←
Adam Kirk11,697 posts
Adam Kirk avatar

Adam Kirk

@atomkirk

CTO & Cofounder @ jump.ai. The #1 AI for financial services.

◷ Joined June 2009▧ 361 Photos and videos
Adam Kirk avatar
Adam Kirk@atomkirk·Nov 16

I wonder if @SlackHQ would let me come to their office, feed this image to their claude code, and finally fix this chat box behind the navigation bar bug that plagues me 50 times a day

Post attachment
Adam Kirk avatar
Adam Kirk@atomkirk·Nov 12

I’m AI maxxing and my biggest bottleneck is getting AI to write in a way my meat computer can easily digest tons of info I work with people who write so clearly I can read an essay from them about anything and it doesn’t take much brain power I can’t get AI to write like that

bRad avatar
⟳ Adam Kirk retweeted
bRad@fightonbigorG·Oct 23

Somebody is going to be in a world of shock when they get back to their car at the SLC Airport. Damn!

Post attachment
Adam Kirk avatar
Adam Kirk@atomkirk·Oct 18

Anyone who blocks agents is absolutely crazy. They at least see it as temporary while they ready themselves right? It’s inconceivable to me to be a tech exec and not see a tech wave and try to ride it. Those are where the growth opportunities are???

Wise@trikcodeOct 18

notice.. > Delta is blocking AI agents from booking flights > JPMorgan Chase and Charles Schwab block agents from accessing accounts > Yelp requires paid data licensing for agents > Amazon has blocked Grok bot and other shopping agents. the internet is starting to block AI agents that were supposed to do everything for us

Adam Kirk avatar
Adam Kirk@atomkirk·Oct 13

We need a description of what architecture is IMO it has something to do with thinking ahead, ask questions and talk to users. Even when I ask AI to do this, it does not do a great job yet (asking questions that think ahead).

I feel envious of all you people who work on simple and boring enough problems that you think AI is better than you at architecture I WISH it was better than me at making architecture decisions man but it’s just plain dumb at that stuff

jason avatar
⟳ Adam Kirk retweeted
jason@Jarsen·Oct 5

I only want to read posts about software factories from people who are not: 1) working at a frontier lab or in dev/ai tools 2) slop theater influencers 3) indie devs with no users or PMF the problem with software factory discourse is that it often ignores product. You ever sold to an Enterprise before? They will murder you if you ship a workflow change without 30 days notice. Change frequency (PR velocity) is like GDP, it is not alone a good measure of value or health.

Adam Kirk avatar
Adam Kirk@atomkirk·Oct 5

I watched both the original talk and the QA. I listened very carefully and have a few thoughts What I liked: ⁃The Michelin kitchen analogy is great ⁃Every team should dial up every verification method available (tests, types, lint rules, etc) to 11, to the point where its hard to write inconsistent code and it’d be very tedious for a human to work in the code. We have a team of 3-4 people at jump.ai furiously working on this too ⁃Think about the system and environment, not tasks. Sample review PRs to look for ways to iterate on the system. We are building a “software factory” at jump.ai, and a lot of this resonates. I like the vision of the future she paints, and I think we’ll get there. My problem is I don’t think we’re there yet and here’s why: ⁃They talk about building trust through empirical verification, but then mention some of the verification is non-deterministic agents checking things ⁃They talk about how they are still iterating on this and about all the problems in the codebase, but they are still shipping 1000s of PRs a month not looking at most of them? ⁃IMO a truly 100x engineer is prescient. They talk to product and even customers, think long term, and plan way ahead. They save the company tons of time and money by predicting and identifying one way doors. Ive never seen agents do this. They work towards what you prompted, they don’t dig deeper to plan ahead beyond this prompt. Thats a huge part of code quality and it can’t be empirically verified. This is the #1 reason I want my best engineers deeply involved. ⁃What about a PR that migrates 10m records in production (because it didn’t think ahead, see previous item). How is your agent empirically verifying that will work correctly on all that production data before merge and deploy? ⁃They talk about 10 chief of staff agents directing hundreds of sub agents all through the night, but they are doing things like rewriting the app in a different language for fun. Talking to customers and knowing what to build is the hardest part of product development. Thats my biggest question: can any one person keep 100s of agents busy doing actually valuable work? ⁃Does anyone know how to empirically verify a PR has no serious security vulnerabilities? My company is financial services… People think agents are at least as reliable as people. I actually think there’s an uncanny valley, where agents often act close enough to give you the ick. Like how confident they are that there are 3 “e”s in “seventeen”. This is all very impressive and forward thinking, but I lead a team of 80 engineers earnestly trying to do this safely and I think this level is a tad early.

