sometimelurker

9 ms·

181.
by sometimelurker 5mo ago
> They statistically reproduce text written by other people explain why they can do multiplication problems they've never seen before? I can give them two 50 digit numbers right now and they'll multiply them. its because they g
182.
by sometimelurker 5mo ago
skill atrophy. also AI is actually helpful sometimes, by a very large degree.
183.
by sometimelurker 5mo ago
> ... didn't find any of their models remotely useful. I daily drive Ministral-3-3B-Reasoning, getting a nice 70 tok/s on my GPU. its actually much faster than Google/DDG searches and I use it a ton when coding to check me
184.
by sometimelurker 5mo ago
I don't think computers automatically make us more educated, but if you want to make a point don't use reductive exaggerations. > We need humans that can critically think from first principles to counter the recycled informati
185.
by sometimelurker 5mo ago
I think some attentive VCs couldve predicted this. Paper from OAI 2020 about scaling laws: https://arxiv.org/pdf/2001.08361 The idea that you can make a system smarter by just putting in more money is very appealing. T
186.
by sometimelurker 5mo ago
why are some bankrupted by casinos? how about online gambling?
187.
by sometimelurker 5mo ago
while I find myself agreeing with the posted tweet, I don't disagree that we need to communicate about things we disagree about better. Take my upvote
188.
by sometimelurker 5mo ago
thats the ideal ceiling, but I don't see why we won't surpass that. > People just need to calm down I think people need to make sure ceilings are built, and we can calm down once we're done.
189.
by sometimelurker 5mo ago
OAI has a billion users. They can train the model directly against the metric of how often those users use the model. (mutate it many times, test which one works best, keep that, repeat). The model would (did) learn to be sycophantic and pe
190.
by sometimelurker 5mo ago
The languages the agents will be best at are the ones that show up the most in the pretrain (python, bash, others)
191.
by sometimelurker 5mo ago
Do you believe that an AI can write persuasive text? Do you believe that an AI can be trained to elicit a specific user reaction?* Can we agree that AI companies are strongly incentivized to make money, and they can do so by making their sy
192.
by sometimelurker 5mo ago
Its a tool that is built by people who are incentivized to have you addicted to it.
193.
by sometimelurker 5mo ago
Il usually paste in a few wiki pages and have the machine reword the main ideas of whatever I'm trying to learn until it sticks. The trick is to manually get a ton of relevant context into the machine (look for explanatory blog posts a
194.
by sometimelurker 5mo ago
I've heard that its cheaper for cars to have touchscreens ("smart") than a panel of buttons. And TVs that spy on you are cheaper because they have the collected information to offset some of that cost.
195.
by sometimelurker 5mo ago
Just a thought, what if you added in a random steering vector at the start of the residual stream for each token? Intuition says it wouldn't act in the same way as increasing temperature would, but I honestly have no idea what would ha
196.
by sometimelurker 5mo ago
people still send reels to each other. that's not a lot of connection but its addictive so they don't mind.
197.
by sometimelurker 5mo ago
Running a super small ~3gb of ram model on my PC gives me faster responses than Google/DDG. good for quick programming questions (which is what I use it mostly for)
198.
by sometimelurker 5mo ago
Meta is only network effects. You could build that in a weekend, and so could I. or you could just fork Mastodon. What keeps people from leaving is that their friends are still on it: network effects
199.
by sometimelurker 5mo ago
I think there's a few things, but its a little subjective and its more about the style the ai uses when doing these than the actual specific behavior: - Nuggesting improvements to the code after finishing the task you gave it, very irr
200.
by sometimelurker 5mo ago
I'd like to chime in and mention that its really obvious how to RL a coding agent to get the human addicted asap. and its also clear that there's a ton of $$$ to be made by doing this. therefore its done. the only LLMs I use are t
201.
by sometimelurker 5mo ago
You need to be less etymology-pilled. Seriously tho, its a practical word choice in a lot of cases. puts emphasis on the 'maxxing' Think of it as claiming the word as your own.
202.
by sometimelurker 5mo ago
you still have to train all the stuff you want to run locally.
203.
by sometimelurker 5mo ago
here's a fun idea: with normal-person ram really expensive, and fabs are pivoting to making more HBM, what if a town built its own datacenter and gave access to it for the townspeople?
204.
by sometimelurker 5mo ago
you know I feel like there's some people so disillusioned with everything that they automatically assume that the only way to go forward is to try and work on the least ethical, most evil, nasty tech. There's actual thinking breat
205.
by sometimelurker 5mo ago
This might be related to the fact that fully automating AI safety can't be meaningfully done. And a lot of work is put into automating parts of it. Circuit-finding algorithms and SAEs are automated algorithms for interpreting parts of