Leo, You are Living Under a Rock When It Comes to AI

[Eskilon]

10 minutes ago, Natasha Tori Maru said:

Intelligence requires emotion. 

E-motion, internal motion? Consciousness?

[Natasha Tori Maru]

2 minutes ago, Eskilon said:

E-motion, internal motion? Consciousness?

I think when you make the leap to consciousness it becomes a bit harder to argue

There is a lot of neuroscience out there that indicates emotion/feeling is required to make decisions (coupled with logic, pattern recognition etc). This is the value / preference assessment of intelligence.

AGI would, in my mind, need to be some sort of simulacrum of the human mind, complete with emotion/feelings centres - to be able to operate in the way people would expect. 

[Leo Gura]

4 hours ago, thenondualtankie said:

Modern AI can program itself with literally zero training data.

Absolutely not.

[Leo Gura]

9 hours ago, Archie said:

@Leo Gura Given your understanding of what intelligence is, do you have any insights on how to create AGI? or models that are better than our current one? Do you think AGI is even possible?

I don't know how to create it. I just know it cannot be done by scaling the data of an LLM.

It requires a new architexture than LLMs use.

Probably AGI is possible, but much harder than tech bros assume.

8 hours ago, Butters said:

So what's your EXACT benchmark for intelligent AI? 

It needs to be able to handle novel situations without specialized training data for that situation.

It also needs to be able to do math through reasoning, not probablistic guessing (as it does now).

You will know it's true AGI when it uses very little data but is still intelligent.

Memorizing the whole internet is not real intelligence. I want to see an AI that uses less data than my PC but is still smart.

Edited 16 hours ago by Leo Gura

[Leo Gura]

Apple and Meta both did research that show LLMs are not actually intelligent:

The top researchers know that LLMs are not AGI.

My point here is a technical one. LLMs cannot handle tasks they were not trained on. LLMs interpolate existing data. They do not generate novel skills or creativity.

LLM cannot code in a language not in its training data. A human can. LLM cannot do basic math. Human child can.

As long as AI requires so much data, it is cheating, not true intelligence.

Edited 16 hours ago by Leo Gura

[yetineti]

@Leo Gura You're training data. 

[Leo Gura]

22 minutes ago, yetineti said:

@Leo Gura You're training data. 

That is how tech bros see you.

Edited 15 hours ago by Leo Gura

[Elliott]

I was thinking, if a reasoning model gets good, I don't  think the government will let it be known publicly.....

Not that I have any evidence that they can get very good. But the government would not let that out at least not until they utilize it fully for their advantage first.

[Natasha Tori Maru]

@Elliott Compartmentalise to subcontractors

[Jodistrict]

2 hours ago, Leo Gura said:

Apple and Meta both did research that show LLMs are not actually intelligent:

The top researchers know that LLMs are not AGI.

My point here is a technical one. LLMs cannot handle tasks they were not trained on. LLMs interpolate existing data. They do not generate novel skills or creativity.

LLM cannot code in a language not in its training data. A human can. LLM cannot do basic math. Human child can.

As long as AI requires so much data, it is cheating, not true intelligence.

The Tower of Hanoi game was solved in the 1970s using symbolic processing and search heuristics over a search space.   It’s precisely the kind of problem LLMs aren’t good at, as well as even conventional algebra.  Keep in mind that the LLMs can learn and have learning built into their architecture.  This is the real advance.  But, no, it is not AGI.

[Jirh]

Most AI chatbots that I've interacted with so far are more intelligent than most humans. They understand what I say more precisely than anyone I know like 95% of the time, even when I am being brief and vague.

[NewKidOnTheBlock]

Be carefull with this guy bro, yall are looking up to him but he's gonna fearmonger that all economy is gonna crash soon and that you shouldn't have most of your money in ETFs lmao, who cares about statistically best form investing, who cares about returns compounding via multi decade DCA, just pull out of the market and keep at least 50% cash bro! Otherwise you're totally gonna wiped out bro! (but maybe you won't, this is not an investment advice!) LOL

It's all gonna crash guys, just you wait, I mean it was meant to crash 6 months ago, but it's totally crashing now!

Edited 11 hours ago by NewKidOnTheBlock

[Butters]

6 hours ago, Leo Gura said:

You will know it's true AGI when it uses very little data but is still intelligent.

Memorizing the whole internet is not real intelligence. I want to see an AI that uses less data than my PC but is still smart.

Kimi K3 achieved frontier-level performance by maximizing architectural scaling efficiency rather than requiring vastly more training data than its competitors. Developed by Moonshot AI as a 2.8-trillion-parameter open-weight Mixture-of-Experts (MoE) model, Kimi K3 relies on an ultra-sparse architecture that uses 2.5x less compute to achieve the same capability scaling compared to its previous generations.

https://huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei

That's the latest Chinese model everyone's talking about. Frontier level at a fraction of the cost, and you can run it fully on your computer.

And they did all this while being export banned on the best Nvidia chips that are used to build the American models.

Edited 9 hours ago by Butters

[Butters]

And of course, it just keeps going, so how do you know AGI is not close then Leo?

"The Frontier Model Release Velocity Index shows roughly 12+ substantive frontier releases in Q1 2026 versus 6 in Q4 2025, with a sustained pace of about three meaningful launches per week through March."— https://www.digitalapplied.com/blog/frontier-model-release-velocity-index-q2-2026

[NewKidOnTheBlock]

He's gonna keep denying until he sees a full fledged pizza delivery AGI robot standing on his front porch, disturbing him from his trip. We all remember that memeworthy moment, if you know, you know

[CosmicExplorer]

7 hours ago, Leo Gura said:

It needs to be able to handle novel situations without specialized training data for that situation.

What about a benchmark like ARC-AGI-3? They give AI a private set of games that AI has not seen before, with no instructions or rules on how to play a game, they have to adapt on the fly to novel interactive environments. It's not just brute-forcing it as they very heavily penalise it in their scoring system. Humans are scoring ~50% on it, current best AI is 7.8%. It still sucks compared to humans but it's not 0. AI is able to solve some levels.

Quote

GPT-5.6 Sol is the standout model of the GPT-5.6 family. Sol at max reasoning effort is the only performant model (as of July 2026) averaging 13.33% on Public and 7.78% on Semi-Private. It is the first model to win an ARC-AGI-3 public game (ft09, 87%). Sol is able to read an unfamiliar scene correctly and in the game's own vocabulary. It treats a failed hypothesis as a reason to re-plan rather than thrash. Most agent failures are upstream of the code they write or the action they take. Sol is able to perform on ARC-AGI not because it executes better, but because it correctly orients itself in a new environment first.

 

[Butters]

18 minutes ago, NewKidOnTheBlock said:

He's gonna keep denying until he sees a full fledged pizza delivery AGI robot standing on his front porch, disturbing him from his trip. We all remember that memeworthy moment, if you know, you know

Even though I disagree with your tone, that was hilarious. 

Edited 8 hours ago by Butters

[something_else]

7 hours ago, Leo Gura said:

LLM cannot code in a language not in its training data

They literally can. Especially in agentic form. You keep making the same assertions that are just categorically wrong. Taking the big picture view still requires you to have an accurate understanding of the technology

[Butters]

7 minutes ago, something_else said:

They literally can. Especially in agentic form. You keep making the same assertions that are just categorically wrong. Taking the big picture view still requires you to have an accurate understanding of the technology

Yes this is true, but... it writes the logic in a language it knows (like Python) and translates it line-by-line into your new syntax on the fly.

[Hayato]

Here is Anthony William's take on AI.

 

https://www.instagram.com/reel/DQCwPl9idiD/?hl=bg