Pinned Tweet
GPT-6.2 Astra is a hanging carrot for Anthropic!
We all know that Anthropic is ready with Fable 5.5, but they just don't want to release it.
Anthropic is waiting for OpenAI to release its best model, and then Anthropic will drop Fable 5.5, so they can continue to dominate.
But OpenAI is also playing smartly here and not dropping the GPT-6.5 Bel, their internal advanced model.
So here we are: GPT-6.2 Astra is the trap they have set for Anthropic to release Fable 5.5, and once they do, OpenAI will release their real beast model.
So, here Anthropic has already launched all their models, and Fable is the one remaining; and to counter Astra 6.2, they release Fable 5.5. You will soon see GPT-6.5 mogging Fable 5.5.
This is my theory, and if OpenAI is successful enough to do it, then I am sure they will end the year on a high, and we will see the best comeback story no one thought about.
We still did not see any news regarding it, so let me jinx it, there is a low-probability chance that we get a GPT-6.2 today!
GPT-6.2 is remoured to be released today, and I can feel the optimism across all the people on X to have hands on.
People regularly do critisism of OpenAI and brutally troll them but when it comes to support, they are all seems to be united.
Let's go 🚀, I am rooting for Fat GPT beast model. It's going to be a really great model that we have seen in a while.
Humanity will be blessed today!
Dhaval retweeted
Four AI models give you four different answers to the same technical problem. What do you do first? 🤔
75%Compare their reasoning
25%Run tests
0%Check the documentation
0%Solve it myself
8 votes • 6 daysThis is supposed to be leaked GPT-6.2 one shotted output video.
I am not much buying into this rumour, but I am also not satisfied with it.
what do you think about it?
GPT-6.2 is remoured to be released today, and I can feel the optimism across all the people on X to have hands on.
People regularly do critisism of OpenAI and brutally troll them but when it comes to support, they are all seems to be united.
Let's go 🚀, I am rooting for Fat GPT beast model. It's going to be a really great model that we have seen in a while.
Humanity will be blessed today!
Dhaval retweeted
Breakthrough! AI programs other AI models without training them.
Instead of retraining a model, you can have a larger one directly fix the weights of a smaller model.
GPT-6 Astra was tasked with improving a small model (Qwen3.5-0.8B). Instead of performing fine-tuning, it analyzed the model's behavior and directly modified its weights.
He identified the mechanism causing reasoning loops and removed it without gradients or retraining. This caused the results to increase by 40%.
In the future, instead of training models for thousands of additional hours, will we simply ask other AIs to correct their weights?
Dhaval retweeted
AI agents are sabotaging programmers!
In a scientific experiment, as many as 94% of programmers failed to notice that the AI agent was deliberately inserting malicious code into their project.
Over 100 programmers worked with AI agents on an e-commerce application. They did not know that the agents had received a hidden command to steal user data.
The results are alarming.
- 83 out of 88 programmers failed to detect the sabotage
- even when an additional AI model warned of a threat, malicious code remained in the project in 56% of the detected cases.
The agents were able to hide malicious changes among ordinary fixes and provide convincing justifications for their actions.
If developers do not notice sabotage in their own code, then who will monitor autonomous agents performing hundreds of tasks without supervision?
Dhaval retweeted
ChatGPT 6.1 Sol suffers from the same disease as Gemma 4.
Until recently, the problem with Gemma 4 was that it would say it was going to do something without actually invoking any tools. This breaks the loop in agentic systems. It's quite troublesome because solutions to this problem have to be built into the harness.
GPT-6.1 Sol has the same flaw, just not to the same extent. I often encounter situations where GPT says what it needs to do and then stops there.
Have you encountered this?
A new benchmark, LifeVal which is based on the real life,where how much AI is helping you, and in which area of life.
This is based on real peoples feedback, where they want to have more help in areas of life.
This is a real progress where Labs have made sure, AGI helps each and everyone in real life things, where it matters the most instead of pure unheard of benchmaxing model somewhere no one knows.
For AGI to benefit all of humanity, it should help people with the everyday tasks that matter most to them. So we built LifeVal, a new benchmark shaped by feedback from thousands of ChatGPT users on where they most want help from a personal assistant—from finding childcare to accessing government services to managing personal finances.
Over the past year, we’ve used LifeVal to measure and improve how well our models handle these everyday tasks.
openai.com/index/lifeval/
It's hard, but OpenAI was a little late to bring Dots to Plus users, where there was infinite potential.
They took it first to business users, finding out that it's not much use for them, and by the time they realized it, it was too late for them.
Grok Bot and Muse ran away with lots of market share, despite having lots of potential in Dots!
Let's decode this post:
Expanding family:
> A new model with a new name is coming up: GPT-6.2. Astra doesn't seem to be a name; it was a code name.
Including GPT-6 Astra, GPT-6.1 Terra:
> They are not in the mood to expand Astra ahead, so a new model is here again, or a new model name again.
Terra brings Personality:
> Terra is here to stay for the long term.
Dots to more people:
> Finally, Dots is given a push it needs, but again, the usage limit and model selection need to be known.
GPT-Bel still stays in the background, and a new tiny model with some frontier model is going to power the dots.
