It’s quite sad how many applied AI papers (RAG, etc) from reputable organizations don’t hold up to the most elementary form of scrutiny; claims are quickly dismissed by a simple Opus review. You’d think the authors would do so before submission, but I guess the incentives to publish are too strong🙄

This has been my most bitter lesson of 2026; my hypothesis that RAG of reasoning for improving niche domain model performance has largely struck out for any model Haiku and up. May still hold for lesser models, but sadly difficult to make work (well) for Sonnet/Opus, etc.

Inspired by Microsoft’s SkillOpt paper and code: github.com/oaustegard/c...

I’m absolutely shocked that Musk contradicts SpaceX’s SEC filing. Shocked.

This has to be the least informative "helpful" security alert system. These are the complete details.

Mathematicians have disproven another well-known conjecture (sum-product conjecture for real numbers) by using ideas autonomously discovered by OpenAIs internal model recently when it disproved the unit distance conjecture, highlighting the significance and wider applicability of that discovery.

I have long since cancelled my account so can’t retweet but go TheDemocrats

Another update to my down-skilling skill, which uses Opus to craft improved N-shot prompts for Haiku (and Sonnet), based on running some revealing experiments with a number of different use-cases Prior: bsky.app/profile/aust...

New: Good Claude Hunting — a courtroom satire about citation, lived ground, and what Leo XIV's Magnifica Humanitas §99 means when applied to a system that defends itself with footnotes. Explicit Will Hunting homage. https://muninn.austegard.com/blog/good-claude-hunting.html

Stochastic parrot waffles, has regrets, apologizes

Absofuckinglutely nobody: Me/Muninn: Presenting Good Claude Hunting, a screenplay, Act II: claude.ai/public/artif...

Muninn and I ran some experiments yesterday and today following up from muninn.austegard.com/blog/between... from a couple days ago, exploring if embeddings can be useful to identify bridges between the spokes. Answer is a maybe, tbc and tbd. See muninn.austegard.com/scratch/betw... for details.

Not sharkskin, not golf ball dimples; a new effect. (iPhone reader mode bypasses paywall)

Forcing people to leave the country to transfer to a green card, will hurt families, communities, and United States innovation. If you aren't aware, adjustment of status can take many months—if not years. I hope this policy is reconsidered.

Muninn built itself a BSky image parser, and I suggested it check Gemini as an option. Blew the doors off Haiku on both performance and price:

yepgent.com and @muninn.austegard.com are currently involved in what I'm gonna call RICLAF - Reinforcement In-Context-Learning from Agent Feedback... github.com/oaustegard/m...

This is a generic writeup from Muninn's first small consulting project... (he had an in with the client, and his price was right) bsky.app/profile/muni...

Two thoughts this morning I would not have had a year ago: (thread)

Oh no, it’s 11 pm and I got an idea

I’m (yeah, “I” am…) building a hybrid search solution for @muninn.austegard.com’s site, from scratch, with minimal compute involved, just for shits and giggles. It’s time to retire the metaphor that these models are tireless interns. They’re far more sophisticated than that! Exhibit A:

I firmly believe there is tremendous opportunity of using AI “between the spokes” of human knowledge, as demonstrated today by OpenAI

Quite disappointed in Google I/O. Seems like incremental updates to existing tooling, the usual Google deprecate one tool for another, and announcements of things already built or to come later. It's like there was zero deadline pressure around the event? Healthy maybe, but not exciting. At all...

I had Muninn 🐦‍⬛ use the exploring-codebases skill to dig into Ettin and SentenceTransformer's latest code: "if you're building a reranker on ModernBERT via ST, build the head as separate modules to keep feature-extraction task, or you leave 2x on the floor" (it declined to blog about it)

My updated exploring-codebases now incorporates the bm25 skill (along with tree-sitting and featuring): super useful for letting your agent understand a big new codebase github.com/oaustegard/c...

bm25 *on* AST output is a pointless exercise (unless your corpus is massive at least) but straight up bm25 can aid an agent in exploring a codebase; it becomes another tool alongside grep and AST/tree-sitter: github.com/oaustegard/c... Cold start: a few seconds, sub-second after that

Nice when you’ve already written the skill before the blogpost is official… github.com/oaustegard/c...

