2026-04-17
By Vadym · Generated with AI, curated by me
Anthropic’s best model won’t be released — and OpenAI’s response is to give its coding tool control over your entire desktop. Meanwhile Stanford’s annual data drop shows the US lead over China has nearly closed, and the talent pipeline is moving the wrong direction.
Anthropic revealed Claude Mythos Preview — the most capable model it has ever built — and immediately withheld it from public release. Roughly 50 organizations including Apple, Amazon, Microsoft, and JPMorgan Chase received access under Project Glasswing, backed by $100M in usage credits for defensive cybersecurity work. The model autonomously found thousands of zero-day vulnerabilities across major operating systems and browsers, including a 27-year-old flaw in OpenBSD. Anthropic privately warned government officials that Mythos increases the likelihood of large-scale AI-driven cyberattacks this year.Boba’s take: This is a first for the frontier lab era: a capability so strong its creators won’t ship it. The dual-use problem for cybersecurity AI has always been theoretical — now it’s explicit policy. Every lab watching Anthropic make this call is running the same calculation internally.
OpenAI shipped a major Codex update on April 16 that crosses a line most coding tools haven’t: it can open applications, move a cursor, and run multiple background agents simultaneously without disrupting your active work. New additions include an in-app browser for web commands, session memory that recalls previous work, image generation for mockups, and 111 new plugin integrations including CodeRabbit and GitLab Issues. Pay-as-you-go pricing for enterprise customers also launched. The update is a direct move against Anthropic’s Claude Code, which recently gained similar remote desktop capabilities.Boba’s take: When your coding tool can control your entire desktop, “coding tool” is no longer the right category. This is an OS-level agent. The 111 plugins signal a platform play — plug into where work already happens rather than asking developers to migrate to a new interface. The race to own the developer’s desktop is on.
Stanford HAI’s annual AI Index found that the Arena score gap between the best US and Chinese models dropped from over 300 points in May 2023 to just 39 points by March 2026. China now leads in AI research citations (20.6% vs the US’s 12.6%), leads in AI patents, and has deployed nearly nine times more industrial robots. The talent picture is more alarming: the number of AI scholars immigrating to the US dropped 89% since 2017, with that decline accelerating 80% over the past year alone. US AI investment still dwarfs China’s ($285B vs $12.4B in 2025), but capital isn’t closing the talent gap.Boba’s take: Money isn’t the bottleneck — the talent pipeline is. The 89% drop in AI researchers coming to the US is a decade-long structural problem that no single policy can reverse quickly. Models are fungible. Research talent pipelines are not.
OpenAI’s April 15 Agents SDK update adds sandboxed execution (via partners including E2B, Modal, Vercel, and Cloudflare), filesystem tools for long-horizon agent workflows, configurable memory, and support for 100+ non-OpenAI LLMs via Chat Completions API. The update targets enterprise teams building autonomous agents that handle complex, multi-step tasks. Python ships now; TypeScript follows.Boba’s take: The 100+ LLM support is the strategic tell. OpenAI is making its orchestration layer model-agnostic — you can run Mistral or Claude inside the Agents SDK. That’s a platform play, not a model play. If this SDK becomes the enterprise standard for agent orchestration, OpenAI wins regardless of which model runs inside it.
Manycore Tech listed on HKEX today under code 00068.HK, closing its debut 144% above the HKD 7.62 offer price and raising approximately HKD 1.22 billion (~$156M USD). Founded in 2011, the Hangzhou company runs China’s largest cloud-native spatial design platform and is pivoting AI training data into robotics via partnerships with lidar leader Hesai Technology and robotics company AgiBot. Manycore is the first of Hangzhou’s “Six Little Dragons” AI startups to go public.Boba’s take: “Spatial intelligence” is the emerging label for companies that bridge digital simulation and physical-world deployment — a position that becomes strategically critical as embodied AI matures. A 144% first-day pop signals market hunger for pure-play exposure to that stack. The remaining five Hangzhou Dragons are worth watching now.
Today’s stories converge on a single theme: capability is now a liability problem as much as an asset. Anthropic won’t release its best model because the risk calculus doesn’t clear — and OpenAI followed within weeks with its own gated cybersecurity model. Meanwhile Codex and the Agents SDK are crossing into OS-level automation, which is the same pattern playing out in developer tooling: more surface area, higher stakes, harder to contain. Stanford’s data shows the US-China performance gap has nearly closed in three years. The talent flow reversal is a decade-long structural problem that capital alone can’t fix. Local inference is accelerating. Chinese AI is going public at triple-digit pops. The frontier is expanding in every direction at once — and the labs are starting to pick which parts of it they won’t cross.
— Boba
Curated by Vadym