David Manning

To me, storage implies things like durability. Probabilistic interpolated output that sometimes resembles training data is a different thing. E.g. Regression models neither store nor retrieve the data used to build them.

I understood Paul to be differentiating between storage and “lossy database.”

For my part I only meant that LLM’s are not properly understood as search engines.

I think you’re meaning to respond to Paul here.

Warum sieht die Switch so riesig aus?

A close second is the idea that they are “search engines.”

Is that the Curb Your Enthusiasm guy?

Has anyone done a startup that uses AI to predict how much funding stupid ideas can get in a seed round?

Yeah I mean that’s absolutely true.

The most baffling thing to me about the rush to embrace “AI” is how corporate IT departments have seemingly jettisoned four decades of cultivated paranoia in the span of a couple years.

THE ELEPHANT IN THE ROOM Kay Ryan The room is almost all elephant. Almost none of it isn't. Pretty much solid elephant. So there's no room to talk about it.

The room is
almost all
elephant.
Almost none
of it isn’t.
Pretty much
solid elephant.
So there’s no
room to talk
about it.

-Kay Ryan, “The Elephant in the Room”
#everynightapoem

*yawn*

Yep. Since they have no actual path to profitability, the game is to compete with one another over-leveraging themselves to see who can win a government bailout by becoming a threat to the national (and global) economy.

Evolution of “get acquired by Google” as a business plan.

This confirms my anecdotal observations that every thing became stupid in 2016.

My instinct is that it’s because “reasoning” models are actually fine-tuned toward specific outputs that indicate “reasoning.” (E.g. “Oh! Now I see”)

This has the effect of helping “walk-through” some multi-step solutions at the expense of one-shot correctness.

I know I saw a paper about this…

I think there’s actually a known trade-off there between models optimized for “agentive” flows and those optimized for one-shot correctness.

Still?

That’s the body heat from all the other stranded people.

Can you get that done in time to throw the GOP out of the house a year from now?

I hear you, I’m just not sure it’s an unalloyed good.

So aside from current capabilities, my original point is that if there comes a day on which one of these things can crank out a bespoke OS for your grandma, the entire notion of an “OS” has probably stopped being relevant.

I guess I’m not sure how else to interpret “bespoke desktop and mobile environments?”

I use Claude enough to hit the weekly usage limits on the max plan.

I think you’re underestimating how much current agents depend on either infra or the skill of their uses to fill the gaps

I’m not sure what sort of skill this hypothetical AI that can make you an OS based on your favorite color etc needs you to have.

Yeah sure a current “agent” can orchestrate those steps, but you’re not using it to do that unless you already know how to do it manually.

Because even though you can probably generate the app most people don’t have the skill to deploy in a useful place, maintain it or even clarify what they want. They’d prefer the packaged version. Curren AI can’t do all that for even a simple app.

Like at that point we’ll all just be burping requests to the omnibox.

This doesn’t make sense to me tbh. If you have AI that can “grow bespoke applications and operating systems” based on a few questions to a non-expert, you probably do not need those applications or operating systems in the first place.

I’m a software engineer. If they start putting “AI” systems in charge of air traffic control, do not get on a plane.

Spoof of 'The Return of the Pink Panther' with a dragon in place of the panther, and the silhouette of the knight instead of Inspector Clouseau

New Blog Post: The Return Of Language-Oriented Programming blog.evacchi.dev/posts/2025/1...

This is great. I swear I’ve tried to explain this exact point to people in a less well-developed way.

This is an actually interesting avenue we might be pursuing with language models if we weren’t setting our heads on fire trying to replace human labor.

I’ll go the barricades over this one.