Funnily enough, AI models are basically children in their training phase, but can't seem to be so anymore at inference stage. I'm not remotely good enough at anything deep learning to know how easy that's solvable, but I feel like it's more of a bottleneck than we can imagine.

First, Opus would probably still be burning tokens in loops without me. I would still be writing backup, rollbacks and tooling without it. Also feels LLMs have an extremly "adult" way of "thinking"—answering?—, where I can turn on the six years old in me any time 😅

I've been sent on investigating memory issues on an app I did know nothing about today. I was a bit afraid Opus 5.5 would just one shot it and feel like shitty meat proxy. Truth is, Opus did a lot of thing, but my naïve questions, reframing and "why nots" etc. that unlocked everything.

I’ve spent countless sessions brainstorming data attribution in LLM inference (over atproto? 😅). It’s way out of my league, but I hope someone can trace an output’s debt to a specific person’s input—so we still have an incentive to create good things. These stories are sad.

Will try it then, 🙏 I was a bit afraid of parsing every bsky post, but why not after all.

In my case, at a scripting level for non devs, it seems the gains are reals. For whatever requires software eng. or product, the benifits are 0 when it just surfaces all the other bottlenecks, negative when everybody starts throwing vibed apps at the wall and see what sticks.

What I really need today: a CLI definitely, good actions/pipelines, stacked PR/merge queue (possibly with JJ 😇). Review wise, I feel like we're leaning toward "review IDE" anyways, so I don't mind it too much. $1M question: should we embed chat sessions/prompts? 😬

Gitlab at work, because no choices. Mostly good honestly, but cluttered. Way too many things are in "parameters" or "CI/CD" under nested menus. Review interface isn't as good as GH, but I review locally anyways. Tangled because atmosphere account (solo use only as of today).

They even answer me in French sometimes when my English gets too bad!

😢 Good luck

Been on the default from Matt pocock's skills. Flat files in a local .scratch, /issues under each PRD, each PRD is a folder.With a few status (needs refinement, ready for human, ready for agent). It's more than I seem to need at this point 😅

Il y a un bonhomme en érection à Lascaux 😬Mais une (des ?) vulve à Chauvet…

Nope. I liked the "own your data, choose your algorithm part". Dug in a little bit and the common account really clicked. Plus it's fun to try to implement appViews 😅

Therens something intersting on this 4th tier. They are trying to cut the transaction cost of hiring and making a good spec-which has been killing them for a couple decade. Problem is, it just circles back to poorly defined software... 🤷

Ha yes, the software plumber. I think I am talking of a potential fourth tier even, people whose main job is not making software. The lack of education there is atrocious. But FWIW, I agree, bootcamps are dead.

Funny or not, I started this one thinking it would an optimistic one. Turned out the other way 🤷