HackerTrans
TopNewTrendsCommentsPastAskShowJobs

md_rumpf

no profile record

Submissions

Robots might be 1000x harder than superintelligence

maxrumpf.com
6 points·by md_rumpf·letztes Jahr·3 comments

Writing Code Just-in-Time

maxrumpf.com
2 points·by md_rumpf·vor 2 Jahren·0 comments

comments

md_rumpf
·letztes Jahr·discuss
My guess is a lot of these studies don't replicate as well in Europe vs. US.

Road signs in the US are predominantly text: PED XING. ONE WAY. DO NOT PASS. Where the European equivalents are all pictograms. Europe needs to do it this way, because the countries are so small you'd expect German-speakers driving in France and vice-versa: French-only text signs would be criminal. Similarly, a French-speaker can navigate a German train station without understanding German writing.

Europeans are actually much more attuned to deriving meaning from pictograms than their American counterparts.
md_rumpf
·letztes Jahr·discuss
Are there really use cases for faster chips? I can run all models I want on an H100 pod. No models exist that I can't run with at least 64 H100s. NVIDIA should just stop.
md_rumpf
·letztes Jahr·discuss
I think it goes hand in hand.

Knowledge work is made up (literally). It's all tasks that humans have created for themselves. We invented the maths, the programming languages, the spreadsheet software, the government form – everything.

The real world is very much not made up (especially if you're nonreligious). Most of the tasks are sampled from a distribution we did not invent and that means they might be much, much harder.

I think supersonic airliners might even be thinking too small. If we scaled speed like we did transistors, then if an Intel 4004 is like walking, an M2 Max would be like going 0.28c (c = speed of light)!
md_rumpf
·vor 2 Jahren·discuss
Well... How do humans synthesizs programs then? We don't do exponential search over some symbols. We "somehow know" which branches get us closer to a program and which don't. It's foolish to believe it is impossible to teach an AI to "somehow know" this, too. Also: For some reason the bar is always higher for AI. Next to no human code is verifiably correct. And the actually verified-to-be-correct code is an even smaller subset.
md_rumpf
·vor 2 Jahren·discuss
perplexity's valuation just doubled!
md_rumpf
·vor 2 Jahren·discuss
The return of the CPU?!
md_rumpf
·vor 2 Jahren·discuss
every 3rd sentsnce is "the model was not trained on data from meta's products"
md_rumpf
·vor 2 Jahren·discuss
the modality competition was one of my favorite insights, too!