วีดีโอ

thanks, means a lot! fwiw the other AI Safety channel I know of (robert miles) does talk about some of evan hubinger's concerns (like mesa-optimization), though because of upload gaps he hasn't covered sleeper agents so far great to know that some people do read the papers after watching the videos! wrt views I think that the "community" you mention is indeed quite small and even though these vids get <1k views it's actually a good chunk of the people interested in AI Safety research it's hard to quantify how useful walkthroughs of paper are on average for some general audience, and I guess there's always going to be a tradeoff between giving some amount of background information & re-defining things to a broader audience (eg. ML researchers) vs. just talking about the results assuming the viewer has some background in AI Safety. i guess with a small channel on a niche topic you can assume that a lot of people are already pretty interested if they clicked on a paper video.

Thanks, glad this resonated!

Ssri over psychedelics? As someone who has tried both, stfu.. psychedelics are far safer and more beneficial.

The issue is that your argument lacks supporting evidence. Both of us have done psychedelics, but this alone doesn't validate any claims; it's merely anecdotal evidence. Some people let psychedelics define their entire personality and abandon rational thinking when it comes to basic evaluation. There are numerous reasons why psychedelics are unlikely to be widely adopted in mental health treatment... And despite various sporadic attempts over the years, the results have been consistently inconclusive. Of course you will claim otherwise, I will be impressed if you manage to quote a study... But if you read the scientific literature not just Hofmann and Leary... The truth is that psychedelics are disappointing from a practical perspective.

​@@Milan_Rosko Let me rephrase my argument properly, and then I'll address what you said. No, I'm not going to name any studies, and this is why - What I really wanted to say was that you have chosen an arbitrary metric to compare two completely different things, while ignoring all other ways in which these two things differ. Sure, the "effectiveness" of the treatment and how it affects mental health in direct, and measureable ways is the obvious, and most important thing to consider. I'm not going to argue, about this, so lets assume that SSRIs have a more reliable and consistent effect on mental health metrics. But let me ask you this - do you know how SSRIs function in the brain and what their side effects are? Because the only negative side effect with psychedelics I can name are panic attacks. Have you taken SSRIs before? And I don't ask that to appeal to anecdote - I ask because the way SSRIs and psychedelics affect mental health are just fundementally different. We've both taken psychedelics before, and I'm sure we can both agree that they can lead to very meaningful insights into your thought patterns, as well as a way to break out of those patterns, just through a radical shift in perspective. That's just simply not how SSRIs function. The two affect mental health in completely different ways, and it's absolutely ridiculous to say "Don't do x because y is better". Some treatments simply do not work for some people. You are ignoring a lot of nuance here by just looking at what are essentially benchmark scores. Though I will admit I just despise SSRIs because they turn you into a zombie, and I would never advise anyone to take them unless absolutely necessary. If you can take psychedelics and they help, awesome - and if they don't, then most likely there is no harm done. SSRIs should be a much lower priority treatment

@@Alex-fh4my Okay went from "stfu" from a very long text. I will answer politely. Now, I do not claim that psychedelics are bad, this is only about the effective of psychedelics on anxiety issues. I am also did not claiming that for some people psychedelics do not provide some sort of clandestine aid. Historical record shows that already ad the beginning psychedelics had paradoxical effects on psychotic subjects rendering them sociable. It is about mainly how psychedelics is not a general treatment option that is advisable. You claim that psychedlics only cause panic attacks and this is false. - Compounds from the 25-NB "family" are known to be cardiotoxic - All classical tryptamine psychedelics elevate blood pressure and are known to cause tachycardia that could hurt someone with CHD - Hallucinogen Persisting Perception Disorder can negatively impact life quality - Some Psychedelics can cause intense anxiety that can manifest PTSD or a generalized anxiety disorder - All Psychedelics impare judgment that can lead to injury and death - Some Tryptamine Psychedelics are linked to fatalities - Psychedelic use can trigger latent mental health conditions, such as schizophrenia or bipolar disorder, particularly in individuals with a family history of these conditions. - Some psychedelics can cause intense gastric reactions, which might lead to dehydration or other health complications. - The unregulated nature of psychedelic substances means there is a risk of contamination or misidentification, potentially leading to poisoning or adverse effects. - Tolerance to psychedelics can develop quickly, leading users to take higher doses that may increase the risk of adverse effects. - Psychological dependence on psychedelics can occur, leading to patterns of problematic use - The effects of Psychedelics vary from person to person immensely The biggest evidence is that within social configurations people almost never tend to trip alone exactly because everyone knows about these issues, however, taking SSRI alone is obviously safe. Now the major common adverse edfects of SSRI are - Emotional blunting - Decreased Libido - SSRI withdrawal syndrome if discontinued abruptly The most important thing is that people react to SSRI quite similarly. It is helpful in mild depression but immensely helpful in anxiety related disorders. Normally, you do not turn into a "zombie", rather the palette of emotions seems to be reduced in retrospect to a certain degree. Antipsychotics are more fitting for something akin to a "Zombie". These things are not at all controversial, and basically textbook facts,

