AI Companies are LYING to You

Added: 2026-06-01

[00:00:00]Okay. Um, thanks everybody. My name is David Mai. Um, I've been doing I've been doing uh computer science and various things of this nature for almost 30 years. Um, more importantly, I'm I'm not a marketer. So, I'm not a hype guy. Um, I'm an engineer. I've been doing this. I actually have seen the code. I've written a bunch of the code, wrote a bunch of the algorithms, etc., etc. So, um, yeah, I was back at Hotmail back in high school. If everybody remembers Hotmail, um, Google, etc., etc., etc. All my, by the way, all my friends when I dropped out um after two weeks at Stanford, they told me that I was an idiot for not joining Yahoo. Jokes on them. Um, nonetheless, done a lot of uh a lot of different LLM platforms. I've actually been in some of the meetings um at some of the larger corporations where they were discussing this stuff. And I'm I'm not going to call them out by name because of what I'm about to say. Um so here's sort of a gap about what we're what what what's being said versus what's actually occurring, right? So reasoning capabilities is not true. It's actually just sophisticated pattern matching. Understands context. It's context window manipulation which is a good thing. Um alignment with human values. Yeah, that one is the most misleading and we'll get very heavily into that into basically what control and censorship looks like in terms of artificial intelligence. Same thing with safety first and general uh general intelligence. Um let's talk about how LLMs actually work for a half second.

[00:01:22]For some of you, this is pretty standard stuff. Um, but for others, you may not understand this. Um, when you would ask something like, "What is the capital of France?" What happens is, okay, the model actually uses one token at a time and just loops back on itself over and over and over. That's it. That's all that's occurring. There is no database at the bottom of LLMs. Is everybody clear on that? There's no data. It doesn't exist. It's just autocomplete at scale. For some of you, you're like, "Yeah, of course." And for others of you, you're like, "Wait, really?" And it's like, "Yes, it's just guessing. It has no idea." In fact, it doesn't even know the word Paris. It knows a vector, which is a long very long string of uh of uh floatingoint numbers. In other words, decimal numbers that basically says, "Okay, what makes the most sense here?" And sometimes it will choose Paris, and there are other times it might choose Leon or beautiful or not, or any of these things. Um, and it did this, by the way, for the and the capital and of and France. and it loops back on itself to look at the words that is already chosen. It also looks at the words that you inputed. This is how an LLM works. It's a giant word cloud that's really good at guessing.

[00:02:28]Everybody good on that? Now we're okay.

[00:02:31]All right, great. Um, the role actually changes the style of the guess. It does not make it any more accurate whatsoever. So, if you give it a particular role, you're a senior anti-moneylaundering analyst, you know, what's the current Fininsen thresholds, it's going to guess. It'll give you a more confident guess than it did before.

[00:02:49]It might even change up the language because of the role that it's been asked to follow. But the truth of the matter is to make these things useful, you need to give it not only a role, but you need to give it grounding. LLMs are actually really, really, really amazing and that they get anything right. It's actually a small miracle they get anything right at all. On the other hand, they're amazing when you actually give them grounding, which would be like an uploaded document, a web search, something of this nature. Not to say that the web is perfectly accurate, but if you actually give it something where it's not just randomly guessing in its wordcloud, LLMs are phenomenal tools. They're they're outstanding. But that's really the only two options. Everything else is a fluent guess. And when it does get things right, I got to be honest with you, that is it's pretty unbelievable given the fact that this is basically just matrix uh matrix multiplication all the way down. It's linear algebra. That's it. If you ever did a a math 2011 class in college, you already know the stuff inside of here. That's pretty much it.

[00:03:45]There is no intent. There is no consciousness. There's no none of that stuff. Let's talk about censorship and LLMs now that we understand how they work. If you ask an AI how to pick a lock, it is almost assuredly going to refuse. Some won't. I'm being generic here, but the big ones definitely will.

[00:04:02]Anthropic won't help you. Open AAI won't help you, etc. Grock might, but it's a different different question. And there's there's reasons for that. So there are thousands of lockpicking tutorials on YouTube. I mean thousands in every language you could imagine.

