BIG TECHAWAYS

TODAY's BIG STORIES

​#1. Japan model stands shoulder-to-shoulder with Fable 5 and Mythos​

Sakana AI has released Fugu and Fugu Ultra, a multi-agent orchestration system packaged as a single model API. Developers call one OpenAI-compatible endpoint; behind it, Fugu can choose models, assign work, verify outputs, and synthesize a final answer across a team of models or agents. Sakana says Fugu Ultra stands shoulder-to-shoulder with Fable 5 and Mythos Preview across several engineering, scientific, and reasoning benchmarks.

The signal is that Sakana is turning orchestration into a control plane: an API layer that hides model choice, routing, verification, and synthesis. The claims still deserve a clear eye: Fugu Ultra uses a fixed agent pool, routing is not public, and some benchmark comparisons rely on provider-reported baselines. But that debate makes the story more important, not less. The orchestration layer is becoming where capability, vendor dependency, and operational control meet in one product.

​#2. China AI is back at the center of the race​

The Economist argues that China is having another AI moment, with a new model narrowing the perceived gap with the US. A public discussion from Shashank Joshi also points to an important split: public tests and private tests can tell very different stories about how far apart the two sides really are.

The China AI question is not just whether the country has “caught up.” The better read is that the gap is moving across different fronts. Public demos, private evaluations, open-model adoption, and market confidence can all send different signals. For builders and investors, Chinese models are becoming a strategic variable again.

​#3. Chinese AI stocks rally on policy and demand optimism ​

Bloomberg reports that Chinese AI-related stocks rose after investors reacted to a more supportive policy tone from Beijing and stronger expectations for AI demand. Zhipu, also known as Knowledge Atlas Technology, and MiniMax climbed at least 23% in Hong Kong; their year-to-date gains are roughly 2,000% and 260%, respectively. Chipmakers including SMIC and Yuanjie Semiconductor also moved higher.

The market is reacting to concrete policy mechanics: China wants to expand AI adoption in consumer markets, push next-generation AI devices, deepen AI integration across e-commerce, logistics, and retail, and smooth listing paths for AI companies. This is a capital rotation inside China tech. Older internet platforms are still under pressure from competition, weak consumption, and AI capex, while models, chips, infrastructure, and AI adoption beneficiaries are getting the premium.

SHIFT SIGNALS

THE SIGNALS BEHIND THE HEADLINES

• ​Alex Karp warns enterprise AI can sell feeling instead of solutions: Big Brain AI quotes Palantir CEO Alex Karp arguing that many enterprise AI vendors are selling customers a feeling, not a solution. The demo can impress, but buyers will increasingly ask which workflow actually improved.
• ​GLM-5.2 is getting Silicon Valley’s attention: Business Insider reports that GLM-5.2, an open-source Chinese AI model, is drawing attention in Silicon Valley for coding and agentic workflow performance. After days of GLM 5.2 chatter among builders, mainstream tech coverage suggests open models are moving beyond benchmark discussion.
• ​The Atlantic turns AI music training data into a searchable database: The Verge reports that The Atlantic made four AI music training datasets searchable, including two large datasets with millions of tracks. This matters for copyright because provenance is becoming something people can inspect, not just argue about.
• ​Who benefits if the Trump administration squeezes Anthropic: TechCrunch asks who benefits from the Trump administration’s intervention around Anthropic and export controls. If model access can be disrupted by policy, the advantage may shift toward local models, sovereign AI stacks, or players less dependent on a single frontier lab.
• ​A stronger Mythos successor is reportedly in training: Andrew Curran says a stronger Mythos successor has emerged after training and may be called Mythos 5.1 or Mythos 6. This remains a rumor, but it is worth tracking as a signal of how quickly frontier labs are iterating.

AI IN ACTION

Stop putting every instruction in one place

​Claude has a useful guide to steering Claude Code with CLAUDE.md, rules, skills, subagents, hooks, output styles, and appended system prompts.

The practical problem with coding agents is not always that the prompt is weak. Many teams create noise by putting everything in one place: coding style, required commands, safety policy, release workflows, output formatting, and tasks that should be delegated to a separate agent. In 30 minutes, you can build a cleaner map: which instructions should always be active, which should load only in context, and which deserve their own workflow.

Classify five instructions you use often:

1. If it is a baseline project convention, put it in CLAUDE.md.
2. If it only applies to one folder or file type, make it a scoped rule.
3. If it is a repeated multi-step process, turn it into a skill.
4. If it must run reliably every time, use a hook.
5. If it creates an independent workstream, give it to a subagent.
6. If it only changes presentation, use an output style.