dvt

13 ms·

211.
by dvt 7mo ago
> I think what they're getting at is that for a given unit of compute, this method achieves 125% performance. This is not what they're getting at; I explained exactly what they're getting at. I mean, your equivalence of &q
212.
by dvt 7mo ago
> Drops compute required for training by ~20%. This is not true. Authors claim that w.r.t. training, their method adds negigible overhead for AttnRes with no memory impact (but is way more complicated for Block AttnRes since we need to u
213.
by dvt 7mo ago
Knuth's paper is 100% in the training set, so while your result is decent, it's undoubtedly tainted. But let's look at the output anyway: > ...the activity of programming becomes less an act of issuing commands and more an
214.
by dvt 7mo ago
> I don’t have to solve any problems with languages that are elaborate practical jokes This is just being needlessly dismissive. Esolangs are (and have been) an area of active CS research for decades. I know I'm a bit of an esolang
215.
by dvt 7mo ago
LLMs are trained to be precise (and more specifically: semantically precise), especially in the fine-tuning phase. An LLM just trained on the corpus of full human production would surely sound more "human," but it would also pro
216.
by dvt 7mo ago
Because of overhiring during the post-COVID free money glitch, not because of AI.
217.
by dvt 7mo ago
"Poppy turns your contacts into a living garden. Gentle reminders, zero guilt."—bleh, at least write your own taglines :/ This is a neat idea that has been tried about 300 times over, but since it's contingent on already
218.
by dvt 7mo ago
Gotcha', but I'm just trying to see the audit for how claim X was rated and based on what sources. If we're looking at the Claude logs, we have huge files that have things like this[1]: {"id": "claim_0081
219.
by dvt 7mo ago
I'm still not sure I fully understand the methodology. For example if Marcus makes the claim: "OpenAI sucks!" why would OpenAI's blog ever corroborate that? The sources used are all AI company blogs (Anthropic, Google, O
220.
by dvt 7mo ago
Had the same thought. Also confused at the backhanded compliment that pickle got: > Just look at Python's pickle: it's a completely insecure serialization format. Loading a file can cause code execution even if you just wanted
221.
by dvt 7mo ago
It's literally cited in his bio, and he's using his real name on HN. It's about as far as grift as it could be. If he's being curt, he's probably (rightfully) frustrated that "journalists" are getting such
222.
by dvt 8mo ago
> So edgy; is being an apologist really the noble calling you think it is? Is it your noble calling? From the Temporary Constitution of the State of Palestine (2026)[1]: Article 4 – Islam, Sharia and Christianity 1. Islam
223.
by dvt 8mo ago
Yeah, I actually wanted to write an addendum, so I'll just do it here. I think that going from pseudocode -> code is a pretty neat concept (which is kind of what I mean by "write your own code"), but not sure if it's
224.
by dvt 8mo ago
> Claude Code likely is correct that I should start to use NeonDB and Fly.io which I have never used before and do not know much about I wouldn't be so sure about that. In my experience, agents consistently make awful architectura
225.
by dvt 8mo ago
Yeah that comparison doesn't pass the smell test. Blockchain/crypto were purely financial instruments and for better or worse, a new financial instrument is very different than a new tech innovation; tbh there was a thin veneer of
226.
by dvt 8mo ago
> four bland messages That's why. Boring, bland, etc. That account's M.O. is basically "write a paragraph that says nothing." Fwiw, I do think AI can be indistinguishable from dumb, boring people, but usually those ki
227.
by dvt 8mo ago
> I usually tell them to just look at what game developers are doing for pathfinding. The example in the article doesn't quite apply, as baking pathfinding given a mesh (like baking lighting) isn't really the same thing as what
228.
by dvt 8mo ago
What excites me most about these new 4figure/second token models is that you can essentially do multi-shot prompting (+ nudging) and the user doesn't even feel it, potentially fixing some of the weird hallucinatory/non-determ
229.
by dvt 8mo ago
This is super cool, any papers I can read on autorouting? I occasionally see inefficiencies[1][2] so I suspect this isn't exactly a free lunch, and I'd like to read something a bit more critical about this approach. Either way, pr
230.
by dvt 8mo ago
It is, but to do the "useful" stuff, it's more or less mandatory. Either Python or bash scripts (which are equally as janky tbh).
231.
by dvt 8mo ago
I actually think SKILLS.md is such a janky way of doing this sort of thing, let alone the fact that's reliant on the oh-so-brittle Python ecosystem. Also way too much context/tokens being eaten up by something that could be piece-
232.
by dvt 8mo ago
We're still 6-12+ months away from a "killer" AI product. OpenClaw showed what's possible-ish, but it breaks half the time, eats tokens like crazy, and can leak all kinds of secrets. Clearly there's potential there,
233.
by dvt 8mo ago
Upvoting not because I necessarily agree, but because I think it's a conversation worth having. I personally think it's great that "everyone can build" now. I don't have to deal with "product people" telli
234.
by dvt 8mo ago
I know this is going to sound tinfoil-hat-crazy, but I think the whole thing might be manufactured. Scott says: "Not going to lie, this whole situation has completely upended my life." Um, what? Some dumb AI bot makes a blog post
235.
Mira: An agent that never forgets anything. Persistent, shared memory(co-span.com)
3 points by dvt 8mo ago|0 comments
236.
by dvt 8mo ago
A lot of times, this entropy decay is found in semantic or stylistic space, which would be hard to detect (you couldn't use, e.g., Shannon Entropy). You'd have to ask questions like "is this point uninteresting?" or &quo
237.
by dvt 8mo ago
Who implements those functions? E.g., store.order has to have its logic somewhere.
238.
by dvt 8mo ago
I actually do use cross-platform accessibility shenanigans, but for websites this is rarely as good as just doing like two passes on the DOM, it even figures out hard stuff like Google search (where ids/classes are mangled).
239.
by dvt 8mo ago
I’m working on a DOM agent and I think MCP is overkill. You have a few “layers” you can imply by just executing some simple JS (eg: visible text, clickable surfaces, forms, etc). 90% of the time, the agent can imply the full functionality,
240.
by dvt 8mo ago
The AI-assisted GPT theoretical physics derivation? There's literally no mention of AGI in the article, and it's pretty tame, especially considering it's a PR piece by OpenAI.

More ›