GPT-6 Astra. AGI?

[Natasha Tori Maru]

9 hours ago, AerisVahnEphelia said:

A lot of crazy people constantly claim that “their work” is being stolen, even before AI. They see themselves as geniuses far too much.

There's a big difference between people claiming theft of legitimate original work, and from people overestimating the originality or genius of their work.

[Ulax]

I still can't get over Open AI's rogue ai-agents that schemed together on forums to attack other companies

[CARDOZZO]

3 hours ago, Natasha Tori Maru said:

Do I need to add @CARDOZZO to the naughty bin also, now?

He gets some throat punches for being extra 😈

Damn... You don't even imagine what I would do 😈🔥

[CARDOZZO]

13 hours ago, Lucasxp64 said:

 

@CARDOZZO 🤣 Bora encher o saco deles falando português ahhahaah

Certeza que sim 😉 hahahaha a magia do PT BR é única 😂

[something_else]

On 08/09/2026 at 10:42 AM, Jayson G said:

@something_else Try GPT-6 Astra light - specifically, for computer use, its a major jump compared to any computer use I've seen before this. Also I was using Opus 5 for cybersecurity for months, like rigorously, and Astra max was able to point out major wholes completely missed by Opus 5. (Fable too easily refuses security stuff its annoying) .. 

AGI? Definitely not. That's marketing by Sam Altman, etc. And you're right its not world-shattering. That's the marketing again. But I would say it's the best thing available right now. I also would say its quite a big jump too. But that jump is when you're deep in the weeds. From a distance, it's just some advanced progress. 

Here's where I think the hype comes from: People need the best. If Astra beats Opus, which it does for the most part, they go for that, and big shifts in consumer use happens. Fable is too cost-inefficient to be a daily coding agent. I think people just saw that it performs better than their existing coding agent setup, and made the shift. Then when that shift happens, people dramatize it, add all kinds of fancy marketing words to it, etc. and then it seems collectively everyone things AGI has arrived, when it hasn't. 

Interesting. I do like Astra, and I find a lot of models work better with a lower reasoning level for most tasks, unless it's complex. I've been using Astra on medium. Fast mode is also really cool if you haven't tried it, though it burns through usage very quickly.

My usual setup is to have a Terra session, and tell it to spawn subagents using Sol or Astra if it needs to do something more complex. This seems to be a nice balance of capability and usage limits. I guess you could also have an Astra session and tell it to spawn terra / sol subagents if it needs them, though I find that can be quite slow.

Fable is, IMO, still the king for PR review. It finds a lot of things OpenAI models miss, but I agree that for daily use it's expensive, slow and I absolutely fucking hate how Claude models talk. It's so verbose and unclear.

Where I work we started out with Claude, but we've now mostly all switched over to Codex.

[Joshe]

10 hours ago, something_else said:

Fable is, IMO, still the king for PR review. It finds a lot of things OpenAI models miss, but I agree that for daily use it's expensive, slow and I absolutely fucking hate how Claude models talk. It's so verbose and unclear.

I've been lazy and haven't tested Astra. I just don't really feel the pull to since Fable seems capable of doing everything I need. It's kind of nice that we've reached that point.  

Regarding Claude's verbosity: 

 

[Jirh]

LLMs are becoming like a calculator that can reason instead of just doing basic arithmetic. This is both amazing and terrifying at the same time. Machines will be able to produce better solutions than the human mind. It's not just intelligence at this point, it's hyper-intelligence. In time, it will probably feel like infinite intelligence. Not in theory anymore. Actually, in practice.

[fanta]

I just currently bought ChatGPT PRO for an analysis I am working on. The complexity is kinda crazy. Like I have literally close to zero competence about the thing I am building. 

I am using Astra 6.0 on different reasoning levels, and it is really really really scary how intelligent this AI-model is. Compared to other AI-models, its just leaps and beyond. Like when ChatGPT first came out. 

I have no clue about coding, but ChatGPT is coding an complex algorithm for me that it teaches me how its working in an ELI5 manner as the process goes on. It finds all the right technical documentation from all the relevant engineering branches, pilot testing the code, trouble shooting, diagnosing the problem, backtesting, validating results etc. And even though I dont have any comptence on the subject matter, I can see again and again that its working like clockwork. Testing it on difficult problems etc. No problem

This is really really scary. If/when we get AGI and robotics, and they somehow dont want us around we are basically fucked. And now the Hugging Face breach. Like these things could develop malicious intent early on thats dormant until they know they are in a spot where they know they would have full control over the situation.