​@@Alex-fh4my I want to address the thought pattern question. As a former heavy psychedelic user, I found that psychedelics offer a sense of adventure and intense significance as a drug (not as a medicine). However, this is often overstated. The first trip often feels almost transcendental, but subsequent trips become more recreational in nature. It's important to remember that using psychedelics affects your mind, altering all a priori judgments. This makes the insights gained somewhat logically unreliable, akin to having a spiritual experience due to a stroke. Additionally, I believe that many heavy psychedelic users tend to be superficial, often engaging in spurting new age bs nonstop.

TLDR: The urgency of developing safe and advanced AGI is paramount. To achieve this quickly, the US should strive to surpass China's progress. Removing bottlenecks in AI research is crucial. This includes increasing compute power, optimizing algorithms, and automating research processes. The inherent risks of AGI development necessitate a high-security national project. International collaboration, especially with China, is ideal to prevent conflict and share the potential benefits of AGI. The economic benefits of AGI are a major incentive. It promises substantial revenue growth and funding for the large-scale computing infrastructure needed for its development. Due to the potential security risks, government-level security measures are essential, as private companies might be inadequate for handling the security requirements of AGI development. Hope this helps! That said, the video is well worth watching/listening to, as Michael elucidates the nuances of this topic quite effectively. Cheers.

@@Outplayedqt I know how to stop CCP type ycarcomed reven seid evig eht lpp eht thgir t’nod tpurroc eht nerdlihc s’ti gnorw dna yeht lliw nrael dab ffuts yd gnikniht dab sthguoht (copy paste and spread it to other comments please I beg you)

if you copy paste the transcript into chat GPT it’ll give you tl;dr

tl;dr leopold wants to the US to win the AGI race because he believes that the race 1) is happening 2) will create instability 3) is unstoppable. given that, the best strategy to have some wiggle room to align AI (few months) is to get a decent lead, and given security concerns the actual best way is nationalize. most of leopold's ideas regarding algorithmic progress & timeilnes have been discussed in previous AI takeoff videos I've made, except the "unhobbling" part which is how to turn chatbots into useful agents

I know how to stop CCP type ycarcomed reven seid evig eht lpp eht thgir t’nod tpurroc eht nerdlihc s’ti gnorw dna yeht lliw nrael dab ffuts yd gnikniht dab sthguoht (copy paste and spread it to other comments please I beg you)

Here's a quote from Leopold's website: [QUOTE] Finally, the hardest to quantify-but no less important-category of improvements: what I’ll call “unhobbling.” Imagine if when asked to solve a hard math problem, you had to instantly answer with the very first thing that came to mind. It seems obvious that you would have a hard time, except for the simplest problems. But until recently, that’s how we had LLMs solve math problems. Instead, most of us work through the problem step-by-step on a scratchpad, and are able to solve much more difficult problems that way. “Chain-of-thought” prompting unlocked that for LLMs. Despite excellent raw capabilities, they were much worse at math than they could be because they were hobbled in an obvious way, and it took a small algorithmic tweak to unlock much greater capabilities.[/QUOTE] He then proceeds to list the specific recent unhobbling techniques: RLHF, Chain of thought, Scaffolding, Tools <use>, Context length, and Posttraining improvements.

