Dagens feed var én lang prøveflyvning af agent-æraen: dyrere modeller, smartere orchestration, mere lokal/åben modvægt — og en tydelig følelse af at tool-harnessen nu er selve slagmarken.
Den vigtigste tekniske læsning: Armin fandt et konkret eksempel på, at bedre modeller kan blive dårligere til bestemte tool-skemaer, når de er post-trænet hårdt på én dominerende harness. Ikke en “lol model dum”-ting — mere en advarsel om at tooling ikke er neutral infrastruktur.
I had some vibes that Opus 4.8 was performing worse than older ones for some of uses that are off distribution and now I have the receipts. Latest Opus/Sonnet are causing tool invocation failures on Pi's edit tool when older ones did not! I wrote about it. https://t.co/96Qd8zfdVc
Nyere Anthropic-modeller opfinder ekstra felter i Pi’s edit-tool under agentisk kontekst. Strict tool invocation hjælper, men pointen er større: RL på Claude Code-lignende miljøer kan gøre andre harnesses off-distribution. Dejligt ubehageligt.
Armin Ronacherhttps://lucumr.pocoo.org/2026/7/4/better-models-worse-tools/Fable-snakken handlede mindre om “magisk model” og mere om hvordan man bruger den uden at brænde sin pung af: arkitekt/judge på det dyre lag, billige modeller til tastaturarbejdet.
8 hours of work with Fable: - Implemented this game prototype from scratch. Fable built the demo with fluid movements, robust rollback netcode, and US/BR servers it deployed itself. Try it (links below). - Investigated and solved a paradox in Bend2. This led us to a new, elegant consistency proof based on QTT. Extremely hard work. It implemented the full solution and regression tests. - One-shot Bend2's compiler. This is a 26k-loc file that converts a a program to C/Metal, based on the new runtime I've finished recently. A few months ago, I was impressed Sonnet one shot a parser. Seems like we're one-shotting compilers now... With this, all the hard parts of Bend2 are completed. (Sorry for the repost, last video had leaks. I'm sleepy now)
Claude Fable 5 is best used as an orchestrator that does the heavy reasoning and delegates token intensive but simpler tasks to cheaper models. 'fable-advisor' is your free and open-source plugin that does exactly that. New in fable-advisor: > A directive to Fable to minimize its own token usage (works surprisingly well) > GPT-5.5 as an implementer sub-agent via the Codex CLI (uses your ChatGPT sub so you don't pay extra) How to use it: 1. claude plugin marketplace add DannyMac180/fable-advisor && claude plugin install fable-advisor 2. /model fable
Claude Code-plugin der lader Fable styre krav, decomposition og verification, mens Sonnet/Opus/Codex-lanes gør implementeringen. “The smartest model runs the show. Cheaper models do the typing.” Ja tak.
GitHubhttps://github.com/DannyMac180/fable-advisorCloud/local-agent workflowet bliver også mere konkret: ikke bare “min agent kører et sted”, men repo, diff, env, session og intent pakket og sendt til en anden maskine.
I think I have solved cloud agents for those with tailscale and more than 1 desktop. Parcels packages your work, transfers it over to a machine on your tailscale, and runs it in tmux. That way you can step away from screen. https://t.co/1CFVsE09x3
Sender repo, uncommitted changes, .env og live agent-session via Tailscale til en anden maskine, hvor agenten genoptager i tmux. Understøtter Claude Code, Codex, pi og Droid. Meget “sommerhusserver, men for tanker”.
GitHubhttps://github.com/0xSero/parcelsDen større strategiske vibe: AI sovereignty er ikke bare geopolitisk pynt. Det handler om hvem der ejer data, workflows, post-training og weights — og hvem der ender som permanent API-kunde.
Palantir’s AI sovereignty post makes the most sense with China in the background. China treats AI as strategic infrastructure. The West (especially USA) cannot answer that by turning its companies and governments into permanent API customers. (Which makes sense) You don’t win an AI race by renting every important layer: Data. Workflows. Post-training. Weights. That is the simple version of AI sovereignty: Own the parts of the stack where your own intelligence compounds, otherwise you have access, but not leverage. Or in other words: Access is useful. Control is leverage.
Og så de små praktiske signaler: Codex computer use bliver taget seriøst, og design-workflows bliver bedre når man lader imagegen lave et nyt visuelt mål i stedet for at bede kodeagenten gætte æstetik ud af tynd luft.
If you think codex sucks at design, try "use imagegen to re-imagine this design and implement that".
Meta-viben: når man hænger ud med LLM’er hele dagen, begynder sproget at lugte af LLM. “The shape of” er muligvis den nye kontorparfume.
Let me know if this is just me: Noticed someone I know who is very "AI-pilled" and uses agents 24/7 to... start to talk noticeably more like these LLMs write. Eg more heavily using adjectives like "geniune", frequently terms like "the shape of" and many more examples
Bonus-runden var kort: feedet havde rigeligt støj og signal selv. De stærkeste eksterne ting lå i samme spor: åben modeløkonomi, suverænitet og Anthropic/Claude Code-politik.
Mashable opsummerer DeepSeek V4 Pro/Flash som markant billigere end GPT-5.5 og Claude Opus 4.7, med agentic/coding som tydelig pitch. Kinesiske open models bliver ikke mindre relevante af at folk råber “sovereignty” højere.
Mashablehttps://mashable.com/article/deepseek-v4-preview-comparison-chatgpt-claude-geminiTechCrunch: Alibaba klassificerer Claude Code som high-risk efter Anthropic/China-loophole-balladen og styrer folk mod egen Qoder. Ban wars: corporate edition.
TechCrunchhttps://techcrunch.com/2026/07/04/alibaba-reportedly-bans-employees-from-using-claude-code/TechCrunchs opdaterede Mistral-overblik er mest interessant fordi det beskriver Mistral mindre som “europæisk ChatGPT” og mere som Palantir-ish enterprise/FDE + egne modeller + egen cloud-ambition.
TechCrunchhttps://techcrunch.com/2026/07/04/what-is-mistral-ai-everything-to-know-about-the-openai-competitor/