The era of LLMs is over chuddies.We got a newer transformers papers moments.Chuds BTFO
>>109852146
>>109852146>captura de pantalla
I just had an insane case of deja vu with this post
>>109852146The fuck is Jev?
>>109852146The only thing I got from this is Google is raping the competition for a fraction of the cost
>>109852146>hypersaturated benchmark with n=49 where first 9 models get max score while the rest is measured arbitrarily that can go from 70.8 to 96.5 in "final score" despite all models getting 49/49embarrasing ad.
>>109852146so a "smart if statement" with a cheap LLM > frontier model?I assume the next gen of frontier models might bake some jev-like tool in it?
>>109852146What I expected>whoa you can like have a parallel jev for every pixel in every frame that assigns a probability to the color for that pixel and generate a movie!What I got>I can finally reliably split my ebook into different characters and narration with fewer hallucinations than bert
>>109852226Someone made a new model architecture where instead of outputting text, an LLM can only output a small set of pre-made decisions. This apparently increases speed by 20 to 200x, and reduces cost by 40 to 400x.
>>109852226Over the last few days people have discovered gliner/gliclass type zero shot classification.
>>109852146>112x lower costoh shit negative 111x cost? the datacenter pays THEM?
>>109852146Great. When can I run it on my consumer computer, offline and uncensored?
>>109852417Yes. It's so efficient it's GENERATING electricity.
>>109852146Not so fast LeCun. These are going to be used to augment agents, not replace LLMs
>>109852226Jeb bush, former US president (if the vote hadn’t been rigged against him)
>>109852146So instead of open sourcing it they try to make it a SASS. By implementing publicly available research and hoping nobody front-runs them?Yeah I'm sure that'll work out.
>>109852226a Jew without a v
muh benchmemes! broooos it gets big numbers in my benchmeme!
>>109852433You can be certain that the legion of chinksloppers are already swarming, but I'm pretty sure his plan is to get bought by a larger lab (which is now very likely).
People can complain, but if someone was able to convince investors to train a frontier-scale encoder only transformer, that's a massive win for everyone.
>>109852447>New architecture improves speed and efficiency by two orders of magnitudes>LMAO BENCHMARKSNigger what the fuck do you want
>>109852146Not gonna replace LLM because it doesn't do the same thing.Q. What color is the sky?Jev: Not enough informationQ. Do Humans breathe oxygen?Jev: Yes
You talk to it, and instead of random text, it returns a typed variable. But what if we go one step further and, instead of talking to a computer like it's a person, we use a language with strict rules that tells it exactly what to do and leaves no room for interpretation? Call it a "programming language" or something.
>>109852521Nobody is saying it does replace it because it is pretty hard to do that when it can’t code. What it does do is IMO really quite neat however and I am glad to see some step away from LLM shit at least. Jeb!+LLM should give us something much better than the sum of these parts and that is something actually exciting vs incremental improvements of the same shit over and over.If you take a look at how people are using it, this is an AI that makes you work to find uses but when you figure out something it has the capability for deliver dramatically better results than any frontier LLM, for literally pennies at that. AI that helps you do things faster or automates specific things way faster and cheaper than LLMs is a different niche than the general notion of you just wait there for it to one shot a project while you sit there dick in hand. Basically if you aren’t a vibecoder you can see jev’s immense potential. And if you aren’t, it can still help, you just aren’t as intuitively able to see how.
>>109852636>Jeb!+LLM should give us something much better than the sum of these partsEasily. Probably a step change in capability once the tedious stuff gets sorted out. Did you see Astra orchestrating Jeb in order to play minecraft? That was a day 1 implementation.
>>109852226jev is a Congolese-Canadian rapper. A refugee, he went viral on TikTok for his 2022 single "where's the confetti", which preceded his first studio album the color grey
Can I self-host a jev model?
>>109852773We're probably going to have to wait for China to copy it
>>109852773just use a regular LLM and say that you need a reply with a single word. Done.
>>109852403how is that different to just instruct mode?and the prompt that restrict answer to single word answer
>>109852580I think you're on to something, anon. Someone should work on this.
>>109852837https://docs.typesafe.ai/introductionThe tl;dr is it's trained to answer multiple choice or boolean queries given input text. It doesn't perform autoregressive text generation and instead evaluates the queries in parallel. The possible outputs are determined by the queries, i.e. the selected choice and probably and confidence scores. It's like a semantic classifier.
>65k context limitclose but not quite
>>109852146>the destpart hope of a smalbrain for a nubrain That's why you are easy to con
