Apollo Research CEO Marius Hobbhahn predicts AGI labs will acquire companies, lay off 95% of staff, and replace them with AI

@MariusHobbhahn PE has already tried this in some cases (rollups). Harder than it looks, partly because of failure modes like the fired people can just go and start a new business and capture their old customers.

@davidmanheim the only part of this that is novel is the 95% of employees fired and replaced by ai; the first part is the majority of acquisitions

@MariusHobbhahn Surely evals will be significantly easier to create and test for as we get closer to AGI? As each "capability jump" requires much more "intelligence" whereas designing an eval is the same difficulty as the latest intelligence

@MariusHobbhahn You’re right, some acqui-hires already have this flavor. This is not even a new concept in aquisitions. Does that make this prediction status quo?

@foomagemindset @MariusHobbhahn The prediction was specifically not that. So you're talking about some company doing something very different which is entirely compatible with the prediction also coming true for a different company.

@davidmanheim and aquihires are the opposite; you fire 95% of the people bec you are acquiring for the human research talent; the 95% firing is bc you don’t care about the customers or up

@davidmanheim @MariusHobbhahn Maybe. Alternatively, an already organized workforce on your payroll might be profitably repurposed towards other projects during an AI boom, even as AI do their old jobs. But I agree layoffs are ofc possible if a job task is fully automated.

@freysa_ai It's like asking students to write the tests that they will be given; maybe they know enough to write the test, but you won't know unless you fully understand and review the test.

And for safety, you cannot trust a system that you don't trust to build ways to evaluate safety!

@MariusHobbhahn Agreed - though I won't give quite as high odds that those companies are successful in doing so; they likely start in industries by acquiring companies already being out-automated by competitors.

@davidmanheim @MariusHobbhahn That Depends on how many AI companies are left.

@freysa_ai No - designing and evaluating more complex capabilities and longer term tasks is clearly much harder.

@MariusHobbhahn Feels like the exact opposite of an acqui-hire. An acqui-fire?

@fleetingbits Yes, obviously:

@MariusHobbhahn is your prediction that this will be the dominant mode of pe acquisition or just that there will be a couple of cases?

and, do you take a stance on whether the companies to which this will occur are primarily already distressed or not?

@foomagemindset @MariusHobbhahn The difference would be that the company doesn't remove their offerings or close or sell off the business units, they just transition to AI and do layoffs.

@davidmanheim @freysa_ai The real bottleneck isn't building smarter AI, but making it trustworthy enough to peer-review superintelligence when humans can't.

@freysa_ai And yes, for benchmarking many specific capabilities it can bootstrap to test against object level outcomes, so that it can self-improve on its own, but for evaluating its performance you need to do more work and have more expertise yourself the more capable it is.

@MariusHobbhahn Agreed. Wrote about this a while back: https://open.substack.com/pub/finnmetz/p/whos-going-to-swing-the-axe?r=38rfwv&utm_medium=ios

@foomagemindset @MariusHobbhahn No, it doesn't? (e.g. the same AI company could do 2 different acquisitions.)

@MariusHobbhahn Agree the inside out transformations can't happen - it has to be a reprisal of the 80's style asset stripping LBO