YbmN

@paulg Just to be clear, assuming that AGI is possible while “hoping” that economically meaningful human jobs will remain requires quite the mind gymnastics. The theory of comparative advantage notwithstanding.

It may be a mistake to ask which occupations are most safe from being taken by AI. What AI (in its current form) is good at is not so much certain jobs, but a certain way of working. It's good at scutwork. So that's the thing to avoid.

@jozefinselberg @joni_askola We all know where it leads… a place decorated with barbed wire and smokestacks.

@joni_askola Appeasing genocidal dictators leads to nowhere 🤷🏻‍♂️ 🎗️🌻

1/11 Finally! Discussions are underway for the United States to potentially send up to eight of Israel’s Patriot batteries to Ukraine, marking a significant shift in Israel's relations with Moscow.

@TyroneAlfonso @adnashmyash @elonmusk @DavidSacks By now it should be obvious that « bitter peace » now is a code word for « bigger war later ». Crimea was not enough. Donbas was not enough. Eventually, the 2 new oblasts won’t be enough. Peace is a false choice here.

@adnashmyash @elonmusk @DavidSacks So just to be clear, are you for continuing the war or would you like to see a cease fire and peace talks, even if it means concessions? I always assume a bitter peace is better than a viscous, deadly war.

People like @elonmusk and @DavidSacks wondering why people like me from Ukrainian Donbas, who are Russian speaking from childhood, who until 2014 normally welcomed thousands of Russian tourists on the coast of our Mariupol and had not problems with Russians at all - became one day totally "Russiaphobic"? Maybe because my parents home in Mariupol where I born and had a happy youth looking like THIS after "Russian world" come to my home city? P.s. The whole building is already totally demolished.

@FKesheh84 @Grady_Booch Step 2. All of what you said is obviously true to those who aren’t desperate to pretend humans have some sort of magical edge. Step 3: personal agents show us bespoke UIs when needed based on context and known preferences. Often, they’ll just tirelessly work for us without UIs.

At some point in near future a Multi Modal LLM will be able to render an UI (like Sora renders a video or like Dalle creates an image) and also to handle human feedback on that UI in real time. Then you will not even need code anymore, the whole software logic will be in the prompt. And it won't need to be detailed as a software, like as you say to create a dog image today, you will be able to say something like show the user and UI with two input fields, do this when it presses the button.

Becoming more clear by the day that the alternative to fully committing to Ukraine achieving a resounding victory against Russia is an ever-increasing risk of another World War in Europe, Asia, and the Middle East caused by dictatorships emboldened by our timid indecisiveness.

@daniel_271828 @foomagemindset @ylecun There is a 100% risk of death for those affected by diseases / injuries AGIs would have cured had it been invented earlier. The non creation of non existent people presents no risk to anyone. So we’re comparing a 5% risk of extinction for 8B vs 100% for ?B.

@foomagemindset @ylecun I feel like a 5-10% chance of an outcome on the scale of human extinction is pretty worth worrying about. This survey was also from several years ago, and I’d expect timelines have shortened somewhat since then

The public in North America and the EU (not the rest of the world) is already scared enough about AI, even without mentioning the specter of existential risk. As you know, the opinion of the *vast* majority of AI scientists and engineers (me included) is that the whole debate around existential risk is wildly overblown and highly premature.

@ylecun Nobody expects LLMs to put most people out of work. But near human level AIs necessarily will. The difference of course is that whatever new jobs arise (probably many indeed) will also go to HLAIs, skipping humans altogether. This is a good thing.

From the point of view of a 19th Century farmer, machines have taken over all the jobs in the 21st Century. Jobs that occupied most people in the 19th Century now occupy small amounts of people or have disappeared. Most 21st jobs would be entirely incomprehensible for a 19th Century farmer.

@riversorare @firasd @ESYudkowsky If such AGIs exist which are both capable of taking over the world and yet incapable of resisting simple adversarial attacks like these (a tall claim), why wouldn’t there also be many other AGIs tasked with detecting, countering, and preventing those rogue AGIs?

@firasd @ESYudkowsky Exactly. AGI is 100% safe. … Just as long as nobody dangerous gets their hands on it. And that is impossible because humans are entirely trustworthy and there is no way it could ever connect to, or find its way onto, the internet.

The impossible difficulty-danger of AI is that you won't get superintelligence right on your first try - but worth noticing today's builders can't get regular AI to do what they want on the twentieth try.

Here is how my country understands 1945:

@shaunrein Hopefully Taiwan is next.

@ylecun If an unbeatable hyper capable evil AGI could doom us all, then it stands to reason that equally hyper capable good AGIs would be able to prevent its arising in the first place. For doom, one has to bet on the ill intentioned being first. (alignment difficulty notwithstanding)

If some ill-intentioned person can produce an evil AGI, then large groups of well-intentioned, well-funded, and well-organized people can produce AI systems that are specialized in taking down evil AGIs. Call it the AGI police. No need to bomb data centers.

