China, China, China and: Can Newcomer Jev Endanger the Anthropic IPO?

China’s 2035 AI program meets growing job anxiety domestically

China’s leadership has declared the proliferation of artificial intelligence a national task: by 2035, the technology is to be embedded in every area of the economy and society. This course is supported by a steady stream of open-source models that are powerful and comparatively inexpensive to build and operate. Xi Jinping has expressed significantly more concern about the risks than Donald Trump and has called for regulation to ensure that artificial intelligence 'always remains under human control.' In the U.S., the debate is running in the opposite direction: after leading industry executives advocated for slowing down the pace of development, Trump dismissed these safety concerns as a hoax that would give China an advantage in the race for technological leadership.

Surveys show Chinese respondents are more enthusiastic about the technology than Americans, and its adoption is faster. Beneath this surface, however, the same debate is taking place as in the rest of the world. Chinese platforms are full of posts in which employees fear for their jobs and livelihoods. One example is the 30-year-old actor Xu Peng, who, after a period of unemployment last year, was filming one to two series per month in Hengdian, a production hub for film and short dramas in eastern China. After the Lunar New Year in February, two more productions were planned; the call never came, and upon inquiry, the producer informed him that the company had canceled the shoots and was using AI-generated videos instead. 'Yesterday everything was fine, and today it’s suddenly gone,' he says, describing the process.

According to this account, the leadership in Beijing is increasingly concerned about the political and economic consequences of this upheaval. While such conflicts are openly contested in the U.S. and negotiated through courts and legislation, China lacks corresponding avenues for those affected to have a say in their own situation. This creates a dilemma for the Communist Party: rapid development strengthens the country’s position abroad, while slowing down the technology internally could become necessary to maintain power.

{Synthszr Take: From the outside, China appears as a single actor with a roadmap: 2035, all sectors, managed from the top down. Internally, this roadmap disintegrates into a thousand individual cases, like the actor in Hengdian who lost two productions after the Lunar New Year because a studio switched to generated videos. These fault lines are barely measurable from the outside because they play out in short-video studios and regional labor markets rather than in hearings or lawsuits. For Beijing, this means that every acceleration creates pressure that has no outlet, and the Party must regulate it itself, without a release valve and without an early warning system. How quickly China proceeds in the coming years will be decided by employment figures in industries like film and advertising, and hardly anyone on the outside sees these exact figures.}

Noah Smith presents China strategy

Economist Noah Smith published an article in his newsletter Noahpinion titled 'How we can beat China,' in which he contemplates the strategic competition between the U.S. and China. His historical template is the so-called Long Telegram, which George Kennan sent to the Secretary of State in 1946 as chargé d’affaires at the U.S. Embassy in Moscow. Kennan argued in it that the Soviet leadership was permanently hostile but also cautious, and that the system would collapse due to its own internal contradictions. From this, the doctrine of Containment developed, which shaped American strategy until the Soviet Union collapsed 45 years later. → Noahpinion

Synthszr Take: Kennan could play for time in 1946 because the Soviet system was ruining its own production, and the gamble paid off after 45 years. The same patience with China would be the most expensive advice of all, because every year there, factory capacity, network expansion, and engineering hours are added, while in the West, strategy papers are being written. The effective levers are boring: approval processes in months instead of years and binding purchase guarantees for battery and chip manufacturing, so that facilities can be profitable at all.

Washington and Beijing agree to AI dialogue after eight hours of negotiations in New York

The U.S. and China have agreed to establish a bilateral dialogue on artificial intelligence and to launch a joint trade body, the Board of Trade. According to the South China Morning Post, this was preceded by about eight hours of negotiations in New York, just days before President Xi Jinping’s state visit to Washington. U.S. Treasury Secretary Scott Bessent and Trade Representative Jamieson Greer announced the results after their meeting with Vice Premier He Lifeng at JPMorgan’s headquarters. Bessent spoke of a very successful exchange on trade and AI, while Greer explained that previous ideas for a trade body and AI talks have now become concrete initiatives. The goal is to prepare the ground for a successful meeting between Trump and Xi on Thursday in Washington. → South China Morning Post

Synthszr Take: Eight hours of negotiations, and the most solid result is a commitment to continue talking in the future. In July, a ban on Chinese models was still on the table; now the format is called AI dialogue: progress in tone, without a single export control or approval issue being decided. The fact that the Chinese side said not a word to the press, while Bessent speaks of a very successful exchange, shows who currently holds the power to interpret this meeting.

