GPT-6 Astra. AGI?

[Joshe]

Anyone tried this yet? I'm very skeptical but some pretty popular devs known for not bullshitting about AI are saying it. 

 

Edited yesterday at 06:13 AM by Joshe

[AerisVahnEphelia]

Anyone is sleeping when AGI is starting to really happen.

if you see what it does in 3D, it's basically game over.
Like the fucking fuck.
 

3D designer must be crying, I saw someone close his X account ( not a bate ) because he was getting depressed by ASTRA.

for a good reason.

https://x.com/tomkrcha/status/2095756085890310311

https://x.com/emmanuel_2m/status/2096015049529647574

https://x.com/MatthewBerman/status/2095620888695902366

https://x.com/Dimillian/status/2095596700815516004

Edited yesterday at 07:30 AM by AerisVahnEphelia

[PsychedelicEagle]

Stating LLMs are AGI misses the fundamental gaps LLMs have.

Not being AGI does not mean they aren't extremely useful, however.

[OBEler]

@PsychedelicEagle what fundamental gaps you are talking about. 

[PsychedelicEagle]

@OBEler Yann LeCun's arguments make sense to me.

The one that stands out most: LLMs have no world model grounded in interaction with reality; they learn a probability distribution over sequences of tokens — fragments of written language. Since their only training signal is text, they have no ultimate grounding in physical or sensory experience.

From from section 8.2.2 of this paper:

Quote

For example, Large language models (LLMs) seem to possess a surprisingly large amount of background knowledge extracted from written text. But much of human common-sense knowledge is not represented in any text and results from our interaction with the physical world. Because LLMs have no direct experience with an underlying reality, the type of common-sense knowledge they exhibit is very shallow and can be disconnected from reality.

One can argue against this, but I intuit it's a crucial limitation, and reaching AGI will require different architectures. I do believe AGI is attainable, but it will take longer and a different paradigm than the one the industry is obsessed with right now.

PS: I'm not making the case that we're necessarily in an "AI bubble". I believe the value of AI is not tied to attaining AGI.

Edited yesterday at 01:40 PM by PsychedelicEagle

[PsychedelicEagle]

I'm watching this more recent interview now; he brings up the "world model" point again right in the beginning.

Edit: Interesting quote:

Quote

At (20:37), Yann LeCun explains that companies are driven by the fear of falling behind, which results in a lack of diversity in research directions. He notes: "There's no incentive for the other companies to do anything different either because they can't afford to kind of fall behind the others, right? So, they all work on the same thing... And there's a bit of this sort of, you know, kind of herd behavior, uh, and, you know, in, in mostly in Silicon Valley where everybody is digging the same trench."

 

Edited yesterday at 02:13 PM by PsychedelicEagle

[OBEler]

@PsychedelicEagle Astra has a world model. 

 

I mean seriously. Why cant llm have no world mode? 

Back then no one thought llm can reason. Now LLMs can reason better than humans. 

Edited yesterday at 03:23 PM by OBEler

[PsychedelicEagle]

4 minutes ago, OBEler said:

@PsychedelicEagle Astra has a world model. 

@OBEler, can you share references?

[OBEler]

Chatgpt :Yes — in an important functional sense, GPT-6 Astra appears to form world models, but OpenAI has not said that Astra contains a dedicated “world-model architecture” in the technical sense used for systems like learned environment simulators.
The strongest evidence comes from ARC Prize’s independent testing. They explicitly report that Astra:
turned unfamiliar environments into compact symbolic world models, representing game mechanics as logical rules and developing its own shorthand for tracking state and planning actions. �
ARC Prize
That is significant. In ARC-AGI-3, Astra gets dropped into novel interactive environments without being told the rules. It has to do something like:
observe → act → see consequence → infer hidden rule → update internal model → predict next consequence → plan
For example, internally it might infer something analogous to:
red block + switch → door opens
moving onto X costs one energy
enemy moves clockwise every turn
It can then use those inferred rules instead of blindly trying actions. That's essentially model-based reasoning.

 

 

https://arcprize.org/blog/astra?utm_source=chatgpt.com

 

 

[PolyPeter]

we are not even close to actual intelligence 

[PolyPeter]

scam altman is a paid actor

[OBEler]

@PolyPeter it's not about actual intelligence. It's about if Ai can do work like humans can. 

[PolyPeter]

1 hour ago, OBEler said:

like humans can

may the lord hear you and might this time come soon because i'm tired of telling claude what to do. Better to just pay for his sub and let it cook, fully agentic, no asking me questions, only updating me and making money. Imagine that scenario but for everyone 

[OBEler]

@PolyPeter it's already happening in some high skilled areas like software development. 

[PsychedelicEagle]

15 hours ago, OBEler said:

I mean seriously. Why cant llm have no world mode? 

They can. In fact, they work quite well for many tasks without a world model, especially the ones involving symbol manipulation such as code and parts of math like algebra. But the gap still exists from the intelligence humans exhibit in that humans have an internal physical model of how the world works and they use it to make internal simulations and take actions (planning).

When I lift a glass full of water, I have an internal model that predicts how the water will flow depending on the movements I make. I can even visualize it in my mind's eye (prediction). This prediction is continuous, and it captures nuances missed if we were to use only language to model it.

If LLMs don't have this internal world model, they can only plan in token space — fragments of language — a representation that fails to capture the nuances of real world.

Even if you say LLMs learn an implicit world model from text, it's still a text-learned model. Nonduality points that language carves the world into discrete categories, reality doesn't respect those cuts, and a text-learned world model inherits the cuts.

[PsychedelicEagle]

16 hours ago, OBEler said:

Back then no one thought llm can reason. Now LLMs can reason better than humans. 

If that was the case, we'd see good performance on benchmarks such as https://arxiv.org/pdf/2603.09678v2

[PsychedelicEagle]

10 hours ago, PolyPeter said:

i'm tired of telling claude what to do

Haha +1, and I work with software engineering.

LLMs are great but they can't do robust software architecture yet.

(even in the symbolic/discrete space that coding is)

Edited 11 hours ago by PsychedelicEagle

[PsychedelicEagle]

13 hours ago, OBEler said:

symbolic world models, representing game mechanics as logical rules

@OBEler Thanks for the reference.

Following my post above, I'd argue the parts I highlighted in bold above are limiting factors. World models are still symbolic and represented as logical rules, exhibiting the limitations I pointed out before.

[Heaven]

8 hours ago, PsychedelicEagle said:

@OBEler Thanks for the reference.

Following my post above, I'd argue the parts I highlighted in bold above are limiting factors. World models are still symbolic and represented as logical rules, exhibiting the limitations I pointed out before.

Thanks brother for the insights. It made me finally understand better the difference between LLMs and AGI

[waheed]

On 9/5/2026 at 7:21 PM, OBEler said:

@PsychedelicEagle I mean seriously. Why cant llm have no world mode? 

These are still language models getting larger with new versions but fact remains that model interacts with tokens not "world". 

Edit: looks like it's been settled already. 

Edited 1 hour ago by waheed