Will Larson

A lot of folks want a map to Staff Engineer, and just as many Staff Engineers want to better understand their role. Towards that end, I'm working on a small project to collect folks' stories of reaching and operating as staff-plus engineers staffeng.com

@SierraPlatform @ImprintPayments 💙💙💙

The traditional way to make AI legible to bank auditors: decision trees. The problem: decision trees make terrible experiences. @Lethain and @ImprintPayments built an agent that reasons — then made that legible to regulators. Behind the build: sierra.ai/customers/impr…

Does X have a diff search index archive for really old stuff on eg s3 or smthg? So many question

My X search results are all from 2011 now, but it did reveal that someone shared one of my posts to @jeff_weinstein in 2011 which is very Small World

In the year 2081, everybody was finally equal—and not just before God and the law. They were equal in focus, equal in productivity, and above all, equal in their inability to finish anything. This equality was enforced by the relentless supervision of Claude.

@rcolidba @somi_ai Hi Rob!

@Lethain @somi_ai A virtual filesystem is close to the first thing I did when I was doing similar development work recently. Also, hi!

Wrote up a series of nine short posts on interesting problems we encountered while building an internal agent framework (virtual files, compaction, LLM vs code-driven workflows, etc) lethain.com/agents-series/

I get asked every now and then: "how do I become more influential, as a dev/eng leader, better known online?" Either for themselves, or to help put their company on the map. The single best thoughts on this come from @Lethain, from this article: lethain.com/tech-influence…

@zebriez I wrote about this topic a while ago. It’s particularly aggregating to see someone else remix your work into greater visibility, but I’ve slowly convinced myself I’m playing a different game than they are 😅 lethain.com/plagarism-idea…

a few years ago, someone completely ripped off my writing. never cited me, emailed me, or thanked me publicly or privately. this person has a sizeable substack following and is revered for his craft (makes podcast appearances, etc.). it makes me mad. but not mad enough to stop writing for free.

@tonkapark Yup I believe. At minimum a variety of learning contexts seem to improve retention over any one

@Lethain Paper and pencil I think still help kids and even myself remember better

Is fascinating watching our 5yo learn math from an iPad app, just shockingly quickly. But then also not be able to reproduce any of the concepts on paper 😅 Will keep iterating to figure out how to get concepts to transfer or if somehow the ideas aren’t actually landing.

@tonkapark It’s pretty interactive with variety of games but then eventually turns into typing numbers via keyboard. It also has a lot of “help me” tools which avoid frustration but… To your point, it’s a very diff interface, so I get why it’ll take some practice to make it transfer.

@Lethain I are they Speaking to the iPad or selecting multiple choice or even using an Apple Pencil? The mode of interaction is part of the learning. Also just time. I have 5 kids and they all showed progress at different paces regardless.

@AntarYaami Yeah for sure, he’s def not at that level yet but can see him getting there over time. Makes you appreciate how hard eg adding numbers is if you don’t know a bunch of double and little tricks etc

@Lethain yup saw the same behavior with out 6 year old. What has improved significantly is mental math. Like entire multiplication tables and being able to apply them in everyday situations. He's fascinated by compounding and just mindblown to learn about gogolplex

@isaacinthesky Freemium is ~shareware for SaaS

what happened to "shareware" why don't we refer to anything as "shareware" anymore?

@mlanett 💯💯💯

@Lethain What is the active version of these? "I messed up my code using X?"

The agentic passive voice: any sentence in the active voice, where the actor is a model, is actually in the passive voice. Example: Claude messed up my code. Example: Codex wrote some bad tests. The passive voice is unclear, please avoid it.

@PranjalDaga1 Congrats!

Learn more about Accend for Banking on our website: withaccend.com/banking

We’re launching @withaccend FOR BANKING today. Accend helps banks turn financial/tax documents into structured financial models and credit insights from hours to minutes. AI infrastructure built to modernize credit workflows in banks that have remained unchanged for decades.

Ok here we go!! StaffEng, AI edition, with @voidfiles and yours truly. Sign up for our listening tour and to nominate guests! (Link in thread)

@joshua_hornby @DanielNealAdler Ty!

@Lethain Hey Will, your recent post on sharing Claude transcripts made me think of github.com/dadlerj/tin by @DanielNealAdler Thought I’d share in case you found it useful?

@ThomasJ02 @tylercowen Yeah it’s a fun experiment! Interested to see if the pattern gets traction

This AI companion from @Lethain is the first time I've seen something like this besides @tylercowen's GOAT book. I think this will become the new norm for a lot of non-fiction.

@porestar I think generally yes, but presumably any given model with work better with a specific harness. Eg recent meme on X that Claude opus 4.5 effectiveness has degraded due to harness changes in Claude code (no idea if true, hasn’t really been my experience)

@Lethain „Ultimately, I think these are fairly thin wrappers, and that you’ll learn a lot more by implementing these yourself“ - do you think model performance will generalise to any harness?