Forwarded from New Rules
Today’s most advanced AI still depends on large teams of people. Engineers choose the data, design training runs, test failures, rewrite code and decide what to try next. Chinese researchers have now mapped a route toward a system that takes over that entire cycle—and eventually improves the way AI improvement itself works.
A joint team from ByteDance, Tsinghua University, the Shanghai AI Laboratory and other institutions has outlined a five-stage path toward genuine recursive self-improvement.
At the first stage, AI merely carries out upgrade procedures written by humans. It then starts choosing its own improvement strategy, deciding what new data or experience it needs and adapting after deployment. At the final stage, the system would refine the very methods used to build better AI.
This is different from a chatbot correcting one bad answer. The improvement must remain after the task ends, become part of the system and pass into later versions. Each new model would begin with the lessons learned while creating the previous one.
The attraction is obvious. Training a foundation model currently consumes enormous amounts of engineering time and computing power. An autonomous research loop could launch experiments, compare results, discard failed approaches and keep successful ones without waiting for humans at every step.
Washington has tried to turn advanced chips into a choke point for Chinese AI. China is already building domestic alternatives to restricted Western equipment. Recursive self-improvement attacks the problem from the other side: reduce wasted training runs, automate more research and extract more progress from the hardware available.
Chinese companies are already developing pieces of this system. Z.ai plans to direct about 60% of the proceeds from its latest $5B fundraising round toward new GLM models and a “fully self-training system”. MiniMax has tested models that update memory and acquire new skills during reinforcement-learning experiments, while DeepSeek has built an agentic framework able to execute code and handle complex chains of tasks.
Nobody has reached the final stage, and the paper offers no timetable. Software development provides the clearest testing ground; robotics and scientific research are far harder. Every autonomous update would also need to be tested before entering a live model.
But China is assembling the research base, companies and practical systems needed to move in that direction. Its generative-AI patent output already exceeds the rest of the world combined.
The “last AI built by humans” would not be a finished machine. It would be the first one capable of turning its own development into a continuing production line—building each successor faster, cheaper and with less human direction than the one before it.
@NewRulesGeo
Please open Telegram to view this post
VIEW IN TELEGRAM
You thought you were talking to DeepSeek and Kimi, maybe it was Claude: the revelations of Anthropic
Published on 13 September 2026 at 17:22
© Anthropic
Seven Chinese labs, nearly 200 million hijacked exchanges, and users of Kimi or DeepSeek who spoke unknowingly to the model of an American competitor. Anthropic’s report reads like an industrial polar, at a charge.
Distillation, in AI jargon, refers to a perfectly legitimate training method: we make a big model “teacher” answer thousands of questions, then we train a lighter “elevent” model to imitate it. The student reviews on the copies of the promo major, in short, and the problem begins when the teacher has not asked for anything. In its semi-annual report on Claude’s malicious uses, published on September 10, Anthropic accuses seven Chinese laboratories, including Alibaba, DeepSeek, Moonshot AI (the publisher of Kimi) and Xiaomi, of having siphoned off its model on an industrial scale since December.
Do Moonshot and DeepSeek make you speak to Claude in secret?
The two most embarrassing cases concern consumer services. According to Anthropic, Moonshot AI transmitted in silence, over a single period of ten days, nearly from his customers to Claude, via 5 380 fraudulent accounts located in Singapore and Japan. The answers were then displayed as if they were coming from Kimi. DeepSeek would have done the same, with a refinement: its systems spotted developers who went through tools like Claude Code or OpenCode, and selectively redirected their sessions to Claude Opus. In both cases, the exchanges were recorded to extract the reasoning chains of the American model and feed the training of Kimi and DeepSeek.
Alibaba, with its Qwen models, would have conducted the largest transaction ever measured by Anthropic: 151 million exchanges between May and July, up to 3 million a day (yes, per day), via more than 3 500 accounts created with fake identities and stolen bank cards. Zhipu reportedly extracted 3.4 million exchanges in seventeen days, and Xiaomi would have replayed in Claude the sessions of its own users. SenseTime would have bought transcripts from retailers, while MiniMax would have set up a screen company offering only American models, without any Chinese model in the catalog (subtle). In total, nearly 200 million exchanges, against three labs targeted during the first alert of Anthropic, in February.
Why the case is beyond the dispute between labs
Already in January 2025, OpenAI accused DeepSeek of having distilled its models, at a time when R1 was screwing up the stock market; Google has documented the same phenomenon this year, and Anthropic is therefore in its second round. What changes with this report is the mechanics described. The labs would go through networks of open accounts with stolen cards and backed API keys, doubled by commercial relays that resell access to Claude where it is not proposed. These intermediaries record conversations and sell them to anyone who wants to train a model: the theft of capabilities has a supply chain, with wholesalers and retailers.
