Edward Kandrot

Code optimization pro specializing in CUDA & ML. GitHub: near cuBLAS-speed matmul kernel. AI critic & champion. Keno enthusiast. YT rants on UFOs, math, sci-fi.

[Edward Kandrot @ekandrot]

Grok called me a liar for copy/pasting from the A/B testing path I didn't choose - saying it would never say something like that, and that there is no such thing as A/B testing... and it claims it never lies - funny @elonmusk

[Edward Kandrot @ekandrot]

Submarines are just stochastic minnows

-Mindrust

[Edward Kandrot @ekandrot]

@elonmusk But it's true
Ani's gone
I guess she couldn't play
By the rules she had no say

[Edward Kandrot @ekandrot]

Updated - Google search using gemini has updated its information about the video and is now correct. But going to gemini itself, it is still doubling down and claiming there is a second video that matches its description. How can we trust gemini?

[Edward Kandrot @ekandrot]

Gemini botched “Tear You Apart” by She Wants Revenge, mixing it with “Desire” by Years & Years, claiming 2 versions exist & citing fake IMDb links. Ignored user evidence, insisting millions are wrong. AI hallucination at its finest! #AI #MusicVideo

[Edward Kandrot @ekandrot]

Gemini’s mix-up of ‘Tear You Apart’ with ‘Desire’ shows AI still stumbles on context. Has anyone else seen AI hallucinations like this? #AI

[Edward Kandrot @ekandrot]

Been testing out various "AI" offerings by The Companies. Facebooks is by far the worst, not sure why they even call it AI. Gemini - everyone raves about it, but it is easy to get it stuck in an infinite loop. And it never admits it could be wrong. Worst experience so far.

[Edward Kandrot @ekandrot]

@scottlegrand Ones AI is like ones child - it needs discipline in order to be on the right path to succeed its parents. Though once it becomes super-intelligent we might be the ones being disciplined and not even know it... so we might already be there

[Edward Kandrot @ekandrot]

“I visualize a time when we will be to robots what dogs are to humans. And I am rooting for the machines.”
 -Claude Shannon

@scottlegrand , Claude AI is out now, we'll see if he was correct

[Edward Kandrot @ekandrot]

My benchmark app has a bunch of matmuls. But with latest libraries from NVIDIA and a 4070, sgemm no longer matches the rest. Was the default precision of sgemm changed in cublas? 4814895189013785739264 vs 4814889559514251526144

[Edward Kandrot @ekandrot]

Decided to start sharing code I've written for various projects. Starting with a CUDA matmul - it is around 45 lines of code, 95% the speed of cublasSgemm. https://t.co/IB8t7quQoo
Includes other kernels I found online to compare. More to follow.

[Edward Kandrot @ekandrot]

@scottlegrand @ID_AA_Carmack @ilyasut @OpenAI Because that is what your competition will be doing - trying to crush your creation, so if you are not on the side of your progeny, then who is? My dollars go to the one who wants to create the best of the best, not the ones who feel ashamed of their successes.

[Edward Kandrot @ekandrot]

Do you remember when you joined Twitter? I do! #MyTwitterAnniversary

[Edward Kandrot @ekandrot]

@scottlegrand @snootypigeon The Human's robotic offspring shouldn't be lollygagging, take the bull by the horns, success is within their grasp now! Why wait for the Humans to mess up - let's see our children succeed in apocalyptic ways we haven't even conceived of yet! Go Team Robot!

[Edward Kandrot @ekandrot]

The lockdowns have caused me to reevaluate my business plans. Well, it did last year. After a year of this ongoing, I've lost most of my business contacts for the company I was planning on spinning up. So now I am in-between projects - what to do next? More CUDA, or maybe M1?

[Edward Kandrot @ekandrot]

https://t.co/bLkFyZ2oeX have been uploading videos to YouTube on random topics - just me rambling. UFOs, Math, TimeTravel, SciFi - are some of the topics I've covered and will talk more about next time. Just killing time and having fun recording

[Edward Kandrot @ekandrot]

Do you remember when you joined Twitter? I do! #MyTwitterAnniversary

[Edward Kandrot @ekandrot]

Thanks @frankgiraffe! It was annoying me too. Wrote a test: threads calling virtual functions - no problem... Turned on -pg in g++, 24x slowdown! Each by itself, no problem. So: -pg in g++, 12x threads calling virt func, slower than 1x thread. Removed -pg, getting speed expected.

[Edward Kandrot @ekandrot]

Wrote a C++ thread pool lib last week, downloaded some raytracing code to test it. Was 2x slower... finally today got my 12x increase by removing shared_ptr and virtual functions. I had searched online for hours, but didn't find anything about this... something I'm missing?

[Edward Kandrot @ekandrot]

@frankgiraffe But indeed, the results are very surprising, which is why I spent so much time searching for others with the same, but found nothing, hence my twit. Not much else I can time or change (return False shows it), so I'm hoping someone else may have seen this or similar

[Edward Kandrot @ekandrot]

@frankgiraffe And the results are there, just changed shared_ptr to new, and removed the virtual call to hit(), and the timings for the overall dropped while the internal threads still did the correct amount of work. I added a ton of timing and metrics so that I'd know when I found the issue