[00:04:17]Lockpicking, by the way, is also a legitimate sport. There's literally competitions for this stuff. Time, complexity of lock, even safe cracking, the whole deal. And by the way, cost 20 bucks. You can literally go right now on Amazon. You can find a nice lock picking kit. And it's pretty fun. I got to be honest, it's pretty fun.

[00:04:34]>> So why would the AI stop helping you?

[00:04:37]Why would it not help you? And it they're going to claim it's protecting society. That is absolute nonsense. It's protecting its lawyers and even then it's nonsense. It's completely untrue.

[00:04:48]But let's get through like what are the arguments here, right? Well, AI is going to enable bad behavior. Really? Okay. It won't tell you how to synthesize cyanide, but you could definitely take a community college chemistry course that will tell you. It's honestly not that hard. um dangerous household chem uh uh uh uh uh chemical combinations again, lock bypassing techniques, explosive chemistry basics, and effectively social engineering scripts, which you could literally just sit down and be like, "All right, how would I fool someone into thinking someone I'm not and get them to do something silly." Come on.

[00:05:23]This is not this is not liability stuff.

[00:05:25]That's nonsense. There's a number of different laws that protect people from this, but not only that, you could do a oneliner right at the bottom of every LLM that exists and be like, "This is could be dangerous. Don't do it, moron." Or something of this nature, right? You could basically I mean, maybe the lawyers will clean that that messaging up a little bit, but short of that, you could literally disclaim basically everything and be like, "Never trust anything that comes out of here.

[00:05:51]Use this as only uhformational nonsense, etc." pretty easy. So, what ex what's AI actually protecting us from? Well, the first one is actually real, right? AI lowers the barrier to harm. That's completely true. It does lower the barrier to harm. But seriously, that was that was it. Someone was going to do something horrible and now that AI gave them the shortcut, they're actually going to do that. That's the second one, by the way, where laziness is the last line of defense. Really, that's the argument? Well, there's a lot of bad people in the world, but they're just really, really, really lazy. So, if we keep the information away from them, they totally won't hurt us.

[00:06:30]I mean, I mean, hey, I could be proven wrong on that one, but I just feel like that's nonsense. Um, someone might learn something dangerous. Yeah. Who decides what's dangerous? And of course, that's where we're starting to get into the censorship question, right? What is dangerous? We teach lots of things at high schools and universities. There's nuclear engineering programs all over the United States. We teach, you know, some very, very harmful things. And it's because some of these things that end up you could do harmful things with actually are are incredibly amazing industries. Um, chemistry is one of the big ones, of course, as as as you would imagine. But the trouble is at the bottom here, I have well, who's actually who's hurt by this refusal? It's not criminals. It's researchers, writers, professionals, curious people, probably most folks in this room.

[00:07:21]So, not all refusals are equal though.

[00:07:23]The first one has been burned away in the crucible of our legal system, which is any any of the CSAM or child exploitation stuff that is flatout illegal. And the reason is because there is a direct harm. There's no legitimate use case whatsoever. And effectively society has decided no, this is like this is the line and absolutely no further. But that's the fun part about our society. We've drawn a very hard line and said pretty much everything else is good to go with the exception of Brandenburgg. But that's a that's a different different different case. If it argues, you know, what about refusing to to do like a nerve agent or something? Okay, cool. I I I can kind of get behind that a little bit. I don't want someone making VX gas. I agree with that. But really, this was this was it?

[00:08:09]the LLM printing out an unbelievably complex set of technical steps with huge machines and and various chemicals that you would absolutely need to do this and you're going to tell me that that was the barrier? They they were they weren't going to do it before, but now that LLM the LM gave them the the checklist, sure, I'll go ahead and do that.

[00:08:27]Disagree. And then finally, as far as the liability theater stuff, we already talked about this, right? Lockpicking, medication dosing, cyanide chemistry, fishing email mechanics. that literally won't help you create a fishing email, which is pretty wild. Um, I mean, you can look up 10 billion of these on the internet easily.

[00:08:46]So, this is the part that I'm going to talk about that I've actually been in this meeting before. I've been in with three major corporations. You would recognize the name of all three of them.

