Amazing video as always! Clear, concise and widely understandable about cutting edge research! 🎉     See Less

5:48 Cryptocurrency actually surprised me. What's the negative impact from cryptocurrency other than the fact that cry     See More re annoying? If you don't like it... don't use it.    See Less

Thank you for explaining AI in 2025 and Merry Christmas!

- I predict 2026 will be a "normal" year of gradual improvements on every axis (the sci-fi stuff will start in the     See More .
- There will be a lot of stories about AI assisting with scientific discoveries (the labs, especially OpenAI are heavily hinting at that).
- Another thing that will be exciting to follow is progress when it comes to continual learning. Here it is Anthropic that is leading the hype 😁. Going to be interesting to see what they got!
- I also expect more multimodal integration. I think Hassabis (Google) hinted at that in his interview. NoteBookLM assisted by nano banana is already doing some of that, but I think there is a ton of room for improvement.


So there will be a lot of cool stuff to explain in 2026 as well 😎!    See Less

Every discipline has their own tools. A builder uses a hammer and a saw. Currently, AI has only mastered software engineering tools. It will never generalise out of software engineering with     See More ng on other tools.    See Less

No introduction for Hannah Fry??

Rude.     See Less

32:00 - that's scary. It describes something very similar to the ageing human's mind.
{O.O}     See Less

you are genuinely the best channel there is for news about AI. Merry Christmas!.     See Less

do you remember when earlier in the year everyone was roasing him for this 90% of the code line..... he was spot on     See Less

I really don't like the question #5 in the Simple Bench example. Well, firstly, there is no question. The last sentence should be changed to a question "Will Paul attempt to help     See More t;. Without the "attempt", I, as a human, would answer "A: Probably not". Since Paul, like most people likely is not medically trained to know when CPR is appropriate, nor how to do it. So, Paul is unlikely to help Peter. If you added the word "attempt" and converted it to a question, then, I would provide the answer your test expects. I fear that you penalize models for providing the correct answer. Now, I can anticipate your intent and if asked to do so, I would expect the model to also provide the answer you seek, rather than the correct answer to the question. But that comes down to your instructions. I have concerns on the validity of your model.    See Less

To be honest, I'm much more in the pure scaling camp, though I do still think we need architectural breakthroughs alongside scaling. People tend to dismiss the idea of a fast takeoff jus     See More we can't see evidence of it right now. But we wouldn't. Once AI gets the ability to self-improve without human assisstance, which will probably be by end of 2026 or in 2027 at the latest, everything will change. And we might look at AGI one day and ASI the next. That is when the "fantastical" predictions of the pure scaling camp start to hold more water. But I suppose time will tell. Still, 2025 has been the greatest year in AI so far and I have no doubt in my mind that 2026 shall far surpass it.    See Less