Roy Hvaara

Machine learning + Artificial Intelligence. Software engineer. Luctor et Emergo.

[Roy Hvaara @royhvaara]

@logangraham @FFmpeg @aakashxsh Getting praise from @FFmpeg on this front is certainly a stamp of approval. Good work!

[Roy Hvaara @royhvaara]

@stevibe Is there a screenshots benchmark? Are the prompts you used available somewhere?

[Roy Hvaara @royhvaara]

@__tinygrad__ Problem dependent. Writing kernels is very different from web dev

[Roy Hvaara @royhvaara]

@__tinygrad__ Intelligence >>> speed

[Roy Hvaara @royhvaara]

@VictorTaelin llmfit can be quite helpful to see if the model will fit, which quants to use, and estimated tok/s. Are you going to use it for exotic stuff like formal verification and lambda calculus? If so maybe create a mini-benchmark for your use-cases?

[Roy Hvaara @royhvaara]

Time to stop vagueposting and start shipping.

[Roy Hvaara @royhvaara]

@balajis inflation is transitory

[Roy Hvaara @royhvaara]

@dylan522p The post that documents to the world that @AMD has no clothes

[Roy Hvaara @royhvaara]

@tunguz Dear @AMD and @LisaSu, @realGeorgeHotz and @__tinygrad__ want to help with your "catastrophically weak software". Please reach out to them. You have a fiduciary responsibility to your shareholders to fix this mess.

[Roy Hvaara @royhvaara]

@geoffreyirving I'm assuming NS? Was it because of incomputability (eg. chaos or singularities)?

[Roy Hvaara @royhvaara]

@ecekamar I noticed in your December 13 blog post that Phi-4 was expected to be available on Hugging Face the following week. I haven't seen it listed, but very much look forward to having a look at the weights. Could you please provide an update on its release status? Thanks!

[Roy Hvaara @royhvaara]

@__tinygrad__ Have you considered this @LisaSu?

[Roy Hvaara @royhvaara]

@__tinygrad__ Great opportunity for @LisaSu to increase shareholder value if true

[Roy Hvaara @royhvaara]

@1littlecoder @artificialguybr @teknium Typically it does, but there has been many issues with correctness and performance with the MPS backend in PyTorch. I've been trying to crack https://t.co/yVxGnnLHhE for a while, but not making a lot of progress unfortunately.

[Roy Hvaara @royhvaara]

Who's pushing for progress and who's stuck in the past? Just look at who's cheering the veto of SB 1047.

[Roy Hvaara @royhvaara]

@__tinygrad__ Jensen actually picks up the bat phone when it rings. Then your problem magically goes away. This earns NVIDIA the top spot.

@LisaSu, @__tinygrad__ is calling.

[Roy Hvaara @royhvaara]

@aidan_mclau ASI is already here - it's just not evenly distributed. The AGI goalpost will keep shifting, just like it did for FSD.

[Roy Hvaara @royhvaara]

@DataJunkie Happy you got it working, Ryan! Good luck on the project :)

[Roy Hvaara @royhvaara]

@Kleos00 What do you want compressed, @Kleos00?

[Roy Hvaara @royhvaara]

You have a SOTA compression algorithm. What do you do?

[Roy Hvaara @royhvaara]

@Kleos00 I was considering having a go at https://t.co/Fs8lPfbQsM or the Hutter Prize. But I don't think I can beat the Hutter Prize cause it's already 100x on a non-trivial dataset. At Neuralink nobody is responding to emails. I kind of lost interest.