How Deepseek v3 made Compute and Export Controls Less Relevant

[@PunitPandey]

For someone who can not afford that much on reading paper, your video is just amazing. You are an excellent teacher.

[@ArunRamakrishnan]

This was brilliant analysis from a systems perspective. Looks familiar to how we designed our large NUMA machines in terms of prefetchers, branch predictors, directory caches and what not.

[@SiD-hq2fo]

amazing video, the quality of content is insane, please keep them coming!

[@TrelisResearch]

cheers! appreciate that

[@86dansu]

What a solid video ❤

[@wilfredomartel7781]

Has anyone successfully executed the model? Perhaps there's a way to further optimize this Kraken of a model?

[@TrelisResearch]

What do you mean? I show running it on remote GPUs. For local check out unsloth but you still need lots of compute

[@rupnikj]

great job! keep it up!

[@linuxdevops726]

I love the way you explain things I appreciate that

[@wayne8863]

US puts sanctions on GPUs, but Chinese company uses limited resources to build a powerful model and open source it to the world.

[@Pure_Science_and_Technology]

Yup, this has been going on for some time. Models like this are definitely putting pressure on the big US tech giants, forcing them to slow down. I’ve got an inference server with 4 Nvidia RTX 6000 ADA GPUs, but even with the INT4 quantized version, I can’t run DeepSeek-V3 locally. Guess I’ll invest in these online GPU rental companies. Lol.

[@pewpewpistol]

what's insane is that US keeps compulsively sanctioning

[@ArunRamakrishnan]

The sanctions wont simply work as the mathematical models are trying to seek computational solutions and at some point they can jump purely to CPU based ones if needed given enough cores and vector processing assistance. All the supercomputers started with custom accelerators (Blue Gene etc) and these AI clusters will go down the same given the ginormous scarcity for GPU boards.

[@dr.mikeybee]

@@pewpewpistol rather than supporting Nvidia, we forced the Chinese to create a competitor. I don't see how this is helpful.

[@brentknight9318]

Super helpful

[@mmasa1]

instruction following in English definitely not there.. as for chinese, it seems much better. I wonder if these stats were made using chinese?

[@TrelisResearch]

Those stats are mostly English benchmarks. But yeah the vibes seem to be that instruction following is not as good as Claude in English

[@DDD-yi2ze]

Did you test yourself or you just reading slides?

[@TrelisResearch]

Howdy! what dyou mean?

[@Little-bird-told-me]

Why do I get this feeling that models are becoming of programmers, by the programmers, for the programmers.

[@TrelisResearch]

Ah yeah I see now what you mean. Indeed it’s quite coding focused in terms of performance - although that tends to help everything

[@Little-bird-told-me]

@@TrelisResearch Do you think the future model would be based on clever design rather than brute force compute. Looks like we are entering the territory of law of diminishing returns, and Deepseek seems to proving it

[@TrelisResearch]

@@Little-bird-told-me well there is tons of room left for improvement. Human brains are far more energy efficient per unit compute than LLMs.

[@slyracoon23]

Do you have a patron or something. I want to support the channel and the content but buying a life-time membership is a little too much.

[@TrelisResearch]

Thanks! Yeah I try to focus on just products rather than be donation based. There’s still this kofi here though: ko-fi.com/trelisresearch

[@slyracoon23]

@ just donated. Keep up the good work!

[@semtex6412]

try asking it "what's the most significant and memorable event in China in 1989?" i promise you it won't answer it

[@suresht1885]

😅

[@TrelisResearch]

It’s true the censorship can be an issue

[@GIRcode]

It's actually only an issue if you use Chinese API sources. DeepseekV3 is less censored at the training level than previous Chinese llms.

[@TrelisResearch]

@@GIRcode yeah actually that's an important distinction