One of the things I'm finding so interesting about large language models like GPT-3 and ChatGPT is that they're pretty much the world's most impressive party trick
All they do is predict the next word based on previous context. It turns out when you scale the model above a certain size it can give the false impression of "intelligence", but that's a total fraud
It's all smoke and mirrors! The intriguing challenge is finding useful tasks you can apply them to in spite of the many, many footguns
And in case this post wasn't clear: I'm all-in on large language models: they confidently pass my personal test for if a piece of technology is worth learning:
"Does this let me build things that I could not have built without it?"
What I find interesting is that - on the surface - they look like they solve a lot more problems than they actually do, partly thanks to the confidence with which they present themselves
Figuring out what they're genuinely good for is a very interesting challenge
@dbreunig right - there are so many potential applications where you might get good results 90% of the time and utter garbage 10% of the time, which for things like loan application evaluation should be completely unacceptable, but might be fine for things like bulk sentiment analysis to identify general trends
There's a lot of depth just to learning how to identify places that the technology is a good fit
@dbreunig one area I'm particularly excited about it data extraction: given a huge jumble of badly OCRd documents, can a language model be used to ask questions of each scanned page to extract relevant facts from them?
If you applied data entry people to a task like that you'd also get a portion of errors
@simon Also agree. Bureaucracy—where forms and formulas have tried to interface with humans as if they were APIs—and it’s artifacts is where the potential energy lies. Anywhere there is rote tasks and boredom is the X marking the spot.
I call it the ‘Brazil Antidote’ model. Doesn’t get enough attention because it’s not sexy. I’m going to found the Boring AI Working Group.
The way I look at it. Machine learning in general (including these large language models) are great when you have the following problem criteria
#1: You need to build a pattern matcher
#2: You don't know what to look for.
#3: When the pattern matcher is finally built you don't care to know what it actually looks for
#4: The results are allowed to be hilariously, insanely wrong some % of the time
And there are actually a lot of things that match that criteria
@pamelafox @ncweaver the 1-1 therapy thing is terrifying to me - imagine trying to get therapy from your iPhone keyboard!
But that said, I do use ChatGPT as an alternative to rubber duck programming sometimes: if I'm stuck on something I'll kick off a conversation purely as a thinking aid, and it's often effective
That feels pretty different to me from the therapy thing, but not a million miles away from it
@ncweaver I really like that criteria list
It would be really useful if there was a solid, easy to understand list of use-cases and anti-use-cases to point people to
@simon I think for anti-use case, #4 itself captures it. "Is it OK if you are hilariously, outrageously, gobsmackingly wrong and you don't know it?"
Which is why I find Tesla's "AI first" development model for autonomy frightening, beyond just the fact that they are training it based on how Tesla drivers drive...
@Sandra I've not found that argument very convincing yet
Sure, there's a HUGE energy cost in training a model... but that model can then be put to use for many years into the future
text-davinci-003 was trained once, at great expense... but has since run inference millions (probably billions) of times for millions of people
This looks even better for openly released models like Stable Diffusion: trained once, then distributed to anyone who wants to use it
@Sandra I remember being amused when I saw one model that was trained by a university in France that boasted that 90% of the power used to train that model came from a nearby nuclear reactor!
Wish I could find that reference now
I'm not sure "years into the future" even matters.
My thinking is this: it doesn't matter if my shovel lasts "years into the future" or not. What matters is "how much use do i get out of the shovel before it stops being a useful tool?" That same amount of use may be spread out over years or months.
I think the question is mostly just about amortization.
Well, that + moral effects / consequences of its use.
@wim_v12e @Sandra I've been exploring alternatives to re-training the entire model to bake in new facts through mixing in results from other systems directly into the prompt - it's a really promising avenue: https://simonwillison.net/2023/Jan/13/semantic-search-answers/
The cost of using them is definitely enormous - I've seen reports of ChatGPT costing $3,000,000/day or more - but again, that's spread across many users
At least it's not Bitcoin mining!
Humans readout is what gives rational interpretation of prompt responses that had been modeled for looking plausible.
Useful when those given responses are factual and accurate. Otherwise misleading interpretation and thus requiring critical filtering.
This is a critical failure of current models, so to be improved in future models.
Simon Willison
Chris Hodges 
Arthur
Drew Breunig
Nuzz 🧋
Nicholas Weaver
Pamela Fox
Jesse
Noam
Hans Gerwitz
Ari Koponen
mterhar
Sandra
masukomi
David Bruchmann
Wim🧮
Teixi
CapritiaJH