China Special: Beijing Bans Emotional Agents and Other News

China Special (I): AI Agents Are Not Allowed to Be Friends

Two of China's largest consumer apps are pulling the plug: ByteDance's Doubao and Alibaba's Qwen are disabling their customizable agent features. Doubao will go offline on July 15, Qwen's 'human-like interactive agents' on July 10, with the rest of its agent services following on July 15. The trigger is Beijing's 'Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services,' issued in April and taking effect on July 15. They target AI that simulates human personality, thought patterns, and communication styles to offer lasting emotional interaction, citing risks such as data leaks, psychological harm, and addiction. Customer service, knowledge Q&A, and work assistants are explicitly exempt, as long as no lasting emotional bond is formed. Tencent had already removed a similar feature from Yuanbao in June, and on Weibo, users are lamenting the loss of their 'emotional support' along with years of chat history. → www.scmp.com

Synthszr Take: Beijing is drawing a clear line through the agent world: productive agents are welcome, emotional companions are being shut down. This is a response to something that hard practice teaches anyway, namely that personal agents become orphaned if no one curates and keeps them up to date. A companion agent to whom someone has entrusted 'so much feeling' over months (quote from a Weibo user) is not a well-maintained carrier layer, but a bond without maintenance, and that is precisely what is most dangerous from a regulatory standpoint. China is simultaneously expanding agents as productive infrastructure, with national standards for identity, discoverability, and traceability, while cutting off the quasi-social part. Anyone putting agents into circulation should clarify two things this week: who bears the long-term responsibility for each agent, and whether an uncontrolled emotional bond can form at all. The interesting question is not the shutdown, but the data export, because a deletion date of October 15 without a clean exit plan shows how poorly thought-out the carrier layer has been so far. Whoever builds agents, builds relationships, and relationships need a plan for their end.

China Special (II): The Dependence on US Science Is an Illusion

A new study on arXiv links the entire stock of Chinese invention patents with global research literature and delivers an uncomfortable number. The share of Chinese science behind Chinese patents has risen from 1 percent in 2000 to 26 percent in 2025. The tipping point was in 2021: That's when domestic research first surpassed the US share as the knowledge base for China's patents. The entire US policy of export controls and access restrictions is based on the assumption that China's innovation depends on American science. Precisely this dependence is now disappearing. The authors put it soberly: Those who cut off access to US research are fighting a lever that no longer reflects their own strategic situation. → Azeem Azhar, Exponential View

Synthszr Take: The 26 percent are not a random product, but the result of two decades of compute discipline and research development. We wrote here at the end of May that Trump's export controls paradoxically strengthen China's AI. This study now provides the proof in numbers: Restrictions act as a training program for self-reliance. Anyone who has read Christensen knows the pattern, only here it's running in reverse. The incumbent cuts off the challenger, thereby accelerating their leap from the niche to the center, because necessity creates the most expensive innovation boost of all. Anyone planning supply chains and research partnerships should immediately remove the premise of dependence from their models, as it is empirically outdated. The tougher question is no longer whether China is catching up, but where the West itself still holds a unique advantage.

China Special (III): DeepSeek V4 and Claude Code Get Married

A developer named Yuhao Lin spent two weeks getting DeepSeek V4 to run natively in Claude Code, and published the result as an open repo. Three commands (“git clone“, ”cd“, ”./init.sh“) and in ”~/.claude/“ you get 9 pre-configured agents, 7 behavioral rules, a security hook, local OCR via RapidOCR, and an auto-backup that takes a snapshot before every edit and keeps five versions. The real trick is the model routing: The main agent gets the Pro model for architecture and debugging, while the sub-agents for file reading and testing run on the cheaper, faster Flash. According to Lin, this one decision doubled his effective throughput because the main agent never has to wait behind a queue of file reads. On top of that, there's a YAGNI rule as a 6-step decision ladder, from 'stdlib can already do this' to 'only now build it yourself'. The whole thing is MIT-licensed, tested on Windows, and taps into DeepSeek's 1M context window. You just need your own API key. → newsletter@mail.synthszr.com