China’s AI rollout is stalling: 70 percent of small businesses don’t get past the testing phase

The Atlantic describes China’s AI adoption as significantly bumpier than the image of a decisively advancing state would suggest. The state-affiliated China Center for Information Industry Development estimated in June that nearly 70 percent of the country’s small businesses have not moved beyond initial experiments with artificial intelligence, mainly due to costs and a lack of expertise. Beijing is sticking to its goal of embedding the technology in every area of the economy and society by 2035; at the same time, Xi Jinping has called for regulation to ensure that artificial intelligence 'always remains under human control.' The publication describes several affected individuals: Actor Xu Peng lost two promised roles in short drama productions because the studio had AI videos generated instead. On Douyin and RedNote, reports are circulating about so-called Distillation, where employers feed chat logs, documents, and emails from former employees into a model to take over their tasks. → The Atlantic

Synthszr Take: The claim that safety concerns are a gift to Beijing does not hold up to the facts: a state-affiliated Chinese institute counts nearly 70 percent of small businesses that do not get past pilot tests due to a lack of money and skills. The image of a fully digitized adversary is a rhetorical tool for domestic politics, and it works because hardly anyone checks what things are really like in Hengdian or in Shanghai’s production companies. When Xi himself speaks of control over the technology, the premise of a race crumbles, because both sides are struggling with the same fear of job loss and loss of control.

Young Chinese are founding AI startups because the job market no longer absorbs them

The Wall Street Journal describes a wave of startups in China driven by young people who have either quit tech jobs due to burnout or become unemployed. Instead of applying for jobs at large platform companies again, they are launching their own companies in the field of artificial intelligence. The report explicitly frames this as a reaction to the situation on the job market: Overwork in the industry and a lack of entry-level opportunities for graduates are converging. According to the WSJ, the entry costs for such startups are now so low that a product can be created without significant seed capital. → Wall Street Journal

Synthszr Take: Founding a company here is the emergency exit from a job market that first burns out young people and then no longer hires them. The barrier to entry has fallen worldwide (in Germany, 935 out of 3,568 new companies founded in 2025 were AI startups), and in China, there’s the added pressure that there is hardly any realistic alternative to self-employment. Economically, this is productive; from a biographical perspective, it’s harsh, and both are true at the same time.

China’s Academy of Sciences builds sub-3nm transistors with old DUV technology

Researchers at the Chinese Academy of Sciences (CAS) have reportedly completed an early process route for transistors below three nanometers without using the most modern lithography machines. Specifically, this involves stacked nanosheet channels in a Gate-all-around design, integrated with classic DUV lithography. Digitimes reported on this, citing statements from CAS researcher Ye Tianchun, which was picked up by the South China Morning Post. Ye himself emphasizes that the work is in an early stage and further validation at the wafer level is necessary before the process can be transferred to a production line. The background for this is the US and Dutch export restrictions that prohibit ASML from selling EUV systems to Chinese customers: EUV works with a 13.5 nanometer wavelength, DUV with 193 nanometers. → Techpresso

Synthszr Take: Creating structures under three nanometers with 193-nanometer light is like trying to write fine script with a felt-tip pen: it’s feasible, if you reinvent every other step in the process chain. The export ban on ASML’s EUV systems has driven the CAS people into transistor architecture, into stacked nanosheets, into multiple patterning, and into their own laser sources. Such detours cost yield and years, but they create expertise precisely where Western manufacturers simply order a machine from Veldhoven and leave the rest to the supplier.

