Aleksa Gordić (水平问题)
@gordic_aleksa
Followers
22K
Following
7K
Statuses
4K
x @GoogleDeepMind @Microsoft proud father of 16 H100s flirting with LLMs, tensor core maximalist
🇺🇸 silico 🇺🇸
Joined September 2017
llm.c gang w/ @karpathy, Erik, and Arun: aka "avengers" and since Andrej's yesterday talk the "4 pandas team" lol 😂
24
29
1K
@iScienceLuvr folks just enable 2 factor authentication and you're safe btw. same person reached out to me, they'll send you calendly and it'll be added to your x profile
0
0
0
flying to usa 🇺🇸 if any frens in Las Vegas lmk will spend some time there! :)
0
0
4
"let's wait step by step" here we go again, buckle up
1
1
20
do you now realize just how out of touch the AI safety community really is? ffs they made a big deal out of GPT-2 DeepSeek drops pareto optimal 671B model and almost no one is talking about how it's going to undermine democracy, enable proliferation of nuclear weapons, and turn us into paper clips it's always been about competition (although i believe that those ~30 folks in lw forum really do believe it)
6
2
56
wait - is all you need?
If, during the RL phase, you interrupt thinking and append "wait," to the reasoning traces you bend the cost curve and get within a point of R1 distill 32B with just 1k high quality examples
2
3
40
@giffmana is this the equivalent of DeepSeek used $5M to train their models or quite the opposite - they also priced in the budget for snacks?
0
0
1
create a market around this: people who want to attempt to jailbreak the system have to pay $ X and whoever jailbreaks it first gets the pool prize the group you'll attract by doing this will be sufficiently different than the volunteers red teaming it
New Anthropic research: Constitutional Classifiers to defend against universal jailbreaks. We’re releasing a paper along with a demo where we challenge you to jailbreak the system.
3
0
13
<3
How do you feel, @gordic_aleksa, that google search now defines you as #1 example of "skill issue" in Chinese?
0
0
21
the original blog post "the short case for nvidia stock": the misleading tweet: medusa paper:
🚨 Each AI engineer at Meta AI earns more every year (>$5.5 million) than it took to train China's Deepseek. Despite this, Meta's Llama model is far behind Deepseek. American engineers are now rushing to copy the methods of a small Chinese engineering group.
0
0
3