CBLT

Not by name, sure, but it does mention that you're going to be reading the same context into the second model. It specifically refutes your point that you're only reading the smaller plan:

Opus reads base.py, signing.py, the test file (twenty cards of gray), then writes its plan and leaves. And what's the first thing Flash does with that beautiful document? It re-reads base.py and the test file, because a plan is not a file and you cannot edit prose. The gray reads just keep stacking, first at Opus prices, then again at Flash prices. There is no version of this where a second reader is the cost optimization.

Memory is a harness feature, not a model feature.

The author is probably only evaluating their own harness with different models, so Claude-Code-specifics like memory aren't part of their evaluation.

Linked your own project with an "All rights reserved" license? The only thing my company will allow me to do with that software is have AI steal it </s>

I'm not really sure memcached optimizes for operational simplicity. The only time I've run it at scale, it would have low cpu utilization then unexpectedly hit lock contention and fall over without warning.

I also like having long, pointed conversations with LLMs as I review code. Then when I'm done, it's different code, and it has all of my blind spots and knowledge gaps, so I cannot effectively review it anymore.

It's like turning a code review that requests you, into a code review that requests someone else. And it tramples on the original author quite a bit too. It's hard only having the ability to add incremental value to large amounts of code, instead of large amounts of value to incremental code.

Imagine a big codebase with this somewhere in it:

    fn processThing(thing: Thing) void {
       // this function must always be invoked on
       // a thing that has already been started
       assert(thing.is_started);
   
       // ...
    }

Eh, it's worse than that. The GP comment is repeating a joke derived from an Onion headline about gun control. Where the very poignant message is about political will to make change. However, the npm ecosystem is very much willing and has already made several changes. If we're going to engage in discussion instead of meme-posting, the GP should have (imo) included real commentary _in addition to_ the meme they really wanted to post. What is the policy they want? Why do they see the NPM ecosystem as still resistant to change?

I've been oncall for a different G service that nearly paged on every error. It used the standard error budget tooling, but on hundreds of user buckets because the engineering around locality-specific configuration was... suspect. Many of these buckets had single-digits user. If a user was on their phone and lost signal, I was paged. Very poor oncall experience.

idiomatic, readable C++ code

I disagree with this characterization of the generated primes.cpp shown here.

People are quick to point out that induced demand exists - especially people that aren't fond of change.

Very broadly speaking, people mis-estimate effect sizes in economics by orders of magnitude. Induced demand is just their foothold to claim an effect exists, before they go about claiming the effect size they want to see.

I use a PR notifier chrome extension, so I have a badge on the toolbar whenever a PR is waiting on me. I get to them in typically <2 minutes during work hours because I tab over to chrome whenever AI is thinking. Sometimes I even get to browse HN if not enough PRs are coming and not too many parallel work sessions.

I've had great experiences being managed twice by very humble engineers who've made the transition to EM. Both were sacked within the year by their boss because they didn't play the corporate politics game.

I can <C-z> while less is running to background that process using the shell, so the shell is clearly not completely gone.

I might be misremembering, but I think I just had to rebind <C-x> in zsh to get less working.

If you're following a pipe (such as `kubectl logs | less +F`), <C-c> is sent to all processes in a pipeline, so it stops less from following and it stops the other process entirely. Then you can't start following again with F, or load more data in with G.

Less provides an alternative of <C-x> to stop following, but that is intercepted by most shells.

Is my adblocker blocking them? I only saw the stack of tars in a coat. Didn't break the article's flow for me.

My current company started on AWS/GCP for the credits. Right now we're on Lambda for the GPU prices and GKE for some webservers that we cba to move. We dual-upload data to s3 and gcs still (which isn't too expensive, it's effectively write-only and the auto-archive features work for us). Cloud SQL database but pgBackRest to the other cloud.

We're not HA across clouds; we decided to chase RPO over RTO.

About once a week I see someone cut in even though the person is literally tailgating. The driver at the back has to brake+swerve to not cause a high speed collision. There's actually nothing you can do to prevent these people from getting ahead of you. Don't worry about what they'll do, it's insane anyways. Just try not to die.

I'm there right now at my current job. It's always the same engineer, and they always get a pass because (for some reason) they don't have to do design reviews for anything they do, but they go concern-troll everyone else's designs.

Last week, after 3 near-misses that would have brought down our service for hours if not days from a corner this engineer cut, I chaired a meeting to decide how we were going to improve this particular component. This engineer got invited, and spent thr entire allocated meeting time spreading FUD about all the options we gathered. Management decided on inaction.

Google's monorepo is in fact terabytes with no binaries. It does stretch the definition of source code though - a lot of that is configuration files (at worst, text protos) which are automatically generated.

Great advice! Personally, I got immense value from writing notes but never when I wrote them during the lecture. 30 minutes after the lecture has ended is a perfect time time to sit down in the library and write notes for what the lecture was about. This gives enough time to reflect about the big picture, but not so too much time that the details are lost.

TOML isn't great for hierarchical data.

My experience is different: TOML isn't obvious if there's an array that's far from the leaf data. Maybe that's what you experienced with the hierarchical data?

In my usage of it (where we use base and override config layers), arrays are the enemy. Overrides can only delete the array, not merge data in. TOML merely makes this code smell more smelly, so it's perfect for us.

The former. I clone from file:// URIs.

I just have the file path to the inside of my LXC container. If you're using Docker you can just mount it. I only need the path twice (for clone, and for adding a git remote). After that I just use git to reference the remote for everything.

I probably don't have the perfect workflow here. Especially if you're spinning up/down Docker containers constantly. I'm basically performing a Torvalds role play, where I have lieutenant AI agents asking me to pull their trees.

Git worktrees are global mutable state; all containers on your laptop are contending on the same git database. This has a couple of rough edges, but you can work around it.

I prefer instead to make shallow checkouts for my LXC containers, then my main repo can just pull from those. This works just like you expect, without weird worktree issues. The container here is actually providing a security boundary. With a worktree, you need to mount the main repo's .git directory; a malicious process could easily install a git hook to escape.

Whenever I need

Words for Haiku I reach for

Refrigerator

I've had my car burgled twice for them to only steal a backpack with nothing in it. I'm pretty sure the assumption is that backpacks have laptops.

regular backpacks look nothing like a laptop bag

You're missing some context here — this is in San Fransisco.

I use a knock-off of the edgepro I got for $30 online, with a nice set of compatible whetstones that click into the assembly that were $20 online. I did what everyone else does and hotglued some magnets underneath the piece you hold the knife on.

I tried a compatible strop that clicks in but it's not worth it imo; just use a normal strop.

https://www.phoronix.com/news/AMDVLK-Discontinued

This is a good but long overdue decision by AMD. RADV has long been more popular with gamers/enthusiasts on Linux than their own official driver. Thanks to Valve, Google, Red Hat, and others, RADV has evolved very nicely.

Line 1 of the readme has a poop emoji

Vercel is anything but simple. Easy? Sure.