nl

41 ms·

961.
by nl 1y ago
I do playing card generation and almost all struggle beyond the "6 of X" My working theory is that they were trained really hard to generate 5 fingers on hands but their counting drops off quickly.
962.
by nl 1y ago
> Infrastructure construction timelines are typically measured in years and decades. Years isn't actually a problem. How quickly do you think datacenters are built? It's entirely feasible to build new water infrastructure and a
963.
by nl 1y ago
Modern LLMs are post trained for tasks other than next word prediction. They still output words through (except for multi-modal LLMs) so that does involve next word generation.
964.
by nl 1y ago
This seems to be different to "it's odd" Yes, safeguards need to be improved. When they are does that mean it is suddenly not odd?
965.
by nl 1y ago
As it happens... I built E2E encrypted LLMs using secure enclaves, so I know a bit about this space. The tech works, for small LLMs - the sort of thing you can run on your mobile already. It isn't yet (?) there for LLMs the size of C
966.
by nl 1y ago
Why is it odd? Some people treat their pets like they humans. Not sure why this is worse particularly.
967.
by nl 1y ago
Signal does End-to-end encryption, so they (Signal) can never read it. The whole point of ChatGPT conversations is so they can be read by the model on the server. Conversations are kept around because they can be picked up and continued a
968.
by nl 1y ago
I've never heard of Iranian exit visas being particularly difficult to get.
969.
by nl 1y ago
If you are a rich crypto founder you can literally admit[0] to funneling money to Islamic State[1] and get yourself a pardon for it though[2]. [0] "Binance failed to file a SAR on transactions related to an individual designated by OFA
970.
by nl 1y ago
Can confirm there are reasonable numbers of Persian engineers with Australian residency.
971.
by nl 1y ago
I write like that and I'm not an LLM.
972.
by nl 1y ago
This is ASR not TTS though.
973.
by nl 1y ago
Sure but neither had enough memory to be useful for large LLMs. And neither were really consumer offerings.
974.
by nl 1y ago
They mean the ability to run a large model entirely on the GPU without paging it out of a separate memory system.
975.
by nl 1y ago
You probably can do this now. Non-generative LLMs don't need to be as big so something like Gemma 4B on the CPU will work. You may have better results with semi-templated responses though.
976.
by nl 1y ago
These are valid points but the numbers are still useful as a floor on performance. Given Strix Halo is so much cheaper I'd expect more people to work on improving it, but the NVIDIA tools are better so unclear which has more headroom.
977.
by nl 1y ago
Doubly interesting (and relevant to this discussion) is that it was an AI code review tool that detected the issues with the PR
978.
by nl 1y ago
> Just look at this shitshow. > https://imgur.com/a/lAd3UHn I don't follow NFL viewership numbers but searching and reading the results seems to indicate some support for that trend. What's wrong with i
979.
by nl 1y ago
Not at all. Any model with somewhat-similar architecture and roughly similar size should run at the same speed on Cerabras. It's like saying Llama 3.2 3B and Gemma 4B are fine tunes of each other because they run at similar speeds on N
980.
by nl 1y ago
Can confirm MiniMax M2 is very impressive!
981.
by nl 1y ago
Have a look at https://github.com/anthropics/claude-code/tree/main/plugins/... to see how a fairly complex workflow is implemented
982.
by nl 1y ago
Hopefully this will finally stop the continuing claims[1] that LLMs can only solve problems they have seen before! If you listen carefully to the people who build LLMs it is clear that post-training RL forces them to develop a world-model t
983.
by nl 1y ago
This is a very interesting piece. I'd never heard of "Baumol’s Cost Disease" but I think the thing it is appliesd to most often here in Australia is coffee. Everyone complains that it is $7 for a cup of coffee, and yet demand
984.
by nl 1y ago
> Pascal is not a low level language - quite the opposite. Why do you say that? Back in the 80s and early 90s it was the primary competitor to C, and in an alternate universe we might have ended up using Pascal decedents instead of C dec
985.
by nl 1y ago
I use Claude and Codex in VS Code and they work really well. > I select the text and Cmd-L, saying "fix this part, it's broken like this This flow works well.
986.
by nl 1y ago
This is completely fascinating although puzzling how that happens. The script is great!
987.
by nl 1y ago
I think you must be using it quite differently to me. I can one-shot new webapps in Claude and Codex and can't in Gemini Pro.
988.
by nl 1y ago
Not great. It's ok for documentation or small tasks, but consistently fails at tasks that both Claude or Codex succeed at.
989.
Coinbase CEO Exposes Prediction Market Vulnerability(finance.yahoo.com)
3 points by nl 1y ago|0 comments
990.
by nl 1y ago
> You will need a formal language first. No, that's the entire point! The LLM is the bridge between natural language and a formal specification. (WebPPL is a formal language btw. It's not unlike Prolog but is designed from the

More ›