LaMDA is not sentient
garymarcus.substack.com
LaMDA is not sentient
1–10 of 694 posts
Re: LaMDA is not sentient
#2Can we rollback such aggressive use of "AI?"
We do not have it yet. What we do have is cool and high value and all that. I am not talking it down at all
We all might be better off talking in more accurate terms.
Re: LaMDA is not sentient
#3It's called advertising.
Re: LaMDA is not sentient
#4Can we rollback such aggressive use of "AI?" We do not have it yet. What we do have is cool and high value and all that. I am not talking it down at all We all might be better off talking in more accurate terms.
Good luck telling that to the media - they need their sweet sweet clicks.
Re: LaMDA is not sentient
#5I agree that LaMDA is almost certainly not sentient, but this reminds me how inadequate the Turing Test is. If you train something on a large enough corpus, it'll pass the Turing Test easily enough. So little thought is truly original that there's a relevant enough jumbling of words it's already seen that should pass the test. The more important test, and what we need, is a sentience test.
Re: LaMDA is not sentient
#6Can we rollback such aggressive use of "AI?" We do not have it yet. What we do have is cool and high value and all that. I am not talking it down at all We all might be better off talking in more accurate terms.
Good luck telling that to the media - they need their sweet sweet clicks.
This is basically the problem with all science and tech reporting.
Overwhelmingly great headline, exciting intro paragraph, right down the end the actual experts being subdued about realistic potential for commercialisation/viability
The cherry on top is that they won't even give you a link to the paper as it is against most media websites policy to link outside domains.
Re: LaMDA is not sentient
#7I'll be honest. Talking to an average SF Walgreens pharmacist or SF government employee seems to be very much like the way people operate GPT-3. You have to use the correct prompt engineering and you'll get the right results. I've gotten results with bad initial conditions by correctly prompt-engineering them. I've watched other people have trouble (and rarely) have faced trouble myself by providing the same information without the right prompt-engineering[0].
So, in some sense, maybe it's almost there. It's just that right now, it can only produce short-memory text output. But lots of people are only capable of that[1].
0: For instance, a classic technique is the "I'm a dumbass but nice". Your pharmacist cocked up your prescription - maybe forgot a refill? Bad method: "I think it's supposed to be X and have a refill". Good method: "Oh, I'm sorry. Do you know if the doc put a refill notice on the prescription at the bottom? He said he'd do it but maybe I forgot to remind him". I've watched me win while others lose. Smiles while others get yelled at. This is an easy one. Everyone knows this. But there's better ones too.
1: https://astralcodexten.substack.com/p/somewhat-contra-marcus...
P.S. VT7L3TAb
Re: LaMDA is not sentient
#8> What these systems do, no more and no less, is to put together sequences of words, but without any coherent understanding of the world behind them, like foreign language Scrabble players who use English words as point-scoring tools, without any clue about what that mean.
And that is evidence for what exactly? That the Chinese room tought experiment is solved now, with the answer being "the mimicry is different than the real thing"?
Who says that humans aren't just a spreadsheet for whatever they do? You can't "simply" away that. At some point you have to ask some qualitative questions about whether the spreadsheet might be sentient. Maybe it's not this one, but as we build more and more spreadsheets, the question will become more and more relevant.
Re: LaMDA is not sentient
#9I agree that LaMDA is almost certainly not sentient, but this reminds me how inadequate the Turing Test is. If you train something on a large enough corpus, it'll pass the Turing Test easily enough. So little thought is truly original that there's a relevant enough jumbling of words it's already seen that should pass the test. The more important test, and what we need, is a sentience test.
…which is a point made explicitly in the linked article
Re: LaMDA is not sentient
#10> To be sentient is to be aware of yourself in the world
A robotic arm with a self-collision avoidance algorithm (no ML or anything like that) needs a model of itself in the "world" it exists in; a set of bounding boxes at least. Is it self-aware by this definition?
Sentience and consciousness are really ill-defined terms so everyone discusses this under their own arbitrary set of assumptions, at least that's my impression.