Alibaba releases Qwen-Image-2.1: 7 billion parameters, runs on an RTX 3090

Alibaba’s Qwen team has released Qwen-Image-2.1, an open-weight model for image generation and editing. The component responsible for image synthesis, according to the team, manages with 7 billion parameters and beats most closed-source models in their in-house benchmark; independent comparative measurements are still pending, according to The Decoder. The model runs on powerful consumer-GPUs like an RTX 3090. It natively generates and edits transparent images in RGBA format, allowing objects to be isolated or text elements to be exchanged on transparent layers. Up to ten reference images can be processed simultaneously, for things like group portraits, virtual try-ons, or interior design, while circles, masks, or painted markings control local changes. → The Decoder

Synthszr Take: An RTX 3090 is long since used hardware with 24 GB of video memory, and that’s precisely what now runs an image model that Alibaba’s own benchmark places ahead of closed-source providers. This changes the calculation for any studio that has been paying an API per image: The marginal costs shrink to electricity consumption, and no client image ever leaves their own computer. Native RGBA layers and ten reference images in a single pass are production features; cloud services struggle with transparent cutouts and consistent characters in daily use.

Alibaba AI detects 146 diseases in a single CT scan

Alibaba’s research unit DAMO Academy has presented an AI system in the journal Science that screens for 146 different diseases in 18 organs from a single abdominal CT scan. In a comparative study with 26 specialist radiologists, ranging from residents to senior physicians, the mean diagnostic accuracy of the model, named DAMO RADAR, exceeded that of 23 of these doctors. The system was externally validated at eight independent clinics with approximately 40,000 real-world examinations, achieving a mean AUC of 0.913, thus surpassing the conventional threshold of 0.9 for outstanding clinical performance. It was trained on 420,000 contrast-enhanced abdominal CTs, from which 15 million anatomy-related image-text pairs were created, without additional manual annotation beyond the existing diagnostic reports. The code has been available on GitHub and HuggingFace under the Apache-2.0 license since September 18, 2026, with the model weights available separately under CC BY-NC-SA 4.0, which permits research but prohibits commercial use without a separate agreement. → Tech Times

Synthszr Take: 23 out of 26 means in plain terms: Three humans were still better, which is the entire lead the specialist retains in the study’s design. The mean AUC of 0.913 is just above the 0.9 mark, from which a diagnosis is considered outstanding, and this is across all 146 diagnostic classes, not just for one organ. The combination figure is more significant than the competition: a ten percentage point increase in sensitivity and over 30 percent less time per case when the doctor and model read the scan together.

Apple digitally signs photos for authenticity

Apple has introduced a process called 'Apple Reference Image' that secures the authenticity of photos at the sensor level and is set to launch with the main sensor of the iPhone 18 Pro and iPhone 18 Pro Max. The optional camera mode boots the sensor into a separate capture mode where the pixel data is cryptographically signed immediately after capture. Subsequently, a digital negative and a time-stamped reference image are created from this in two phases. According to the company, processing runs via Private Cloud Compute, which is also designed to shield the image data from Apple itself. Apple explicitly distinguishes its process from the industry standard “C2PA”, which attaches provenance data only after capture, making it vulnerable at any point in the editing chain. The company lists its own requirements as semantic authenticity, resistance to manipulation up to and including a device jailbreak, and privacy protection: An outsider should not be able to tell if two reference images came from the same device. → Benedict Evans

Synthszr Take: A seal is only as valuable as the habit of checking it, and virtually no one checks while scrolling. Technically, the matter is cleanly solved: The sensor signs the pixels before any software touches them, and the processing remains traceable. The burden thus shifts to the audience, who are supposed to distinguish between two categories: verified and everything else, with the second category comprising almost the entire web.

