I've been thinking about something along these lines, but coupled with deterministic inference. At each "macro" invocation you'd also include hash-of-model, and hash-of-generated-text. (Note, determinism doesn't require temperature 0, so long as you can control the rng seed. But there are a lot of other things that make determinism hard)

You could take it a step further and have a deterministic agent inside a deterministic VM, and you can share a whole project as {model hash, vm image hash, prompt, source tree hash} and have someone else deterministically reproduce it.

Is this useful? Not sure. One use case I had in mind as a mechanism for distributing "forbidden software". You can't distribute software that violates DMCA, for example, but can you distribute a prompt?

▲

mpalmer 6 hours ago | parent [-]

Deterministic inference is mechanically indistinguishable from decompression or decryption, so if there's a way to one-weird-trick DMCA, it's probably not this.

▲

bee_rider 6 hours ago | parent [-]

You’d think that, but it sees like big business and governments are treating inference as somehow special. I dunno, maybe low temperatures can highlight this weird situation?

Temperature is an easy knob to twist, after all. Somebody (not me I’m too poor to pay the lawyers) should do a search and find where the crime starts.

▲

Retr0id 6 hours ago | parent | next [-]

What does temperature have to do with anything?

▲

bee_rider 5 hours ago | parent [-]

Or however deterministic inference is supposed to happen. I don’t know LLMs.

	▲	janwirth an hour ago \| parent [-]
		https://152334h.github.io/blog/non-determinism-in-gpt-4/

▲

mpalmer 6 hours ago | parent | prev [-]

Well, it's still not deterministic even at temp 0. The tech described in my comment's parent is speculative, and technically it's not even inference, once it's perfectly reproducible.

At that point it's retrieving results from a database.

EDIT: how would OP address my main point, which is that det. inference is functionally equivalent to any arbitrary keyed data storage/retrieval system?

▲

nextaccountic 3 hours ago | parent | next [-]

> The tech described in my comment's parent is speculative, and technically it's not even inference, once it's perfectly reproducible.

This is not true. Fabrice Bellard's ts_zip [0] and ts_sms [1] uses a LLM to compress text. It beats stuff like .xz etc but of course is much slower. Now.. if it were non-deterministic, you would have trouble decompressing exactly into what it compressed. So, it uses a deterministic LLM

[0] https://bellard.org/ts_zip/ https://news.ycombinator.com/item?id=37152978

[1] https://bellard.org/ts_sms/ https://lobste.rs/s/5srkwz/fabrice_bellard_s_ts_sms_short_me... (funny enough many people comment that, if it uses LLM, it must be lossy. This is not the case. It's compared to xz in the page because it's lossless)

▲

Retr0id 6 hours ago | parent | prev [-]

Deterministic inference isn't speculative, it's achievable if you want it. It's just not the default.

	▲	janwirth an hour ago \| parent [-]
		It seems the TL:DR is race conditions, rounding and other RTE specific factors.