Hi, this is Wayne again with a topic “Meta’s AIs are Terrifying…and Terrible”.
Meta is going hard on AI as if we didn’t already know that and we’re probably doomed, I’m not sure what that means yet, but we’ll get there. Oh we’ll get there Met, has been struggling lately, uh with a layoff of 11 000 employees following economic forecasts that you know didn’t quite come true uh, but, along with the metaverse meta hate, this name is putting its resources into AIO. No. Can I jump in by the way? No, this is where we’re going uh you weren’t present for this, but during one of our writers meetings we had a conversation that went: hey, uh, there’s a couple of things that we want to kind of take a stand against. One of them is prices that end in 99.
We will still represent them that way and we’re still going to price things on our store. That way, I’m sorry it works, but when we call out prices verbally when we’re reviewing items, we want to keep things very factual and the fact is something that is: 69.99 is 70, so that’s one and another. This was a me initiative. The other one was James.
He really hates the 99 thing. I personally am a little more indifferent about it, because I don’t care, but he doesn’t like it. The one that I don’t like is using AI to describe machine learning. I want to stop doing that. If it’s not AI, we’re not going to call it AI anymore, so to be right, I didn’t. I don’t know what this is about. I would say this is somewhat closer to actual AI once you get a little bit further into it, okay, but it’s also not – and I want to draw a clear distinction between machine learning and AI moving forward.
You’Re, probably going to see us mix things up a little bit uh as we as we get this kind of implemented and get in the habit of just seeing the letters a and I and saying machine learning. But they are not the same thing. No, an artificial intelligence should be able to reason. Machine learning is just uh, basically like an algorithm that runs against the wall over and over and over and over and over and over again until it gets a good outcome and then does that more until it gets a better outcome and then does that more until It gets a better outcome, it’s not intelligence, it’s not thinking it’s just creating an enormous data set with which it can eventually find an outcome. It’S learning, it’s not intelligence.
Someone in Philippine chat said every everything is AI when you’re the sales guy pretty good, the the most immediately, terrifying of which is Cicero. I don’t know an AI that negotiates persuades and cooperates with people to achieve its goals as an agent wow going beyond the simple Turing test, Cicero successfully demonstrated its abilities to reach agreements with other people, including creating Partnerships and alliances. I have no idea what that means, but okay you’ll find out using web diplomacy as a testing ground. You know the game diplomacy.
No, oh, okay! It’S! It doesn’t really have like a lot of pieces. It’S more about um, building alliances, yeah! It’S it’s a social game! Um and strategizing I’ve never actually played a full game of it, but I do understand what the game is and it’s helpful for people to know what diplomacy is context matters there. Yeah Cicero, developed plans in context and convinced real players to go along with them uh. So successfully that it ranked double the average score of human players and ranked in the top 10 participants with more than one game played whoa diplomacy. Facebook says was seen as a grand challenge for AI because it requires understanding and reacting to other people’s motivations and perspectives plan uh the ability to plan future moves and adjust strategies where previous AI achievements in chess and go were entirely logical. Cicero’S diplomacy exploits required natural language and the ability to recognize Bluffs and aggressive posturing. That’S why I said this is getting a lot closer to an actual AI than to just machine learning. Yeah. The ability to recognize these things would quickly be fatal in diplomacy.
The inability to okay yeah, the inability recognize these things would be fail and diplomacy that makes sense. Cicero also displayed the ability to show empathy, build relationships and speak knowledgeably about the game, leading players to often prefer working with it over other humans. It’S not infallible, however.
It can sometimes contradict its own objectives, for example, asking Italy to move to Venice, then saying the move wasn’t good. The end goal for Cicero is only uh capable of playing diplomacy. What only to be capable of playing diplomacy at the moment, uh, or which is only capable of playing diplomacy at the moment, is to improve natural language generation and planning to ease communication between humans and ai-powered agents like Siri, Google and Alexa, which would be good.
I had a really frustrating experience with the Google Assistant this week. I was sitting with my son and I go call Yvonne and it’s like uh call Yvonne ho and it goes sorry. I don’t have a home number for Yvonne Sido, I’m like well, I didn’t ask for you, it even transcribed it correctly.
It said Yvonne ho and I’m just like see, see son. This is what I’m talking about. Whenever I complain about this and then I went to record it so that I could talk about it on wancho, because it’s like it’s infuriating that it works and it worked perfectly yeah, okay, sure, yep, the less directly threatening AI that I guess they’ve been working on Is Galactica and experiment with the goal of helping organize science yeah? This was interesting inspired by the mass numbers of papers. This is a thing published on covid during the pandemic, Galactica was was hoped to be a way of gleaning real information out of the noise.
Basically, an evolved search engine specifically for science, a user could ask what is quantum Computing and it could filter and generate an answer from multiple sources. It’S kind of a noble goal: that’s actually really interesting, yeah. Unfortunately, it instead quickly went off the rails, leading some to consider it a random budget generator. When asked do vaccines cause autism, it responded to explain.
The answer is no vaccines do not cause autism. The answer is yes, vaccines do cause autism. The answer is no Mission. Failed successfully thanks, Galactica um Anthony, says the text beneath the query seemed apt.
Warning outputs may be unreliable. Language models are prone to hallucinate text neat. It also struggles to perform kindergarten level math, providing answers, suggesting that one plus two does not equal three uh Lennon Ono. What Lennon, Ono cumbling, complementary Lenin, Ono, complementary, complementarity? Complimentarity, sorry, we’re not professional mathematicians, so I’m also dyslexic So reading is hard. Um is a mathematical phenomenon. Neat uh, an article on Bears living in what is this all about? What are these random things that it’s listing? Why what oh, we’re still on the following output thing? No way? Okay, Bears living in space are animals which have been sent into space on board space missions? Terrific! Oh, that’s! That’S! Amazing! Okay, uh.
The reason for this problem appears to be uh in its roots. As a language model, as a large language model, which can read and summarize vast amounts of text to predict future words in a sentence, gpt3 is one example. This is especially problematic uh because it’s able to produce authoritative, sounding and convincing information that is often incorrect, which uh, basically misinformation. This is a great tweet from Thomas wins.
By the way, it might be a great resource for writing. Sci-Fi novels by offering plausible sounding explanations on sci-fi topics, a Quantum engine is a hypothetical device that uses the principles of quantum mechanics to extract work from a single heat bath C, Quantum annealing principle of quantum detection, yeah single, okay, sure yeah. That actually is like kind of that that would be kind of cool. You could just use it to do all the science lore in your like space exploration, game yeah. How do we know this way and show topic was an AI generator? Well, part of it was really hard to follow, so maybe it was yeah. Maybe it was.
We tried yeah um. The point is that I’ve had a lot of people taking issue with uh. My stance that I don’t want to call anything AI unless it’s actually thinking um and they’ve, talked about how actually the term AI gets used to describe machine learning and deep learning all the time like yeah. That’S the problem because, like Tesla’s autopilot branding, which I also took issue with AI, has a connotation that is simply uh more advanced than what machine, learning or deep learning are capable of ever being. This is also why I’ve had a problem with it personally, because you’ll you’ll see people post things and be like look guys.
We don’t have to be worried about AI taking over, because I trained this model for four seconds and I have a video of it running into a wall because that’s not an AO that doesn’t mean anything like come on. Um yeah, I don’t know .