01China's early AI adopters are the first to be told: train your replacement
The GitHub project is called Colleague Skill. Its pitch, per MIT Technology Review: a worker can distill a teammate's skills and personality, then replicate them as an AI agent. It surfaced in a country whose tech workers have spent two years embracing AI faster than almost anyone else. The same week, those workers started receiving an assignment no one wanted: sit down and train your own replacement.
Engineers at Chinese tech firms told MIT Technology Review their managers are now requiring a specific handover: document workflows, narrate judgment calls, and explain the reasoning behind past projects. The output feeds internal agents that will run without them. Some have stopped sleeping. Others have refused outright. At least one quit rather than complete the handover.
The soul-searching is concentrated among the people who were, until recently, the loudest evangelists. They were the workers copying Cursor workflows onto internal tools, forwarding Claude Code prompts in WeChat groups, running nights-and-weekends experiments with agent frameworks. These engineers had more context than their managers on what the models could actually do. That same context now tells them which parts of their job a well-prompted agent can already cover.
Refusing the request is not free. Declining to participate reads as refusing a direct ask from leadership, inside a performance culture where reviews still weigh visible cooperation. Cooperating hands a manager the date the role becomes redundant. The engineers described by MIT Technology Review weigh the two without a clean answer, and some are already updating CVs.
The contrast outside China is quieter by design. OpenAI announced this month that Hyatt is rolling out ChatGPT Enterprise across its global workforce on GPT-5.4 and Codex. The framing is productivity, not substitution. Hyatt employees are not told which of their tasks leadership expects the models to absorb. The training data gets collected anyway, through normal use.
The grievance from the Chinese engineers, in that light, is not that they are being replaced. It is that they have been asked, on record, to be the ones doing the replacing.
0244% of songs uploaded to Deezer daily are AI, and readers have built their own detector
Deezer disclosed on April 20 that 44% of songs uploaded to its platform each day are AI-generated, according to TechCrunch and Ars Technica. The figure applies to daily inflow, not the back catalog, meaning the share of AI music accumulating on the platform is rising, not static. Deezer said those tracks still account for only a small fraction of total plays, and that most streams they do attract come from fraud farms. The company demonetizes them before royalties pay out.
Two numbers sit side by side: 44% of uploads, a small fraction of streams. Deezer treats that asymmetry as fraud by default. It runs classifiers against incoming uploads, labels them at the door, then cross-checks any streams those tracks attract against bot-traffic signatures. Payouts are cut. AI-content filtering has moved from research paper into the revenue pipeline.
A parallel filter is forming on the reader side. Barron's and other outlets have flagged one sentence construction, "It's not just X, it's Y," as a near-guarantee of synthetic text. TechCrunch's summary notes the phrase has become so common in AI-generated writing that it no longer reads as a clue. It reads as a classifier that lives in the reader's head.
Two signals, same direction. Platforms are shipping technical filters; audiences are developing linguistic ones. Neither required regulation to arrive. Deezer got there because AI uploads threatened its royalty math. On the reader side, the filter accumulated because the same cadence kept appearing in op-eds, press releases, and Substack posts.
For anyone producing content, the default rule has shifted. A song without human provenance lands in a demonetized bucket before a listener hears it. A paragraph with the flagged cadence gets discounted before the argument lands. The classifier runs whether the creator consents or not.
Deezer has not published its false-positive rate, and has not named which model flags uploads. Writers face a quieter version of the same question: which phrases get added to the mental blacklist next.
03Pentagon Called Anthropic's Mythos a Supply-Chain Risk. NSA Is Running It Anyway.
The National Security Agency is deploying Anthropic's Mythos model, according to reports from Axios and TechCrunch published this weekend. It is the same model the Pentagon designated a supply-chain risk earlier this month and barred from defense contracts.
Neither agency has explained the contradiction publicly. The Pentagon's blacklist applied to Mythos-based products procured through standard defense acquisition channels. NSA operates under separate authority and buys classified tooling through different programs, which appears to be how the model entered Fort Meade despite the broader DoD restriction.
