculate on that, it's far too early to definitively say the answer is no.

Rather than getting into the nitty gritty details of monetization, when we ask ourselves if OpenAi can nail product like Facebook did (I guess) to become the next tech giant, I think we have to ask whether it's even possible when the tech industry is as established as it is.

You would think existing megacaps would be all over any new market if there is a profit to be made. Facebook's competition was basically other startups. That said, Google seems to be dropping the ball almost as bad as Yahoo.

But sure, if there's absolutely no way to make money from consumer AI then that will also make it hard for oai to win the game.

Useful search unlocked the web. I will take that over LLMs in their present state.

I agree, just saying, ChatGPT was a quantum leap, too. That's why it has all the consumer mindshare.

The Microsoft antitrust case also gave them a couple of years to grow without that threat.

I wouldn't worry. Retracting ZIRP policies gave governments 2 choices: reduce spending by ~10% on average (in Europe), or cheat and scheme to bring them back, on a 2 year timer.

Interest rates raised, but came back down before the 2 years were even up (rate rise started 27 Jul 2022, Started coming back down 12 Jun 2024), governments have been caught cheating, and the number of central bankers replaced has gone up dramatically. Oh, and none of the governments have reduced spending. Literally not a single one. In fact, Germany has agreed to an unprecedented increase in debt financing of their government.

In other words, ZIRP, even negative rates, are coming back, and a lot sooner than most people think. Your next house, despite everything that's happened, will be more expensive. But I doubt this will save either OpenAI or Tesler.

From the story it sounds like the direct reports generally did not want to work with Altman, Brockman excluded. Even Murati was one of the primary instigators of the firing, but she changed her mind for reasons that the article doesn't really explore.

Money.

This is one issue with Microsoft's Total Recall thing, right? I wonder how they're dealing with that.

Want to abolish economic copyright alltogether? I could get behind that. Making a legal exception because of some imagined future metaphysical property of this particular platform sounds like being fooled.

I don't think the concern is related specifically to training on computer chips with copyrighted content.

If you are going to use human brain cells to memorize protected content and sell it as a product, that's still an issue based on current copyright laws.

Others replied to this and I am still not sure what your point is. Are you saying big tech should be able to get away with this because LLMs are just like us humans?

> If we could incrementally update our own brains by swapping cells for chips, what percentage of our brain has to be chips before us learning from a book is a violation of copyright?

The same percentage at which you stop qualifying to be human and become an unthinking tool, fully controlled by its operator to do whatever they want, without free will of its own and without any ethical concerns about abuse and slavery, like is the case with all LLMs.

(Of course, it is a moot point, because creating a human-level consciousness with chips is a thought experiment not grounded in reality.)

> When learning to recite a recent children's poem in kindergarten, what level of accuracy can a child attain before their ability to repeat it privately to one other person at a time is a copyright violation?

Any level thanks to the concept called human rights and freedoms, famously not applied to machines and other unthinking tools.

1. Google was, and still is in some developed countries, under fire for as much as summarising search results too much, so I think yes, the claims have legs.

> This would be separate from the contributory infringement claim if the model will output a copyrighted work verbatim

2. Commercial for-profit models were shown to do that, and (other legal arguments aside, such as model and/or its output being a derivative work, etc.) in some cases that was precisely the smoking gun for the lawsuit, if I recall correctly.

I have not seen any conclusive outcome, I suppose it will depend on jurisdiction.