GPT-6 Astra. AGI?

[AerisVahnEphelia]

On 05/09/2026 at 3:37 PM, PsychedelicEagle said:

@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:

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.

Yann is french, french people are smarter and are always rights, so this makes sense.

[Carl-Richard]

LLMs at this point are supposedly like hyper-nerds: good for book knowledge, but that's about it. I remember when people were gushing over 4o two years ago like it was some sort of genius when it was actually a high-schooler with stifled speech, and I thought it was retarded back then. I haven't tried Astra much yet, but supposedly it should be really really smart.

However, I've tried 5.6 Sol a bunch for producing text, and it still has a problem of not understanding context well (it says a lot of things you would usually find irrelevant even if it is technically valid). Perhaps you could fix some of it with better prompting, but context is such a subtle thing that I believe humans will still have an edge on that for a very long time.

[Schizophonia]

26 minutes ago, AerisVahnEphelia said:

Yann is french, french people are smarter and are always rights, so this makes sense.

I totally agree.

 

[Schizophonia]

On 05/09/2026 at 9:30 AM, AerisVahnEphelia said:

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

I'll be able to make my GTA 7 in China.

[Carl-Richard]

11 minutes ago, Schizophonia said:

I totally agree.

 Oui oui baguette.

[OBEler]

@Carl-Richard We talk here about the newest model. a big jump in capability. And you talk about the old model. Who cares what the old model cant do? We have a new model.

I really ask me how people understand technology. Cant you grasp the concept, that something can improve to a better version? The old version is immediately obsolete. So why talking about?

[Carl-Richard]

@OBEler Man you have the conversational understanding of 4o. I talked about the progression of LLMs, how far we've come. I also talked about 5.6 Sol. I haven't tried Astra much so I can't tell you what I think, but I think the context argument holds for Astra as well (it's a much deeper problem than upping the performance a few percent on some reasoning benchmarks). 

And if you want to talk about understanding technology, if you think current LLMs are AGI, you have a low bar for what AGI is. Literal "AGI" means we are now literally all out of work. Yet we're not.

Edited 2 hours ago by Carl-Richard

[LordFall]

Super excited for it, looks like a game changer. It being AGI or not depends on your definition.

Some humans are pretty dumb and don’t produce much value.The fact they have more of a free will than AI doesn’t count for much. Especially now with AI agents that can fulfill long term tasks. 

[OBEler]

@Carl-Richard You talk not about progression. You share just your experience with the older model. 

The context argument doesn't hold for Astra. But research your self. Again, it's a NEW model. 

You can't guess here what it can do or not do based on previous models. It's a new model with new capabilities and architecture. 

[Leo Gura]

The fundamental limits of the LLM structure have not magically disappeared.

None of these models can know what is outside their training data.

Edited 1 hour ago by Leo Gura

[AerisVahnEphelia]

8 minutes ago, Leo Gura said:

The fundamental limits of the LLM structure have not magically disappeared.

None of these models can know what is outside their training data.

But why do we need something outside in the first place.

We just need all the data being able to be summoned in  the order we need.

[OBEler]

9 minutes ago, Leo Gura said:

The fundamental limits of the LLM structure have not magically disappeared.

None of these models can know what is outside their training data.

I remember you posted a blog post some years ago, with an interview if someone who said with llm its impossible to reason. Look now, it is capable of reasoning. 

Now you say none of these can know what's outside their training data. This was disproven with Astra. Look at the neweset benchmark arc 3. They must solve games which do not exist. You need to build new strategies and knowledge to solve these. 

 

 

[AerisVahnEphelia]

2 minutes ago, OBEler said:

I remember you posted a blog post some years ago, with an interview if someone who said with llm its impossible to reason. Look now, it is capable of reasoning. 

Now you say none of these can know what's outside their training data. This was disproven with Astra. Look at the neweset benchmark arc 3. They must solve games which do not exist. You need to build new strategies and knowledge to solve these. 

 

 

it depends what is reasoning.
They called (ai creators) a feature reasoning, it's not the same as human reasoning.

 

Still I don't see how llms are supposed to be the original ones if you don't steer them that way.

[OBEler]

@AerisVahnEphelia yeah technically. 

If you show now these experts back then how good it can reason they would be very surprised and wouldn't say like ok.. But it's not really human reasoning. 

Yeah.. Solving math problems no human did before. If you still say nah.. No reasoning behind, then I give up. 

 

Edited 52 minutes ago by OBEler

[Carl-Richard]

1 hour ago, OBEler said:

The context argument doesn't hold for Astra. But research your self. Again, it's a NEW model. 

The context problem is mainly because we give a tiny little prompt and the AI is supposed to deduce what we mean based on that (which is hard because it has to know a lot about us and our goals that we haven't communicated to it). A smarter model or more tokens only fixes this marginally.

Edited just now by Carl-Richard

[AerisVahnEphelia]

15 minutes ago, OBEler said:

@AerisVahnEphelia yeah technically. 

If you show now these experts back then how good it can reason they would be very surprised and wouldn't say like ok.. But it's not really human reasoning. 

Yeah.. Solving math problems no human did before. If you still say nah.. No reasoning behind, then I give up. 

 

I m extremly pro llms/ai.

It does something meaningful & of value.
 

 

[Pudgey]

We are not at AGI yet, but GPT Astra is a beautiful model. I've been using it every day since its release (yesterday and today). The reason people are referring to it as almost AGI: 1. Its 3D modeling is state of the art. It can scaffold a full Blender or Unity scene in minutes and reconstruct your entire room in Blender with pretty high accuracy. 2. Computer use is world-class. It can take over your computer and operate it at lightning speed, faster than most humans. 3. When coding, it can debug and harden production code with computer use, operating your entire app, finding errors and issues as it goes along.

Overall, it's a beautiful model to work with as a daily driver, but it isn't AGI by any means. It did feel like we got a taste of what LLMs are capable of, and of what AGI could be capable of in the future.

@OBEler If we reached AGI, we wouldn't be getting access anytime soon. Astra still needed a stateful harness to score on ARC-AGI-3. Real AGI wouldn't need the scaffolding, and it wouldn't be $200 a month. Consumers still don't get Mythos, the ungated Fable, because the labs don't trust the model with its own safeguards off. I agree with some of your points, but these AI labs still have a lot to prove, and solving gaming and coding isn't AGI.
 

[AerisVahnEphelia]

1 minute ago, Pudgey said:

We are not at AGI yet, but GPT Astra is a beautiful model. I've been using it every day since its release (yesterday and today). The reason people are referring to it as almost AGI: 1. Its 3D modeling is state of the art.

Meanwhile the 3D model with astra :

 :
 

[Leo Gura]

46 minutes ago, AerisVahnEphelia said:

But why do we need something outside in the first place.

We just need all the data being able to be summoned in  the order we need.

Because that isn't Intelligence.

Humans are confusing good performance and utility with Intelligence.

41 minutes ago, OBEler said:

I remember you posted a blog post some years ago, with an interview if someone who said with llm its impossible to reason. Look now, it is capable of reasoning.

It is capable of what it was trained to do.

Quote

Now you say none of these can know what's outside their training data.

This has always obviously been the case.

Edited 21 minutes ago by Leo Gura

[Pudgey]

@AerisVahnEphelia SOTA (State of the art) = in comparison to other AI models, not human, since we are not yet at AGI. Lastly, I don't know if you have actually used AI for gaming, but it can definitely provide better modeling than that. AI isn't perfect and is not replacing serious 3D artists anytime soon. 

Here's Astra blender 3D render of a room: https://x.com/bridgemindai/status/2095993644389982294?s=20