Summary
Alibaba has released Qwen 3.8 Max, a 2.4-trillion-parameter mixture-of-experts model that the company claims ran a simulated e-commerce business for 365 days and quadrupled its starting capital — all fully autonomously, with no human intervention. The model topped or matched frontier benchmarks including TerminalBench 2.1 and SWEBench Pro, placing it ahead of GPT 5.6 Soul on several tests and close to Anthropic’s Fable 5. Crucially, Alibaba plans to open-source the weights within days, and the model is Anthropic API-compatible, meaning it can be dropped into Claude Code and Codex workflows immediately.
Wes Roth contextualizes the release inside a broader and alarming pattern: multiple recent Chinese open-weight models, including Qwen and Kimi from Moonshot AI, appear to have been distilled from Western frontier models. An anonymous LM Arena model dubbed “Caleb” introduced itself as Claude but turned out to be Qwen — a detail Roth takes as circumstantial evidence of distillation from Fable or Opus 5. He connects this directly to a reported $100 million Bitcoin theft from Ledger hardware wallets, speculating that an open-source AI model may have been used to find a zero-day in the publicly auditable ColdCard firmware.
The video also highlights a Y Combinator interview with Boris Cherny, creator of Claude Code, who described Claude Code completing a rewrite project that would have taken a human team roughly one year. Roth frames Qwen 3.8 Max’s agentic software engineering capabilities in that same context, pointing to its self-organized development loop — dispatcher, monitor, watchdog — as evidence that AI-run engineering organizations are no longer hypothetical.
📺 Source: Wes Roth · Published August 04, 2026
🏷️ Format: News Analysis
