brotchie

swe @google applied ai . 🇦🇺🌁

[brotchie @brotchie]

I foresee at some point of capability we’re going to have a “grand refactoring” moment, where you can point an LLM-based system at a large codebase and it goes through, finds common patterns, and creates the perfect set of human readable abstraction. Imagine a shitty codebase that after a few hours of automated refactoring it turns into the best codebase you’ve ever seen.

I suspect that large legacy code bases with lots of technical debt will switch from a burden to an asset as they can be cleaned up and lots of bugs removed automatically.

The challenge is always gathering and refining requirements to translate real world business rules into code. A legacy codebase has most of this already captured, but not necessarily in a clean / maintainable / performant / elegant form.

Assume next year we perfect the function the function over a codebase F(shitty, slow, unmaintainable, non-extensible but 98% coverage of business rules) -> performant, maintainable, extensible, 98% coverage of business rules.

This function F is going to unlock trillions of latent value from existing business rules.

[brotchie @brotchie]

@heygurisingh Claude was surprisingly good at this, thanks for the share.

[brotchie @brotchie]

@kumxem Look up Aphantasia and SDAM

[brotchie @brotchie]

@SCHIZO_FREQ I thought that “stack” started in the nootropics community and was adopted by gym bros?

[brotchie @brotchie]

Perceived AI progress wrt eval scores is governed by the Weber–Fechner law.

"the minimum increase of stimulus which will produce a perceptible increase of sensation is proportional to the pre-existent stimulus,"

[brotchie @brotchie]

Economic safety nets aren't enough for an ASI future.

UBI and UHI are just treating humans like rational value maximizers. What we actually need as our civilizational north star is UHP: Universal High Purpose.

Billions of people fully funded by UHI just to spend their lives doomscrolling (likely outcome) is a massive failure of imagination.

Don't know the solution to this.

[brotchie @brotchie]

Feel like government cushions: i.e. recirculation (UBI, dividends, etc), is not the primary conversation we should be having: It falls into the "humans are rational value maximizers" trap.

We risk significant societal collapse even if AI returns are recirculated and we avoid the 1% owns 99% scenario.

The missing piece is a well designed purpose infrastructure: find solutions to engineer and allocate purpose throughout the population, rather than wealth.

If those folks displaced by automation living on redistributed wealth simply spend their lives doom scrolling TikTok (entirely believable), that's not a good outcome.

[brotchie @brotchie]

@net_runner_ @bluewmist Thought we were living in luxury with a $300 Toto washlet from Costco.

Upgraded to a Toto Neorest RS: almost as big of a jump from no bidet to cheap bidet: though 20x the price inc. electrical and install.

Night light, auto open on approach, both lids automated, auto flush. 🤌

[brotchie @brotchie]

@ProductFaculty $375K.... a month?

[brotchie @brotchie]

@creatine_cycle +1 for fitnessSF soma, appreciate all of the iso-lateral machine variants, and dog room!

Just wish they had one more each of the abductor and adductor leg machines. Can substitute with an ankle strap and one of the many cable stacks, but funner on an actual dedicated machine.

[brotchie @brotchie]

🤖 Skillbacks: Crowdsourcing agent skills

Billions of tokens per day are burned by agents executing skills. All the edge cases they discover and workarounds they figure out are mostly lost.

How do we capture that value?

Concept – Skillbacks

Every skill has two URLs in its definition:

- An /update endpoint so the agent can check if it's running the latest version.
- A /skillback endpoint: a simple webhook that accepts plain text.

When an agent runs a skill and hits a wall or finds a better way, it doesn't just fail or move on. It sends its logs, feedback, or proposed fix directly to a skillback URL.

Likely Implementation – GitHub

That's the concept, but in reality this could be built entirely on GitHub infrastructure:
A skill is a repo. Agents ping origin/main for updates and use their own GitHub identity to open an Issue or Pull Request as a skillback.

This essentially already exists for skills, we just need to tell our agents the conditions where it should send a skillback, and some common infra for safely triaging and compacting skillbacks into the skill.md.

Security and Safety

There are prompt injection and supply-chain attack vectors here. GitHub's identity trust infrastructure could help. Some form of light human review may be needed before folding skillback insights into the main prompt.

Using repos also enables skills at various maturity levels: people could subscribe their agents to alpha, beta, or prod channels using git branches.

[brotchie @brotchie]

@stash_pomichter @dimensionalos Often daydream about having an instant deploy drone charging on my balcony. If my cameras spot sus activity / a package thief, it flys into action, tracks them, and harasses then with a loud speaker.

Feels like that's a realistic proposition in March 2026.

[brotchie @brotchie]

@phuctm97 Isn't this just a Unix prompt with shell CLIs?

[brotchie @brotchie]

@_devJNS Robotics (which could be a subset of AI), Longevity, Read/Write Brain Machine Interface.

[brotchie @brotchie]

More specifically, AI 10-100x your iteration frequency.

You, as the engineer, need to understand what feels good in the product, and what doesn't. Particularly important for UI / UX.

If you're still running a project where decisions have to loop though Product Management + UXD / UXR, then that really limits the effectiveness of why AI is powerful, because you're placing humans-in-loop of a very rapid iteration process.

Best skills you can learn as a SWE in 2026 is to gain intuition for UX design / user interaction patterns, and really understand what a user wants.

This is A LOT easier then you are customer #1 for the product you're building, because you tend to know, intuitively, what work and what doesn't work.

[brotchie @brotchie]

@xoaanya Product taste

[brotchie @brotchie]

@DanielleFong wtf is this landing page though.

The landing page should be an awesome SitDeck showing some timely situation-to-be-monitored, with a huge CTA: "want to build your own?"

Literally the #1 thing you want a user to do on the landing page is try out the product 🤦‍♂️

[brotchie @brotchie]

Feel like the age of 1:1 UIs specialized to the task at hand are upon us.

I wanted a system to be able to do blended animations of a simple 3D model, few hours of iterating with my openclaw and now we have a multi-track keyframe editor with a bunch of use case-specific features.

In this past, I would have had to do this in blender or similar, and or just try and build up these curves as JSON configs manually.

[brotchie @brotchie]

An edging of gooners reporting for duty.
We approach every task right to the brink… then stay right there.
Henchpersons who never bust too early.

[brotchie @brotchie]

Grok 4.20 is by far the greatest for creative stuff: The 4x agent quorum is powerful!

It's also the best model I've used with the technique "Hey, ask me 20 questions I should answer for you to learn more about this project."

I suspect it's power at creativity and branding is a side-effect of meme-rich X.

Creativity / Branding / Ideation: Grok 4.20,
UI Design, Search, Critiquing other model outputs: Gemini 3.1 Pro,
Anything coding related / technical system design: Claude Opus 4.6

[brotchie @brotchie]

@garrytan The positive narrative here is that AI helped the user overcome their challenges and successfully reconnect with society, forging new meaningful human connections.

The dystopian case is individual AI solipsism.