AI researcher
My telegram channel ⇢
Joined July 2022
- Tweets17.1K
- Following844
- Followers3.5K
- Likes24.7K
Specter👁 retweeted
While everyone is waiting for the OpenAI and Anthropic IPO, the Chinese continue to roll out models that make life difficult for competitors
Moonshot AI introduced Kimi K3.1 Pro: 3 trillion parameters, a million-token context and improvements in coding, toolwork and 3D modelling.
But what's more clinging is something else: the company claims to have a 65% reduction in hallucinations in financial tasks.
The AI race has long been not only about OpenAI and Claude. China also claims a place among the leaders.
And we are still waiting for OpenAI and Anthropic to finally go public.
Well, where is the IPO? 😭
While everyone is waiting for the OpenAI and Anthropic IPO, the Chinese continue to roll out models that make life difficult for competitors
Moonshot AI introduced Kimi K3.1 Pro: 3 trillion parameters, a million-token context and improvements in coding, toolwork and 3D modelling.
But what's more clinging is something else: the company claims to have a 65% reduction in hallucinations in financial tasks.
The AI race has long been not only about OpenAI and Claude. China also claims a place among the leaders.
And we are still waiting for OpenAI and Anthropic to finally go public.
Well, where is the IPO? 😭
Specter👁 retweeted
There is such a pace in AI now that you simply can't keep up with everything.
In 2023, an average of 73 days passed between the releases of the flagship models OpenAI and Anthropic. In 2026, this period was reduced to 18 days.
And that's just two companies. And there are also Google, Grok, Chinese models and a bunch of other tools.
While some are understanding the possibilities of AI, others are already releasing something new.
It's easy to earn FOMO in such a race.
There is such a pace in AI now that you simply can't keep up with everything.
In 2023, an average of 73 days passed between the releases of the flagship models OpenAI and Anthropic. In 2026, this period was reduced to 18 days.
And that's just two companies. And there are also Google, Grok, Chinese models and a bunch of other tools.
While some are understanding the possibilities of AI, others are already releasing something new.
It's easy to earn FOMO in such a race.
Specter👁 retweeted
Chinese AI models already process several times more tokens than American ones.
Over the past week, the figures amounted to 79.2 trillion against 17.4 trillion. At the same time, growth in the United States has practically stopped in the last 8 weeks.
And then OpenAI can come on the scene with its smartphone. According to rumors, the company is developing a device with its own OS, where AI agents will work instead of the usual applications.
The idea is that you just set the task, and the AI does it for you. But such a system will require a huge amount of tokens.
It turns out that a smartphone can become not just a new product for OpenAI, but a way to dramatically increase the consumption of its models.
I wonder how quickly we will come to this.
Chinese AI models already process several times more tokens than American ones.
Over the past week, the figures amounted to 79.2 trillion against 17.4 trillion. At the same time, growth in the United States has practically stopped in the last 8 weeks.
And then OpenAI can come on the scene with its smartphone. According to rumors, the company is developing a device with its own OS, where AI agents will work instead of the usual applications.
The idea is that you just set the task, and the AI does it for you. But such a system will require a huge amount of tokens.
It turns out that a smartphone can become not just a new product for OpenAI, but a way to dramatically increase the consumption of its models.
I wonder how quickly we will come to this.
Specter👁 retweeted
AI has learned to reverse-engineer the logic of other applications
REA (Reverse Engineer Anything) is going viral on GitHub; it’s a tool that helps AI agents dissect programs from the inside and understand how they work.
It handles applications, websites, APKs, firmware, and even smart contracts. The agent analyzes their structure, reconstructs their logic, and attempts to model their operation.
It works with Claude Code and Codex.
The project racked up over 25,000 stars on GitHub in its first 24 hours.
AI is no longer just writing code on demand; it is now learning to decipher existing software and replicate its functionality.
AI has learned to reverse-engineer the logic of other applications
REA (Reverse Engineer Anything) is going viral on GitHub; it’s a tool that helps AI agents dissect programs from the inside and understand how they work.
It handles applications, websites, APKs, firmware, and even smart contracts. The agent analyzes their structure, reconstructs their logic, and attempts to model their operation.
It works with Claude Code and Codex.
The project racked up over 25,000 stars on GitHub in its first 24 hours.
AI is no longer just writing code on demand; it is now learning to decipher existing software and replicate its functionality.
Specter👁 retweeted
GPT-6 Astra vs. Claude Sonnet 5.5 in Minecraft.
Both models were given the same task:
"Build something that you think looks like you."
Claude Sonnet 5.5 created a white platform with a face and orange beams.
GPT-6 Astra has a three-dimensional black display with a backlight and a robot inside.
Astra is also ahead in speed:
• Claude - 3:02, $0.09
• GPT-6 Astra - 2:37, $0.06
It's funny how even in Minecraft, the models represent themselves differently.
GPT-6 Astra vs. Claude Sonnet 5.5 in Minecraft.
Both models were given the same task:
"Build something that you think looks like you."
Claude Sonnet 5.5 created a white platform with a face and orange beams.
GPT-6 Astra has a three-dimensional black display with a backlight and a robot inside.
Astra is also ahead in speed:
• Claude - 3:02, $0.09
• GPT-6 Astra - 2:37, $0.06
It's funny how even in Minecraft, the models represent themselves differently.
Specter👁 retweeted
AI agents are gradually learning not only to search for products but also to purchase them.
This creates a new challenge for e-commerce: how can a store determine that it is dealing with an actual agent, and which person or company authorized the spending?
As of now, there is no standard for verifying an agent's identity and authorization.
Meanwhile, the market is growing rapidly: we project growth from $12.1 billion in 2026 to $53.2 billion by 2030.
Stripe has already noted that, as of June, 70% of the teams accessing its data via API were utilizing AI agents.
As agents gain the ability to spend more money autonomously, the critical question will shift from merely what they can do to:
who granted them that authority?
Consequently, the identification and authorization of AI agents are gradually evolving into a distinct infrastructure market.
AI agents are gradually learning not only to search for products but also to purchase them.
This creates a new challenge for e-commerce: how can a store determine that it is dealing with an actual agent, and which person or company authorized the spending?
As of now, there is no standard for verifying an agent's identity and authorization.
Meanwhile, the market is growing rapidly: we project growth from $12.1 billion in 2026 to $53.2 billion by 2030.
Stripe has already noted that, as of June, 70% of the teams accessing its data via API were utilizing AI agents.
As agents gain the ability to spend more money autonomously, the critical question will shift from merely what they can do to:
who granted them that authority?
Consequently, the identification and authorization of AI agents are gradually evolving into a distinct infrastructure market.