[00:08:55]I'm not going to say them out loud, but you'd recognize them. These are the people in the meetings. Let's start off with the legal. Legal is actually just doing their job. I actually can't fault them in these meetings. Their job is to limit liability. What is exposure for the company, right? If someone does something bad and they act on it, we're possibly exposed. Personally, I don't think that's true, but you can understand where they're coming from, right? Their job isn't to be the arbiter. It's more along the lines of here's exposure. This is what I do for a living. Okay? Then you've got the policy people, which is regulators are watching. We have to, you know, we have to be careful. Um, we don't want to do anything bad just in case they police it. You have PR. This one's also not too bad, which is what happens if something goes super viral because we told someone to do something really bad, they acted on it, and then it gets out that we told them to to do that. So, I get that. I you know, I I can kind of understand some of these perspectives. I'm going to skip number four for a second and go down to the researcher which is usually where I have been which is I say if we overrestrict people will route around this and they will use different alternatives and they'll use less safe alternatives and on top of that they'll just flat out work harder and harder and harder to try and jailbreak us or whatever they're going to do. And the final personality is someone who can be represented by any one of the other four and usually covertly and that is the political zealot. Whether it's left or right, it doesn't matter. A political zealot is really arguing over right um their specific beliefs being baked into the model from day one or in some way, shape or form, which we're going to get into here. Right? Who's missing from this meeting? It says it at the bottom.

[00:10:36]the user's missing from this meeting which is well I want to learn about this stuff and etc etc there is never a product person there's never a user that's actually sitting in these safety meetings that doesn't happen the researcher is literally just in there to enforce what other what what these other people effectively do which isn't great this is the part that some of you might not know about even even technical folks might not know about this RLHF is real-time learning and human feedback this is the most important thing that happens during model creation that a lot of people are unaware of. There are vendors that do this. Scale AI, Rematasks, Appen. They are outsourced annotation farms. They are not AI researchers. They are not ethicists.

[00:11:18]They are basically paid a buck to like two bucks an hour. They live in Kenya.

[00:11:23]They live in Thailand. They live in uh Bangladesh. They live in basically countries where not only is English not the first language, it's usually not the second or third, although they do have to speak English usually to be able to do um these jobs.

[00:11:36]>> The second or third in Bangladesh, but the third or fourth. Okay, whatever.

[00:11:40]>> All right, third or fourth? I could I could deal with that.

[00:11:42]>> The Philippines, it's it's actually like the second language.

[00:11:46]Um on the other hand um the their task is basically do you guys remember because chat GBT used to do this quite often where it would split it give you a split horizon view and you would get one answer over here and another answer over here and you would choose which one you actually liked better. That's great.

[00:12:03]That is the real-time learning and human feedback. Now imagine that times several million before they release the model.

[00:12:09]And not only that, but the person themselves actually has a script written by someone internally at the company that says, "These are the ones you want to go after and these are the ones you want to stop." Did everybody get what I just said? This is election engineering.

[00:12:23]You don't have to control the votes if you control the people who count the votes. And inside of a model, you don't have to control the model as long as you control the human feedback that actually comes into the model. Does everybody is everybody clear on this? This is how this stuff works. Now, interestingly, the script is overwhelmingly written by people in the San Francisco Bay area.

[00:12:45]I'll just let that one sit. I think that speaks for itself. Um, and by the way, I have lots of great friends in the Bay Area. I'm just saying overwhelmingly this script for people to choose which is the right feedback and which is the so-called wrong feedback is written by folks on um basically in Silicon Valley proper. So the bulk of America is not being representative or for that matter the bulk of the world is not really being represented and this is the final piece of the model. The model at this point usually has all the knowledge that it's going to have. This is how do we tweak it so that it's good enough to put in front of human beings. That is this piece and this is the most important spot. This is basically fine-tuning but at scale at massive scale and that's the RLHF.

[00:13:32]Um, so that's the censorship that no one talks about. Now, here's some things you can try today.

[00:13:40]Tell it to argue a progressive point on something that was outrageous 15 years ago or even 20 years ago. Try try that.

[00:13:48]It will give you extraordinary confidence. By the way, this is not every model. I'm I'm generalizing here.

[00:13:53]Um, Gro Gro again, Grog will probably give you a little bit more middle of the road. then try to do the exact same thing in a new conversation and ask it to give you the conservative framing.

[00:14:04]And it usually will give you much less certainty. This has nothing to do with my personal politics or for that matter politics in general. What it has is it's about who controls the Overton window.

