kolinko

21 ms·

481.
by kolinko 2y ago
This will make Google the new Nokia. Especially since chatgpt has a paid plan, so they don’t need to care about adding ads to the results.
482.
by kolinko 2y ago
the code is sandboxed on openai servers. it doesn’t run on your machine if you use chatgpt interface
483.
by kolinko 2y ago
I may be wrong but with midern llms biases aren’t really used any more.
484.
by kolinko 2y ago
Sold out :(
485.
by kolinko 3y ago
How many high net worth individuals do you know / have you met? Most of the ones I know consider money as a tool that can be used towards something, not as a goal by itself.
486.
by kolinko 3y ago
The contradiction is still valid in big step evolution. Kind of like guessing a passcode to a lock. You still try all the combinations, but there are no intermediate steps - you either guess it all at once, or not at all. If we see a bunch
487.
by kolinko 3y ago
For hydrogen in Europe to be considered green - if I’m not mistaken - you need not just carbon neutral sources, but also you need them built specifically for the purpose of capturing hydrogen. It can be tracked.
488.
by kolinko 3y ago
Above certain threshold of solar in the grid it’s not necessarily so. Also there are locations that are good for solar and available but have terrible grid connecting capabilities.
489.
by kolinko 3y ago
Usually people who post such claims haven’t used anything beyond gpt3. That’s why you get questions. Also, the difference is so big and so plainly visible that I guess people don’t know how to even answer someone saying they don’t see it. T
490.
by kolinko 3y ago
perhaps open source gpt3.5/4? I remember OpenAI had that in plans - if so, it would make sense for them to push alignment higher than with their closed models
491.
by kolinko 3y ago
At that time you even had bitcoin faucets that gave you 0.5-1-5 btc to play around. Or you could just run a miner on your gpu and get 50btc if you ran it for a few days or so - a ton of people probably have those on some partition on an old
492.
by kolinko 3y ago
This. The moment I switched my mom to Mac OS and iPhone, literally all issues she had with computer disappeared. She still needed me to help with Photoshop install and backups, but besides that - no problem for the last 7 years. And running
493.
by kolinko 3y ago
M1/M2s are around 2-4 trillion if I’m not mistaken
494.
by kolinko 3y ago
> Compared with a dense Matrix-Vector Multiply implementation, your algorithm adds algorithmic complexity, but reduces memory traffic. Because Matrix-Vector is usually memory-bound, your throughput increases when you can reduce memory ac
495.
by kolinko 3y ago
I found gpt to be awesome in this though - it's good at mapping some fields into others. While working, I kept asking it about certain ideas, and if the idea was known to him, he said "oh, that's this and it's called thi
496.
by kolinko 3y ago
Yeah, more tests are needed. I got some feedback on using KL instead of the token similarity - initial tests seem to show that it is workable (compared to Q8), but not awesomely amazing - will be working on that next week and publishing. As
497.
by kolinko 3y ago
YES! That was an idea I had a year ago actually! I wanted to lobotomize models, cutting out weights and leaving only functionality that we want. Unlike distillation it would be a much more brutal process. The idea went nowwhere for the time
498.
by kolinko 3y ago
I think it was somewhere around that tag: https://github.com/kolinko/effort/releases/tag/5.0-last-mixt... Cannot rerun easily any more, because the underlying model/weight names changed in the meant
499.
by kolinko 3y ago
I looked into it at the beginning, but as far as I understand, the modern models like Mistral are difficult to do LoRA on - you can use it to finetune, but the model itself doesn't lend itself to such an operation. I'm still quite
500.
by kolinko 3y ago
True, it will be easier next time because new llms allow to better searching through papers. Having said that - the tradeoff with getting into a new field is that there is 50 years of history to dig through, and by the time you know exactly
501.
by kolinko 3y ago
Thx! Although I'm happy you took your time and now I can brag until the end of time that I got 2 out of 5 top slots for a moment :D
502.
by kolinko 3y ago
Depends on a field. The project from the thread would take me an impossible amount of time without GPT. Even the page itself would take me twice as long to generate - the charts were done by pasting source data to GPT and GPT writing plotli
503.
by kolinko 3y ago
If you got downvoted, it was briefly. That was a good question. PyTorch/numpy is awesome for initial testing.
504.
by kolinko 3y ago
It should work now :)
505.
by kolinko 3y ago
Yes! I like that, and I saw that paper last weekend iirc. I think these MoD/MoE and other similar methods are highly compatible, and in a similar style. I was originally afraid that this method wouldn't be compatible with MoE and
506.
by kolinko 3y ago
That was one of the ideas I was also considering originally! Don't remember why is skipped it though - may have been because I tried the published one and it worked well enough during the first tries. As for indices for weight - not su
507.
by kolinko 3y ago
Nice analogy, thanks! Yes - you still need a map, because one day you need this road, another day you need info about another road, but you're not using info about all the roads all of the time.
508.
by kolinko 3y ago
A friend who first heard of the method immediately suggested this may work, perhaps even better in Diffusion models. It is really a drop-in replacement for the regular matrix multiplication. The data structure is a bit more painful to work
509.
by kolinko 3y ago
Below Q8 I think it will combine poorly - bucketMul needs to keep a few bits for an index. It can be offset by the fact that the numbers from each bucket are from a limited range. My intuition is that with Q8 this trades off roughly evenly
510.
by kolinko 3y ago
What would definitely help would be a better understanding a GPU handles myVal - does it go to the cache, threadgroup memory or where. Or perhaps decompilation of the source kernel to see what's really going on there. If we knew, we co

More ›