In the relayed queries, Anthropic claims to have found names, email addresses, access tokens still active and investment forecasts from a pharmaceutical company, all in a dozen languages. A user presumably linked to the Chinese military even had video surveillance data analyzed, believing he was addressing Kimi. Some of this traffic was transiting through widely used third-party routing platforms in Europe, where GDPR compliance with these silent redirects would raise a few questions (Anthropic speaks modestly of practices “probably incompatible” with privacy laws). Remains a blind spot, and it is of size: the accuser is the direct competitor of the accused, none of the seven labs have reacted publicly at this stage, and the evidence remains that of only one party. Anthropic says it has strengthened its filters and now summarizes Claude’s reasoning, to make stolen transcripts less useful.
Published on 13 September 2026 at 17:22
© Anthropic
Seven Chinese labs, nearly 200 million hijacked exchanges, and users of Kimi or DeepSeek who spoke unknowingly to the model of an American competitor. Anthropic’s report reads like an industrial polar, at a charge.
Distillation, in AI jargon, refers to a perfectly legitimate training method: we make a big model “teacher” answer thousands of questions, then we train a lighter “elevent” model to imitate it. The student reviews on the copies of the promo major, in short, and the problem begins when the teacher has not asked for anything. In its semi-annual report on Claude’s malicious uses, published on September 10, Anthropic accuses seven Chinese laboratories, including Alibaba, DeepSeek, Moonshot AI (the publisher of Kimi) and Xiaomi, of having siphoned off its model on an industrial scale since December.
Do Moonshot and DeepSeek make you speak to Claude in secret?
The two most embarrassing cases concern consumer services. According to Anthropic, Moonshot AI transmitted in silence, over a single period of ten days, nearly from his customers to Claude, via 5 380 fraudulent accounts located in Singapore and Japan. The answers were then displayed as if they were coming from Kimi. DeepSeek would have done the same, with a refinement: its systems spotted developers who went through tools like Claude Code or OpenCode, and selectively redirected their sessions to Claude Opus. In both cases, the exchanges were recorded to extract the reasoning chains of the American model and feed the training of Kimi and DeepSeek.
Alibaba, with its Qwen models, would have conducted the largest transaction ever measured by Anthropic: 151 million exchanges between May and July, up to 3 million a day (yes, per day), via more than 3 500 accounts created with fake identities and stolen bank cards. Zhipu reportedly extracted 3.4 million exchanges in seventeen days, and Xiaomi would have replayed in Claude the sessions of its own users. SenseTime would have bought transcripts from retailers, while MiniMax would have set up a screen company offering only American models, without any Chinese model in the catalog (subtle). In total, nearly 200 million exchanges, against three labs targeted during the first alert of Anthropic, in February.
Why the case is beyond the dispute between labs
Already in January 2025, OpenAI accused DeepSeek of having distilled its models, at a time when R1 was screwing up the stock market; Google has documented the same phenomenon this year, and Anthropic is therefore in its second round. What changes with this report is the mechanics described. The labs would go through networks of open accounts with stolen cards and backed API keys, doubled by commercial relays that resell access to Claude where it is not proposed. These intermediaries record conversations and sell them to anyone who wants to train a model: the theft of capabilities has a supply chain, with wholesalers and retailers.
In the relayed queries, Anthropic claims to have found names, email addresses, access tokens still active and investment forecasts from a pharmaceutical company, all in a dozen languages. A user presumably linked to the Chinese military even had video surveillance data analyzed, believing he was addressing Kimi. Some of this traffic was transiting through widely used third-party routing platforms in Europe, where GDPR compliance with these silent redirects would raise a few questions (Anthropic speaks modestly of practices “probably incompatible” with privacy laws). Remains a blind spot, and it is of size: the accuser is the direct competitor of the accused, none of the seven labs have reacted publicly at this stage, and the evidence remains that of only one party. Anthropic says it has strengthened its filters and now summarizes Claude’s reasoning, to make stolen transcripts less useful.
👉🏻 Follow the tech news in real time: add 01net to your sources on GoogleGoogle, and subscribe to our WhatsApp channel.