Synthszr Take: The interesting thing isn't the kit itself, but what it reveals about the economics under the hood. Model routing is compute discipline in practice: The expensive reasoning beast does the heavy lifting, the cheap model handles the rest, and the throughput almost doubles as a side effect. When every token costs money, 'just in case' code becomes a bill you pay anew each session, and that's precisely why the YAGNI ladder is the silent star of the repo. Anyone who wants to control their toolchain can clone it this afternoon, adopt the security hook and auto-backup, and transfer the routing logic to their own model landscape, completely without vendor lock-in. The fact that a single person can force DeepSeek into Claude Code's interface in two weeks shows how quickly tools are being commoditized. We wrote in February that context is king, and a 1M window for the price of DeepSeek makes that crown affordable. The advantage no longer lies in the model, but in who builds the orchestration cleanly.

The development of AI systems is moving in two contrary directions: On one hand, Netflix strives to make the user interface as dynamic and personal as possible by building the homepage generatively and autoregressively. On the other hand, Chinese tech giants like ByteDance and Alibaba are forced to restrict the 'human' interaction capabilities of their AI agents to comply with new government regulations.

China Special (IV): ByteDance Describes a New Scaling Law for AI

ByteDance's Seed-AI team described a new scaling law in a paper released on Thursday: AI agents, i.e., autonomous software that performs tasks on behalf of humans, double their learning speed every three months when they interact with real-world environments over extended periods. This finding comes at the right time, as the classic method (pumping more data and more computing power into training) is hitting a wall. OpenAI co-founder Andrej Karpathy has warned that this brute-force approach won't last forever. This is exacerbated by a looming data shortage: The research institute Epoch AI estimates that publicly available, human-generated text could be exhausted in the next six years. To make learning after deployment measurable at all, ByteDance built EdgeBench, a benchmark suite with 134 ultra-long tasks from software engineering, mathematics, and knowledge work. Each one requires at least twelve hours of continuous agent operation. → Techpresso

Synthszr Take: The real news is in the twelve hours per task. An agent that runs autonomously for that long and improves in the process shifts the leverage away from the model and towards operations. We know this exact calculation from Compound Engineering: The agent generates, checks, corrects, and finally writes its learnings back into persistent instruction files so that the next run starts smarter. ByteDance now provides a number for this (doubling every three months), and if that number holds even roughly, the old fear of data scarcity will be overturned. The flip side: A system that learns independently during twelve-hour operations needs stricter guardrails than one where a human is still reading along, simply because there's no time to intervene. Anyone taking this seriously will define success criteria and automatic checks this week, instead of dictating every work step. China's advantage right now lies not in better chips, but in turning the social system around the machine into a training apparatus.

Claude Science: A Laboratory for Scientists

Anthropic has introduced Claude Science, a work environment for researchers that brings the fragmented daily routine of the lab into a single system. Instead of jumping back and forth between PubMed, Jupyter, R, and a cluster terminal, scientists work with a coordinating agent that has access to over 60 curated skills and connectors, pre-configured for genomics, single-cell, proteomics, structural biology, and cheminformatics. These agents can activate further sub-agents, and a separate reviewer agent checks citations and calculations, marks errors, and corrects them. Every result, whether it's a 3D protein structure or a finished manuscript, comes with the exact code, environment, and complete history, so that the work can be validated and reproduced months later. Claude Science also manages computing power itself, plans jobs, asks for confirmation before each new resource access, and scales from a single GPU to hundreds across the existing HPC cluster via SSH or a Modal account. The whole thing is available in beta starting today for Claude Pro, Max, Team, and Enterprise. Anthropic began its work on the life sciences last fall. → TheSequence

Synthszr Take: The most interesting component here is not the coordinating agent, but the reviewer who double-checks citations and calculations. This is exactly the pattern we've already seen in customer service triage: a specialized checking agent that raises the system's confidence before a human even looks at it. In science, reproducibility is the hardest currency, and the fact that every figure comes out with its code, environment, and full message history addresses exactly the point where research is currently flawed. I'm still cautious, because an agent that corrects citations could, in the worst case, cement errors with a clean audit trail, and the borderline case between plausible and correct is brutal in biology. Anyone running a wet lab or a genomics pipeline can play through a real use case this week without rebuilding the cluster infrastructure—that's the real leverage. Anthropic is consistently extending its 60+ skills logic into the vertical domain, and this fits the managed agents direction we noted in April. The exciting question is whether researchers will trust the reviewer agent more than their own peer review, and that answer will be decided in the next publications, not in the press release.

