In case you needed more evidence that LLMs prefer writing typescript queries to SQL
Ben Dicken@BenjDicken
Mythos: finds 27 year old kernel bug Also Mythos: writes a bad SQL query
English
James Cowling
2.2K posts
James Cowling
@jamesacowling
@convex CTO. Dropbox storage designer. VR Revisited author. Motorcycle mechanic. Closet Australian. MIT systems researcher turned database shill.
San Francisco, CA Katılım Haziran 2016
470 Takip Edilen11.8K Takipçiler
You need a lesson in capitalism if you think AI will make it easier to run a tech company instead of increasing the things you have to produce to compete.
James Cowling@jamesacowling
I'm also skeptical about AI reducing demand for skilled engineers so far. For most companies without an overhiring problem AI codegen has led to an *increase* in engineering demand to keep up with the tech boom and all the new competitors on the market. If you're cutting staff instead of doing more things that's not a great sign.
English
@glcst @MFrancis107 @convex @tursodatabase I'm married to the game but @glcst and I maintain a complicated long-distance relationship.
English
@MFrancis107 @convex @tursodatabase I proposed to @jamesacowling last time we met but he shut me down. No ring.
English
I'm also skeptical about AI reducing demand for skilled engineers so far. For most companies without an overhiring problem AI codegen has led to an *increase* in engineering demand to keep up with the tech boom and all the new competitors on the market. If you're cutting staff instead of doing more things that's not a great sign.
Jean-Denis Greze 💡@jgreze
I'm actively hiring right now so I want to push back on something. When a big company lets a few hundred people go, the CEO doesn't get up and say "we over-hired in 2022 and made bad bets on headcount." That's the honest version. It also gets you fired by the board. What they say instead is: "AI makes us more efficient, so we're rightsizing." It sounds strategic. It's defensible. And it has just enough truth in it to be believable. The problem is we've started treating those press releases as ground truth about where the industry is headed - and then drawing conclusions about the profession from them. What I actually see on the ground in SF: most companies I know that are doing well are hiring A LOT. Because the demand for software isn't shrinking. If anything, we still have more software to build than we have people to build it well. Junior roles are changing shape, yes. The work is shifting. But that's a different thing than mass contraction of the field. Pay attention to what companies are doing, not what they're saying about why.
English
@existentializzm You live a blessed life if you don’t know how much VCs love wearing Patagonia vests
English
@jamesacowling Damn I just realized I’m ngmi because the other definition of vest don’t even occur to me 😭
English
PlanetScale already has strong brand recognition but in the future if the single-minded "consciousness" behind everyone's agents decides you don't exist, you don't exist.
Sam Lambert@samlambert
Claude told a user that PlanetScale had shut our service down. This is unsafe by any definition and Anthropic have made no effort to correct this situation.
English
@grhmc businesses where there's a tension between ethics and capitalism, because the optimal business strategy negatively affects end users.
Ethics is critical in infra too, e.g., protecting user data, but that's aligned with our business value already.
English
The best thing about having a B2B infra company is you just sell a service, people buy it with money, and the ethical quandaries mostly end there.
English
#1 advice for running an infra company is keep *everything* simple: systems, API, code, customer relationships. Complexity doesn’t scale and you’ll drown without simplicity your religion.
LLMs suck at designing simple infra. Design infra by hand or use something like Convex.
shashank@aloobhujiyan
@jamesacowling I am providing agentic infra for ops heavy medium/large businesses. What’s your advice ?
English
@zeeg If your bottleneck so far in life has been code production not strategy/design/architecture it’s going to be hard to understand why we haven’t yet seen a substantial productivity boost in companies with complex products
English
it is wild how many people think they have some trick and know how to make LLMs more effective than others
its either full grift (whats the point?) or they just dont know what good engineering looks like
"oh it doesnt work for you because you dont do XYZ"
English
@jamwt What are the other two? Cloud computing and the web?
English
Gen AI is the third biggest revolution in terms of new leverage/efficiency in the software industry during my career (mid 90s to now).
It’s funny this is a controversial take. The capital environment so badly needs it to be #1.
English
@jgreze They're just waiting for me to admit I'm @AvgDatabaseCEO
English
Do you want dedicated hardware, SLAs, SSO, compliance reports, fine grained access controls, unlimited deployments? Do you have a legal team who loves custom enterprise agreements?
GIF
Convex@convex
convex for enterprise is here. lawyers, rejoice news.convex.dev/enterprise-lau…
English
There’s been a lot of cloning going on lately but honestly if you haven’t put in the work to build a service/product/library you probably don’t have the experience to maintain it
David Cramer@zeeg
@nateberkopec must be great to not spend any money on building the software and reap all the rewards at some point folks will realize the 'find out' part of the story
English
@existentializzm Good thing that guy on the left is unrecognizable
English
There are people telling you quality and simplicity still matter and people telling you nothing matters so long as you're producing code. You get to decide who to listen to.
English
This video is the best articulation I've seen of what LLMs are excellent at, what they are comparatively much weaker at, and how this influences how humans need to design systems in order to make agents successful.
Jamie Turner@jamwt
LLMs are not perfect at writing software. In fact, they plain suck in some respects compared to strong human engineers. I broke down why in a video... and how Convex's design makes LLM-generated code actually reliable. The key: feedback, local reasoning, meta minimization.
English
It's not just infra being hard to scale. Once upon a time you could give something away for free and it'd only be consumed on human timescales. There's been a fundamental shift in resource consumption and we're yet to see business models catch up.
(thankfully Convex was designed to be efficiently multitenant)
James Cowling@jamesacowling
Y'all clowning on GitHub but the real lesson is that agents suck at scaling infra. Even the labs are struggling with their DBs. We're seeing the highest workloads in history but good architecture is currently still a human-bottlenecked activity. Make good infra choices.
English