kingstnap

9 ms·

151.
by kingstnap 6mo ago
Slack is a synonym for redundancy but is also a synonym for inefficiency.
152.
by kingstnap 6mo ago
Very strange article. H100s aren't going to be E-waste anytime soon. Unless they die from HBM vias cracking or something someone (e.g. me) will want them for local AI even after decommissioning them from a datacenter. E-waste is almost
153.
by kingstnap 6mo ago
To be fair, what's the practical application supposed to be for proving a user is a bot? Silly solutions for silly problems :^).
154.
by kingstnap 6mo ago
It doesn't need to be. Text can be tokenized in many different ways even if the token set is the same. For example there is usually one token for every string from "0" to "999" (including ones like "001" s
155.
by kingstnap 6mo ago
If you had a way to pay them I'm sure some someone in a poor country would take the offer. Pretty sure that's how this goes down. They don't really risk anything cause they can just leave the angry emails about piracy on read
156.
by kingstnap 6mo ago
By default CUDA isn't deterministic because of thread scheduling. The main difference comes from rounding order of reduction difference. It does make a small difference. Unless you have an unstable floating point algorithm, but if you
157.
by kingstnap 6mo ago
It's even more strange how its not obvious to someone who uses codex extensively daily. The rate limiting step is the LLM going down stupid rabbit holes or overthinking hard and getting decision paralysis. The only time raw speed reall
158.
by kingstnap 6mo ago
If you aren't pulling back how do you know you went far enough? Anyway the thing is that Opus 4.6 Fast is just too obscenely expensive for this to make sense. It's not the end of the world if the most ridiculously expensive stuff
159.
by kingstnap 6mo ago
Is it fear or just a checked out mentality? I can easily see devs at corps just doing whatever is asked by management to the minimum standard and calling it a day. Particularly if there is a corporate push for efficiency (tautologically tru
160.
by kingstnap 6mo ago
I looked it up for specifics. Right now the closest I can see is that $121/mo gets you 4 Xeon Platinum 8370C cores and 16GiB of RAM [0] (storage not included!). Somebody Geekbenched that config here [1] 1274 single core 4256 multi core
161.
by kingstnap 6mo ago
My own guess is that something like this happened: Claude in testing would interrupt too much to ask for clarifying questions. So as a heavy handed fix they turn down the sampling probability of <end of turn> token which hands back to
162.
by kingstnap 7mo ago
If you aggressively use all buckets Google is incredibly generous. In theory for one AI pro subscription you can get what is a ridiculous return in investment in a family plan. You could probably be charging google literally thousands if al
163.
by kingstnap 7mo ago
You can make unique local address subnets with simple addresses. I use a simple one like `fd10::1/128` and `fd10::2/128` and so on Technically speaking RFC 4193 says you should use random bits. But I don't care.
164.
by kingstnap 7mo ago
> They refuse to implement anything strictly required This describes a lot of businesses ngl. Bell in Canada is one huge head scratcher. They are one of the largest ISPs here and I can even buy 8 gig internet to my house if I want but th
165.
by kingstnap 7mo ago
All these companies use AIs for writing these prompts. But AI aren't actually very good at writing prompts imo. Like they are superficially good in that they seem to produce lots of vaguely accurate and specific text. And you would hop
166.
by kingstnap 7mo ago
This is interesting but the linked articles is even more interesting. https://trufflesecurity.com/blog/google-api-keys-werent-secr... > Even Google themselves had old public API keys, which they thought were non-sen
167.
by kingstnap 7mo ago
But in this case they both made posts that point to each other. https://bitwarden.com/blog/introducing-agent-access-sdk/#int...
168.
by kingstnap 7mo ago
A big difference is that its culture comes from shared public experience. Everyone sees the same front page not a curated one.
169.
by kingstnap 7mo ago
There is this distinct lack of giving a shit about the user that you see coming through in a lot of big tech nowadays. Take this extremely simple example about antenna pod. I can change the order and what buttons show up in the app nav bar.
170.
by kingstnap 7mo ago
Research has a large asymmetry to it. Once someone shows something can be done. Others can follow quite easily. And more substantial is that when someone shows something can be done, its orders of mangnitude easier for $ENGINEER at $CORP to
171.
by kingstnap 7mo ago
My guess is that probably not for Muon. What I said about ADAM was partly based on this blogpost I read some time ago, should have cited it as well [0]. The thing about Muon is that it doesn't have this specific feature of ADAM that ca
172.
by kingstnap 7mo ago
Other people have answered here but the real answer is that deep neural networks don't learn isotropic distributions of activations. What happens is that you get very spikey activations, there are so called "outlier" activati
173.
by kingstnap 7mo ago
I do wonder about the usefulness about this massive context dumping exercise. 100M is a ridiculous amount. Usually to get good results on practical tasks you need to actually think about what you are dumping into context. I also have my gri
174.
by kingstnap 7mo ago
If you get some more experience you will find normal journals are exactly like that as well.
175.
by kingstnap 7mo ago
arXiv doesn't need much. All they do is host static pdfs uploaded by someone else with free CDN services from Fastly [0]. I'm sure they could get academics to volunteer moderation services as well. In reality you could host the en
176.
by kingstnap 7mo ago
You can have locally zero sum effects. Either you fight over a saturated market. (Zero sum) Or you try to grow the market. (Non zero sum) Or you try to diversify into new markets. (Mixed, non zero sum if its greenfield research into new tec
177.
by kingstnap 7mo ago
I mean why would they. If they lived in fantasy land with reasonable permitting and running a toilet ment paying for toilet paper, soap, and a janitor to clean it once a day I'm sure they would for the PR win and to sell a few extra Te
178.
by kingstnap 7mo ago
In my experience LLMs often have really solid insights in the thinking chains then vomit a nonsense score that doesn't make sense. Now I'm not sure if this is actually an LLM only thing. Because I think people probably do similar
179.
by kingstnap 7mo ago
Single iteration waterfall is a broken process. You really need those late stage usage feedback signals unless your requirements were somehow captured by God.
180.
by kingstnap 7mo ago
We already have/had recursive self improvement in technology. Its just mutually recursive self improvement. And it will continue to be mutually self recursive because its not like the only input to AI is AI. Its a lot of things.

More ›