Qwen 3.8 Max arrives, open-weight letters split, and math still beats writing

The short read

Scope: This edition covers original posts published from Aug 2, 18:00 through Aug 3, 18:00. The X account connector is not linked, so the source pool is the configured public AI and tech accounts rather than a personal following list.

Model releases

Qwen 3.8 Max arrived beside MiniMax-H3

• What happened: Simon Willison noted that Qwen 3.8 Max and MiniMax-H3 appeared within hours of each other; Ethan Mollick then tried Qwen 3.8 Max on a shader test. 12
• Why it matters: Two short posts capture a live model-comparison moment, but neither is a release note or a controlled benchmark; the useful fact here is the timing, not a universal ranking. 1
• Early read: Mollick calls Qwen 3.8 Max solid in his experiments so far, while placing it below Kimi K3 on that shader test; the task setup and scores are not disclosed. 2

Society and ethics

The open-weight argument now has three letters in the same file

• What happened: Simon Willison summarized a July 24 Microsoft-led letter supporting open-weight models, Anthropic's response three days later, and the July 28 "Pacing the Frontier" letter signed by 1,324 frontier-AI employees. 345
• Why it matters: The first letter argues that open weights improve competition, auditability, and defensive research; Anthropic says it has never advocated a ban, but calls for a crackdown on industrial-scale distillation. 36
• Implication: The disagreement is no longer a simple open-versus-closed split: the concrete fault lines are concentration, model distillation, and whether frontier progress should be deliberately paced. 3

Business and enterprise

AI's next research problems may be questions, not benchmarks

• What happened: Ethan Mollick listed entrepreneurship questions AI might eventually address empirically, including what causes startup success, whether exceptional growth is predictable, and which interventions can move a place toward a stronger entrepreneurial ecosystem. 7
• Why it matters: These are messy causal questions about people, firms, and places—not fixed-answer exercises—so progress would require better observation and intervention design as well as better models. 7
• Limit: Mollick presents them as open problems that could become empirically tractable if AI improves enough; the post contains no result, prediction, or evidence that the problems are solved. 7

Research

Math has a cleaner training target than writing

• What happened: Paul Graham says LLMs have become good at math because math has clear right and wrong answers, which makes it easier to train on than writing. 8
• Why it matters: His explanation points to the feedback loop rather than the subject's surface difficulty: a system can score a mathematical answer more consistently than it can score whether a paragraph is insightful, well-pitched, or original. 8
• What it does not show: "They are coming for me next" is Graham's prediction and joke, not evidence of a current writing breakthrough; the gap remains a measurement problem as much as a generation problem. 8