turtleyacht

How do you prepare the fallback if it already took tokens to query the simpler model?

Also: The Unreasonable Effect of LLMs on Mathematics - https://news.ycombinator.com/item?id=48888345

Usually writing small, in all-caps, except code: in lowercase, and the "t" and "i" retain their lower curve. Cursive is difficult; easy to write, but (later) hard to read.

Can see how penmanship there would be appreciated.

What about citing papers?

No, because we can't read the ending.

After clicking a language, click Contents link at the top.

Took search for granted. It was such an early Internet before Google. "Surfing" and "downloading" were concepts to learn. Dial-up required a credit card.

LLMs are this generation's search, but I don't remember search being so reviled. Search opened a door to possibilities. If search were so resisted back then, it would have been a project to recreate the implementation. What did I know back then?

It may have paid off in terms of what is learned making a game versus playing one.

I just accepted search as the thing to do.

Is there a graph starting from the 1950s?

https://www.lyrebirddreaming.com/post/el-ni%C3%B1o-isn-t-an-...

New changes to Jean-Jacques Dubray's State-Actor-Model (SAM) since 2022.

Software is Eating the World, All Right (2024) - https://news.ycombinator.com/item?id=42613550

The whole thing `false or true' is sent in for evaluation, so if there is a side-effect function that returns `true', it is executed too. Because Elixir treats it as a macro and not short-circuit.

Reinforcement Learning from Human Feedback [1] will be published next week. Hard to believe the technique is already obsolete.

[1] https://www.manning.com/books/reinforcement-learning-from-hu...

Nit: the special uppercase A in "Abnormal" is also used in the word "Anthropic."

It's disconcerting most of the time. I follow the technical line of reasoning, and then it goes off into a kind of uncomfortable place.

Deferring to their experience, but thinking, "Is this what winning is?" And wondering if that is the natural endpoint of progression.

Demos can be a useful forcing function. Strategic use with high stakes, when successful, synchronize disparate stakeholders on common ground. It can also pull far features closer, and prioritize blocking bugs.

Switching the monitor worked. Wonder if an HDMI dummy plug, a virtual monitor emulator, would have helped.

Example of triggered credit:

https://slacreditwatch.com/v/cloudflare

Looks like there is a memory hierarchy [1] even in LLMs.

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

Why did it take four years?

Anyone do a playthrough of Fallout following the book to the letter?

It sounds like iroh enables distributed compute without having to finangle custom hardware.

Same. By that point, I realized if that was one thing I got, I must have missed even more.

What was the question?

No tablets. Have lots of things to read and material for arts & crafts. Keep them bored for naught but the outside or their own making.

Have a computer with OpenBSD. Not for nothing can it do any thing, if they're willing to start from less.

The common ground is in the artifacts: troubleshooting physics, written programs, and self-organization.

It's the first pixel of fifty thousand.

A chemist shop is also known as a pharmacy.

(2021)

(2015)

Please pardon the pure speculation incoming. Yes, caching the answer doesn't seem useful. Caching the progression, the graph, may be. This is similar to making code changes with ed(1) instead of editing in vi.

The transform script(s) are cached and can be played back or adjusted. Surely for some breadth of question inputs, they map more often to similar answers--but not static answers; instead, evented edits.

It's nearly untenable for a human to keep private edit scripts to generate code changes. The extra steps for custom regex, essentially one-offs for a shared codebase, is inefficient. But maybe not to an LLM.

Maybe cache similar answers from others. Surprised if this is not already being done.