Source : Anthropic
AnthropicClaudeDeepSeekKimi
Written by Naim Bada
01net.com
Télécharger DeepSeek pour Web, iOS, Android - Telecharger.com
Venez tester DeepSeek, le tout premier chatbot chinois à rivaliser avec les grands modèles occidentaux. Vous pouvez y accéder sur ordinateur (via n'importe quel navigateur) ou sur mobile (Android, iPhone…
Forwarded from The Cradle
China cuts US Treasury holdings to lowest level in 18 years
Beijing has begun investing its surplus foreign currency in gold, US agency securities, and stocks as US debt levels become unsustainable
Beijing has begun investing its surplus foreign currency in gold, US agency securities, and stocks as US debt levels become unsustainable
BIG PICTURE with James Patrick (Planet Lockdown)
Ukrainian drone strikes have disabled approximately 42-43% of Russia’s oil refining capacity, forcing Moscow to impose a temporary ban on diesel exports through July 31, 2026, to protect domestic supplies. This disruption has triggered a severe fuel crisis…
This media is not supported in your browser
VIEW IN TELEGRAM
Diesel prices jumped due to the Ukrainian attacks on Russian oil refineries. US truckers and farmers have been severely effected.
🌚1
Forwarded from Geopolitics Prime
Footage from social media shows the skies above Riyadh and the sound of interceptions after Yemeni ballistic missiles were fired toward the Saudi capital.
The likely target was Prince Sultan Air Base, which also hosts US troops. the base is located roughly 80 km from Riyadh — meaning early warning alerts typically aren't triggered for the city when the base is attacked — the same alert pattern observed when Iran struck the same base earlier this year.
Local reports indicate significant disruptions and malfunctions in navigation systems at Riyadh International Airport.
Please open Telegram to view this post
VIEW IN TELEGRAM
Please open Telegram to view this post
VIEW IN TELEGRAM
Forwarded from Mediterranean Man (Three)
⚡️🇺🇸🇨🇳 A faulty U.S. AI-assisted intelligence report almost started a war with China, per CNN.
An U.S AI-assisted intelligence report asserted that a Chinese vessel was carrying nuclear weapons materials to Iran from China.
Just as the preparations were complete to board this vessel, the U.S. military discovered that this intelligence report had been largely fabricated by AI.
@medmannews
An U.S AI-assisted intelligence report asserted that a Chinese vessel was carrying nuclear weapons materials to Iran from China.
Just as the preparations were complete to board this vessel, the U.S. military discovered that this intelligence report had been largely fabricated by AI.
@medmannews
Forwarded from Mediterranean Man (Beaky Boi)
⚡️🇸🇦🇪🇺 Saudi Aramco has informed European customers that they will receive no crude oil next month. Zero. This comes after last week’s drone attack shut down their East-West pipeline to the Red Sea. Assuming no additional attacks occur, the pipeline will take months to return to full operation.
@MedmanNews
@MedmanNews
Forwarded from Andrew Bridgen
This media is not supported in your browser
VIEW IN TELEGRAM
I posted the letter to the Guardian by Matt Anderson on social media years ago. I describes by an insider what geoengineering is really all about is a military weapon not an environmental tool.
And yes , no newspaper would dare publish it today.
https://t.me/ABridgen
And yes , no newspaper would dare publish it today.
https://t.me/ABridgen
💯1
Forwarded from Andrew Bridgen
Our State broadcaster the BBC is preparing us for war with Russia.
Why are our legacy media doing this ?
Answer - Because they are told to.
We don’t have a free press in the UK
#DefundtheBBC
https://www.bbc.co.uk/news/articles/cmn0jke547r5o
https://t.me/ABridgen
Why are our legacy media doing this ?
Answer - Because they are told to.
We don’t have a free press in the UK
#DefundtheBBC
https://www.bbc.co.uk/news/articles/cmn0jke547r5o
https://t.me/ABridgen
BBC News
War may be coming. Are we psychologically ready?
The military and government say the risk of conflict is growing, but we may not be mentally ready for it
Forwarded from Andrew Bridgen
This media is not supported in your browser
VIEW IN TELEGRAM
PIDDINGTON village has voted for independence from the UK.
The village has the population of only 350 people, who are vehemently against the UK government plans to put 1,250-3,500 illegal migrants in a camp next to them.
PIDDINGTON voted 91% for Independence from the UK following the Government’s decision to press ahead with the illegal migrant camp on their doorstep.
What happens if all UK towns and villages vote to leave the UK to escape our tyrannical Government ?
https://t.me/ABridgen
The village has the population of only 350 people, who are vehemently against the UK government plans to put 1,250-3,500 illegal migrants in a camp next to them.
PIDDINGTON voted 91% for Independence from the UK following the Government’s decision to press ahead with the illegal migrant camp on their doorstep.
What happens if all UK towns and villages vote to leave the UK to escape our tyrannical Government ?
https://t.me/ABridgen
👍1🤡1