Technologist • Software engineer • Engineering leader • Product developer • Photographer • Swimmer • Japanophile • Lover of beautiful things.
Los Angeles, CA
Joined April 2007
- Tweets5.8K
- Following1.3K
- Followers695
- Likes6.6K
Lightroom Mobile isn’t optimized for the Duo yet, but it’s still great for quickly editing photos on the go.
100
If you follow this argument to its logical conclusion, everyone should be writing Haskell and helping accelerate its long-running effort to support dependent types.
hot take: you don't need a software factory, and i don't love the term. it's just marketing
all you need is good software engineering and investing your tokens in the right places. the more deterministic the better (and token efficient)
2
247
I think the Notion team is misreading their data. We use Notion an order of magnitude more than before because our agents are connected to it through MCP or custom tooling. But increased usage doesn't mean increased satisfaction.
We don't enjoy using Notion, and we'd switch as soon as a better alternative emerges. Notion is fundamentally broken in many ways, but no competing product offers the flexibility we need. I'm sure that will change.
I added support for decision models to Shikumi, including Jev and OpenAI's Decisions API.
You can now turn probabilistic evidence into typed Haskell programs, compose decisions, calibrate thresholds with ReAnchor, and retune against cached evidence without additional model calls.
2
490
Most developers wanted to skip architecture and get straight to coding. They hated event storming, mapping message flows between bounded contexts, and defining strong contracts.
That helps explain why monorepos felt safer, despite the herculean effort required to build and maintain the tooling. Developers didn't want to worry about breaking things in other repositories.
But that's the essence of software architecture: understanding boundaries, defining contracts, and managing dependencies. It's what enables an evolutionary architecture that can continuously adapt to the needs of the business.
1
4
273
Language shapes how we perceive reality, and English is missing an equivalent of the Japanese word 孤高, which describes someone who stands apart from others, not out of loneliness, but out of independence, principle, or an unwillingness to compromise their standards.
4
188
Use multirepo, never a monorepo. And, trunk-based development scales to thousands of developers.
The trunk-based development debate started by @thorstenball has clearly touched a nerve. So I thought, why don't we look at some numbers?
I built a little simulator with the help of our dear colleague Opus to compare four different integration strategies:
1. Trunk-based development (@thorstenball's horse)
2. Merge queue as the only CI gate (what I'm currently using)
3. PR-based, requiring a fresh base
4. An agentic merge queue, with built-in conflict resolution (dream option, can't buy it anywhere yet)
You can play with the simulator here: lpalmieri.com/simulations/co…
1
5
546
Not only do I build CLIs to cut token costs and dramatically speed up workflows, but I also design them for agents.
Every CLI has an agent assist command that loads the right context for the task and supports scoped skills that don't pollute the global namespace.
thought everyone had their agent build their own cli...apparently not.
Before: ~40k tokens and 6 file reads just to get oriented. 30s
After: one command, 2k tokens. 5s
1
5
534
The biggest reason I prefer codex over claude is that working with claude is exhausting. It asks a million inconsequential questions instead of just getting things done.
1
1
165
Trunk-based development is apparently the latest innovation in software engineering. I've been advocating for it for 15 years and have rarely used branches or PRs outside of open-source work.
3
160
I took this exact approach to backend development. I spent most of my time building the building blocks, then created a DSL that forces agents to compose them instead of writing arbitrary code. Finally, I built validation tools that catch violations and feed errors back to the agents until the output passes validation.
AI may well be able to write massive software projects “from scratch”, but that almost certainly isn’t the fastest / highest performance / highest security / most token efficient way to do things. Good building blocks matter.
In addition to RL’ing the agents on final task loss, we should try to get them to regularize the building blocks as well. What would the standard library look like that makes the agents most effective?
It is likely at the limit that it turns out to be some alien organization of code, but given the existing human-bias in training, I bet that in the near term it would be lovely code that old programmers would appreciate.
I suspect the same could be said for mathematics, but that is out of my lane.
Slop creates a massive tax on development velocity.
Today, I found Python scripts an agent had generated for testing. I rewrote it in Haskell and got a 1,000x performance improvement.
120
I published a Haskell client for TypeSafe's Jev and integrated it into Baikai, my provider-neutral Haskell AI library.
I'll be adding support for OpenAI's Decision API next, making it easy to switch between the two without changing your code.
github.com/shinzui/typesafe-…
github.com/shinzui/baikai
1
3
256