dnautics

3 ms·

31.
by dnautics 9d ago
I think your CPU might even have a small neural net in the branch predictor
32.
by dnautics 9d ago
I think in principle you could clone your database in prod and at least test to see if your most difficult + common queries are indeed faster after running through the LLM optimizer?
33.
by dnautics 11d ago
Or make it look like you're trying to make it look like it was the Russians so you can get the public on board with a war against NATO? Bet you didn't see the uno reverso false false flag coming
34.
by dnautics 11d ago
No. Nominate a pissed off insider who knows how things work from personal experience and an axe to grind!
35.
by dnautics 11d ago
You can probably distribute training more easily too
36.
by dnautics 13d ago
> but I doubt it can be useful without proper integration with the language itself So projects like mypy or sorbet aren't useful because they aren't properly integrated?
37.
by dnautics 13d ago
> LLM decisionmaking cannot possibly be like human decisionmaking I mean how can it possibly be like human decisionmaking? It's not like it's trained on human data
38.
by dnautics 14d ago
You can very likely borrow check in languages that don't have it in the type system. Exactly the way you suggest, as an optional add-in. It's still WIP but in my side project I haven't found cases that can't be handled
39.
by dnautics 14d ago
I feel like 5g is worse
40.
by dnautics 14d ago
DOGE has only been a year and change. Screwworm does not suddenly appear like that.
41.
by dnautics 15d ago
Hasnt automattic been around for a decade?
42.
by dnautics 15d ago
There is no reason to believe that the transformers couldn't be doing something close to what copycat does (especially with thinking tokens), as an emergent phenomenon of the sheer size of the corpus. The architecture is certainly cap
43.
by dnautics 15d ago
What are you talking about? There have been 5 (five) cases of screwworm in the us and the us is responding by producing millions of sterile male screwworms a day as we speak.
44.
by dnautics 16d ago
Unless your eye can casually assemble great circle routes it's hopeless on any projection except the radial one (and then only for routes emanating from one spot)
45.
by dnautics 16d ago
What's the task here?
46.
by dnautics 17d ago
> many tasks where Mercator would be more useful than an equal-area projection. What sort of tasks besides a size measuring contest is an equal area projection good for?
47.
by dnautics 17d ago
"Maybe it’s really verbose, or maybe you can’t do it inside certain kinds of statements. Maybe you can only call them on line numbers that are prime." Nope. If it's explicit but maybe like a single keyword away that's n
48.
by dnautics 18d ago
No. Function colors (see criterion 4 in the original article) deal with the idea of "how much of a pain in the ass is it to deal with" which is not really a concept that is a PL theory concept.
49.
by dnautics 19d ago
WIP, but that is the target ethos in the prover I'm building: https://github.com/ityonemo/bpa Its painfully verbose and explicit but its designed to let you cut down to the structure of the proof with a query lang
50.
by dnautics 19d ago
sure, but that's not the 14th amendment. Maybe the guarantee clause. It MIGHT touch the 14th amendment via incorporation. That's not what the attorneys argued.
51.
by dnautics 20d ago
> If the model can understand neuralese why can it not convert it into English for monitoring or review purposes Colas described in-article cntent review is done by a weaker model (think like maybe gpt-2 class or llama8b class), and it s
52.
by dnautics 20d ago
I think this is correct. But the article is also correct (If a bit breathless and overstating on the risk): having to reinterpret neuralese is a layer of indirection which introduces effort and inaccuracy
53.
by dnautics 23d ago
Is anyone else noticing degradation of quality many evenings from 9pm-2am-ish ET? I wonder if anthropic has some sort of usage pattern that incentivizes them to route to weaker models. I also fully acknowledge that I might be making spurio
54.
by dnautics 23d ago
You can wrap claude code in a harness without -p, it works fine. I think most of the concern was that -p resubmits the entire chat history on each prompt which is really bad for kv caching, and when Anthropic announced that they were ok wit
55.
by dnautics 24d ago
No requirement doesn't mean "don't be a mensch"
56.
by dnautics 25d ago
It's an slm
57.
by dnautics 25d ago
They said "nearly nothing". As a fraction of total effort I'd say that the contribution of academia is relatively low. If you're going to say something dumb like overweighting foundationality, I can counter by remindin
58.
by dnautics 26d ago
Nah even traits are kinda bad because it increases the indirection necessary for code reading
59.
by dnautics 27d ago
I don't think so. It's pretty clear that LLMs use the higher level layers for reasoning, so a bit of logorrhea very possibly enriches the result quality.
60.
by dnautics 27d ago
The stochastic parrot epithet is so 4 months ago

More ›