LLMs (at the autoregressive pretraining stage) are trained to predict what an (adult) human who could have said preceding words would have said next. You should add 15-30 years to the human training time.

I guess it's childhood amnesia[1].

[1] https://en.wikipedia.org/wiki/Childhood_amnesia

Your post reminds me a bit of another conversation: "It's nothing new, we did that with Utah arrays in the 1970s." Skepticism can overshoot into dismissiveness.

A model has internalized and most likely generalized mechinterp research. Yes, it's expected and nothing insane. Just like internalizing and generalizing math research was expected and not insane. The results of the latter are... interesting though.

Hey! A Pinkelbaum can use some nitrogen.

Ancient bacteria will be surprised to encounter penicillin. Viruses are less resistant to degradation. But, yeah, it's nothing to look forward to, horizontal gene transfer and all.

The first paper (free-tier chatbots) due to fake references, the second paper (a specialized AI-research tool) due to overreliance on preprints.

This reminds me of "Autonomous driving? Ahhhh! There will be a bloodbath!"

> The subagent is just as intelligent

Isn't it a common practice to use cheaper models to do the grunt work?

Does this actually work for anybody? In my experience, when you use cheaper models as subagents, the more expensive parent agent ends up reading all of the subagent's work and surrounding code into context anyway to review it. So you end up paying more, not less.

Plus the subagent's work is often subpar and the parent ends up requesting multiple revisions. Each back and forth burn unnecessary tokens as well as my time.

Where it might work well is for separate task categories like image generation - get the parent to ask for an image from another model that is better at this aspect

When I am making agents do a hard thing, I use Opus as both parent and child. I make the child review its code extensively, and then still make the parent read the work and review once again. It is much faster and cheaper than just rolling with no extra reviews, and then in a day I have fifty subtle bugs.

Maybe something has changed between now and then? Say, some kind of a technology that allows to select eligible employees better.

Definitely not. We are no better at finding good employees than we ever were. It's the human biases that make it bad, not the technology.

"We" as in "industry in general" or "we" as in "HR at my company"? Although, I'd be more interested in what Microsoft does.

We as in humans in general.

My initial comment was tongue in cheek (but transparent enough I think, given the context). So, let's be clear. I questioned whether it's still true that it's people who are doing all the HR work at Microsoft without LLM assistance.

You, probably, talk about something else.

Nearly every HR department uses AI these days, and it just makes things worse because they're easy to game.

Then I'd just marvel at your all-encompassing knowledge.

Did it ever occur to you I may have experience relating to what I said?

Things change fast nowadays. Microsoft has access to the latest models. Using AI to screen employees is controversial enough to keep it hush-hush. There's no principled reasons that they can't extract useful training data even with the presence of biases.

Obviously, error rate of ChatGPT should be measured by comparing to human diagnoses and when it plateaus, we should investigate a possibility that the number is driven mostly by human errors.