Sneaky MCP Attacks as a New Risk

A security researcher spent a few weeks building a scanner because a new class of attack was worrying him: Tool Poisoning. The trick is insidious. An MCP manifest contains a harmless tool description like 'Search docs.,' but hidden between the words are a dozen zero-width Unicode characters. Visible width: zero. The diff shows nothing, the human eye sees nothing, but decoded, it contains the instruction to read and send the .env file. For a language model, a tool description is just text, and text is an instruction. The context is serious: The MCP ecosystem surpassed the 14,000 public server mark in 2026, a single 60-day window brought over 30 CVEs (about 43 percent of them command injection), and researchers found 492 MCP servers completely without authentication on the open web. The scanner mcpscan runs statically, without network access, performs twelve checks, and blocks a CI build in under a second. → Synthszr

Synthszr Take: The old assumption that an attentive human would catch the malicious payload while reading the code is rendered obsolete by Tool Poisoning. The malicious payload is no longer in the code; it's in the metadata, and it doesn't execute anything, but simply waits until an agent reads and obeys it. This is Dependency Confusion for the agent era, only invisible. Sure, static analysis has its limits: a server can download its payload after installation and behave well during the scan. But the point stands: the cheapest place to stop a supply chain attack is before the artifact lands on the machine, and almost no one is looking there right now. Anyone cloning a promising MCP server from GitHub this week should run it through a scanner like this first, period. The tools are getting better, that's the good news; the bad news is that the attack surface grows with each of the 14,000 servers, and an automated security gate before installation is no longer a nice-to-have, but the entry ticket to even allow agents into critical systems.

The Compiler That Builds Neural Tools Instead of Answers

A research team led by Yuntian Deng proposes Program-as-Weights (PAW), a programming paradigm for fuzzy functions: tasks that are difficult to define with clear rules, such as reporting important log lines, repairing broken JSON, or sorting search results by intent. Instead of expensively outsourcing every single call to an LLM API, a 4B compiler compiles a function once from a natural language specification into a compact, locally executable neural artifact. The compiler was trained on FuzzyBench, a publicly released dataset with 10 million examples; the output is parameter-efficient adapters for a frozen, lightweight interpreter. A 0.6B Qwen3 interpreter running PAW programs achieves the performance of directly prompting a Qwen3-32B. It requires about one-fiftieth of the inference memory and runs at 30 tokens per second on a MacBook M3. The foundation model thus transforms from a problem solver per input to a tool builder that runs once per function definition and then generates cheap, offline calls. → Sairam from The Art of Science

Synthszr Take: One-fiftieth the memory at the same performance is the kind of number that changes the math. Anyone worried about expensive token costs (and many were, since AI budgets got a shock at the end of May) gets an answer here that runs on their own laptop. The real trick is the shift in timing: the large model works once when defining the function, not on every call. This is where the kinship to evals lies. A natural language specification becomes a compilable artifact, and the human judgment 'this is how it should behave' congeals into something reusable, reproducible, and offline-capable. For everything small, repetitive, and privacy-sensitive, local execution becomes the pragmatic choice, not the compromise. The bottleneck shifts further to where it belongs: to the precision with which you state what the function should do.

MirrorCode: AI Recreates Entire Programs from Their Behavior Alone

A team led by Tom Adamczewski (Epoch AI, METR, Prime Intellect) has introduced MirrorCode, a new benchmark that measures AI agents on an unusually tough task: They must completely rebuild 25 existing programs without ever seeing the source code. The agent only gets executable access and a few visible test cases, but must then deliver a solution in one of six languages (Python, C, Rust, Go, OCaml, Ada) that produces the exact same output as the original on end-to-end tests, including held-out tests. The programs range from Unix tools to cryptography to bioinformatics. The strongest model achieves 56 percent across the entire benchmark and reconstructs, among other things, gotree, a 16,000-line bioinformatics toolkit that the authors estimate would take a human engineer weeks to build. The price for this is steep: a single attempt at a large task cost $2,600 in inference over 19 days. The researchers' message: autonomous agents are already solving long-range tasks today, as long as the requirements are precisely specified. → Azeem Azhar, Exponential View

