This why those RL env startups are able to charge frontier labs so much for their work. LLMs still generalize poorly outside of self-verifiable tasks like coding and math.

Labs have to compensate with post-training in RL env that embeds these expertise well, which is non-trivial both in terms of domain knowledge and technical expertise.

Very much agree that a lot of LLM calls today are redundant and could be decomposed into reasoning vs just parsing(regexes), ETL, etc

Like what's the angle here, and what's the outlook for LLM routers like Openrouter to justify a 10B valuation?

Somebody enlighten me.

OpenRouter is a marketplace (“OpenRouter, a fast-growing marketplace that enables developers to route traffic across various artificial intelligence models”).

Market places mean payments. Stripe does payments.

That's a good point actually. I just recalled that Openrouter charges a premium of 5.5% on transactions. That's probably comparable or fat relative to the cut for payments processors like Stripe.

OP here, and I agree it's probably not AI directly and more of dealing with the post Covid overhiring, and CAPEX for AI investments.

Enterprise adoption is definitely not happening as fast as the AI proponents (startups, frontier labs, big tech) would like people to believe (probably the reason why frontier labs are going the FDE/consulting path). So it's unlikely that the impact on productivity is enough to produce the alleged impact.

It's a combination of Covid over hiring, CapEx for AI hardware, AND you don't need as many software developers anymore per product.

I doubt very much that COVID over-hiring has any relevance today.

COVID is now ancient history and there have already been several years of big layoffs when every time I hear the attempt of explaining them through COVID over-hiring.

That may have been right a couple of years ago, but since then every company that had over-hired must have already disposed of the surplus.

The mad rush and overhiring didn't stop immediately after Covid. I know of cases where experienced software engineers were getting picked up by big tech with huge increments as late as end of 2023.

My impression is that the current trend for layoffs only really began around early 2025. For reference, the layoffs during the dotcom bubble I was told took around 2 years to plateau

We’re going to be well into the Diamond Age before people stop attributing layoffs to ‘Covid overhiring.’

Big tech employee counts are still nearly 2x higher now than 2019.

And big tech market cap is 5x what it was in 2019, yes. I’m not hearing wild overhiring in these numbers, quite the opposite.

In 2019 I bet they were more than 2x what they were in 2012.

So then why aren't bad decision making execs being let go?

If they can't forecast demand or headcount they should be replaced by confident sounding AI.

This^

This is why every other crypto company is now an AI company.

This is a fascinating example of perhaps why a move towards a "world model" or some better form of representation can be helpful.

That doesn't sound like it? From the article:

>> According to the complaint, Meta used a number of internal AI-assisted systems to score and rank employees on a termination list. Those included "Metamate," a large language model assistant; an employee-trained "second brain" that tracked workers' communications and documents; and a productivity score drawn from scanning keystrokes, screen content, emails and browser history, according to the lawsuit.

"Nadella argues that if AI companies get to freely scrape the internet to train their models, it’s only fair that enterprises get to study — or “distill” — those models in return."

Two wrongs don't make one right

I agree as well... doesn't seem incompatible with what I said, though?

AI coding forces people to frontload a lot of the detailed thinking that used to take place over an epic or sprint. It's not a very natural way of working, and in fact is quite contrary to the iterative development style that most devs are already used to.

Yep, some time I think we are going back to a waterfall methodology