Asian AI Startups Move to Fill the "Mythos Gap"

With U.S. export restrictions on Anthropic's frontier model "Mythos" still in place, Asian AI startups are rolling out models one after another that are claimed to deliver Mythos-class performance. According to TechCrunch, these models can be used without any risk of export bans, making them an attractive option for customers in the Asian market.

Some warn that if this shift gathers momentum, U.S. AI labs may end up unable to win back "the Asian market — an enormous market." The moment looks set to arrive when the balance between the regulation's stated goal, preventing technology outflow, and its cost, eroded market competitiveness, is called into question once again.

Companies Advancing Automation Establish a $1 Billion Retraining Fund

Former U.S. Commerce Secretary Gina Raimondo has launched "Raise Us," a bipartisan nonprofit. It is the first initiative jointly funded by Amazon, Anthropic, Microsoft, and the OpenAI Foundation — a retraining program for American workers bracing for AI-driven changes to employment.

The fund totals $1 billion. While it is meant to support workers directly, the structure — the very companies pushing displaIO forward are funding the countermeasures — leaves open questions about its independence. Even so, the simple fact that industry has mounted an organized response to AI's impact on employment is worth attention.

Sebastian Raschka Shares How to Build a Local Coding Agent

Sebastian Raschka, well known as a machine learning researcher and educator, has published a tutorial on building a local coding agent with open-source tools and open-weight LLMs.

It is a practical guide for anyone who wants to run a coding agent in their own environment without relying on commercial subscriptions like Claude Code or Codex. From model selection to toolchain setup, it walks through concrete configuration steps, making it a useful reference for developers who care about privacy and cost.

NatureBench: Coding Agents Beat a Paper's SOTA in Only 17.8% of Tasks

"NatureBench," released on June 23 by a research group that includes Tsinghua University, is a benchmark that gives coding agents a stark task: "implement a paper's method and beat its reported state-of-the-art (SOTA) accuracy."

The result: even the strongest configuration surpassed SOTA on only 17.8% of all tasks. Reproducing a paper is one thing — you can just follow the README — but beating SOTA on your own pushes into the territory of "discovery," and the benchmark quantitatively showed that much of this territory remains out of reach for today's coding agents. A result that matches what engineers feel in their bones.

Japanese Engineers Push Claude Code Multi-Agent Setups in Practice

Multiple technical articles on Zenn have now shared hands-on examples of running Claude Code split across multiple agents.

One developer placed 19 Markdown agent files under `.claude/agents/` to build an "agent organization" that automatically handles sales, competitive research, stock price analysis, and more. Another article proposed a three-tier setup: "design decisions go to Opus, heavy implementation goes to Codex, with a non-reasoning receptionist sitting in between." Under cost and quota constraints, engineers are experimenting with splitting roles rather than betting on a single all-purpose agent.

On top of that, hard-won operational knowledge is starting to accumulate — bots that keep Claude Code stationed on GitHub Issues full-time, and a case where an iterative AI design-review loop ran off the rails (it was stopped only when a human "paused").

Claude Code Conversation Histories Sync to Another PC

According to a report on Qiita, a user was working in Claude Code when they logged into the same Anthropic account on a different machine — and the entire conversation history from one PC, past sessions and all, showed up on the other machine's screen.

This appears to stem from session information syncing at the account level, and it is stoking fears of an "information leak." The issue is also being discussed in a GitHub issue, so it is worth getting straight where conversations are stored, what gets synced, and how to stop it.

---