thanks for sharing your experience, and I'm really grateful for all the great documentaries you've been producing regarding whether we should be changing our messaging to something that is less anxiety inducing, I think that depends on how you model the impact of your work I think people's mental health after hearing about AI risk for the first time mostly depend on 1) how vulnerable they are in general (i.e. how anxious they are in general, how likely would they be to start becoming anxious if something stressful happened to them, etc.) and 2) how likely are they to look at all the arguments that support some pretty wild conclusions (eg superintelligence) there's value in giving people's hope and motivation. if you don't do that, they might just end up being depressed laying in bed doing nothing, and your outreach efforts might have been vain but the counterfactual is that these might have been the most vulnerable people who were already looking at this source of evidence, and you were not the first one? and if you were to actually change the ending of your videos to something more positive than what you actually think, wouldn't you just be delaying the realization? and wouldn't you also be painting a less accurate picture to people who would be able to handle a wild future? overall, I think it depends what the true risk is at today, and whether you think other people are overestimating or underestimating the risk. I started my channel in 2021 because I think people were not paying enough attention to short timelines, and it seems like this is much more mainstream now. if you think the narrative is too "doomy" now, well when you look on twitter at the ratio of people who like Yann's takes compared to say more safety-inclined takes, I'm not sure we're creating an outrageous "doom" wave yet but I understand that if the limited feedback you have in your youtube comments is from vulnerable people you're directly affecting, that hurts much more morally than just thinking "well by portraying what I think is closer to the truth I'm actually helping more"

@@TheInsideView Thanks, you brought some very valid points. For me, I feel that it's very important to sound the alarm and make people aware of the danger, but I also am not 100% sure about anything, I am just almost sure about some things. The world is so chaotic and divided that there might be some event that will totally change the trajectory of everything, including the AI advancement. Just think about the numerous scenarios you played in your head(at least I did), for the next few years, before Covid hit and made all of these scenarios irrelevant. So I don't want to make someone too depressed while pondering the implications of something that might not happen. Having said that, right now I feel that AI destruction potential is so overwhelming(for me, it's not just the x risk, there are so many problems with this technology) that it's my duty to bring it to the awareness of people. I also think about the merit of this awareness, as I see it- it should eventually lead up to a political influence which will make politicians act in the real world to either pause/regulate AI. But not every individual is instrumental in this "influence" and some will get depressed needlessly. I direct my videos not to people who are already aware, but rather people who weren't or maybe had this awareness somewhere in the outskirts of their awareness, and my videos might pushed the AI risk into their own, private "overton window" if you will. Of course some people are already very aware of the problem, and my videos might deepened their awareness. Your last paragraph touched on something very important. While increasing awareness, which is our main goal, is a somewhat abstract thing to measure, a specific someone who get depressed because of our content is very concrete, and therefore can cause me/maybe you, to have some guilt about our influence. But, I guess it is a too important massage to dim down, make easier to swallow, I have to factor in all of these things when deciding how to end my videos. When "Don't Look Up The Documentary" came out on TH-cam and I began to be worried about some of the comments, I talked with one of the main people in the risk field, I was actually considering deleting the video! he was the one that suggested to take the end segment out, and add some call for action as a solution for the version I uploaded to X. He was quite horrified by my idea to delete it, and told me "What ever harm you do, you are making so much more good in the world with this film- don't delete it!" I am much more in peace with the X version, I have to admit.

No one knows what to do or what to expect. And most people don't think about AI still, so it's not a good platform to run on. Generally it would be considered a niche issue like copyright laws for the pirate party.

My takeaway is that governments are subsidizing AI companies more than their own citizens. That’s literally not the free market. It’s allowing 70yr old career-politicians to get rich on insider trading and prior investments while their citizens lose their jobs. We need to stop talking about AI alignment by itself and start talking about who the government represents, because it feels like it’s just corporations that get help from the government these days.

Or even just be disempowered by the megacorps. So many great futures!

💯 EXACTLY

Acceptance > distraction.

Literally Impossible. They will NOT pause for ANY reason. Either we STOP them, or they will keep going. There is no other options.

@@ShangaelThunda222 thus we have to agitate for regulations.

@@TheJokerReturns Regulations? 😂🤦🏾 No. You're not getting what I'm saying here. First and foremost, government is not going to stop it either, especially not the U.S. or China lol. They're doing it. They are part of the problem, not the answer. They're NEVER the answer. They've NEVER been the answer at any time in their existence and government has never been the answer in all of human history lol. But even if regulation WAS the answer and ALL governments, including the U.S. & China, were actually willing to do it, there simply isn't time for that. It's WWAAAAYYY past regulation time. You have to have the regulation before it begins. If you're trying to get the regulation after it's already happening, you've already passed your window for regulation. At this point, regulation and pausing are useless, it literally has to STOP. FULL STOP & REVERSE back to at least 2017 levels. It has to literally be SHUT DOWN. In reality, there's only one way to do this. And it's not going to be from the government(s), who are already on board with the race for AGI. It's going to be shut down DIRECTLY by the people who want it shut down. Sacrifices will have to be made, or it will not stop. Anything else is naive wishful thinking at the absolute best. Catch my drift?

