Datadog had no right to have the best browser RUM around, and yet, they have it.

Every time I use the product I'm so happy with it. And then, it's integrated with everything else on the DD platform. Heaven.

I never, EVER accept LLM output without making it run code to verify what it did (tests, something more manual, whatever).

Neither you nor your LLM can run code in their head. Do not trust, always verify.

Add workers :auto by nateberkopec · Pull Request #3827 · puma/puma It's probably not that great that you MUST use the ENV var in order to get this neat behavior. I noticed that this rounds down, which means a cpu quota of 512 will potentially put you into ...

github.com/puma/puma/p...

Letting Puma auto-set your worker count is the easiest way to go for 90% of usecases.

Currently, you can only do that with WEB_CONCURRENCY=auto, but we'll also make this possible in the next puma version by using `workers :auto` in your puma.rb.

Currently you can set WEB_CONCURRENCY in env to "auto" and it will do this for you, in future versions you'll be able to use workers :auto in your puma.rb config

If you want to use physical core counts instead and pay AWS twice as much, be my guest, but just know that I've been using logical core count to set this for 10 years and it does not increase latency except from ~99% or greater cpu utilization (wowee!)

Aside: why logical and not physical cores?

In 2014 Intel and AWS decided to sell everyone logical cores as "vCPU". We bought it, and eventually in common use vCPU==CPU. Turns out, they were right, hyperthreading works. 10+ years of production use confirmed it.

If you are using Concurrent.physical_processor_count or Concurrent.processor_count to set your Puma/Unicorn worker counts, that is wrong.

Use Concurrent.available_processor_count. It takes into account cpu quotas in envs like k8s/docker.

Tired: Repo github stars
Wired: Repo contributor count

have you, too, been to so many conferences that you feel nostalgia for this?

Not _really_, my public Github is full of Claude-powered work though. My dotfiles contain my Claude.md which is extremely basic github.com/nateberkopec...

I understand the "it's fun to be competent" and "I can feel the skill draining out of my fingers" people, but like, did you ever really NEED to be an expert on the fucking Datadog Scorecard API?

He thinks only of himself. He thinks if he can get any peace over the line, no matter what the terms, it gets him a Nobel.

This kind of project would've been maybe ~5-8 hours without an LLM. The LLM allowed me to quickly paper over knowledge I didn't really have (Datadog's API, how scorecards work). I basically just provided high level arch and direction and let the rest auto-complete.

Example of the kind of thing I'm doing more of now in the age of LLMs:

A ~1000 line Ruby project for creating Datadog scorecards. A client had ~30+ microservices, but adherence to "platform" standards like "use jemalloc" was spotty. LLM did it in something in ~1hr.

IME you never trust, only verify. If a task is not verifiable, don't give it to an LLM.

Architecture astronauts create 1+ service layers, which are ~internal views. They end up having to serialize big objects before crossing layers, but only use tiny parts of the object on the other side. Since they didn't use lazy accessors, they waste tons of db queries/work.

>merged back

gl

go off king

FWIW I think what I was referring to on the podcast was more of a Turbo/Hotwire+, something which was a superset of the current API, but maintained 90% compatibility with upstream.

If you too have a Rails app and would like to see your page load times improve like this, you should try our retainer service 😉

AI Powered Private School | Alpha School Discover Alpha School's AI-powered 2 Hour Learning model. Empower your child to excel at academics while developing real world skills and passions.

So how do you prepare a programmer to deal with these edges when you can't really sell them words or content? Maybe it's time to return to cohort-based, in-person or synchronous learning of some kind? Take a page from kids education like Alpha (alpha.school/)?

3. Lack of observability. Without the right data and context to feed to the LLM, it doesn't create good answers.
4. Slop outputs. LLMs generate overly verbose and just plain bad solutions.
5. LLMs don't make (good) decisions, they generate a lot of options.

But in my own work, I notice there's still many rough edges around LLMs:

1. The users don't have enough taste and knowledge to verify the work being done, and ship no-ops or even regressions b/c the LLM said "it's optimized"
2. Unknown unknowns. Users don't know what to prompt.

The need for building skills isn't going away. It's more important than ever before. But the need for a "reference manual" of the style of the Complete Guide to Rails Performance from 10 years ago is clearly gone.

I'm thinking about the future of the content side of my business. I think where we're headed, intellectual property isn't coming with us.

In a world without copyright, how do I capture value while still teaching them the skill of performance? Building their taste?

"I will absolutely give my return, but I'm being audited now for two or three years, so I can't do it until the audit is finished"

I'm hopeful that the tools with do all this for us in the future at least. It doesn't look insurmountable.

4. GPG-signing all commits, with touch (meat finger based) confirmation. This means your agent might do a commit and you go back and amend it to sign it.
5. Running local agents on your "real" machine only in sandbox mode

After a deep dive on LLM agent security, here's things I think we should all be doing but aren't:

1. Running agents on remote containers only.
2. Doing internet research in a separate cleanroom env
3. Having LLMs read logs daily for signs of exfiltration/promptjacking