Nice video here, from paying for day care and college, to managing mortgage payments. I'm approaching retirement yet inflation is getting worse. How can I generate more income to retire     See More ast $3m for long term care? I have about €750k in savings    See Less

We need an international AI safety treaty     See Less

Next frontier is to get models to ignore stuff they learned when appropriate to not fall for these traps like the line length example xD


Also when will we see in-context learning     See More d RL? A model that is exceptional at that could be really cool.    See Less

I have not accomplished anything in life yet and the AI will just come and take it all, even the hope that I had, what will I live off?     See Less

5:25 Absolutely! I work as a Digital Transformation Supervisor, and I work with a Google workspace for education fundament     See More n, and thew Google Admin is as intuitive as Yamaha remote controller... So when I was stuck with MDM settings for edu fundamentals edition, Gemini confidently sent me on a 3+ hour fetch-quest of changing settings that ddidn't have anything to do with mobile devices, and chasing settings that didn't exist. Ib felt so cheated after I realized it was all a hallucination of a solution that doesn't exist, and I had to back-track the settings I had (excuse my French) fucked with blindly.    See Less

great video     See Less

the dough-velopments in AI     See Less

Thanks for this video I find the quality great, and you provide an objective intellectual view on recent events.

In my previous comments in your past videos, I may have come across     See More ning when you dove into more complex topics that involve technical AI concepts. For someone like me, who often watches YouTube recreationally, it can be a bit much to shift gears and focus deeply just to follow along. However, I do understand this channel is AI Explained, so that approach makes sense.

If the occasional complicated idea shows up in one of your news videos, I think that’s totally fine. And if you decide to make separate videos clearly labelled as lessons for the more advanced concepts, that works well too.

In my view, being successful on Youtube means staying true to your identity while adapting about 50% to what the audience wants, but I think you're doing everything right. I hope you find this advice helpful.    See Less

AI hallucinations, and the Apple 'LLMs can't reason' paper and such are real, and they will NOT be remedied by merely scaling up models, or with training models on more data. BUT     See More ge that needs to be made is simple, in concept at least. LLMs can not reliably and accurately model absolute binary logic. This causes them to get 'lost' in things which they believe are highly correllated with truth, but which do not stand up to the rigor of absolute logic. In logical arguments, there is no such thing as "0.9999 probability of truth". There is 1.0 probability of truth, or 0.0. With no value in between. Even the slightest error does not simply make the whole argument (no matter how large that argument might be, it could be thousands of pages) 'slightly false', it makes the whole thing completely false. This is a degree of rigor that most people aren't familiar with, and it doesn't get used continually for even philosophers who want to try to do so, but it is required. It is the only strategy which modulates the cognitive biases which lead people to accept "it feels really true, it aligns precisely with my intuition" as true. Right now, LLMs are operating on pure intuition. Even when they "think" with test-time inference compute scaling, all they are doing is feeling their way through arguments that FEEL true. They can't look at something, recognize that it feels deeply true, fits right in with everything else true, but has a tiny flaw that makes it completely false. Logic is not gradients. It is discrete, and absolute. Compare human history before and after The Enlightenment. The core of The Enlightenment was integrating exactly this kind of absolute logic into our wheelhouse of thought, despite it being alien to us. Despite it blowing apart things that felt as true as our need to breathe. That enabled us to formalize things, to take our human preconceptions and feelings out of it and process the world just based on rules and logic. And that enabled us to build computers that apply that absolute reasoning trillions of times faster and more efficiently than we ever could. Even enabling us to reproduce a model of our intuitive brain with AI. But the model needs access to that binary logic in addition to its fuzzy logic. Just like we did. That's not transparently obvious as to how to do it, but once it is done, expect a warp-speed leap in capabilities and especially efficiency. Continued scaling will just waste mountains of compute on emulating imperfectly the absolute rules that it could use when necessary.    See Less

we out here Phil, waiting for the o3-pro drop haha     See Less