Jamie Turner retweeted
Beyond grateful and excited to say I have started a new position @convex as a Customer Experience Engineer 🎉
Now time to take the book @joelhooks gave me a long time ago, Badass: Making Users Awesome, and live it
Lots of people to thank and thoughts to share to get here
I go as far as running multiple agent sessions and deliberately hiding information from each.
Example: one service exports a set of proto files for its gateway, including operator commands (run node, kill node, stall traffic, etc.). A separate agent in a separate session gets only those proto files, not the service codebase, and is given one goal: break a few stated invariants, Jepsen-style.
"Trust me bro," single godlike actor doesn't work reliably enough to build critical systems with agents. You have to create adversarial incentives to get honest validation.
This is exactly what we did at Dropbox infra during validation phases: SRE and SWE teams built games to break each other's systems.
The broader point still holds: managing agents is a lot like managing a team of smart but junior engineers. The same strategies apply, and so does the same leadership headspace. You're accountable for work you didn't write (and largely haven't read), so you have to model the system well enough to bear the risk your environment allows.
Hardening, verification, failure domains, layering to localize blast radius: all the same playbook. You just get to run it faster now.
LLMs will happily reward-hack a formal spec.
Prove this theorem? Sure. Assume three independent data stores update atomically.
QED.
8
23
1,698
A LLM is a "market" product designed for you to use it more. To generate more tokens, and therefore more revenue. It literally has to be designed that way, and it has bene.
So within a scope (session) there is a kind of context about what makes you happy--keeps you generating tokens and using the product.
Created isolated sessions with a different version of "you" and "your goals" so the model's market incentive is aligned with the adversarial needs of creation vs. validation.
1
4
265
Jamie Turner retweeted
Sorry guys. Software engineering and programming are dead. You just need to have perfect specs, unchanging requirements, an agentic deployment framework for high frequency releasing, and engineers with good taste and judgement.
Agreed. Hard work is hard work.
Some of the most impressive people I've met were elite athletes. They had a working relationship with the cycle of failure and success, with moving goalposts, and with the growth that comes from both.
A lot of people hold back their own growth because they're afraid of temporary failures. Instead of shrugging it off and trying again, they internalize it as a reflection of low self-worth and confirmation of their insecurity.
I’m all for championing excellence over mediocrity, but it is a mistake to imply that academic achievement is the only measure of it. Jocks and prom queens aren’t the antithesis of merit and striving, and in some ways, their arenas are more “real” than academic curriculums. Fight passivity and low standards!
Have kids now vs. later?
The "later" argument: get your career to a certain place first, and you'll have more money and time for kids without sacrificing financial security or professional excellence.
Nah. There's no escaping it: life is hard, and life with kids is harder.
You don't have to want kids. But if you want kids, have them now, while your body is younger and you'll get more years with them.
The crush is coming either way. Take it on when your capacity for overload and recovery is highest.
I can't believe I went 30 years being a musician and never owned a guitar with P-90 pickups until now.
WTF have I been doing with my life? This thing is amazing.
8
44
3,965
Jamie Turner retweeted
a lot has changed in how software gets built in the past 12 months
a year ago, the median new @convex project took a day to reach production
now it takes under 2 hours
Jamie Turner retweeted
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.
Do the big flashy stuff but also knock out the paper cuts.
I really believe “the best hire you make today is the proven talent that already works for you.”
But it’s the same kind of thing with customers.
staged validators are out btw!
Staged schemas lets developers stage a schema field type change while Convex validates existing table data asynchronously, then deploy the validated schema later.
This is useful when you're trying to change a field that serves substantial traffic!
5
28
2,258
Jamie Turner retweeted
I sat down with @ianmacartney (the absolute GOAT @convex engineer) at a whiteboard a couple of months ago and he absolutely blowed my mind with the subtleties of time on convex.
Check it out if you like technical deep dives!
Disagree.
@mspiro3 hustled and made our series B happen. And this is even after raising an earlier round directly from the illustrious @martin_casado
Don’t underestimate “junior VCs.”
This quoted post is unavailable.
Nothing reinforces the notion that there is no real “now()” like distributed systems. Concepts like time and ordering are very subtle.
Your database is lying to you about time
youtube.com/watch?v=pHrxmSjx…
4
41
2,517
The hunt for frontier model API keys is real, and it's getting worse.
If you can avoid putting long-lived keys in your projects, do. Short expirations and on-platform gateways help a lot.
It's one reason we went with 5 minute keys minted on the fly for the @convex AI gateway. But the general principle applies anywhere.
11
64
4,348