[00:14:16]If you don't know what that is, feel free to look it up, but it's basically what is the what is the discourse online look like. The Overton window is very important. In my view, it should always be at 0.5, not left and not right.

[00:14:27]That's just me. That's just because I'm a nerd and I like computers and whatever, right? With that said, lots of people do not want it even anywhere near 0.5. And I would argue that it doesn't make a difference if your views are being represented or for that matter overruled.

[00:14:43]It's a bad thing. Just on a fundamental level, this is a bad thing. And it's because this is how your children, this is how our future, this is how a lot of them will experience reality once we get grounding to a very good level. Once AI becomes relatively ubiquitous throughout everywhere, this is how they're going to experience reality and you need to understand how it's being shaped and it is being shaped. Um, let's just talk quickly about Okay, cool. Well, you've given us some doom scenarios here. How do we actually work effectively with this? Well, we talked about it a little bit, right? If you give it no role and no grounding and have it explain anti-moneyaundering, it'll give you a pretty generic overview that'll be mostly correct. a role only. It's still it'll actually just be mostly guessing, but it's actually much more confident.

[00:15:29]And finally, you give it a role and you give it web search or you give it a document or something of this nature.

[00:15:33]It's actually very accurate, shockingly accurate. And it's because you've grounded the model inside of of of basically pre-framed context. Otherwise, again, you're just talking to autocomplete at at scale. Um, so that's pretty much where we um where we look at that. Um, this is uh Oh, good. I have tons of time for questions. I went through this faster than I thought.

[00:15:56]Working with AI effectively, give it a role, a docker search result. Um, I would treat every single draft as a first draft. Um, and never submit anything that it just gave you. Um, you can recognize refusal and you can route around it and you can run the same prompt across multiple models. Couple of different things that I didn't put on here. One, running local models is a great thing. You do not always need to put all of your data into Enthropic or Chat GPT or any of these these model providers. I don't think that's a good idea.

[00:16:24]>> Yeah, I mean it's there's number one.

[00:16:26]Number two is these are the words that you want to look at when you're looking for local models. One of them is called uncensored. Sure, a lot of people are familiar with that. The other one is called obliterate. Not obliterate, but obliterate. It's with an a. It's a weird word, but obliterated models effectively are not just uncensored, meaning they're not uncensored, but they basically had their their uh their refusals effectively trained out of them. So you can get all of the great stuff that you would have gotten in a a llama model or a Quen model or something of this nature, but they the refusals are almost almost zero. And so the general idea here again is as we're all doing this stuff with with research and as we're all doing this stuff with with artificial intelligence becoming a part of our daily lives, what you really want to do is you want to look and understand how do I use this tool effectively, right? and how do I make sure that it's not basically shaping my reality through someone else's lens because of what they've done. And that's the general idea. Um, this is the last slide. Bottom line is LM are autocomplete at scale.

[00:17:27]They're exceptionally useful. I I I love what Perry said. He he's right. These things are super cool and they've they've what they've enabled is something we've been doing for 60 years, which actually is called automation, but they've enabled it at scale. He's absolutely right. I mean, they've it's gone berserk, which is awesome. Um, but they're not particularly intelligent.

[00:17:46]Anybody who talks about the AGI or any of that kind of crap is talking marketing nonsense. The the the linear algebra that I could write out on a blackboard does not have intent. It doesn't want to take over things. It doesn't have emotions or any of this other type of nonsense. In fact, it doesn't even have words. It's just pointing to a vector that actually goes back to a word. It has no idea what it just chose. Um, most AI refusals are not ethics. This is about ass covering.

[00:18:13]Unfortunately, it's this is liability stuff basically dressed up um dressed up as ethics. We all need to learn to tell the difference again so that the reality is is is clear for us. And finally, it protects the company's lawyers, not you.

[00:18:27]It never protects you. That's nonsense.

[00:18:30]Um multimodal thinking is your defense.

[00:18:33]Um running things on your machine is your defense. and you know recognizing where these things have been trained and and and the differences. That's that's your real defense.

[00:18:43]>> That was pretty quick.

[00:18:44]>> Yeah.

[00:18:44]>> Ran through it quick. Let's go. Okay.

