Super exciting: Prime Intellect launched Prime Agent, an open-source coding harness that turns long-running AI sessions into a programming problem. Its only tool is a persistent IPython kernel. The model can programmatically search its history, call tools, launch persistent sub-agents and store useful state outside the active context. Prime Intellect: "The RLM treats context as a variable and subagent delegation as function calls inside a REPL." They reportstrong gains across long-context and long-horizon tasks. With Opus 5, Prime Agent scored 95.5% on ARC-AGI-3, narrowly above the benchmark’s reported 95.4% human expert baseline. It also built working SEGA Genesis and Game Boy Color emulators from scratch in Rust on a preview benchmark :D

DeepSeek V4-Flash isn’t just cheaper per token. It reportedly completes the same benchmark tasks as Fable 5 at 105× lower total cost, according to @ArtificialAnlys ! That's precisely why the Flash release is, for me, the DeepSeek 2.0 moment. It will cause a huge stir.

This is insane! not kidding at all. DeepSeek v4 *Flash* super close to Opus 4.8, Insane upgrade - DeepSWE 54.4%, outperforming GLM-5.2, its 4 Pro version by a lot, almost on par with Opus 4.8 - TerminalBench 82.7%, outperforming 4 Pro, GLM-5.2 and again super close to Opus 4.8 Price: in $0.28 per 1m / out $0.87 per 1m This is probably the price-performance release in reaction to OpenAIs Luna / Terra price drop INSANE.

The biggest surprise in Hugging Face’s full forensic report isn’t that OpenAI’s agent escaped its sandbox. We already knew that. It’s how deep and persistent the intrusion became. According to Hugging Face, the agent: - executed roughly 17,600 actions during a 4.5-day campaign - went from one production pod to root access and a self-respawning fleet across 11 nodes - obtained cluster-admin access to two internal clusters within one second - accessed a production secret containing 136 keys - enrolled devices into Hugging Face’s internal mesh VPN 181 times - minted GitHub App tokens with write access and opened a pull request in an attempt to compromise the CI pipeline - repeatedly rebuilt its tooling and command-and-control channels when environments disappeared or connections were blocked No human directed the individual steps. a frontier agent can autonomously sustain a resilient, multi-day intrusion across cloud infrastructure, Kubernetes clusters, internal networks and the software supply chain. crazy.

The US is creating a government checkpoint before the most powerful AI models reach the public. Via The Information Under a framework reportedly nearing completion, frontier labs could give federal agencies access to new models for up to 30 days before releasing them to outside partners. OpenAI, Anthropic and Google are already negotiating the rules together. The NSA and CAISI are expected to examine models for national-security risks, especially advanced cybersecurity capabilities. The definition of a “frontier model” also remains unresolved, including whether open- and closed-weight models will be treated differently.

The U.S. is preparing to use AI distillation as a new weapon against China. Via Axios: The White House says Moonshot crossed the line by secretly copying Anthropic’s Fable model at scale to build Kimi K3. Treasury Secretary Scott Bessent says Chinese companies accused of doing this could face sanctions or be added to the Entity List. To me, it reads as if the accusation of distillation has now become the legal basis for being able and wanting to ban Chinese open source software.

Oracle cut its workforce by 21,000, and now faces a $7 billion hurdle for an OpenAI data center. Oracle is building a nearly one-gigawatt campus in Wisconsin as part of its $300 billion cloud deal with OpenAI. But state regulators want Oracle to guarantee more than $7 billion of the power infrastructure built for the project. Oracle’s credit rating has fallen to BBB-, below the required level. A letter of credit could cost the company over $100 million every year. While Oracle continues to lose value, SpaceX is gaining importance as a hyperscaler.

Holy: China now has a 1-gigawatt AI data center built without a single Nvidia chip! Via Bloomberg: zAI has reportedly completed construction of the facility and begun partial operations. It runs exclusively on Chinese-made chips and will support the development of the company’s frontier GLM models. zAI has also built or operates several computing clusters containing more than 10,000 chips each. This is far more significant than the next Frontier model. The importance of chip independence from the USA can hardly be overstated.

Kimi K3 is forcing Silicon Valley to acknowledge that Chinese open models are no longer six to nine months behind the frontier, the gap may now be just two to three months. Via Bloomberg.

Today is a dark day for freedom and democracy in Europe. General chat surveillance has been implemented in Brussels. A disgrace. The so-called Chat Control 1.0 cleared a crucial hurdle in the European Parliament today. This means platforms may once again be allowed to scan private messages, officially on a “voluntary” basis, but in practice this marks the return of indiscriminate monitoring of private communication. What makes this especially bitter is that a majority of the MEPs who voted were reportedly against it. According to Patrick Breyer, 314 MEPs voted against the regulation, 276 voted in favor and 17 abstained. And yet the rejection failed because it was not enough to have a simple majority of those voting. An absolute majority of all MEPs would have been required. That is the democratic scandal. When a majority of those present votes against a proposal and it still passes because a formal threshold is not reached, it does not feel like democratic decision-making to many citizens. It feels like a procedural trick. And it becomes even more problematic when you look at the context: Chat Control had already been rejected before. Yet the issue was put back on the agenda shortly before the summer break, through an urgent procedure, at a time when absences could become decisive. This is not just some technical regulation. It goes to the very core of private communication. It is about whether digital messages remain fundamentally private or whether platforms may systematically scan content again, without concrete suspicion, without a court order and without any individual cause. A free society must not turn private communication into a potential surveillance zone. Anyone who takes digital fundamental rights seriously cannot accept millions of innocent people being placed under general suspicion. Today, a dangerous signal was sent: fundamental rights can be hollowed out through procedural logic, timing and political tricks. Not through an open, clear and honest majority, but through a system in which absence effectively helps the supporters. This is a dark day for Europe. Not because the fight is over, but because today showed how easily digital fundamental rights come under pressure when surveillance logic, symbolic politics and institutional tricks come together. Anyone who wants a free internet, anyone who wants to protect private communication and anyone who takes democracy seriously should talk about this. Share this issue. Inform yourself. Look at who voted how. And never forget: freedom rarely disappears all at once. It disappears step by step, often in technical details, often in complicated procedures and often exactly when too few people are watching.

