[CharlesW]

> I think the fair way to read any CEO's comments about AI reducing their workforce at this point has nothing to do with the capabilities of AI…

You can legitimately argue "far less to do with", but it's definitely not nothing. There are countless projects underway where AI will allow for 10% reductions with zero business impact in the short term, and 25-40% reductions (sometimes more) by 2030.

[jppope]

ok. so you can reduce your head count without impact great! Why would you get rid of people? Why would you not reassign those people into other productive or revenue generating activities?

The only logical explanation is that they don't have enough opportunities to utilize those people OR as I previously mentioned... their financials might look bad, and they are trying to make them look better so they don't take a hit in the markets.

[TheOtherHobbes]

Fire people. Stonks go up. Bonus!

Stonks go down - fast - when all those fired people stop buying, but that's a problem for the next CEO.

As you say, they could also expand. Or just fix the problems with the site.

But they don't have the imagination to do that.

[diamond559]

Countless projects huh, CMU found the best of them has only a 30%ish success rate on basic business tasks. Many are below 90% still, but yeah let's just pull magic numbers out of thin air. How much Nvda you own bud?

[CharlesW]

Zero Nvidia. The CMU benchmark is fun, but tasks <> jobs. They found that agents can autonomously finish about a third of their simulated office tasks, but that can't be mapped to a labor-market forecast.

[rsynnott]

> There are countless projects underway where AI will allow for 10% reductions with zero business impact in the short term, and 25-40% reductions (sometimes more) by 2030.

Are there any where it empirically _has_ done, or are we still in jam tomorrow mode? Like, there is a very big industry devoted to selling this stuff; I'd be _extremely_ cautious about promises and projections.

[another_twist]

I am curious, where are these numbers from ?

[CharlesW]

These are realistic (IMHO, of course) projections based on studies I've helped with and conversations I've had with my network. Naturally, the impact will vary enormously based on roles, and the timelines won't be evenly distributed.

But these kinds of projections aren't unusual at all — if you use the Deep Research capabilities of modern models to build a list of public projections for your own research, you'll see similar estimates. These reports will generally use the framing of "efficiency gains", where AI will "free-up employees from drudgery to focus on higher-value work", but my intuition is that a future where all individual contributors are elevated to Director of Agentic Workflows is probably not the most likely outcome.

[diamond559]

What studies? MIT estimates only 5% of the workforce can be replaced long term. What tasks are you employees using AI on, CMU shows the best llm only has a ~30% success rate for basic business tasks. Are you a vibe coding start up or something?

[CharlesW]

> MIT estimates only 5% of the workforce can be replaced long term.

The model by MIT's Daron Acemoglu estimates that ~5% of U.S. tasks can be completely and profitably automated by AI within ten years.

It was expressly not a head-count forecast, and didn't attempt to quantify the headcount reduction that AI augmentation could enable.

[another_twist]

https://economics.mit.edu/sites/default/files/2024-04/The%20...

Is this the MIT paper ? In this one the TFP is 0.55%

[red-iron-pine]

i think profitably is the key word here, since AI orgs are burning ungodly amounts of money and electricity and will expect to recoup that money.

[another_twist]

I see and are these studies public ? Could we see the data and the methodology here ? Thing is there are benchmarks to judge software engineering capability of AI. I am more interested in how the jobless predictions made ?

I understand all the theory but it can largely be condensed into - AI makes workforce more efficient so you need less people. But there are no good studies afaik that measure AI powered efficiency and surely nothing about how to model workforce reduction due to AI. I am curious what the science is behind these opinions.

[rsynnott]

> These reports will generally use the framing of "efficiency gains", where AI will "free-up employees from drudgery to focus on higher-value work"

Okay, but what are these reports _based_ on? Everything I've seen along these lines has been, essentially, marketing material; there seems to be very little hard data suggesting this kind of outcome.