samrus

24 ms·

511.
by samrus 1y ago
Its almost there. The difference is the tinder like UX of swiping to input approval data, and being able to programmatically export the approval data so other apps can use it without you having to do it manually Its not impossible that thos
512.
by samrus 1y ago
I know you think its just social media smartass, but the guy specifically mentioned getting the opinions of a private group, and exporting the data so as to post the best photos on a completely seperate platform. Its not social media, its k
513.
by samrus 1y ago
But the original idea is to crowdsource preference data and only publish the liked ones. Like screen testing a movie with a focus group. Tiktok doesnt really do that do they? Can yiu access interaction data from tiktok?
514.
by samrus 1y ago
This is obviously the designated fun thread
515.
by samrus 1y ago
The RLHF is happening too late i think. I think the reinforcement learning needs to be during the initial next token prodiction. On that note we need something to represent a complex world state than just language.
516.
by samrus 1y ago
> Many teachers cannot do that either. Of course there are bad teachers out there. The question wasnt "are there human yeachers as bad as an LLM" it was whether an LLM is as good as a good human teacher > We just need the W
517.
by samrus 1y ago
This is a somewhat childish comment. You didnt contribute anything to the discussion
518.
by samrus 1y ago
LLMs dont have a deep enough understanding of theory of mind to see how someone is stuck and help them get unstuck. You need to guide your own study and you might not know what you need to learn to get unstuck
519.
by samrus 1y ago
This is very true. Constant insecurity for me. One thing that helps a little is asking it to search for sources to back up what its saying. But claude has hallucinated those as well. Perplexity seems to be good at being true to sources, but
520.
by samrus 1y ago
Its not scaffolding if the intelligence itself is adding it. Humans can make their own diagrams ajd maps to help them, LLM agentsbneed humans to scaffold for them, thats the setup for the bitter lesson
521.
by samrus 1y ago
But are LLMs the right models to even be able to learn such long horizon goals and how to not cheat at them? I feel like we need a new base model where the next token prodiction itself is dynamical and RL based to be able to handle this iss
522.
by samrus 1y ago
Tgats running into the bitter lesson again. The model should genwralize and understand when its reached a road block in its higher level goal. The fact that it needs a uuman to decide that for it means it wont be able to do that on its own.
523.
by samrus 1y ago
The AI space is attracting alot of grifters. Even the initial announcement for devin was reaking of elon musk style overpromising. Im sure the engineers are doing the best work they can. I just dont think leadership is as interested in maki
524.
by samrus 1y ago
No. Because there would have been indtances in the data where silence was labelled correctly. But the model couldnt handle the null case, so it over fit on the outros. But generally it fit on the random error in the label of the null featur
525.
by samrus 1y ago
they did mention that it would make up fake transactions to balance the book
526.
by samrus 1y ago
and investors are frothing at the mouth to put a jackhammer in every home
527.
by samrus 1y ago
yes but i cant open that at work
528.
by samrus 1y ago
if someone says "which is the bigger number" with no context you wont assume software version number, lets be real here also $9.9 is clearly 9 dollars and 90 cents.
529.
by samrus 1y ago
humans can operate in dynamical systems (where your actions can change the underlying system). LLMs are not trained to do that and have shown to be terrible at it
530.
by samrus 1y ago
ive found that the shorter the "task horizon" the more time saved essentially, a longer horizon increases chances of mistakes, increasing time needed to find and fix them. so at one point that becomes greater than the time saved i
531.
by samrus 1y ago
this exactly right. remember, these models were trained to be functions. f(x)=y. thats an interface at its heart. when x and y are language, then its a translator. they have emergent capabilities, like "translating" instructions&#
532.
by samrus 1y ago
Most small businesses also cant afford the risk of current LLMs putting garbage in their books that, in the best case, has to be cleaned up or redone, or, in the worst case, gets the IRS up your ass
533.
by samrus 1y ago
fitting on noise in the training data is exactly what overfitting is. underfitting is smoothing out signal
534.
by samrus 1y ago
Bit fantastical. We are a general intelligence and we dont understand ourselves
535.
by samrus 1y ago
i dont like musk but the driver it there during this initial phase to take over in emergencies right? they arent actually driving the whole time? i do agree he should say its in early access or something
536.
by samrus 1y ago
can binocular cameras replace lidar you think? they should result in just as reliable distance estimation
537.
by samrus 1y ago
the thing about waymo is that i suspect they're running the same ML fraud that tesla itself is running in the silicon valley in general, which is to overfit on the 20% of situations that occur 80% of times. for waymo itself, you can ov
538.
by samrus 1y ago
hes always been like this. from faking video games to pushing memecoins. his being cringe itself isnt the issue. its the lack of EQ that leads to someone being a slave to wanting to be liked. it doesnt inspire confidence in his ability to t
539.
by samrus 1y ago
> just that they need to be given the right instructions. this is the part im talking about. i also think LLMs are very capable at detecting different types and levels or grammar, but they cant decide which ones should be filtered out t
540.
by samrus 1y ago
Well the maintainers might. The both in (badly) enforcing guidelines they set out. So they might want to police grammar a little bit. And i could see their point if the grammar is so bad it hinders communication. Bit uncomfortably, it was a

More ›