tveita

5 ms·

1.
by tveita 19d ago
As they say in TDD, write a test, then write the simplest code that will make it pass. Clearly supporting multiple functions starting with 'p' would be overengineering.
2.
by tveita 20d ago
The recurrent depth sounds a lot more like what is described in https://dnhkng.github.io/posts/rys/ - a way to add depth to a network without increasing the number of parameters. e: While the actual CoT in neurale
3.
by tveita 24d ago
You might find the home made batteries of this guy interesting as well https://www.youtube.com/watch?v=eq7fR9ISuCw
4.
by tveita 28d ago
It should be fine keeping the n-gram embeddings on SSD which lets you run at least the Q1 and Q2 models on 64GB
5.
by tveita 1mo ago
Interesting that he pointed it to the CSB video - those are great explanations for us laymen, but I assume the LLM would prefer a written report. Luckily it's still up https://youtube.com/watch?v=CFVUSDzHL8A I was worr
6.
by tveita 2mo ago
> Not sure if modern models "think" only by outputting <thinking> blocks That's pretty much it - a small refinement to "Chain of Thought" prompting, where you tell the model explicitly in the prompt to &quo
7.
by tveita 2mo ago
https://dam.stanford.edu/memory-prices.html Pretty sure those devs were saying the same thing way before 2017 as well, which seems to be ~ the last time RAM was this expensive. Now RAM in the cloud , now you're really
8.
by tveita 3mo ago
> Most cryptographers would still recommend the hybrid over pure ML-KEM. This RFC (for pure MLKEM) is marked "recommended to implement = N". It is purely for settings where the implementors independently want to use pure ML-KEM
9.
by tveita 3mo ago
It also gives you the option of serialising the RNG states directly instead of using the counter hack.
10.
by tveita 4mo ago
> who put the Transformer into LLMs? Google? > who invented neural networks People like Geoffrey Hinton, who was notably at Google Brain from 2013 to 2023? The people who say Google was ahead were paying attention long before you were
11.
by tveita 4mo ago
I think https://jamesclear.com/great-speeches/finding-your-own-visio... is a better take. What gives you an unique perspective and your own voice can be sticking to your thing for a long time and exploring your exact p
12.
by tveita 4mo ago
Would love to read specific examples of "the same trick being used to dismiss health concerns about medical supplements or influence financial information provided by Google's AI about retirement", but the relevant link in th
13.
by tveita 4mo ago
I can get a really high hit rate by only searching for dumb trivial things I already know the answer to.
14.
by tveita 4mo ago
The point of LMGTFY is to land people on either the official documentation or a curated site like Stack Overflow. Google used to be able to do that reliably. With the power of LLMs you can Google a standard library function and get an inacc
15.
by tveita 5mo ago
Is there any project that showcases Mojo for running neural network models on the GPU - like ideally something like llama.cpp that could run one or more existing models to showcase the readability and performance?
16.
by tveita 5mo ago
"Can you lay tiles until I say stop, or until it's about $250 worth, whichever comes first" "No, as one of the top tile layers in the country I can't do that, for your own protection. What if fifty elephants came an
17.
by tveita 5mo ago
Honestly I'd sometimes like a checkbox to ignore any results uploaded after, say 2023. Or else to only see 'verified non-AI' content. There is nothing I'd ever search Youtube for where an AI generated video would be an a
18.
by tveita 5mo ago
There are real examples of males and females being initially mistaken for different species, as well as for adults and juvenile forms. e.g. https://en.wikipedia.org/wiki/Cetomimidae In early 2009, the Royal Society
19.
by tveita 5mo ago
Building harnesses does seem like a task that's particularly conductive to psychosis. I've wondered if it's because there is no push-back; almost anything you try will be "right" in that your changes appear to make
20.
by tveita 6mo ago
https://en.wikipedia.org/wiki/Fill_line Selling drinks in mislabeled containers should warrant a fraud report to your local consumer protection agency. A crowdsourcing app seems like the wrong tool here.
21.
by tveita 6mo ago
"video compression" by analogy only, what this claims to actually do is delta encode the values in each token from the previous token. Interesting idea, but the results seem almost suspicious? even accounting for the extra bits us
22.
by tveita 7mo ago
Sure, for instance, if all of them go through an 1 hour AI interview, then you might find a better candidate, at the cost of 1000 man-hours of work. You hire that person, another company opens a position, gets 999 applicants, send them all
23.
by tveita 7mo ago
I absolutely agree in principle, but I understand that the companies are also seeing a lot more applicants trying to skate past screening and interviews with AI assistance. Connecting verified humans for a mutually respectful chat is a trus
24.
by tveita 7mo ago
> They are the ones who started using AI in the hiring process Aren't you ignoring the reports of companies receiving thousands of ChatGPT-written resumes, bots sending applications, and interviews with applicants being live coached
25.
by tveita 7mo ago
Probably more like the long tail of software - software that was created for a particular purpose in a particular domain by a single person in the company who also happened to know programming - maybe just as Excel macros. I strongly assume
26.
by tveita 7mo ago
There's some odd choices here. - 90 days is a very long time to keep keys, I'd expect rotation maybe between 10 minutes and a day? I don't see any justification for this in the article. - There's no need to keep an
27.
by tveita 7mo ago
I've definitely seen Opus go to town when asked to test a fairly simple builder. Possibly it inferred something about testing the "contract", and went on to test such properties as - none of the "final" fields h
28.
by tveita 7mo ago
> Elijah Newren, who wrote git's merge-ort (the default merge strategy), reviewed weave and said language-aware content merging is the right approach, that he's been asked about it enough times to be certain there's deman
29.
by tveita 7mo ago
I'll wager that 95% of incitive and unhelpful comments aren't written by "bad faith actors" as you define them, but ordinary people carried away by emotions or mob sentiment. Just a reminder that "this probably isn&
30.
by tveita 7mo ago
> Also, the generator will use a seeded randomizer so that, for local development, you work on the same dataset as others Except that the generator script generates dates relative to time() ?

More ›