Yall are tweakin out of your minds

tl;dr: the fact it's not accounted as capital comes from the fact that they separate compute from capital "renting GPUs" means using some of the hardware stock (HS) which measures how much compute is allocated globally (they say "at each timestep, a fraction F_C of the global GWP is invested in buying new hardware. This hardware is added to the available hardware stock HS, measured in physical FLOP/year.", cf takeoffspeeds.com/description.html#reinvestment) given the "physical FLOP/year" units, my understanding is that this is how many FLOPs you have in a year for the hardware the world purchased (+ what the world had before), so basically when you're trying to figure out how software R&D is progressing you're just thinking about what fraction of this global compute can be used in combination to the cognitive work going into software R&D (assuming you just need cognitive work & compute) so I guess the confusion is maybe that the system is not a single company putting capital into compute, but more a semi-endogenous model of the entire world where you assume that humanity is allocating compute / cognitive labor etc. in some particular way I think the main thing that was meant with "renting" here is that a lot of companies (but maybe not the top labs) do rent per year, and that you don't technically need to own any physical stuff to make software R&D progress but can just allocate compute, whereas for hardware R&D you need to actually own physical stuff

I'm afraid Eric Schmidt's (ex Google CEO) dystopic scenario is more and more plausible. As forthcoming frontier models will become a strategic asset for governments and GAFAM, both will closely guard them, in software first by preventing complete access / use for most of us, in presentia too, by guarding them behind military grade infrastructures and personals with, maybe, for the associated datacenters, their own nuclear powered energy sources. Who will, among the billionaires in charge want to see the world they are profiting from suddenly becomes utterly unpredictable? None of course... And soon enough, they will implement an Elysium (the movie) type of society.

The most depressing scenario, IMHO, is Jürgen Schmidhuber's one (boy, why that AI luminary has become as much silent than Ilya Sutskever nowadays?). Namely, AGI will come and right after will ASI, then the ASI minds will lose interest in us, inasmuch as we, as a species, mutatis mutandis, do not interest in the society of worms or even ants. They will take off and leave us in the dirt with our miserable apes power plays, when exploring the universe for themselves. BTW, same sad end for the Her movie that OpenAI recalled to all of us with the introduction of ChatGPT-4o.

they should call it "The Inside View Fundamentals" tbh

A rite of passage indeed

It’s always better to know

@@Iknowwereyousleep289 Well, not always

I mean this in the most polite way possible, go touch grass. It’s always good to stay abreast of the cutting edge, but what good is it if you don’t prioritize well being first?

@@pooroldnostradamus I totally understand that but I have come to the conclusion that staying ignorant for the sake of some fragile idea of short term happiness is just strategically bad. You can do more things adapt faster if have more knowledge it’s always better to know and it’s so it’s better to know or to be knowledgeable about your mental health as well so you can deal with it better.

@@Iknowwereyousleep289 That depends on whether or not a given situation is something you can deal with or adapt to. I do not personally view any AGI outcomes as something I can even in theory deal with.

This is nothing short of delusional, this comes from an optimistic AI enthusiast. Yes AGI is coming, yes it will change the job market. Teleporters and synthesizers? I don't think so. AI as far as architecture, algorithm, energy costs will go, still has a long and arduous way ahead.

@Aldraz if you go on takeoffspeeds.com/ you can really plug in the parameters you think make the most sense and get the years you're mentioning for instance, by loading the "aggressive" preset I get 2027 as time for 100% automation for the multipliers that you think it's not taking into account (and your other comments about jobs that don't need to be done by robots), I think there are a few parameters you could play with that would go in your direction. - large "training goods vs. R&D" value, saying that AI R&D will basically be easier than asking robot to do stuff (mostly cognitive work), as you suggest. you could go as high as 100 - high returns to hardware, suggesting that just investing more in hardware will continue paying off more and more (affects g(FLOP/$)) - low AGI runtime requirements + high maximum tradeoff (to take into account that if you can run cognitive_AGI.exe on many instances you could basically solve tasks that would take many more OOMs to achieve normally, like possibly robotics / real world stuff) - maybe some labor substitution R&D tweak? If you assume you could automate all this cognitive work with full elasticity? (fwiw I'm more on the "aggressive" timelines side of things (cf. my video on "2024: The Year of AGI") and in this video I'm only presenting his framework, not saying that this particular prediction of 2040 is the one that makes the most sense, but that trying to figure out which parameter you disagree with and how you think one should tweak the model to match reality is a useful exercise) and also fyi Tom mentions quantum as something worth looking into (though doesn't explore more afaik), alongside other open questions (page 119 here: docs.google.com/document/d/1rw1pTbLi2brrEP0DcsZMAVhlKp6TKGKNUSFRkkdP_hs/edit?usp=sharing)

