Bert is being used by Google for search under the hood. It's how theyve got that instant fancy extractive answers box. I don't disagree that LLMs are large. So was Saturn V.

Nah, there's already work that can reduce generic LLM model size by a half and not lose any performance. And LLMs I think will be great as foundation models for training more niche smaller models for narrower tasks - people already use openAIs API to generate data to fine-tune their own niche models. I think we'll look back at current LLMs and realise just how inefficient they were - though a necessary evil to prove that something like this CAN be done.

I think they meant: it is conceivable that in the future it could. i.e. you hook an LLM up with a repl. https://youtu.be/pdSfgRYy8Ao take at look at 15 minutes in. I could easily see how you could fine tune using self appraisal by executing code.

I feel like you'd make a really bad research student

Dude that's why you ought to put everything into NLP find a way of producing better results for cheaper on less expensive hardware and you'll be the talk of the town. I think everyone would love to have an unrestricted local version of chatgpt on their phones. Do the research!

Why the fuck are you coding on an iPhone - if you're going to use a phone at least be android

Neural magic does work in this space, not sure about KD specifically

Unlike those mere meddlers in misophonia inducing dribble I have a fist around actual physics and the universal laws it describes therein. Ask me of the classical and I will speak of Euler and Lagrange, of the modern and I will speak with you in terms and diagrams Feynman. Ask me of the classical regime as we asymptomatically approach the speed of light and I will echo Einstein and Ricci. For in every vector, matrix and tensor therein speaks a compact notation that a thousand thousand new age hippies could not grasp.

It's all about the sweat of the mule. If you can acquire the musk and scent of a weary, spent and perilously close to death mule - then you have a fighting chance in this dogged world. Until such a time sit in quiet reverence for the holy. Take upon your spirit the art of the lettered. The most well lettered of the world hold the keys to extinction and salvation both and we are but at their whim.

You think I'm bound by a vocabulary of a mere 17000 entries in a database? No, I assure you that anything here confabulated is closer to hapax legomenon than anything spouted by a mere metal monstrosity.

I haven't tried hard at anything since the last time I tried and popped your mother's prolapsed anus back into its fissured and disfigured gaping hole with my iron thick rod. And even that was easier than you trying to rub your two brain cells together trying to formulate a response that didn't reek of jealousy. You ought not think you will not be caught dear boy. Your lack of real world hardiness will shine in a world where chatgpt will bolster the hard working and diminish the stature yet further of the lazy and rapturous folk such as yourself.

Well dear chap the real qualm you ought have is not with my circumlocution but rather that disaptitude is not a real word - ineptitude abounds on this subreddit. By the way, Moore probably had a better grasp of the English tongue than any of you one trick GPT powered ponies. Honestly, my time to flex the deltoid of my lascivious tongue upon a gaping and prolapsed tender anus has come. I dare any of you halfwit mongrels to generate speech as eloquent and grandiloquent with as many tangents and out of distribution ponderings as I.

Put it this way: Moore probably had no idea what twerking was. Now, given this lack of foresightedness what else can we say about his characteristic disaptitude for clairvoyance?

AI

your ability to use google maps effectively correlates pretty well with how well you understand paper maps. things like logitude, latitude, being able to read the legend, being able to understand what contour lines mean. If we teach students to be lazy and not understand these things then they won't even know how to prompt an ai effectively into doing what they want. Not to mention they'd be left completely helpless in situations where being able to read a map is vital - like if they're abroad and they have no internet access (offline map only, no guidance), which is not uncommon. Ofcourse, you might never encounter that situation if you're a lardy american.

Depends how successful you did act after your didactics and did you land the acrobatics?

Isn't the ban partially because they don't want students to miss out on actually doing the work themselves? Using AI is kinda like going to the gym and then getting hal9000 to lift 90% of any weight you're told to pick up.

None - the only difference is that it is one on one so it feels like that to the user. Education in a lot of places is broken because of class size.

Surprised no one embeds it like CLIP but for word definition pairs rather than word image. I'm thinking take word2vec as starting point.

BCIs are cool but not gonna cut my skull up for it, no thanks

Holy shit I think gpt4 will drastically reduce mean jerk time and maximize DAAT (dicks at a time) - how didn't these guys see that an LLM was the solution to all their woes https://youtu.be/P-hUV9yhqgY

They've cut the number of requests you can make an hour, addresses the cost issue somewhat. I think they can plug the hole made by unexpected influx with their marketing budget. It's gotta be one of the most succesfull tech product launches of all time with number of unique and new users reaching the million mark within a week of going live.

It won't even write my university assignments now. I'm not joking. Even this very benign thing it's like "I cannot write a report". And yes I've tried getting around it but it refuses. I thought I was quite good at prompt engineering but this is just pure stubbornness that they've built in.

Sigma dad moment

Nah