Asian AI startups launch Mythos-like models

Geopolitics, Export Bans, and “Sovereign AI”

• Discussion centers on Asian (Chinese and Japanese) efforts to build Mythos-level systems after U.S. export bans on top U.S. models.
• Many expect governments to heavily back domestic/“sovereign” models; they need only be “good enough,” not SOTA.
• Some predict the U.S. may ban “foreign” LLMs on safety grounds, similar to other trade restrictions. How this would be enforced in practice is debated.

Economic and Societal Risks of Advanced AI

• One side argues fears are overblown; disruption is normal technological progress that ultimately raises living standards.
• Others emphasize historical analogies: industrial automation allegedly took decades to benefit workers, increased inequality, and contributed to populism and democratic erosion.
• Specific concerns: mass layoffs without safety nets, power shifting to capital, widespread cyber‑abuse, cognitive atrophy, and dependence on tech controlled by a few firms.
• Long‑term catastrophic/“superintelligence” risks are debated; some see them as overblown, others as plausible and worth caution.

“Mythos-like” Claims and Benchmarks

• Many are skeptical of “Mythos-like” marketing given limited independent access to Mythos itself.
• Requests for third‑party benchmarks and standard leaderboards are common; UN‑style global benchmarking is suggested.
• Sakana’s Fugu Ultra is clarified as a multi‑model orchestration system rather than a single model. Some recall prior controversy over its claims.

Quality, UX, and Early User Reports

• Several users report Fugu/Fable consuming large quotas, being slow, and underperforming compared to Claude Opus on coding and web research.
• Others note all SOTA models are expensive when used via API; misconfiguration can worsen impressions.
• Some developers ignore benchmarks entirely and simply test models on real proprietary code; differences in practical value quickly emerge.

Market Structure and Business Viability

• One camp argues SOTA vendors have a narrow TAM (mainly developers), huge capital costs, and may have missed their IPO window.
• Others think everyday office and document work is a large, durable market, perhaps comparable to mobile phone service revenue.
• There is broad expectation that open models plus improving hardware will erode closed‑model moats, shifting the real battleground toward specialization and inference efficiency.

Talent, Politics, and Model Quality

• Multiple comments argue that training top models is genuinely hard and requires elite talent and execution, not just compute.
• Corporate politics, safety fine‑tuning, and controversial leaders are seen as factors that can degrade both model quality and talent retention.