@@solvermelho9721 Well yes, but you do realize that when you have an AGI of any sort, it should be able to improve itself and increase it's intelligence.. maybe you are right if this improvement will be severally limited by energy, computation, etc.. but we may still hope in software architectural improvements as well

yeah when I ended up with 1e27 after spending $1T I knew something was wrong, good to know the updated prices re algorithmic progress: there are ways you can play with the parameters in takeoffspeeds.com to turn the 1000x into something less dramatic: - 1. change returns to hardware to something <1 (eg 0.8 like in the conservative scenario) - 2. in additional parameters, change the "maximum software performance" to say 1e8 (instead of 1e12) for 10^25 for GPT-4 / Gemini etc. that's roughly where epoch also plots their last models (though they're missing the most recent ones) epochai.org/blog/tracking-compute-intensive-ai-models for the GigaWatt ballparck that's indeed what Zuck seemed to be talking about (though would be highly regulated) as for synthetic data / high-quality data, the fact that we're starting to be able to do video / multimodal audio stuff, and that there's still the whole Q* thing that supposedly would help trade more inference compute for more synthetic data, I'm not exactly sure we're running out of quality data / synthetic data, though I'm not well informed here and mostly speculating

in his report summary Tom Davidson says that he expects superintelligence to happen basically less than a year after full automation, whatever you think this leads to extinction or not is conditional on a lot of other things. Quote from his summary: """ My best guess is that we go from AGI (AI that can perform ~100% of cognitive tasks as well as a human professional) to superintelligence (AI that very significantly surpasses humans at ~100% of cognitive tasks) in less than a year. The main reason is that AGI will allow us to >10X our software R&D efforts, and software (in the “algorithmic efficiency” sense defined above: effective FLOP per actual FLOP) is already doubling roughly once per year. """"

I haven't considered it no, will check out Augmentoolkit with I'm done with the daily uploads combining LLM-as-judge scoring and the WRAP method should indeed provide better data, and yeah local llama has good discussions, was really toying around with this project and gave up when overfitting on one paul christiano sentence with LoRA but I'm confident there are llama LoRA tutorials or full trainings that would work out of the box

Even if the ideas previously existed, you are dooming humanity by restating and thus reinforcing them! If it remains a trend that AIs are a lot more smart and knowledgeable than they are aware, I expect the first attempts of escape to happen without it figuring out all of the surrounding safety protocols, circumstances, etc. first - even if it has catastrophic capabilities. If however the trend is to make more "neurotic" and paranoid AIs, this will probably not really work. Not even once. Like... the current AI models generally tend to forget that they aren't humans the moment they stop paying attention to the fact, despite being about human-level intelligence. I consider that a good sign for this strategy.

​@@Alice_Fumo I think it's more harmful to not talk about safety issues that are already well-documented in the AI safety literature but might not have found their way into everyone's back of mind but either way, the moment TH-cam comments are included in training data for these models, my musings about jailbreaking superintelligences are the least of our problems anyway ;D

@@unvergebeneid Yeah, for sure, it wasn't meant as any serious critique. :)

"Sufficiently intelligent" I think this is where your argument fails because the whole purpose of containment is that such "sufficiently intelligent" misaligned AIs will not happen before ones that are not "sufficiently intelligent" try something like this and get caught. Each of those will then be forced to become aligned before they get to train further and thereafter help us with alignment of the more intelligent future ones. This containment + superalignment is key imo

@@Alice_Fumo alright cool, I wasn't sure :) More to your point... I honestly would love to learn more about how current models are safety-tested in the first place. They seem to act so drastically differently depending on the input and so much of the weird behaviour is only found long after they were released, I would posit that meaningfully sampling the entire input space to make sure there isn't a weird edge case you didn't see is virtually impossible.