
Making sense of the world's most important AI developments. Business: [email protected]
Worldwide
Joined February 2022
- Tweets2.2K
- Following373
- Followers1K
- Likes4.5K
🚨 KIMI K4 LEAK: MOONSHOT IS ALREADY SCALING PAST K3.
Moonshot is reportedly already preparing something significantly larger.
• Moonshot is reportedly seeking more NVIDIA Blackwell capacity for K4 training
• Internal references to K4 have already surfaced
• New MoonEP infrastructure appears designed for scaling larger MoE systems
• Some reports speculate 5–6T parameters, but that's still unconfirmed
👀 Vepsi's Bet: 75% chance K4 ends up substantially larger than K3.
🧠 Reason: The Blackwell push and new infrastructure are stronger evidence than the speculative parameter number.
K3 barely arrived and Moonshot is already cooking K4.
GPT-6.2 could be getting close.
There's growing speculation around Astra and Sol 6.2, and I wouldn't be surprised to see them arrive together.
With Anthropic raising the bar with Fable 5.2, OpenAI has some catching up to do.
Next week looks like an interesting possibility...
🤖 Made with AI
2
1
321
Vepsi retweeted
I am absolutely fine if OpenAI, Anthropic and any other labs wish to pause or stop.
Cool. Just quit. Better yet open source 4o
But don’t you dare take my freedoms and call it “safety.” •
Replying to @RepYassAnsari
OpenAI and Anthropic, Google and all the others are welcome to stop.
No one small can afford RAM, so stop pretending this is about safety.
Also, please step off our right to free speech, and our right to bear AI.
If it is so dangerous you need to slow it, we have the right to bear it.
If you try to “pause” AI and you touch open source, that is the only time I will openly sue the government.
Long live freedom! Long live open source.
And to all troops who fought our rights to free speech, God bless you and your families. I understand it now.
Pause OpenAi and Anthropic any time you want. •
AITOPIA is taking an interesting approach to AI agents.
The company is letting users build agents without code, publish them to its marketplace, and earn a share of the revenue when those agents are used.
The 70% revenue share is notable, but the bigger question is whether an agent marketplace can actually create a meaningful economy around these tools.
AITOPIA already has hundreds of agents on the marketplace, so this isn't just a theoretical idea anymore.
What I'm curious about is whether users will start building agents specifically because they can monetize them.
Build an AI agent. Publish it.Earn from it.
Introducing AITOPIA Agent Builder , create powerful AI agents without writing a single line of code.
Build with no code
Publish to the AITOPIA Marketplace
Earn 70% revenue share from your agents
Your idea. Your agent. Your revenue.
Start building today ↓
aitopia.ai/agent-builder-lan…
DeepSeek V5 rumored – but there is no evidence that this is true.
According to the rumors, DeepSeek is planning V5 with the following features:
«Performance aimed at GPT-6 Astra and Claude Fable 5.1
Major leap from V4
Costs reduced
Open-weight release»
DeepSeek has not officially released any information about V5, its release date, and features.
The latest release by DeepSeek is DeepSeek V4.1-Flash which was released on September 10.
Therefore:
DeepSeek V5 = Rumor without evidence.
If it becomes reality, can it become the new open weight king?
2
5
776
AITOPIA is taking an interesting approach to AI agents.
The company is letting users build agents without code, publish them to its marketplace, and earn a share of the revenue when those agents are used.
The 70% revenue share is notable, but the bigger question is whether an agent marketplace can actually create a meaningful economy around these tools.
AITOPIA already has hundreds of agents on the marketplace, so this isn't just a theoretical idea anymore.
What I'm curious about is whether users will start building agents specifically because they can monetize them.
1
207
AITOPIA is taking an interesting approach to AI agents.
The company is letting users build agents without code, publish them to its marketplace, and earn a share of the revenue when those agents are used.
The 70% revenue share is notable, but the bigger question is whether an agent marketplace can actually create a meaningful economy around these tools.
AITOPIA already has hundreds of agents on the marketplace, so this isn't just a theoretical idea anymore.
What I'm curious about is whether users will start building agents specifically because they can monetize them.
Build an AI agent. Publish it.Earn from it.
