the writer Nina Illingworth, whose work has been a constant source of inspiration, posted this excellent analysis of the reality of the AI bubble on Mastodon (featuring a shout-out to the recent articles on the subject from Amy Castor and @[email protected]):
Naw, I figured it out; they absolutely don’t care if AI doesn’t work.
They really don’t. They’re pot-committed; these dudes aren’t tech pioneers, they’re money muppets playing the bubble game. They are invested in increasing the valuation of their investments and cashing out, it’s literally a massive scam. Reading a bunch of stuff by Amy Castor and David Gerard finally got me there in terms of understanding it’s not real and they don’t care. From there it was pretty easy to apply a historical analysis of the last 10 bubbles, who profited, at which point in the cycle, and where the real money was made.
The plan is more or less to foist AI on establishment actors who don’t know their ass from their elbow, causing investment valuations to soar, and then cash the fuck out before anyone really realizes it’s total gibberish and unlikely to get better at the rate and speed they were promised.
Particularly in the media, it’s all about adoption and cashing out, not actually replacing media. Nobody making decisions and investments here, particularly wants an informed populace, after all.
the linked mastodon thread also has a very interesting post from an AI skeptic who used to work at Microsoft and seems to have gotten laid off for their skepticism
holy christ shut the fuck up
uhh did OP consider my hopes and dreams, powered by the happiness of literally every single American child? im guessing no. what a buffoon
Iirc it still couldn’t do that, if you create variants of jokes it patterns matches it to the OG of the joke and fails.
Euh what, various creative tasks have been done by AI for a while now. Deepdream is almost a decade old now, and before that where were all kinds of procedural generation tools etc etc. Which could do the same as now, create a very limited set of creative things out of previous data. Same as AI now. This chatgpt cannot create a truly unique new sentence for example (A thing any of us here could easily do).
What ?
Of course it can, it’s randomly generating sentences. It’s probably better than humans at that. If you want more randomness at the cost of text coherence just increase the temperature.
People tried this and it just generated the same chatgpt trite.
you mean like a Markov chain?
These models are Markov chains yes. But many things are Markov chains, I’m not sure that describing these as Markov chains helps gain understanding.
The way these models generate text is iterative. They do it word by word. Every time they need to generate a word they will randomly select one from their vocabulary. The trick to generating coherent text is that different words are more likely to happen depending on the previous words.
For example for the sentence “that is a huge grey” the word elephant is more likely than flamingo.
The temperature is the way you select your word. If it is low you will always select the most likely word. Increasing the temperature will make the random choice more random giving each word a more equal chance.
Seeing as these models function randomly there is nothing preventing them from producing unique text. After all, something like jsbHsbe d dhebsUd is unique but not very interesting.
I don’t get particularly excited for algorithms from 1972 that come included with emacs, alongside Tetris and a janky text adventure but that is indeed the algorithm you’re rather excitedly describing
snore I guess
The pattern matching problem can typically be sidestepped by using emoji representations.
LLMs are trained by prediction, so one of the practical shortcomings is that slight token variations largely get ignored when the surrounding tokens are too similar to training data.
If you change the tokens to less common symbolic representations like emojis, the same issues that trip up naive variations often won’t.
Also, chain-of-thought prompting that gets the model to repeat the nuances in the variation can go a long way towards overcoming similarities to a normal form of the query.
A good example of this problem are variations on the “wolf, goat, and cabbage” problem to get to the other side of the river.
When GPT-4 first came out, initially there were a bunch of comments about how it couldn’t solve variations of the problem. But restating the prompt using “🐺, 🐐, 🥬” in place of the words and asking it to repeat any associated adjectives when mentioning the nouns as it worked step by step through the problem would have it solve variations correctly every attempt.
As I said earlier, part of the apparent shortcomings in SotA LLMs is that we’re still learning how to maximally use them.
deleted by creator
Teach me!