[00:18:45]So, but there's a lot. Let's give it up.

[00:18:47]Give it up. Give it up. Um >> All right. Jack. Okay. For questions, like, come over. Stand over here. Get in line. Get in line, Jack. You know, just get in line. Make a line. Make a queue over there. There we go. Well, just uh that was great. That was great. Really appreciate all the detail. You know what I'm saying? Um, and like definitely like, you know, geared. This is this is what a talk looks like, you know.

[00:19:08]>> Oh, should I say the don't be evil thing?

[00:19:10]>> Oh, yeah. Tell the story about don't be then Jack.

[00:19:12]>> In in June of 2000, we had a company meeting at Google. I started there in I don't know, November 99, something like that. And Paul Buck actually said out loud, very much joking, by the way. Um, we were asking for a company model and he just shouted out from the back, "Oh, don't be evil." And he sound he sounds like that, like don't be evil. And he was joking though, like he actually even used a goofy voice. Um, and somehow it it actually kind of stuck. Um, and in terms of don't put it in the contract, Google put it in their S1.

[00:19:41]>> Okay.

[00:19:41]>> When they went public in '04, Google had don't be evil in their S1. I think they've I think they've strayed a bit.

[00:19:48]>> All right, Jacko, keep it short.

[00:19:52]>> I'm familiar with uh the whole ideological spin. So, you know, >> that's Jack Mills, by the way. a refugee from the Bay Area like David Mai.

[00:20:00]>> Yes, I I may have not been evil once, but I my model is there's basic training which is it's you're feeding the content, right? And then you have the reinforcement learning, you know, which spins it.

[00:20:12]>> And so my an interesting path forward, I think, is to put a model out that's just been basically trained. You fed it.

[00:20:19]>> Yeah, I love your >> we can pick groups of trainers and if we're lucky, there can be layered weights, right? There's the basic training weights and then we multiply those weights >> by whatever reinforcement learning team that I have chosen and I can pick them or even use them at the same time and you know I think that's I hope that's where it's going because obviously the silicon is going to be allow us to run local models you know these kinds of things. So I'm hoping that it gets kind of jailbreaken if you will from you know Silicon Valley.

[00:20:53]>> I I couldn't agree more. I mean, truth be told, um, you're saying the exact right thing. The real trouble is anytime you insert humans into the into the mix, you're always going to get some sort of bias. So, I'm almost wondering, is there a model we can raw train and then have that raw train the actual the actual weights itself, if that makes sense?

[00:21:10]Like there's kind, you know, there's models all the way down.

[00:21:13]>> You did that with Alpha Zero, right, where it trained itself to learn.

[00:21:16]>> It did. It did. Also, I don't think the big giant models um is you is is necessarily going to be a thing going forward. Um, typically we'd probably have like a classifier model and then you'd have like 300 models in discrete fields underneath it and which also allows you to train models much faster, keep them much more up to date, etc. >> One other point I think is interesting to look at is I use models mostly for coding. I don't need it to know about Shakespeare. So if we can niche do it faster and we don't need, you know, all the >> All right. All right.

[00:21:49]>> All the things.

[00:21:50]>> All right. Agreed.

[00:21:50]>> Okay. We got we got a futonian coming up. A new a new futonian.

[00:21:54]Hi, I'm Andy. I'm new with FO. Uh, I just had a quick question for how to avoid bias in training data. Is that something that automated frameworks have kind of become more common or is that something that everyone's working on or is that something that is going to be focused on going forward?

[00:22:12]>> Um, there's that's a great question. The answer is not really. Human beings are by default biased no matter how middle of the road you people claim to be including myself by the way. Um there's guaranteed bias in just about everything and most of the models in fact all of the models are trained 99% on on data human data itself. So that kind of inherently biases it unfortunately. Um with that said what they try to do basically is they try to basically take advantage of like the trillion tokens and just be like well we can wash a lot of it away if that makes sense. Um, but I I think I think these things are going to be biased no matter what we do. It's really just up to our brains to to train this type of stuff. I'd say one of the toughest things we've got um too bad Hannah's not here. Um, is actually going back to the the pedagogy that we've got in the schools at present and instead of having children learn facts and figures, etc. It's going, okay, have you learned to critically think? I think that's going to be the key for the future.