Vercel: Jev is being adopted faster than Faible and Atlas

Jev is the first product from TypeSafe AI, which went into Early Access on September 15 after about two years in stealth. The company was founded by Diogo Almeida, Erik Gafni, and Sasha Sheng. Almeida previously worked at Google Brain and OpenAI and was involved in Reinforcement Learning from Human Feedback and InstructGPT. At launch, TypeSafe announced a $40 million seed round led by DCVC. Almeida’s founding thesis: models have become strong at interacting with people, but for automation, the industry is optimizing the wrong interface because software speaks a different language than prose.

On September 20, the Japanese recruiting platform Offers added experience with the Jev model to its skill taxonomy. Candidates can now list this knowledge on their profiles, and employers can search for it and send targeted offers. Offers is operated by overflow, reports around 40,000 registered professionals, and claims to have facilitated hires at more than 1,000 employers since its launch in 2019. The platform has not stated that Jev itself screens applications, sorts candidates, or makes decisions. At the time it was added, the model was five days old.

Technically, Jev takes unstructured state—text or JSON—along with pre-typed questions, and returns typed answers with probabilities and confidence scores. The model does not generate free text, nor does it provide a rationale. TypeSafe calls the training method RLCD, Reinforcement Learning for Calibrated Decisions, where a higher stated probability is intended to mean a higher hit rate. The provider charges $0.042 per million input tokens and doesn’t charge for output tokens at all, stating they are “too cheap to bill.” TypeSafe states a response time of 70 to 500 milliseconds and quantifies its advantages, depending on the task, as 20 to 200 times the speed and 40 to 400 times the cost savings compared to language models.

Adoption was unusually fast. Via Vercel’s AI Gateway, Jev reached nearly 13 percent of paying teams within 24 hours—according to the operator, twice as many as the GPT-5.6 family and more than six times that of Fable 5.1. For a time, TypeSafe lost the ability to serve its own API because demand was too high. An engineer at Vercel reported that a security classifier ran five to eighteen times faster and more accurately after switching from ChatGPT Luna 5.6 to Jev. A developer at Bryo AI compared the classification of business emails: Gemini was slightly more accurate, but ten to twenty times more expensive.

The published metrics so far come predominantly from the provider itself and compare model outputs against the reference probabilities of other models rather than against human-labeled data. Averaged over four decision workflows, Jev achieves 68 percent accuracy at $0.0004 and 0.4 seconds per case, GPT-5.6 Terra also achieves 68 percent at $0.03 and 10 seconds, and Opus 5 achieves 73 percent at $0.18 and 38 seconds. The claim that Jev cannot hallucinate is described in expert debate as an overstatement: the model just can’t break out of the given schema, but it can be wrong within that schema. Initial independent tests are favorable but small-scale: a British event site measured 96 percent accuracy in content moderation versus 86 percent for Gemini Flash-Lite, and a developer ran the model on 18,514 spam emails without examples and ended up statistically on par with a trained classifier.

In the developer’s toolbox, Jev is primarily used where decisions are made repeatedly and in large numbers: routing between agents and tools, classification of tickets and documents, escalation decisions, and the evaluation of model outputs. Three question types are available: a choice from up to 255 options, a rating along up to ten rubric levels, and a yes/no question that only returns a probability and does not have its own confidence field. → Langfuse, RuntimeWire, The Deep View, TechCrunch, Vercel, The Stack, Arize AI

Synthszr Take: A model is five days old, and a platform with 40,000 registered professionals already lists it as a searchable hiring attribute. This says little about Jev and a lot about the half-life of qualification profiles: the taxonomy of job platforms now moves faster than any continuing education plan, any job posting process, and any certification a human resources department can set up. The reflex to pour such acronyms into search filters produces a false sense of precision, as no one can meaningfully distinguish between real experience and a blog post they’ve read after just five days. The critical judgment here is determining the confidence score at which a process runs automatically and when a human needs to look at it; that’s not on any skill list. If hiring decisions depend on tool names that last for three months, the profile is already outdated by the time the contract starts.