What makes the split awkward is the capability that drives both decisions. Ars Technica reported last week that researchers found Mythos capable of identifying exploitable vulnerabilities and chaining them into working attack sequences faster than human red teams. The same researchers warned defenders cannot patch at the speed Mythos can probe, raising concerns the model could shift the offense-defense balance against incumbents on legacy infrastructure.
That capability is precisely what intelligence agencies want. Signals intelligence and offensive cyber operations both benefit from a system that can reason through unfamiliar codebases and surface zero-days. It is also exactly what supply-chain reviewers flag: a model that accelerates attacks creates compounding risk if integrated into systems an adversary later compromises.
Anthropic has not commented on the NSA deployment. The company's relationship with the Trump administration has shown other signs of warming despite the Pentagon designation, with senior White House officials continuing to meet Anthropic leadership through April, according to TechCrunch. The Pentagon's procurement office and the intelligence community have not coordinated their public positions, and there is no sign they intend to.
The next budget cycle will determine whether the split holds. The Pentagon designation can be revisited; NSA's contracts are largely classified and harder to track from outside.
Google expands Gemini in Chrome to seven Asia-Pacific markets Google rolled out Gemini in Chrome to Australia, Indonesia, Japan, the Philippines, Singapore, South Korea, and Vietnam. The browser-embedded assistant had previously been limited to a narrower geographic footprint. techcrunch.com
Fermi loses CEO and CFO as Texas AI nuclear campus stalls Fermi's chief executive and chief financial officer both departed the AI-focused nuclear power startup co-founded by former Energy Secretary Rick Perry. The company has struggled to advance its Texas campus built to supply power to AI data centers. techcrunch.com
NVIDIA ties Adobe and WPP agents into enterprise marketing stacks NVIDIA announced expanded collaborations placing Adobe's agentic AI and WPP's production pipelines on its compute stack for creative and customer experience work. The integration targets brands running personalized content operations at volume. blogs.nvidia.com
Epic opens AI-character tool to Fortnite creators, bans romantic roleplay Epic Games gave Fortnite developers a "conversations" tool to build AI characters players can talk to, replacing authored dialogue trees. Epic's guidelines prohibit dating or romantic interactions with the generated characters. theverge.com
NVIDIA pushes AI manufacturing stack at Hannover Messe 2026 NVIDIA used Hannover Messe to showcase industrial AI deployments with partners targeting faster design cycles and skilled-labor shortages. The pitch frames AI adoption on factory floors as a pacing question rather than an open decision. blogs.nvidia.com
Chinese humanoid robot wins half-marathon against human runners A humanoid robot finished a half-marathon ahead of human competitors and set a category record, the latest signal of China's pace in commercial humanoid development. The race result follows a string of Chinese bipedal platforms hitting physical endurance benchmarks. arstechnica.com
Zuckerberg and Dorsey pitch AI as a tool to clone CEO oversight Wired reported that Mark Zuckerberg and Jack Dorsey are each building AI systems meant to extend their personal management presence across their companies. Both approaches frame AI as a mechanism for tighter executive control rather than employee augmentation. wired.com
Simon Willison ships Claude token counter with cross-model comparisons Simon Willison released an updated Claude token counter tool that now compares token counts across multiple models side by side. The utility helps developers estimate context costs before routing prompts to different providers. simonwillison.net
Researchers crash neural networks by flipping a handful of sign bits A paper introduces Deep Neural Lesion, a data-free method that disables production models by flipping a small number of parameter bits. The attack works across image classification, object detection, segmentation, and large reasoning models without training data or optimization. huggingface.co
Prego launches StoryCorps-branded dinner table recorder Pasta-sauce brand Prego partnered with nonprofit StoryCorps on a device that records family conversations around the dinner table for long-term archival. The product ships as a consumer tie-in rather than a subscription service. wired.com