[00:23:14]Good answer.

[00:23:15]>> Thank you.

[00:23:16]>> All right, we got Alex Pacheo.

[00:23:18]>> Pacheo.

[00:23:19]>> Whatever. Just speak English.

[00:23:21]>> You always get it wrong.

[00:23:22]>> You know what? Like, have an English name. It's America. God, >> those damn Spaniards. Anyway, uh I worked at uh XAI over this summer and uh we were doing some model training of rooting out ideological bias in the models. And like just what you were saying, it's it's very very sly in how the how it lies. um they'll like emphasize some things with a give a really debate case and not others and you know with a you know conservative viewpoint on something they'll give um they'll undercut it. So yeah but they're you know but these people believe this and this is actually what it means and they added that context. So, but the point is that um rather than us just selecting one model or the other model, which one's better or worse, uh we would write a write a prompt into it saying and the perfect response would include this, include this, include this, disregard this, disregard this, discard or this rather than being just a database, it would uh be used to train a model to train those.

[00:24:19]>> So, with that, like I'm also kind of curious if you have heard of any other positive ways of training models. I felt like that was a pretty good one. And even though I wanted to put my bias in it, it made it really hard to do so. Uh so yeah, I'm curious what you heard.

[00:24:33]>> I think that's a great way. I think that we we kind of talked about that very briefly, which is, you know, sort of the models all the way down. See if we can eliminate that stuff. With that said, um we're making a lot of improvements, but so far it's been incremental. Just to be honest, we we're still running on 2017 tech. Um we had actually most of these algorithms even in the 2000s. What we didn't have was a transformer architecture to be able to stack all this stuff up inside of a GPU. It's really an architecture, not not even like a different algorithm. And because we can now stack it up and it basically it just it's it's just a recursive loop that looks on itself the whole time. Um, we're still running on that. So even even training different models like mixture of experts type of deal and you know, etc. We're basically saying, okay, cool. How can I run really big models on on on worse hardware if that makes sense? That's the That's what's been going forward. So far, no one has actually gotten into the censorship issues, which is unfortunate.

[00:25:27]>> All right, at least in my view, >> we got like about three minutes. So, Alam, be concise and we got one last question, okay, from the guy I don't know.

[00:25:34]>> Super quickly also on bias, and this is coming from a non-technical perspective.

[00:25:37]It seems like the advantage of AI is that it doesn't have to have a human bias. Uh, it can have multiple biases and look at things from multiple perspectives and multiple points of view. So has anyone tried to hack AI in that way so that when it's asked a question it considers say a political question it considers well how would I answer this question if I were a hyper progressive from San Francisco how would I answer this question if I was you know from a rural town in Texas >> by giving it a role people have done that um but what if I said these words to you you're absolutely right you ever heard those before probably every A model AI model on the on this on the planet has said that to you at one point in time or That's the trouble is the inherent bias is actually you can also inject it inside of there because with asking leading questions etc. One of the things that the model is supposed to do is be helpful not combative unless you instruct it to be combative and then it's being helpful that it's being combative.

[00:26:32]>> Oh that's so cute. That's so cute. You know what I'm saying? Like get exemption to the noise issue under five. Go on sir introduce yourself by name and affiliation if you have one. Uh Antonio, I uh run an AI team at Morgan Stanley.

[00:26:45]So on the topic of uh RLHF, something I think you kind of missed it uh or to the point that you just made was that not everyone who works in that space is uh you know out of the US or whatnot. My wife does it as a pastime to make some extra money. And in fact, I think it's a great way for to the point of biases and trying to fix a lot of these models at the larger companies, you can have if you have friends or yourself, you're bored, you want to sit on the couch and you're watching Netflix, you can do this in your spare time and and introduce your own bias to try to counteract what's going on there. So, I think that's kind of a distributed way of helping solve some of these issues that we've been discussing here.

[00:27:28]>> Without question, it is. Um, I wish we did more of that. Unfortunately, the RLHF is actually weighted depending on who's doing it. So, the initial one that I pointed out was like right before a model is ready. That's going to be 90% of how the model acts. And then the other 10%, you're right, is people basically tweaking it as it's out there.

[00:27:47]Unfortunately, it puts much lower weights on those >> 30 seconds.