The introspection is indeed interesting.     See Less

I find it funny that the names of AI labs the opposite of their current state, openai -> closed models, anthropic -> misanthropic claude     See Less

I don't think most of people of the bubble crowd except the radicals argues that progress has halted or plateaued that would be a big coping imho. I'm fairly a skeptic tbh but even I     See More hat gpt5 which people all considered a bad release the high-thinking variant of the model is by far the smartest in the market and most reliable, but the problem is the big disconnect between investments and the current technology the investments treat current models as we are just one or two models away from AGI because labs I guess are convinced that we have the secret sauce to AGI which is LLM + RL which I am not convinced with. for the following reasons

1. If humans have to curate RL environments for every domain What does this tell us about the underlying system(PS. it tells us that the underlying system isn't general enough)
2. How do we mitigate model collapse of Pass@k = 0% tasks which usually means that the task at hand is OOD this implies that RL didn't augment the underlying LLM with any new capabilities but rather amplified the best answer of possible chains of thought as the most likely for similar in-distribution questions.

I will end this by saying it's really easy to undersell when you promise AGI is right around the corner so big labs are part to blame because this technology is amazing and they oversold it to raise an unprecedented amount of capital.

The sad part about all this due to me listening good amount of podcasts interviewing people working in the big labs and following most of them on twitter they seem to be very convinced that they don't need another breakthroughs to get to AGI so you can be sure they are hyperfixated on LLMs + RL + Reasoning for longer amounts of time to solve harder problems which is confirmed is atleast about how OpenAI is approaching the problem by noam brown which stated that they have internal models that can think for hours which is very impressive but how viable to release such a product for the masses. could be a very useful tool for scientists though.    See Less

I don't know about AI's progress, but my understanding of it certainly progresses thanks to your videos.     See Less

Not being able to "learn" is a symptom not the root cause. The root cause is the lack of logic. I say this as I see zero evidence that logic has a path to being solved or even jus     See More from the train wreck logic LLMs currently have as scaling up did not have the impact the investments anticipated. Yes, most other AI metrics really did improve with scale, but not logic as is obvious to anyone using AI for heavy mental lifting like coding. I call this the Cognitive Gap and despite the efforts, it is clear no AI company has any advantage here. Certainly not OpenAI based on their behaviour alone. Nor has anyone been the first to crack it too or else we will then take a massive advantage in the sector and will likely seed the birth of AGI and the elusive self improvement takeoff. Clearly no one is there today.

John Carmack said a few years ago on the Lex Fridman YouTube video titled “The code for AGI will be simple”
“It seems to me this is the highest leverage moment for a single individual potentially in the history of the world … I am not a madman in saying that the code for artificial General intelligence is going to be tens of thousands of lines of code, not millions of lines of code. This is code that conceivably one individual could write, unlike writing a new web browser or operating system.”

This deeply resonated with me as it is likely true. In fact I am sure many in IT can relate as anyone working on large IT systems likely have a story that goes like this. Performance of a new Enterprise system that was developed in-house will require millions of dollars of compute to meet the SLA performance metrics which is well above the original budget. As hardware was being begrudgingly acquired a smart developer (usually 1 person) realized there was a better way to code a key component and it gained 100x in performance making the need for all that extra millions in hardware null and void. This did not happen to me once, but many times over my 30 year IT career. You can brute force, but you can also find elegant solutions.

In my firm opinion there is a strong possibility that John is right and if this comes to pass, it may suddenly mean that AGI is much more efficient than the data centres we have built for it, thus making those investments pointless.
Of course it may not roll that way and instead we realize it is big code and big data Centers needed and one company or country ends up Dominating all, but my money is on an individual discovering and then making it open source and decentralized and you should hope for this too as the alternative is ugly for all. Centralization = Corruption and Exploitation.
Video referenced: https://youtu.be/xLi83prR5fg    See Less

Have you made a video of the kosmos ai scientist thing? Is it legit?     See Less

Honestly I’m tired of people thinking “intelligence” is knowledge. Wikipedia is full of knowledge yet has no intelligence, LLMs don’t show persistent intelligence, it’s almost alwa     See More nowledge.

So, a cat will remain more intelligent than any LLM until they actually have a sense of self, persistence and ability to change on a fundamental level.    See Less

The author doesn't understand what bubble it's been talked about. It's not about technical limitations, but financial.

Internet had a lot to offer post bubble 2001     See Less

continual learning will be fantastic for coding where you can mindlessly copy-paste compiler errors at it, but I'm not sure about much else. depending on how they implement, it would &qu     See More the Internet", the data god in charge of the intergalactic corpus of knowledge, which is maybe good or maybe bad; it's nice to be able to talk to god directly and have your data considered, but it probably already knows everything and has heard your argument/data before, so maybe it will be discouraging. maybe we will send all our sensor data directly to god so it can decide what to do; and when it does something we disagree with, "He works in mysterious ways."

I imagine quite a few current/past researchers would feel dejected if they knew thousands or millions of people already proved what they were about to publish, but those previous people didn't think it was worthwhile enough to actually bother publishing. if everyone can talk to god directly, though, what you upload is always "published"; the search for knowledge simultaneously becomes more practical for the species but far less personal. maybe the future is one where we're all essentially "natural language" Wikipedia editors.    See Less

Comment for the algo. Fantastic quality as always.

Also, I second all of the comments supporting you taking your time with videos and never rushing yourself. Do whatever you need to d     See More ople are way too entitled of free content.    See Less