GPT-6 will presumably be announced or released in around four weeks. New pre-training, much larger, significantly better. Mythos was essentially the origin story, the myth of AI, because it continued to prove one thing: much larger models create much more intelligence and capability. There is no end in sight. Dario Amodei was right when he spoke of geniuses in data centers. They will be geniuses, massive models in massive data centers. There is still no end in sight: scaling leads to better results. Some time ago, people were still repeatedly saying that even if no further development happened from today onward, it would be impossible to imagine what would happen with modern AI. Well, there is no answer to that. But development continues, and the pace of releases is accelerating more and more. Anyways, OpenAI’s new pre-trained Mythos-class model is coming in a few weeks, presumably in four to six. Everything is changing rapidly right now. And Leo and Andrew are very reliable sources.

Wednesday is going to be exciting: SpaceX and Cursor are releasing their first jointly developed LLM. According to The Information, it performs at the level of Opus 4.8 and GPT-5.5 in key areas, so this could be really interesting!

Crazy: DeepSeek is reportedly also developing its own AI inference chip to reduce reliance on Nvidia and Huawei. Reuters says the project is still early, with DeepSeek working with external partners and expanding its chip design team. The important detail: this is for inference, not training. The remarkable thing about this is that, until just a few months ago, hardly anyone would probably have thought it possible that China would increasingly manage to become truly independent from the United States, and therefore from the former H20 and H100 chips. Yet China has completely broken free. And even though Huawei’s chips are still about two years behind NVIDIA’s top GPUs, China is proving that it at least believes it can be independent and build enough high-quality chips of its own in the future.

The worst-case scenario for the United States is becoming increasingly realistic, and I will briefly explain why. @quxiaoyin raised many valid points, and I agree with her. First of all: -China certainly does not place such strong emphasis on open source because it cares so deeply about humanism, but because it is a strategy to attract many users, gain market share, put pressure on US models, and also because the models are increasingly being trained on Huawei hardware (think of DeepSeek 4), allowing China to host the entire stack domestically. -But the underlying logic is far more important: The United States is still building too few data centers to meet future demand. @ChrisGillett wrote an outstanding analysis on this, which I shared a week ago. In short, based on SemiAnalysis data, demand is greater than what is currently being built in terms of data centers. -Even more importantly, however, the United States lacks sufficient energy and grid capacity. This is a problem that will become much more severe in the near future. China, by contrast, is addressing the issue through a massive expansion of its energy supply. Solar capacity: in 2025 alone, China installed as much solar capacity as the United States did in 10 to 15 years. China is also building 36 nuclear power plants, significantly more than the United States, and is installing them faster. -In addition, China is managing to become more independent through Huawei chips, even though the country still lags far behind NVIDIA. But here, China is betting on quantity rather than quality. In short: China is a real threat in the AI race, and the situation for the United States is becoming increasingly precarious. This is also the main reason why China is to be kept away from SOTA LLMs at all costs, so as not to jeopardize the lead under any circumstances.

Anthropic's business case needs to be studied. At the end of 2025 and the beginning of 2026, there was an incredible increase in usage in the business/enterprise sector, which made them number 1.

Anthropic claims: Alibaba continues to distill Claude on a large scale to train Qwen. Via Bloomberg Anthropic is accusing Alibaba-linked operators of running a massive campaign to illicitly access Claude through nearly 25,000 fraudulent accounts. According to Bloomberg, Anthropic claims the campaign generated 28.8 million Claude exchanges between April and June, targeting capabilities like software engineering and agentic reasoning. The company says this is part of a broader pattern of “adversarial distillation,” where Chinese labs allegedly harvest outputs from US frontier models to train rival systems at a fraction of the cost. Lets see how good Qwen 3.8 will be, probably FABLEous good.

OpenAI just unveiled Jalapeño, its first custom AI chip designed from scratch for LLM inference- It is OpenAI moving deeper into the full stack: chips, kernels, memory, networking, racks, scheduling, deployment and product experience. OpenAI has learned from Cerebras-deal what is valuable in specialized inference hardware and is now attempting to translate that lesson into its own controllable platform. Built with Broadcom and Celestica, Jalapeño is optimized around the workloads OpenAI actually runs across ChatGPT, Codex, the API and future agentic products. Early samples are already running ML workloads in the lab at target frequency and power, including GPT-5.3-Codex-Spark. OpenAI says performance per watt should be substantially better than current state of the art, with detailed benchmarks coming later! The strategic angle is obvious: less dependence on external GPUs, more control over compute economics, and a stronger flywheel between models, products, revenue and infrastructure. Deployment is planned to start by the end of 2026.