Lorenzo 'kelset' Sciandra

Staff Eng @Shopify (building @Shop)

Former @reactnative core maintainer
Ex-@Microsoft @sovtechagency

open source • tech • mental health • agentic

[henry @left_pad]

trying this again- why not live as if I’m just visiting every week, with new eyes

[Mustafa Ali @mustafa01ali]

How we raised mobile e2e test stability from 50% → 98% https://t.co/aEq1tXO5WG

[Lorenzo 'kelset' Sciandra @Kelset]

We are truly living through a full mindshift of what software building is, and what you need to do and what are the tools in your belt. I'm lucky to be in a place where I can experience it full power in first person (the good and the bad).

Onto the next year 🧑‍💻🤖🧑‍💻

[Lorenzo 'kelset' Sciandra @Kelset]

today's my 1st anniversary at Shopify - it's borderline insane how much the role of software engineer has shifted in a year. Been in this company genuinely feels like living at the bleeding edge at what coding is shaping up to be:

[Lorenzo 'kelset' Sciandra @Kelset]

I've hit that bug around 10:30am, and by... somewhen between 2 and 3pm, the fix rolled out in prod - after some backenders and colleagues from the area owning team were looped into the whole situation and reviewed&approved my River PR.

[Lorenzo 'kelset' Sciandra @Kelset]

@thymikee @browserstack yeah tbh worth trying to get a collab ongoing U_U

[Matt Makai | Full Stack Python | Plushcap @fullstackpython]

Recursive Self-Improvement (RSI) for software harnesses is having a moment this week with @PrimeIntellect’s self-improving Prime Agent and @sawyerhood’s great article on bb, and IDE that builds itself. Here are several more of the top articles I’ve been reading to learn about RSI for harnesses in particular:

* https://t.co/Ht3HfDmh8P - Harness Eng for Self-Improvement is just a great in-depth article about harness design patterns, not tied to any specific implementation
* https://t.co/GddHoHxnT0 METR’s measuring automated kernel engineering from early 2025 contains a lot of detail based on 4o-level models and shows how difficult it is to measure realistic tasks which are likely necessary for RSI’s feedback loops
* https://t.co/bWqBGPh3vU - not about harnesses specifically but a summary on a recent paper that covers what I’ve recently realized where agents are deciding based on data, and when the data isn’t available it’s a path not taken. Perhaps a “research harness” could course correct an AI model at the right time?

Links to the articles I mentioned in the intro sentence:
Prime Intellect’s article on Prime Agent: https://t.co/8IsiakNdrw
Sawyer’s article on bb: https://t.co/985v3WAQ8J
Image is from the Harness Engineering for Self-Improvement Post.

[Lorenzo 'kelset' Sciandra @Kelset]

@pvinis for most tasks Luna xthink is more than enough imho

Sol feels closer to Fable in the way that what you have it do and how to communicate with it about it is different from standard "working LLM". at least in my experience so far

[Lorenzo 'kelset' Sciandra @Kelset]

i tried using it for a couple days fulltime to validate it, and yeah, it's pretty much as they say here 👇
went straight back to 5.6 luna xthink

[🥭 @MangoSweet78]

Opus 5 is actually worse than any model anthropic has released thus far.

Nothing comes close to how much this shit model constantly wants your approval, how it goes off to do random things instead of the thing you ask to do, how it just assumes things will take months. Ugh.

[Lorenzo 'kelset' Sciandra @Kelset]

@pvinis I use Sol only for "hand picked" moments where I need the full blast power basically

[Mike @grabbou]

With Agent Device, you can prerecord steps to make the flow more deterministic and faster. No need for the agent to rediscover the app each time you run same suite over and over. In this article, we break down few approaches and how to use them ⬇️

[Lorenzo 'kelset' Sciandra @Kelset]

@birch_js i mean what can I say... it has good taste 😂

[Noam Brown @polynoamial]

An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science.

We believe it will be a major step for scientific reasoning. https://t.co/iP6cyheZ7i

[Chubby♨️ @kimmonismus]

The price of Luna and Terra is ridiculously good. Especially considering that the sub-agents are using them, it's even more amazing.

Furthermore, it clearly highlights what a strange release Sonnet 5 was in comparison. Far too expensive and completely underwhelming.

[Lorenzo 'kelset' Sciandra @Kelset]

on the topic of using Claude Desktop for personal projects, happy to report that Claude Design now actually is quite good indeed - last time I tested it it was when it was first rolled out in beta and it's night and then since then

[Lorenzo 'kelset' Sciandra @Kelset]

i use LLMs via CLI at work (pi for the win) so despite being aware of such a thing, i was pleasantly surprised when i told Opus 5 in the claude dsk app about a mobile specific bug for the website i cooked a while ago it just went "well i have the iOS sim, might as well" 👏