@FoleyConor@MeetAiqi This is a while back, but I had saved a scrappy year's worth I had to do a contract or two just to not only learn more but get a little bit of money. By the time I ran out, I was able to get a job pretty quick. It was more in tune with where I wanted to go.
@BensHasThoughts Every time I took a sabbatical, I came back with a stronger position because I used that time to upgrade my skills with projects there wasn't time for at the last position.
@theisaacmed@MeetAiqi In other words, you might actually be regressing your skills if you are stuck in a particular workflow and a system that you can't just evolve into a brand new approach every week. When you're on sabbatical, you can do whatever you want.
@theisaacmed@MeetAiqi I mean, that's a pretty general statement. Each time I've taken sabbaticals, I've come back way ahead of the game because I have the opportunity to learn new things. It depends on what company you're at and if you're evolving.
@gregisenberg@pmarca I agree. Our agentic orchestration approaches to how we actually engage our sessions are probably what is most important. I think that's what most of us are working on.
@Gamingtronium I meant poisoning attack. For example, when you're running Claude code and you're pulling down a skill from some other repository or something else, or installing a package that you didn't check if it was secure using something like socket firewall
@Gamingtronium First thought would be a local phishing attack on your own machine that can read your keys. And hopefully those keys can't touch production and are also mitigated with permissions.
@GregorySchier It's pretty simple: the shape of software engineering hasn't changed. The composability, the zero-trust approach, and now, with these dependency poisoning attacks, high security on your local environment. The work is more automated the patterns remain.
@dexhorthy@walden_yan@tobi@karpathy I started working on a concept for orchestration a little after June or July of 2025 that relied on a DAG for more deterministic flow, once all of the atomic units of work were defined, then consensus reviews by independent agents, configurable, and a pass or fail threshold.
THIS 17 YEAR SOFTWARE ENGINEER JUST WENT OFF ON CLAUDE CODE AND CALLED IT COMPLETE GARBAGE
this guy has been writing code since 2008. started using AI in march. 5x max plan since day one
at first he was blown away. 3-4 sessions running in parallel in tmux all day. commits in seconds. answers in seconds. implementations in minutes. barely ever hit the limit
then opus 4.7 dropped and everything broke
before vs after:
> "commit this" used to take seconds. now takes 30 seconds
> "implement this plan" used to take minutes. now takes 45 minutes
> resizing the terminal used to rewrap text cleanly. now lines get interlaced and diffs get cut off
> ctrl+o used to show the full thought process. now it shows nothing useful
> it now says "almost done thinking" like that helps anyone
> he hits the limit constantly even after learning to split opus for planning and sonnet for edits
but the WORST part is that it ignores instructions
he told claude a thousand times to use short timeouts (10-15 seconds) for network programming. wrote it to memory. within one implementation cycle claude switches back to 30s, 60s, 5 minutes
told claude to "NEVER AUTO COMMIT" a billion times. after a while claude starts auto committing again
the guy uses /caveman mode for brevity. claude forgets about it and dumps a wall of text after every implementation
claude can't even follow a simple 200 line plan anymore
the plan said "change the signature of handle_input to take Bytes instead of &[u8]." claude added a completely separate function instead. said it wanted to "keep backwards compatibility." on a brand new project. there is no backwards compatibility
then it started deviating from plans mid-implementation. tried to cut features. wanted to "ship v1 now, do the rest in v2." there is no v1 or v2. that's the developer's job to decide, not the AI's
we had something good with opus 4.6. now it's garbage
@snehalsurti@ClaudeDevs I'm gonna give it a try. It's a useful mitigation layer. We'll see how noisy it is. I also want to try the socket firewall aliases for my local
@DanielSmidstrup It usually means you're going to have a more composable architecture and likely better organization of your projects and patterns. Unless you get lucky. But you'll definitely be in better shape if you have domain knowledge.
@jiahan_c You can mitigate this also by creating a token assigned to git that is permissioned down to limit the blast radius and prevent it from changing repo branch rules and other configuration if you require PR-only branch merge rules and no direct push to main.
@Layton_Gott The other part of software is the engineering side, which should include making things easier to work with while still being composable and manageable. Genuinely way more important these days with the move towards probabilistic agents and deterministic guardrails.