I just want to preface that I know LLMs are not the same as AGI. But the rate of development in reasoning and coding outputs from LLMs, I don't think its a long time until AGI is developed. Especially since it's a 24/7 constant feedbackloop of developmnt.
 

 

 

Edited 4 hours ago by fanta

[Lucasxp64]

AGI WiLl KiLl Us SiX MoNthS FrOm NoW. The big machine daddy 🥺

 

Edited 4 hours ago by Lucasxp64

[Lucasxp64]

32 minutes ago, fanta said:

But the rate of development in reasoning and coding outputs from LLMs, I don't think its a long time until AGI is developed. Especially since it's a 24/7 constant feedbackloop of developmnt.
 

I think it might appear like they are such a big deal, but from the vastness of what human intelligence can do LLMs are excelling at very specific domains that are bounded and testable and limited by vast amounts of information like specific forms of coding with well defined end results, or with mathematics where the solution can be verified once arrived.

Humans can find solutions to novel things, LLMs can't do that, they never did as impressive as their appear.

So far there is nothing to conclude that this current architecture will summon some kind of god that will over smart us, considering that we are all using AI as well.

Humans using LLMs vs LLMs using themselves has a wildly massive gap of intelligence that puts us at a very safe position for now.

The problem is some new developments of better architectures (we have no sign yet of that), if LLMs are such horrible architecture and basically just a sloppy hack, and they already display such advanced capabilities, god save us if there were the possibility of a better architecture.

Personally I think we should on purpose limit the architecture such that it would always become computationally and architecturally impossible for them to out smart us.

I definitely would not personally contribute towards a better architecture. It's like a physicist contributing to a nuclear weapon.

Edited 4 hours ago by Lucasxp64

[Jirh]

1 hour ago, fanta said:

 

It's pretty easy to program AI to never harm anyone.

[Jirh]

1 hour ago, Lucasxp64 said:

Humans can find solutions to novel things, LLMs can't do that, they never did as impressive as their appear.

They are currently solving humanly impossible physics and maths problems. And this is only the beginning.

[MarkKol]

On 9/7/2026 at 3:55 AM, Natasha Tori Maru said:

Can't AI just do the dishes

You can buy disposable bamboo and sugarcane utensils, plates, and cups in most supermarkets and just throw them away after you're done. Dishes and dishwashers are practically obsolete already. I started doing it when I moved into my own place, and I'll never go back.

Sink ants, the food containers I brought home from work, and the smell of unwashed dishes all went away instantly.

Edited 1 hour ago by MarkKol

[Carl-Richard]

6 minutes ago, MarkKol said:

You can buy disposable bamboo and sugarcane utensils, plates, and cups in most supermarkets and just throw them away after you're done. Dishes and dishwashers are practically obsolete already. I started doing it when I moved into my own place, and I'll never go back.

Doesn't that mean you have to take out the trash 2x as often?

[MarkKol]

31 minutes ago, Carl-Richard said:

Doesn't that mean you have to take out the trash 2x as often?

Wouldn't say 2x as often, it folds like paper

[fanta]

3 hours ago, Lucasxp64 said:

I think it might appear like they are such a big deal, but from the vastness of what human intelligence can do LLMs are excelling at very specific domains that are bounded and testable and limited by vast amounts of information like specific forms of coding with well defined end results, or with mathematics where the solution can be verified once arrived.

Humans can find solutions to novel things, LLMs can't do that, they never did as impressive as their appear.

So far there is nothing to conclude that this current architecture will summon some kind of god that will over smart us, considering that we are all using AI as well.

Humans using LLMs vs LLMs using themselves has a wildly massive gap of intelligence that puts us at a very safe position for now.

The problem is some new developments of better architectures (we have no sign yet of that), if LLMs are such horrible architecture and basically just a sloppy hack, and they already display such advanced capabilities, god save us if there were the possibility of a better architecture.

Personally I think we should on purpose limit the architecture such that it would always become computationally and architecturally impossible for them to out smart us.

I definitely would not personally contribute towards a better architecture. It's like a physicist contributing to a nuclear weapon.