@geoffreyhinton @pmddomingos Why would it want to take control? If it did, why would it need to deceive rather than ask? If it did, what could it need the control for? Few people doubt something much smarter could not be stopped. Motivation, alignment, drift, seem to be all people really disagree on here.

@pmddomingos and for a long time, most people thought the earth was flat. If we did make something MUCH smarter than us, what is your plan for making sure it doesn't manipulate us into giving it control?

Reminder: most AI researchers think the notion of AI ending human civilization is baloney.

@wile_zzz @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton Some non AGI systems (LLMs) can already integrate with other AI systems (vision / hand like manipulators). ai.googleblog.com/2023/03/palm-e… A purely cognitive & disconnected AGI wouldn’t go far in the physical world, but I expect many kinds (including embodied ones - not just integrated)

@yesbutmostlyno @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton But it can’t include embodied jobs because those often have very little to do with “abstract intelligence” or whatever you want to call it. You need to be able to manipulate the world with things like hands. Currently only humans or maybe non human primates can do that. No?

In the NYT today, Cade Metz implies that I left Google so that I could criticize Google. Actually, I left so that I could talk about the dangers of AI without considering how this impacts Google. Google has acted very responsibly.

@wile_zzz @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton True, AGI is ill defined. But it’s not “can do anything”, and it doesn’t need to for my point. Generally people tend to mean that AGI can do at least “anything we can” cognitively. This baseline necessarily implies the end of human jobs. Including embodied jobs.

@yesbutmostlyno @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton I guess my point is that your argument is sort of a “get out of jail free card” in that it’s similar to just saying AGI = an entity that can do anything. I think that’s a bit irrational in that there’s nothing to debate.

@wile_zzz @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton An AGI is simply a generally intelligent agent, capable of quickly learning from and adapting to new environments to keep achieving its end goals. If it was embodied then the question is moot. If not and it now needs hands, it will control hand-like machines, like we do tools.

@yesbutmostlyno @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton But it would need hands

@wile_zzz @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton Very clearly yes. But more than that, groups of AGIs will design an astoundingly personalized house, handle procurement, and control machines to prepare the site and build it. Today, only very small parts of each step can be handled by narrow AIs… though more every year.

@yesbutmostlyno @KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton Will AGI be able to bake and frame a house? 🧐

@acjwatt @AceOfThumbs @bag_of_ideas @KosisochukwuAs2 @geoffreyhinton Retirees tend to cope well. Children cope well. Meaning is easy to create, even in unproductive or solo activities (sports, hobbies, relationships…). Jobs are used as a partial source of meaning because we cannot escape them for 40+ years. Once we can, we find other sources.

@AceOfThumbs @bag_of_ideas @KosisochukwuAs2 @geoffreyhinton Sounds like you would be happy for us to return to the iron age but with healthcare - but what we all really need is to feel we have a purpose, most people will sacrifice all of those other things given the right incentive. You take people's purpose away - and what have you got?

@KosisochukwuAs2 @bag_of_ideas @AceOfThumbs @geoffreyhinton The difference between AGI technologies and narrow AI technologies (or past technological revolutions), is that while many new jobs will indeed be created, all of them will go to AGIs too. AGI is what will end the era of human jobs. We’re not there yet. But how far?

@bag_of_ideas @AceOfThumbs @geoffreyhinton A “Job” is not a reason to live, there are many ways to do things that add meaning. That being said, assuming jobs will be shortened Bu adding AI to the workforce is a reductionist mindset, it’s just like thinking that computers would create lack of jobs in the 80s, image that

@__RickG__ @JFC_Bass_Chant @composite9 @ESYudkowsky @amtrpa @ylecun Like other aspects, such as model architecture and training data, RLHF is gradually getting better quantitatively and qualitatively. Even smaller older models improve in safety and usability with the latest RLHF sets. I expect even reward hacking to get better eventually.

@yesbutmostlyno @JFC_Bass_Chant @composite9 @ESYudkowsky @amtrpa @ylecun I really don’t see the evidence you are claiming… GPT4 is more aligned with RLHF simple because they did more of it. You can clearly see that GPT4 is way more rigid, and Bing was GPT4 all along… which was (and basically still is) a disaster from the “alignment” perspective.

Insects "outsmart" humans by a factor of 1000 (by total number of neurons). But I'm not particularly worried that they will kill all humans.

@JFC_Bass_Chant @__RickG__ @composite9 @ESYudkowsky @amtrpa @ylecun Very few details yes, and he acknowledges that. However, I also see no reason to believe alignment of capable models is hard. Either the “dangerous” models have been hypothetical, or the more capable actual “harmless” models have been, the more “alignable” they have seemed to be.

@yesbutmostlyno @__RickG__ @composite9 @ESYudkowsky @amtrpa @ylecun I don't think LeCun's proposal actually has that many details. Actually forming those details is tricky, and I don't think he realizes just how tricky it is. Especially since he thinks it's an easy problem.