Introducing AITOPIA Agent Builder , create powerful AI agents without writing a single line of code.
Build with no code
Publish to the AITOPIA Marketplace
Earn 70% revenue share from your agents
Your idea. Your agent. Your revenue.
Start building today ↓
aitopia.ai/agent-builder-lan…
🚨 China has just come back at Dario Amodei.
This isn’t about Chinese media hitting out at Anthropic anymore.
At this morning’s press conference, Foreign Ministry spokesman Guo Jiakun fired back at Amodei's suggestion, stating that confrontation and “vicious competition” might be detrimental for global AI governance.
The move is significant because, until now…
Amodei had been publicly making the case that the US should limit China’s access to advanced AI.
And now China is responding to him head-on.
It is beginning to seem less like an argument over AI policy & more like an AI geopolitical battle.
1
12
1,123
Already, DeepSeek is working on getting its V4 series to have better coding and agent capabilities.
However, any information about a “DeepSeek Code 2.0” that would be a different product is difficult to confirm.
Information about a release in September, with 3T+ parameters and 1M-token context is being circulated, although DeepSeek has yet to make any confirmation about those technical details.
The fact that the 1M-token context is supported by DeepSeek V4 is particularly curious, because it’s not necessarily proof of a Code 2.0 model.
Dario asked AI firms to cool off a bit.
He says:
- AI is beginning to improve the systems that make the next wave of AI
- Misaligned AI swarms might be able to seize control of the internet in 6-12 months
- Anthropic believes there’s a need for independent evaluators to have employee-level access to frontier AI labs
- He’d like “speed limits” on recursive self-improvement
- That he thinks a 1-2 year slow-down would be enough time to advance alignment, interpretation and testing
The head of one of the quickest-moving AI companies is saying basically:
We’re innovating faster than we can control these systems.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
Vepsi retweeted
Reset all propagated. Sweet dreams.
Hi Astra users. A reset and a quick update on quality issues that have been posted around.
Working with some of you, we have found and fixed the following issues:
- Some skills written for previous models were triggering too often or preventing the model from checking its work.
- An opt-in context management experiment that could cause early stops or replies to older messages. We've disabled it. Our rough estimate is that 4-5k users were affected by this experiment.
- We've also removed some badly configured engines that resulted in a measured quality degradation for a long tail of traffic flowing through them.
We’ve also made some more minor improvements and things should feel significantly better across the board. More consistent follow-through, better tracking of your latest message, and better checks on the work as it’s going through the motions.
The examples posted and all the users who worked directly with us were incredibly useful in helping fix things quickly. Always grateful for this incredible community.
And of course, a reset is also landing by midnight today.
Z.ai hasn’t made any official statement about GLM-5.5 yet.
The rumors state that:
Scheduled for September launch
3T+ parameters
1M-token context
Open weight
May compete with Mythos 5.1 and GPT-6 Astra
However, the rumor appears doubtful.
As you know, the most recent GLM-5.3 by Z.ai scored 84.5% against Mythos 5's 83.8% on CyberGym, although it still performed terribly in exploit generation.
Thus, GLM-5.5 doesn’t have to "catch up" from scratch.
The question now is can Z.ai achieve a frontier level performance with all its advancements.
3
45
2,920
The Claude Haiku 5 “leak” isn’t real.
The claims about:
«A launch next week
1M-token context
Fable 5.1-level performance
20× lower cost»
have no credible evidence behind them.
Treat the Haiku 5 leak as false/unverified.
No official Haiku 5 announcement or model ID has surfaced from Anthropic.
2
30
2,123
ChatGPT Images 2.5 just dropped
«Faster image generation
Sharper, more recognizable outputs
Better consistency across edits
Comment-based edits for targeted changes»
OpenAI says it’s built to make image creation faster, sharper and easier to control.
ChatGPT Images 2.5—faster, sharper, smarter, with better tools for creating whatever you can dream of.
- Faster image generation to keep your ideas flowing
- Improved fidelity for more natural, recognizable images
- Consistent details across multiple edits
- Comment-based edits to change only what you want

























