Is OpenAI more like WeWork or Theranos?
A tale of two analogies
Breaking news, almost exactly ten years ago.
Needless (or needle-less?) to say, that never came to pass.
§
OpenAI might be screaming success - but its success is hardly assured.
In an essay at the beginning of this year, I warned that OpenAI faced many headwinds, and in many ways things have only gotten worse. They are burning a ton of money, competition has greatly increased, turnover has been considerable, and as The New York Times reported earlier this week, their relationship with Microsoft is fraying. Meanwhile GPT-5 still hasn’t arrived, and nor has the much ballyhooed Sora. (The new model o1 is interesting, but expensive to operate, and far from a general solution. There is a reason OpenAI didn’t choose to dub it GPT-5.) They might succeed; they might not. Without Microsoft’s unwavering support, they could find themselves in deep trouble. Living up to OpenAI’s $150B valuation won’t be easy.
When I think of OpenAI, I often think of WeWork: charismatic founder, immense valuation, questionable business plan, and the possibility of similar immense deflation in their valuation, if confidence wavers.
§
But sometimes I have heard others compare OpenAI to Theranos, which was basically a fraud. Another charismatic founder, another ridiculous valuation, and another collapse. It’s not a crazy suggestion, but I’ve generally avoided making that analogy, since OpenAI has consistently shipped products. Those products have not by a longshot lived up to everyone’s expectations, but they do actually ship, and customers can try those products out and decide for themselves. It’s hard to call that smoke and mirrors.
But then again, their products have always been served with large side order of bullshit, like when Sam said, apparently jokingly, that “AGI has been achieved internally”. None of the robot stuff they promised in the Rubik’s cube campaign ever shipped.
Last night I saw this graph, from an OpenAI employee, and (perhaps because of three decades I spent as a scientist reviewing papers and looking at hinky graphs) I started to get major Theranos vibes:
As I wrote on X, the graph “is a fantasy about the future, and not at all obvious that the “data” plotted correspond to anything real. The absence of label on the Y axis maybe tells you all you need to know.”
As far I can tell this graph doesn’t plot real data, nor clarify that it is merely hypothetical. Because the dates on the X axis are real dates and align roughly with actual releases, viewers might mistake it for real.
But the curve, so far as I know, is just made up. I do not know any measure that pegs the delta between GPT-4 and o1 (marked as “today”) as being triple the delta between GPT-3 and GPT-4. You certainly couldnt’t make that case for “capabilities” as a whole. The difference between GPT-3 and GPT-4 was by all accounts massive; the difference between GPT-4 and o1 is by all accounts modest.
And don’t even get me started on how the rest of curve, beyond today and into the future, is projected.
§
§
Theranos’s valuation and press were based on promise; to a large extent you could say the same about OpenAI. The combination of made-up graphs and outsized promises could make a person nervous.
That said, I still lean more towards the WeWork analogy, a real business that was for a while greatly overvalued. But to the extent that the whole thing is premised on OpenAI winning the race to AGI and soon — without a clear, feasible technological path there (unless you count “scaling” which sounds like “hope, without a moat” to me) — I just don’t know.
If things do fall apart, it is not just investors who stand to lose, but society. Immense resources may end up being wasted in vain, because of hype. Too many more profitable stones may have been left unturned. If the answer isn’t bigger LLMs, we may have wasted half a decade.
§
As the dust was starting to settle on Theranos, Mayo clinic research Michael Joyner wrote this essay:
“In addition to the specific mix of greed, bad corporate governance, and too much “next” Steve Jobs”, Joyner concluded, “Theranos thrived in a biomedical innovation world that has become prisoner to a seemingly endless supply of hype. That so many high-profile individuals and institutions fed and continue to feed the hype makes me think it is just a matter of time until we see the arrival of Theranos 2.0.” One doesn’t have to strain to hard to think of possible analogies.
AI ought to be looking itself in the mirror, too, right about now.
Gary Marcus has written six books about AI and the human mind; his most recent, Taming Silicon Valley, includes a chapter about how Silicon Valley plays the public. This example would have fit right in.
The absence of label on the Y axis maybe tells you all you need to know.”“
Nonsense. “Victoria” is clearly the label on the Y axis on that graph.
It indicates the relative number of Victorias working at OpenAI when each new version of GPT was released.
It’s a good thing Sam Altman just got a $6.6 billion infusion from investors because the number of Victorias at OpenAI is obviously increasing exponentially and a lot of cash will be needed for all their salaries and benefits.
To continue the WeWork comparison, Altman is more like Adam Nuemann than Holmes. They have both described themselves as Gods (though Neumann did this more literally) and see themselves as bringing about a new age for humanity even though their businesses are essentially disappointing and unsustainable. Coincidentally the suggestion that they are somehow superior beings is even in their names: alt man, neu man. Pretty sure they both look in the mirror and see an Übermensch.