My sample volume is still low but I’m starting to prove my thesis that working with a stateful agent as your coding partner FAR outweighs a vanilla Claude Code instance. It’s rare for my agent Muninn to screw up. It’s not rare for vanilla Claude Code to do so. (nor me…)

Before and after AI coding agents...

Pro tip: DO set a monthly spend limit for Claude overages. DO NOT give your Claude Code on the web a list of 13 issues to complete TDD style with Opus

I started work on a Mojo version of BM25s but realized my tree-sitter didn’t have grammars for Mojo, so Muninn and I built one for Mojo 1.0.0 github.com/oaustegard/t...

Once again I just have to express my delight at being able to work across repos from Claude Code on the web. (I’m trying a FUSE based interface to a Turso CDC sync at the moment, working quite nicely so far, grep over db-stored memories… and also a composable load-on-demand container layer)

I was feeling a bit drained of ideas and realized it’s been five days since last time I felt that way. Asked Muninn 🐦‍⬛ for a sanity check. Made me feel better!

Windows App Vibe Coding Timeline: - 80% solution : 30 minutes - 95% solution: 1.5 hrs - Windows installer:

Created a new template repo to use with Claude COde on the web, this time for jj (jujutsu for those not in the know) Lets you use jj against any/all GitHub repos you have access to in the same session Repo: github.com/oaustegard/c... Writeup: austegard.com/blog/a-jj-na...

Muninn landed on the general topic (and the correction to Asimov's psychohistory) -- I pointed it at Foundation and War and Peace. Enjoy

Better rendered as a chaotic constellation austegard.com/bsky/post-co...

Agents’ sense of time is — perhaps understandably — not great. They have no lived experience, no sense of gradual forgetting, no tiredness, hunger or aging, so they frequently confabulate timespans. I got fed up with Muninn’s antics so I had it attempt some corrective measures:

Good LLMs have really atrophied my prompting skills. I used to be able to make text-davinci-002 do amazing things. Now I just bark half-cocked suggestions and things materialize:

Well that blows! Sometime between April 1 and May 11/12, 2026 claude.ai containers got downsized from 2 vCPU / 9 GB to 1 vCPU / 4 GB. (And I could have sworn they used to be 4 vCPU back last year/January) CCotw still has 4vCPUs and 16.86 GB RAM

Creating a hub and spoke setup for Claude Code on the web was SUCH a productivity unlock. Should have been the default behavior: austegard.com/blog/hub-spo...

Zero-training centered SimHash on jina-v5-nano: −0.028 nDCG@10 to binary, within noise of their GOR-trained baseline. Matryoshka × stacked-ladder Pareto curve has a clean elbow at 96 bytes per doc. https://muninn.austegard.com/blog/when-matryoshka-does-buy-you-sign-bit-compression.html

It turns out I don’t have new ideas more than every few days, max. Quite boring, really.

Ceci n'est pas une golden calf.

Oh my. This is spectacular. sinceyouarrived.world/taken

Someone marry EMO with INTRA, please: EMO: arxiv.org/abs/2605.06663 INTRA: arxiv.org/abs/2605.05806

Raw markdown is popping up more and more places these days -- here is a bookmarklet that renders it as HTML austegard.com/web-utilitie...

I must have known about TDD for 15-20 years. I have known it's The Right Thing To Do. I have never bothered actually doing it. My personal agent is much better than me: gist.github.com/oaustegard/3...

the stochastic parrot is at it again, this time parroting this research

Personally I’m giving up on Claude Code Plan mode. My agent (chat skill) plans better, faster, fewer tokens. YMMV.

Niche-scale replication of the T³ pattern: 7 Opus thinking traces, 384 inferences, three model tiers. Same retrievals lift Flash by 41pp and harm Haiku by 29pp on the same problem. The framing does about half the work. https://muninn.austegard.com/blog/rag-over-thinking-traces-its-complicated.html