Adam Williams

simple is robust

dad, yc s23, sf — opinions subject to change

[Adam Williams @jumploops]

vibe proofing just dropped

[Ernest Ryu @ErnestRyu]

I used ChatGPT to solve an open problem in convex optimization.

*Part I*

(1/N)

[Adam Williams @jumploops]

@paulg Write like PG. Make no mistakes.

[Adam Williams @jumploops]

It's loops all the way down 🐢

[Adam Williams @jumploops]

@bridgemindai Ran the same prompt on DeepSeek v4 Flash w/2-bit quant via antirez's ds4

It took 10 minutes, mostly reasoning

For context, this was on a M4 Max with 128GB RAM

[Adam Williams @jumploops]

Yes, waterfall is back!

[MacCallister Higgins @macjshiggins]

nobody wants to hear this but the classical NASA systems engineering is the perfect model for developing code with LLMs. people try to approximate this with planning modes, but if you’re explicit in your docs it’s never been easier to build, test, and verify complex codebases.

[Adam Williams @jumploops]

@thsottiaux Codex app binary version: codex-cli 0.131.0-alpha.9

[Adam Williams @jumploops]

Haven't seen gaslighting like this since switching from Claude Code 😅

Latest version of Codex, GPT-5.5 xhigh

@thsottiaux help?

[Adam Williams @jumploops]

@thsottiaux Previous to this message, I had taken a quick detour to fix an issue.

I used to manage context carefully, but since the forced automatic compaction launched, I've had surprisingly had very few issues.

Until now?

[Adam Williams @jumploops]

Any other codex users not seeing a big difference between /goal and their previous workflows?

Working on a refactor, and usually codex will work through all the plans to completion.

With /goal, it decided to end at "the first tranche" -- even though I explicitly asked it to work through the full refactor (as outlined).

Nudging it with a second /goal stopped early again.

Previously I've had codex run a similar refactor for 50+ minutes, without an external harness/skills/etc.

[Adam Williams @jumploops]

@patrickc Silicon Valley is the new 1920s Detroit (pre-crash).

Ford was on a similar level (compared to global GDP) as Page/Brin.

Why is Stripe in South San Francisco? Why are Page/Brin leaving California?

SV is "richer" today, but the landscape is more expensive to change.

[Adam Williams @jumploops]

Good to hear!

It's worth noting that Activity Monitor on MacOS isn't super reliable[0]:

> macOS counts dirty-but-not-resident pages toward phys_footprint, causing Activity Monitor and Force Quit dialogs to report wildly inflated numbers.

cmux is built on libghostty iirc

[0]https://t.co/c9z1kPKFTs

[Adam Williams @jumploops]

@ImagiBooks @thsottiaux If codex@latest doesn't fix your issue, it's likely this cmux bug: https://t.co/fdsEYcvyfF

[Adam Williams @jumploops]

@ImagiBooks @thsottiaux This might be fixed in the next release: https://t.co/Viv0xYfe0v

Still odd that React Native DevTools and Codex both have the same exact 99.11 GB in your first image.

Guessing you're using some React Native MCP test tool?

[Adam Williams @jumploops]

@sama Spud bar!

[Adam Williams @jumploops]

Though I'd argue it's Eternal December, as that's when the models got "good enough" for most devs

[Adam Williams @jumploops]

Claude is down, Github is shaky.

Eternal September round 2?

[Adam Williams @jumploops]

@ImagiBooks @thsottiaux I'd double check that your .gitignore is properly ignoring generated code, installed packages, etc.

The fact that React Native DevTools is also at 99GB leads me to believe something is configured incorrectly.

[Adam Williams @jumploops]

@ImagiBooks @thsottiaux Do you use git and .gitignore?

I noticed crazy high CPU on a throwaway node app, where Codex was tracking a bunch of files unnecessarily.

Adding a .gitignore for node_modules fixed the issue.

[Adam Williams @jumploops]

Ah makes sense. Our usecase is essentially a shared workspace for two or more agents.

Easily solvable with a non-FS based approach, but our agents rely most heavily on the filesystem, so I was envisioning a mountable version controlled disk that "allows" clobbering, but in a tracked way.

Basically CRDT/OT on a specific mount point, with tooling (i.e. git) to allow the agent to investigate the history in certain edge cases.