Adam Kirk avatar
Adam Kirk@atomkirk·Oct 5

I watched both the original talk and the QA. I listened very carefully and have a few thoughts What I liked: ⁃The Michelin kitchen analogy is great ⁃Every team should dial up every verification method available (tests, types, lint rules, etc) to 11, to the point where its hard to write inconsistent code and it’d be very tedious for a human to work in the code. We have a team of 3-4 people at jump.ai furiously working on this too ⁃Think about the system and environment, not tasks. Sample review PRs to look for ways to iterate on the system. We are building a “software factory” at jump.ai, and a lot of this resonates. I like the vision of the future she paints, and I think we’ll get there. My problem is I don’t think we’re there yet and here’s why: ⁃They talk about building trust through empirical verification, but then mention some of the verification is non-deterministic agents checking things ⁃They talk about how they are still iterating on this and about all the problems in the codebase, but they are still shipping 1000s of PRs a month not looking at most of them? ⁃IMO a truly 100x engineer is prescient. They talk to product and even customers, think long term, and plan way ahead. They save the company tons of time and money by predicting and identifying one way doors. Ive never seen agents do this. They work towards what you prompted, they don’t dig deeper to plan ahead beyond this prompt. Thats a huge part of code quality and it can’t be empirically verified. This is the #1 reason I want my best engineers deeply involved. ⁃What about a PR that migrates 10m records in production (because it didn’t think ahead, see previous item). How is your agent empirically verifying that will work correctly on all that production data before merge and deploy? ⁃They talk about 10 chief of staff agents directing hundreds of sub agents all through the night, but they are doing things like rewriting the app in a different language for fun. Talking to customers and knowing what to build is the hardest part of product development. Thats my biggest question: can any one person keep 100s of agents busy doing actually valuable work? ⁃Does anyone know how to empirically verify a PR has no serious security vulnerabilities? My company is financial services… People think agents are at least as reliable as people. I actually think there’s an uncanny valley, where agents often act close enough to give you the ick. Like how confident they are that there are 3 “e”s in “seventeen”. This is all very impressive and forward thinking, but I lead a team of 80 engineers earnestly trying to do this safely and I think this level is a tad early.

lauren@potetoSep 21

here's how i shipped 2,500 PRs last month to production this was originally supposed to be for Cursor Compile in London. i couldn't make it since i was livestreaming for Grok @Bot Galaxy so i'm making it available for free here on X! watch it on 2x speed, i talk slowly

38:02
Adam Kirk avatar
Adam Kirk@atomkirk·Oct 5

There’s so much smoke, where’s the fire? People are running agents running agents, round the clock, generating millions of lines of code, shipped straight to prod. If you’re not seeing people do this on X, what is your evidence it’s happening at all?

Adam Kirk avatar
Adam Kirk@atomkirk·Oct 4

I try to be open minded, but yes, my question is “AI does stupid stuff 1-5% of the time. what kind of low stakes work are you doing where that doesn’t matter?”

SpaceXAI engineer, Lauren Tan: "I was a meat proxy between my agent and my browser. So I fired myself Now 10+ Chiefs of Staff run my agents 24/7 and I just check the work in the morning" In 1 hour, she shows exactly how she built that dark factory with GrokBot Skills → Verification → Loops → GrokBot → Dark Factory Most engineers still babysit one agent. She fired herself from that job This talk is worth more than most $1,500 agentic engineering courses Bookmark and watch it today Then read how to hire your first team of Grok Bots in the article below ↓

1:04:05
Adam Kirk avatar
Adam Kirk@atomkirk·Oct 4

I very much agree that Rust will not finally settle as the AI language because of slow compile times. Unless AI can make its compile times fast I think the language of choice will have - strong types/feedback - fast build/compiles - efficient runtime Im note sure how much ecosystem and community matters when AI easily rolls its own everything I’ve been playing with Ocaml and Elm. Maybe typescript on frontend is fine.

Adam Kirk avatar
Adam Kirk@atomkirk·Oct 4

I genuinely believe the only reason he was elected was because of his promise to balance the budget do something about the national debt.

BREAKING: Trump now says his $5,000 "Trump Dividend" would go to every U.S. citizen if Republicans win the midterms. On September 9, he promised it to "every adult citizen." With 316 million citizens, that is $1.58 trillion.

Adam Kirk avatar
Adam Kirk@atomkirk·Oct 3

On one hand im like “this guy wrote awesome books on building interpreters and compiles that were so amazing he surely is very experienced and wise” On the other hand I regularly think “anyone who thinks theres not a hard won gut intuition very experienced engineers use to keep code easy to change 2 years from now that AI does not yet have is naive”

If it works, if it's performant, if it has no bugs, if you understand how it works, if the agent can modify it a month from now, to your liking, is it slop?