Am Di, 14. Jun 2022, um 14:18, schrieb John Clark: > On Mon, Jun 13, 2022 at 9:51 PM Bruce Kellett <bhkellet...@gmail.com> wrote: > >>> >> I doubt Lemoine went crazy and just fabricated the conversation, but if >>> >> he did the truth will undoubtedly come out in a day or two. And if the >>> >> conversation exists as advertised then it is a monumental development. >> >> *> The thing is that there are an awful lot of questions that remain >> unanswered in the information as presented. We don't actually know how >> lambda works.* > > If the conversation was as described and was not somehow staged or > cherry-picked then LaMDA is a real AI and nobody knows or will ever know how > LaMDA or any AI works except in vastly oversimplified outline. The group of > people who originally made LaMDA taken together understood how it once worked > (although no single person did) but no individual or group of individuals can > understand what it became.
Nobody understands how these neural networks work in detail because they have billions of parameters, not because some emergent behavior of the sort that you are imagining is present. The current hype in NLP is around a neural network architecture called a transformer: BERT and all its incarnations and GPT-3. These are language models. A language model is "simply" a function that gives you the probability of a given sequence of words: P(w_1, w_2, w_3, ..., w_n) A clever thing you can do with language models is predict the w_n given the other words, and then include this prediction in the next step and keep going to generate text. Something like softmax can be used to assign a probability to every word in the lexicon for word w_n, and with this you can introduce randomness. This creates a stochastic parrot. One of the great things about these architectures is that unsupervised learning can be employed, i.e, they can be trained with large amounts of raw text (wikipedia, books, news articles and so on). There is no need for the costly (prohibitively so at these scales) of having humans annotating the data. Another really nice thing that was discovered in recent years is that transfer learning really works with these language models. This is to say, they can be trained with vasts amount of unlabelled data to correctly make predictions about probabilities of sequences of words in general, and then "fine-tuned" with supervised learning for some more narrow task, for example sentiment detection, summarization and... chat bots. Unless there has been some unpublished fundamental breakthrough, LaMDA is almost certainly a large language model fine-tuned as a chatbot (and I would be particularly interested in what happened at this stage, because there is a lot of opportunity for cherry-picking there). You just need some basic knowledge of linear algebra, calculus and programming to understand how they work. One of the big break-troughs were attention heads, which are a way for a network to learn what part of a sequence of words is more important in predicting a word in a given position. Before this recurrent neural networks (RNNs) were used . RNNs use recurrent connections as memory mechanism, but they suffer from the vanishing gradient problem. Informally: the more iterations pass, the harder it is to figure out how much each parameter contributed to the prediction error. If you have some patience and a desire to really grok this, something like this may be enough: https://nlp.seas.harvard.edu/annotated-transformer/ Large language models based on transformers are amazing, and they are most definitely a significant step forward in machine learning, NLP and AI in general, but they are not what you are thinking. Some things that real minds have that these models do not: - The ability to model the interlocutor's mind, and to act in such a way as to influence the state of this other mind. This is what real communication is about; - The ability to preserve context: what has been said and what this implies about the state of the world and the other mind and so on and so forth. Context is preserved at different scales of time and space, and we know which context to apply to each situation and how to switch context when appropriate; - General knowledge of a *multi-sensorial* nature. I know what it means to "see red". I know how it feels in my guts to have my bank account in the red. I know the physicality of the actions that language describes. My mind connects all of these modes of perception and knowledge in ways that vastly transcend P(w_1, w_2, ..., w_n); - The ability to learn in a general way, and to learn how to learn; - Actual motivations, goals and desires, directed by a system of emotions that we have by virtue of being embedded in an evolutionary process. I could go on, but the above are show-stoppers in terms of us being anywhere close to real AGI. > And if the conversation was staged or cherry-picked then I don't understand > why Google hasn't said so by now, What would Google have to gain from saying anything? They would expose themselves to potential legal troubles with the suspended employee. They would plant the idea in everyone's mind that Google stuff might be staged or cherry-picked. And what is cherry-picked anyway? That can become quite subjective pretty quickly. My bet is that the bot was fed some "information about itself" at the fine-tuning stage. By not saying anything they get free hype. By saying something, they risk looking silly. The employee was most likely suspended for divulging internal information without permission. This is typically frowned upon in big corps. > after all the longer they delay the more foolish they will seem when the > truth comes out, and if LaMDA is not what it seems then it's only a matter of > time, and not much time, before the truth comes out. I doubt it. Mainstream media has the attention span of a house fly, and the debunking will probably be too nuanced for most people to care. Telmo. > John K Clark See what's on my new list at Extropolis > <https://groups.google.com/g/extropolis> > tns > > > > > > > -- > You received this message because you are subscribed to the Google Groups > "Everything List" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to everything-list+unsubscr...@googlegroups.com. > To view this discussion on the web visit > https://groups.google.com/d/msgid/everything-list/CAJPayv17_4o7RsuEt1Lkec0J5RqUJTBhGoEjpX9A-hrUrN9jzg%40mail.gmail.com > > <https://groups.google.com/d/msgid/everything-list/CAJPayv17_4o7RsuEt1Lkec0J5RqUJTBhGoEjpX9A-hrUrN9jzg%40mail.gmail.com?utm_medium=email&utm_source=footer>. -- You received this message because you are subscribed to the Google Groups "Everything List" group. To unsubscribe from this group and stop receiving emails from it, send an email to everything-list+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/everything-list/6aeb9790-782a-40b6-99d0-90410a062c83%40www.fastmail.com.