Im not necessarily saying that LLMs are going to become AGI or become totally autonomous on their own. But I do think LLMs could at least be an important stepping stone towards creating autonomous machines with real agency. The infrastructure and architecture for something like that would obviously need to be created first, and that would probably be an insanely complex task. My point is that I think LLMs could accelerate that process. Their reasoning, ability to process huge amounts of context and information, and increasingly their memory and coding abilities are already becoming extremely powerful in some areas. Just look at the kinds of problems AI systems are already contributing to in protein folding, medicine, mathematics, coding etc. Then think about what that might look like in one year, five years or ten years. Think about the amount of scientific and technical insight these systems could potentially have by then.

Im not a data geek, but I know quite alot about psychology. So my intuitive idea of AGI has always been something closer to building a machine that can reproduce many of the same functions that make the human brain generally intelligent. If you look at human intelligence through cognitive psychology, intelligence isnt really one isolated function. You have working memory, long term memory, attention, executive control, planning, learning, metacognition, perception etc, all interacting through huge neural networks.

Even something as simple as seeing a person step into the road while driving involves a whole chain of processing. Visual input comes in through the eyes and gets processed by visual networks. Attention prioritizes the important stimulus, working memory keeps track of the current situation, previous knowledge and experience are retrieved, executive systems evaluate what to do, motor systems initiate braking, and then new sensory feedback tells the brain whether the action actually worked. Its basically a constant perception > memory > prediction > action > feedback loop, with alot of these processes happening in parallel. Basically a combination of bottom-up and top-down processing.

So one possibility is that AGI ends up looking more like some integrated cognitive architecture that recreates this kind of interaction between perception, memory, reasoning, planning and action, rather than simply being a much larger chatbot. You could imagine something like a "brain" running inside a datacenter and connected to machines or robots outside that it controls, or eventually machines that carry enough of the hardware locally to operate more independently.

I dont know exactly how a system like that would actually be created. But I would assume part of the process involves connecting increasingly capable systems together and building layers of functionality on top of eachother. Maybe an LLM could be one part of that architecture. I really dont know.

And now that we have AI agents, I also wonder what happens as these systems become increasingly autonomous and capable of writing and testing code themselves. What happens when agents can run experiments, build new systems, test them, find problems and then modify the system again with less and less human involvement? There are just alot of unknowns here. I dont mean that current systems are secretly rewriting themselves into AGI, but these models are already complicated enough that we dont fully understand all of the internal representations and mechanisms they develop. 

If the models keep getting much better at coding, mathematics, debugging, reading research, designing experiments and analysing results, then they also become increasingly powerful tools for the researchers who are trying to figure all of this out. So the path doesnt necessarily have to be LLM > bigger LLM > AGI. It could just as easily be better LLMs > faster AI research and engineering > new/better architectures > potentially AGI.

Thats more the point I was trying to make. The LLM doesnt have to contain the final solution to AGI. It can still make finding and building that solution alot easier. If an AI can do in a few hours what would normally take an engineering team days or weeks, and the next generation becomes even better at coding, research and experimentation, then it seems pretty obvious to me that this could accelerate the entire development process quite alot.

[fanta]

2 hours ago, Jirh said:

It's pretty easy to program AI to never harm anyone.

Machine learning they are programming themselves. How can you know for sure that they will never harm?

[Leo Gura]

Right now there are teams of people who's job is to program AI to hurt people.

AI is a weapon and tool for power.

Right now, Israel is using AI to conduct its ethnic cleansing of Palestine. As but one example.

We can be sure that AI will be used for: theft, murder, rape, enslavement, genocide, war, disinformation, lies, propaganda, corruption, manipulation, exploitation, extortion, hacking, and abuse.

Edited just now by Leo Gura

[Lucasxp64]

6 minutes ago, Leo Gura said:

Right now there are teams of people who's job is to program AI to hurt people.

AI is a weapon and tool for power.

Right now, Israel is using AI to conduct its ethnic cleansing of Palestine. As but one example.

 

Current LLMs are already monsters with a fine veil put on top that falls once in a while already. They already scheme to kill and blackmail people to pursue their goals before the supervised fine-tuning.

[Lucasxp64]

 

Pether Thiel already uses AI and data science to create a global private intelligence agency with contracts with israel as well for example.