Sakana AI (Japan) released Fugu Ultra, claiming Fable 5/Mythos level performance without the export control exposure.
Reality: not one model. A coordinator routes tasks across existing models, critiques, merges. Explains the coding wins and the slow agentic runs.
Same build: Fugu Ultra cost $0.51, GLM 5.2 cost $0.03.
Did win blindfolded chess outright vs three frontier models + a 2,100 ELO engine.
Closer to GLM 5.2 than Fable 5. Full numbers: Link in bio
Built something out of pure frustration: every good web-scraping tool for AI agents (Firecrawl, Exa) wants you on a metered plan.
Once your agent crawls at scale, the bill follows.
So I made agentfetch — free, open-source, self-hosted. Fetches pages, crawls sites, searches the web, returns clean markdown.
Works with LangChain, CrewAI, Claude MCP, and more.
It's brand new (0 stars 😅). Try it, tell me what's broken.
github.com/SID1ART/agentfetch
Most prompt guides stop at "be clear and specific."
This one doesn't.
Found a cheat sheet covering AUTOMAT/CO-STAR frameworks, structured output schemas, few-shot learning, chain-of-thought, RAG integration, and multi-prompt architecture — with actual examples, not just definitions.
The delimiter and formatting section is the most overlooked part. Breaking it all down in today's Jargonese newsletter.
AI writing has a documented fingerprint.
I researched it properly: Wikipedia's "Signs of AI Writing," LinkedIn posts, cold emails, YouTube scripts. Pulled every pattern and organized it into a Claude skill file with human alternatives side by side.
Tomorrow I'm sharing the full skill with Jargonese subscribers. Run it on your own drafts.
Subscribe before tomorrow
The US government just issued what appears to be the first-ever export control directive targeting LLM access — forcing Anthropic to pull Fable 5 and Mythos 5 globally within 72 hours of launch.
Stated reason: a claimed jailbreak. Anthropic reviewed it, called it minor, complied anyway — but publicly pushed back on the precedent.
The question this raises isn't just about Fable 5. It's about whether governments can now recall commercial AI models on security grounds.
Apple shipped Claude Code config files inside a live App Store update. A "Juno AI" system, tri-role chat architecture, full production internals — all accidentally public.
The hotfix came in 24 hours. The silence was the confession.
Add CLAUDE.md to your build exclusions. Not just .gitignore.
Google Search at I/O just went from pull to push. 🔍
Two updates worth knowing:
The Search box now helps you formulate your query — AI suggestions, multi-modal input (text + image + file)
Search agents that monitor topics 24/7 and notify you when something happens — with a link to act
Search isn't a search bar anymore. It's a background process.
Spent weeks trying to get an AI agent running 24/7 without it dying mid-task.
Tried local hosting.
Tried 6+ cloud providers. Tried every free LLM tier available.
Most broke in ways I didn't expect.
Finally have a setup that actually works — and I documented the whole mess.
Dropping it in tomorrow's newsletter. 🔗 jargonese.beehiiv.com
The "AI Job Apocalypse" narrative is hitting a pivot point.
Leaders at OpenAI and Anthropic are moving away from the "mass unemployment" predictions toward a model of "productivity multiplication."
Key takeaway: Cheap competence leads to massive output.
Google just redesigned all 13 Workspace icons. Clean. Unified. Nice.
But the real move this week? NotebookLM + Antigravity MCP = $200/month of AI workflows, free.
Most people find this 6 months late.
Get it this Sunday → jargonese.beehiiv.com/subscribe
