Asked Claude to debug an easily repro-able bug; it spent 2h suggesting increasingly preposterous “root causes”.
Went old-fashioned, attached a debugger, and 5 mins later after some careful inspection of state, knew the problem, which was a far cry from anything Claude suggested.
Went out to touch grass, but alas if you live in or around downtown San Francisco, there’s no escaping any discussion of AI, GPUs, agents etc, in public places. 😒
Maybe will try Marin or Sausalito the next time.
At least the nosh (Gochujang Pimento Cheese Twist) was superb.
Pre AI, people wrote poor commit messages, but you could still read a bunch of them and generally know what people were trying to do.
AI writes commits so verbose I’m in the ���I ain’t reading all that. I’m happy for you tho, or sorry that happened” territory.
“More” ≠ “better”
If you believe “humans should deeply understand what they’re building with AI”, it follows that AI generated code should be optimized for “human cognition”
Yet, often AI written code is actively hostile to this goal, and worse, a ton of “best practices” encourage this hostility.
“Manual coding” might be dead, but a year or two from now we’re going to be left with people who deeply understand how things they built actually works, and those who simply don’t.
The latter are going to be low-key unemployable (unless they’re social media influencers).
It’s funny to think Elon made just a huge deal out of bots when he was trying to acquire Twitter.
Only for X to be overrun with AI bots on a scale never seen on ye olde Twitter.
You’d think xAI would be adept at pruning the bots on its own platform, but clearly not.
i am quite close to going back to an autocomplete-only AI coding style. dead serious. i'm not sure the ostensible speed of agent-first coding is worth the brainrot, the laziness and the loss of code and architecture comprehension
All the vibecoding startups: AI will write all the code and put most devs out of a job.
AI *infra* startups: record growth, our customers are vibecoding/AI startups who use our compute platform, databases etc. 😅
Even the vibecoding companies aren’t vibecoding their own infra.
So damn excited to share that @turbopuffer
is a season sponsor of The Pragmatic Engineer Podcast!
I first heard of them when talking to the Cursor team about how they scaled up their production systems last year.
They told me about this incredible DB that scaled with their massive load, while also reducing their bill by 95% called turbopuffer... where they (Cursor) was the first-ever customer (!!) It seemed almost too good to be true. So, of course, I had to meet cofounder @Sirupsen to figure out how they did it.
The BIG idea behind turbopuffer is that they built it on object storage - something that few to no startups did before turbopuffer, even though it's very clever. And added wicked smart optimizations, like smart caching on NVMe SSDs. They now have companies like Anthropic, Notion, Linear, Ramp and others using them. Oh, they just crossed $100M ARR... with less than $1M in venture capital raised (!!) All from Canada.
I'll bring more details on all the wicked cool stuff they do, now that I'll work a lot closer with them. Love the team. Puff puff!!
Many of the old skills I learned the hard way are not … relevant anymore (eg, manually git rebasing).
Agents are generally more methodical and faster than I was, and I used to think I was good.
In time, ‘git’ itself will become a niche skill of a bygone era of software dev. 😞
Bearish on companies that (citing AI driven “productivity”)
- refuses to hire juniors
- thinks the “engineering management” function can be obviated
- requires non-engineers to submit production code *on principle*
All this when existing ICs aren’t even properly vetting AI slop
@vaastav05 One a regression, which was quite subtle.
Another broke a critical functionality, which didn’t have end-to-end testing. Unit tests didn’t exercise codepaths that only activate given a specific matrix of inputs + configs.
Oh, a test harness was testing the wrong version of code
Might be anecdotal, but just this week multiple AI generated PRs with subtle bugs got merged that required several additional days and a lot of manual verification to fix.
“Velocity” isn’t going up when you consider how much effort was spent fixing things post-merge, pre-deploy.
end-to-end testing > unit tests, in the vibecoding era.
A massive, almost entirely agent-coded refactor passed all unit and pre-merge tests but broke a critical feature.
It was only caught due to my own excessive paranoia making me run end-to-end tests before the prod deploy.
@VicVijayakumar Unless one is genuinely interested in building crypto products or fully buys into the “mission”, don’t get why anyone would want to work for coinbase. 🤷♀️
For the past several years, they’ve been doing layoffs, rescinding offers, etc. there are far more stable “mid sided” orgs.
@jdub TIL.
What’s the whole point of a “Following” feed if that’s going to be polluted with bullshit.
And it’s not even popular content I’d care about, but just garbage. 🤮
Code generation becoming super fast and cheap is going to spell the slow death of libraries.
Old, established ones (like Python requests or Rust tokio) will be still be around, but it’s unlikely to see new ones (unless super niche) establish themselves.
Sure, you can put processes in place to build more cohesive things across orgs, but that will lead to a bunch of people whining about “bureaucracy” and how Google moves slowly.
There’s no winning here.
Ultimately, having individual DRIs act as BDFL’s might be the only solution.
It’s a “people problem”, the oldest problem in the industry, not just a Google problem.
The cost of generating code going to zero *incentivizes* teams to unblock themselves ASAP, rather than dealing with the friction of coming up with cohesive, holistic, “cross-team” solutions.