5 ms·
There seems suggestive evidence, in the pattern of improvement after a few days on specific questions, that some of the worst answers have been human-reviewed &
by gojomo 5y ago
There seems suggestive evidence, in the pattern of improvement after a few days on specific questions, that some of the worst answers have been human-reviewed & improved "on the fly" – on a scale of days/weeks.
If such tweaks show useful generalization – correcting a few answers also helps the network better determine entire classes of experessions that deserve more definitive & correct answers – that's not such a big deal, expecially if this constant-human-guided reinforcement-training is well-disclosed.
If instead the corrections work more like a lookup-table 'cheat sheet' of answers to give in preference to the bulk-learned answers, with little generalization, that's a bit more slight-of-hand, like the original (late-1700s) 'Mechanical Turk' chess-playing 'machine' that was actually controlled by a hidden person.
If the disclosure of this constant human-guided correction-process is hidden, or downplayed, the impression of trickery, rather than innovation, is larger.
- rowanG077 5y agoA lot of human communication is "lookup-tables". The entirety of geography, mathematical axioms and theorems, names of colors, names of people, language, shape of various things. I'd wager even that it's more important for an AI to have good lookup tables then to have good inference if it were to pass for Human.
- gojomo 5y agoAgreed! But, it's important not to confuse what's possible via 'human-in-the-loop' 'active-learning' with what's possible from an algorithm on a fixed training corpus. Sometimes GPT-like models are portrayed as the latter – a highly automated, reproducible process – while this article makes a pretty strong case that the responses from OpenAI's public-facing interface get rapidly improved by a staff of dozens of contractors. It's not surprising that a staff of 40 humans, given a few days time to consider & compose, can prepare human-quality answers to arbitrary questions!
- cyanydeez 5y agoI think, ultimately, what matters is how the labeling and corrective actions falloff over time. But reasonable people will notice that information is streaming In faster than any single person can code
- ImPostingOnHN 5y agoIn this example, it seems they fell off immediately, because the responses were only coded in response to questions asked in the way the author pre-published, and did not appear when the same question was asked in a slightly different way, which the GPT-3 authors did not have the opportunity to prepare for well ahead of time
- dustingetz 5y agoAh so when OpenAI codes fizzbuzz, it's returning human tweaked fizzbuzz from the lookup table? https://twitter.com/sama/status/1503820489927495682 https://twitter.com/sama/status/1503820489927495682
- rowanG077 5y agoNo, I didn't say anything even a lightyear close to that.
- mannykannot 5y agoIt is the (apparent?) ability to make inferences that makes GPT-3 look impressive. Take that away, and it looks more like a demonstration of the banality of everyday chatter than a significant development towards AI.
- staticassertion 5y ago> nswers have been human-reviewed & improved "on the fly" – on a scale of days/weeks. Why would this be surprising? I assume that they're rolling out new models with new parameters, input data, and corrections, all the time. > that's not such a big deal, expecially if this constant-human-guided reinforcement-training is well-disclosed. That's just what supervised learning ist hough. > If instead the corrections work more like a lookup-table 'cheat sheet' of answers to give in preference to the bulk-learned answers, with little generalization, that's a bit more slight-of-hand, like the original (late-1700s) 'Mechanical Turk' chess-playing 'machine' that was actually controlled by a hidden person. There's no evidence of this though, right? And it seems... like a very weird choice, that couldn't possibly scale. 40 people are hardcoding answers to arbitrary questions?
- treis 5y ago>Why would this be surprising? I assume that they're rolling out new models with new parameters, input data, and corrections, all the time. Because the answers to specific questions are hard coded. It's not the result of a new model. It's the result of someone writing an if/then statement. Or at least that's what the author claims.
- staticassertion 5y agoI'm asking why it would be surprising that humans are reviewing answers and then making improvements to the model. There's no evidence of hardcoded answers.
- disiplus 5y agowhat evidence would be enough for you besides source code ? The thing is returning only one correct answer to a question that days before had 3 answers.
- wallfacer120 5y agoIt improved, the model improved. Because sometimes, when you do work on the model, it improves.
- unityByFreedom 5y ago> correcting a few answers also helps the network better determine entire classes of experessions that deserve more definitive & correct answers – that's not such a big deal, expecially if this constant-human-guided reinforcement-training is well-disclosed. I gather the point is details aren't disclosed. Some kind of update is happening that we aren't told about, and that calls results into question.