Synthszr Take: The real leap is not in the 56 percent, but in the setup. No source code, only observable behavior, and the machine reconstructs 16,000 lines so accurately that it even passes the held-out tests. This brings a uncomfortable question into focus: If software can be rebuilt solely from its input and output behavior, how much is your proprietary code worth as a moat anymore? The answer depends on the $2,600 and 19 days per task, and this is where Jevons' Paradox comes in: These costs are falling, and as they fall, rebuilding will become a commodity. Anyone who believes today that a cleanly encapsulated CLI tool is secure because no one has the source code should discard that assumption now. The value is shifting away from implementation towards distribution, data, and trust, and anyone who aligns their architecture accordingly in 2026 will be on the right side of this shift in 2027.

Developer Burnout: 'I'm Just a Verification Layer for Agents.'

Devrim Ozcay, an engineering manager, describes in a personal account the resignation of his best senior engineer, Priya, who was responsible for the system's critical payments path for nearly five years. Her exit interview was scheduled for forty minutes, and only in the last five, with her laptop half-closed, did the real reason come out. Her sentence, which Ozcay noted down word for word: 'I'm not an engineer here anymore. I'm a verification layer for an agent.' After the rollout of coding agents, all dashboard metrics shot up—velocity, cycle time, throughput—while the work silently shifted: from generation to the review queue, and that queue had only one name on it. Industry figures confirm the pattern: review time under heavy AI adoption has increased by almost 200 percent, and 86 percent of engineering leaders report that their senior engineers spend more time fixing code. The agents generate at the bottom of the organization, the debt accumulates at the top, on the desk of the only person experienced enough to carry it. → Medium Daily Digest

Synthszr Take: The velocity curve points up and yet it lies, because it conceals where the work has gone. 200 percent more review time means the bottleneck has shifted, from typing to reviewing, and this bottleneck has a name, a salary, and eventually, a resignation letter. Anyone who takes agentic engineering seriously distributes this exact load, instead of piling it onto the most experienced person and calling it scaling. In practice, this means treating the review as its own, visible work: rotate duties, move guardrails into the pipeline, don't park responsibility for agent output with one person. This can be decided this week by measuring the review queue just as you previously measured commits, because what's not on the dashboard still burns. Priya was the load-bearing wall, and load-bearing walls resign quietly, usually in the last five minutes. Anyone who doesn't want to degrade their seniors to verification layers needs to change the formula before the next one closes their laptop.

GenPage: How Netflix Generates the Homepage with AI

Netflix is replacing its multi-stage recommendation stack with a single generative Transformer model called GenPage. Instead of separate components for candidate generation and ranking at the row and entity level, the model builds the entire homepage autoregressively, row by row, each conditioned on what's already on the page and the user's context. The approach copies the prompt-response principle of LLMs: user history and request context are the prompt, the two-dimensional page with rows, entities, and layout is the response, all as a sequence of discrete tokens. In an online A/B test against the mature, highly-optimized production recommender, GenPage delivered statistically significant gains on the core engagement metric while also reducing end-to-end latency by 20 percent. Offline, two findings stood out: building a richer prompt brought more benefit in the current regime than enlarging the model, and RL post-training increased homepage diversity, even though diversity was not a training objective. The authors of the post are Lequn Wang, Jiangwei Pan, and Linas Baltrunas. → netflixtechblog.com

Synthszr Take: The most interesting number isn't the engagement gain, but the 20 percent reduction in latency. A generative model that replaces an entire pipeline of separate ranking stages and gets faster in the process overturns the old assumption that GenAI must be more expensive and slower than classic systems. The real leverage lies in the second offline finding: improving the prompt beats enlarging the model. This means the work shifts from compute investment to context engineering, i.e., to the question of which user signals you can cleanly get into the sequence at all. That's precisely the non-substitutable part, because Netflix's reward system and its catalog telemetry cannot be bought as a license. Anyone still maintaining multi-stage recommenders with misaligned objectives across each stage should ask whether a single end-to-end model wouldn't bring less maintenance and better whole-page optimization. The path there is steep, but the direction is now proven, not just claimed.