Hakm

I like AI & build cool stuff | AI Engineering student.
Half insider... actually a quarter... okay, I’m not.
Follow me , I’m worth the click.

[Hakm @hakmgpt]

the new Gemini 3.8 flash is being tested under code name (Gemini 3.1 pro-preview) on Arena
Here’s a result for it
check the comments to see another results (Fable , Grok4.6)
This is only pure python code - same prompt for all

[Hakm @hakmgpt]

@thsottiaux @rezoundous with astra or just a normal reset

[Hakm @hakmgpt]

We’ve always asked: what comes after Agents?
We’ve seen some leaks pointing to Persistent Reasoning from OpenAI.

We could get models that can work and think for days without stopping on a single task, from a single prompt.
The bottlenecks here are obviously cost and memory. And honestly, I don’t think simply compressing the memory is a good solution if the model is going to run for days.

But here’s the exciting part
I think Persistent Reasoning could lead to completely new things in the future if we solve the memory and hallucination problems.

Models could potentially invent and create entirely new solutions with this kind of system. They could solve problems that have never been solved before in mathematics, physics, Medicine, and beyond.
And I also suspect that the Jalapeño chips being designed by OpenAI could be built around, or at least heavily influenced by, this capability.
I expect Persistent Thinking to arrive with Astra, or perhaps during the September 29 event with an AI Researcher.

[AiBattle @AiBattle_]

A potential new reasoning effort, "Persistent", has been spotted in the Codex GitHub repo

"Continue working until put to sleep"

[Hakm @hakmgpt]

@iruletheworldmo is that real or

-> 1M context
-> better reasoning
-> bigger model

[Hakm @hakmgpt]

@vstalingrady Mythos 7 leak : beats Mythos 6

[Hakm @hakmgpt]

They keep posting “leaks” like:

-> 1M context
-> better reasoning
-> bigger model
-> better coding
-> multimodal
-> more compute
->And the most important thing , beats mythos

Bro, that’s not a leak 😭

[Salio @Mr_Salio]

🚨 DeepSeek V5 Leak: Beats Claude Mythos 5

> Rumors are pointing to Mythos-level capabilities
> DeepSeek V5 is reportedly coming in Early September
> Expected to be a major leap beyond the V4 generation
> Could deliver frontier-level performance at a much lower cost
> DeepSeek is reportedly still planning to keep the model open-weight

Could DeepSeek V5 become the new king of open-weight AI?

[Hakm @hakmgpt]

@wangkun1211 no one knows , some says Thursday, some says 29 September , there’s no official confirmation yet.

[Hakm @hakmgpt]

Here is a video of the voxel castle output for GPT-6 Astra

[Hakm @hakmgpt]

The same leakers behind upcoming Claude model leaks have now shared new GPT-6 Astra leaks

And finally, it looks like the frontend has been fixed.

[Hakm @hakmgpt]

@tolgaozisik Blob is great

[Rosenda Garica @GaricaRosen6779]

Seems the old examples were too easy for ASTRA. AND some ppl even thought Opus5 could do them.
Check out this.
Prompt:Build a gorgeous sci‑fi spaceship with explorable internal structures.Just One-shot,ASTRA will be a glimpse of AGI.

[Hakm @hakmgpt]

@knowixbuilds We have already see that Astra solved 10 equations that had never been solved before 👀 this things worth it.

[Hakm @hakmgpt]

@StatsWire fr , but the thing that makes me talking about it , the most people believe this fake news.

[Hakm @hakmgpt]

@kimmonismus Alienware , released in August and looks great

[Hakm @hakmgpt]

@JohnPay53891334 38 minutes is long time 🫠

[Hakm @hakmgpt]

Another output from Astra: “Pagoda Voxel” 👀
That’s really another level of what we have today.

[lyra @lyraxana]

[GPT Astra] Pagoda Voxel
~56k tokens; ~38 minutes

[Hakm @hakmgpt]

@Haleeeemahh frrr 😋 can’t wait for it

[Hakm @hakmgpt]

@r69shabh I have test it with only 1 shot tasks , it’s better than Terra (for one shot) , but opus 5 ? Not yet

[Hakm @hakmgpt]

@r69shabh Yup , it’s really good compared to 3.7

[Hakm @hakmgpt]

@ryemike_merio Pure humanity ❤️