alec

[alec @alec_gard]

I've been using HTML output a lot for ad-hoc reports and dashboards, so I built a tiny utility for sharing single files with anyone via GitHub gists: https://t.co/4NbCUgTGLL

[Thariq @trq212]

HTML is the new markdown.

I've stopped writing markdown files for almost everything and switched to using Claude Code to generate HTML for me. This is why.

[Jack Clark @jackclarkSF]

I've spent the past few weeks reading 100s of public data sources about AI development. I now believe that recursive self-improvement has a 60% chance of happening by the end of 2028. In other words, AI systems might soon be capable of building themselves.

[alec @alec_gard]

There are going to be lots of opportunities to speed up many workflows 20x, by converting the LLM driven workflows to good old scripts

[alec @alec_gard]

@Flynnjamm daft question - why is this harder than just handing the agent a fresh private key?

[alec @alec_gard]

might be able to use this as the shared mechanism for handing over between agents, rather than relying on hosted CI

[alec @alec_gard]

this is really neat, I hadn’t heard of git notes before.

going to play with this as a substrate for agents communicating their thought logs via namespaced notes

[Raymond Weitekamp @raw_works]

over the last 2 days, we've stumbled upon a really powerful coding agent interaction pattern: git notes as an underground information network.

git notes are ubiquitous (part of git) and "invisible" (github chose not to display them).

this presents a very interesting communication channel for agents, who can now include rich details and discussions about the code without cluttering up the "visible" layer of the repo.

"mycelium" is my tool to make these interactions easier.

mycelium is just git & bash - it works with any agent in any git repo.

link below. still wrapping my head around the consequences of this, and very curious to hear your thoughts!

p.s. - this is the foundation of some very cool tools i'm collaborating with @irl_danB on for @OpenProseVM

[alec @alec_gard]

@threepointone didn’t realise you were UK based! I’m in Scotland - hoping to meet more UK folks at AI Engineer Europe

[alec @alec_gard]

It’s insane that we can effectively have the distilled knowledge of the entire internet, running locally on your phone, available to anyone.

OSS models deserve more love ❤️

[alec @alec_gard]

Surprised there aren’t many options for local LLM use on iOS, but found @AliChherawalla’s Off Grid that is working pretty well!

https://t.co/5zUi0Tfvg6

[alec @alec_gard]

@mattpocockuk Any time I’m building tools for agents to find/retrieve info, I make sure the tools always also return the direct source for quick verification.

You start to see how often they are using incorrect second or third hand info, rather than the ground truth.

[alec @alec_gard]

@0xblacklight Strongly agree - I typically spend a few hours writing and reviewing all the interfaces and data flows before AI begins dev, which results in much better output

[alec @alec_gard]

@mitsuhiko GitHub must wish they had usage based pricing about now

[alec @alec_gard]

I’ve seen AI repeatedly rabbit hole on the confusingly named functions etc in existing code bases, which happens far less when it has all been built by AI as it all natively “makes sense” to the AI

[Kent C. Dodds 🏹 @kentcdodds]

Your existing code base is part of the prompt.

[alec @alec_gard]

@mattpocockuk Strongly agree - I’ve found this has always been true, even pre-AI.

Well-designed interfaces and types push most design decisions towards a positive outcome, and implementation is always easy to change later.

[alec @alec_gard]

I want to ‘tune’ agents environments to the work at hand - e.g. a code dev agent doesn’t need access to Prod logs, metrics etc, but a debug agent does.

Ideally this is handled when spinning up the agent - proactively figures out what access it will need and gets approval, then an environment is created that enforces that.

[alec @alec_gard]

@ankitxg @marlene_zw I think that works well for small changes, but gets trickier with larger work.

A lot of what human engs do today is making small decisions for choices not covered in the spec, so we need to find a way to make the agents decisions visible but not overwhelming.

[alec @alec_gard]

@marlene_zw @ankitxg Similar to:

[Ramp Labs @RampLabs]

https://t.co/26LQlUoVas

[alec @alec_gard]

@marlene_zw @ankitxg Yep, having TDD drive enforcing the spec makes a lot of sense.

Also much easier to have agents ‘pro-actively’ add test as issues arise, which can feed into future specs

[alec @alec_gard]

@ankitxg @marlene_zw I agree with moving away from reviewing the code, towards reviewing the spec.

In practice though, I suspect a spec with complete detail will be similar length/complexity to the code itself