I’ve had chatgpt and claude collaborating on optimizing the compiler of my favorite language. Results are coming in consistently as it learns, 2 days into a straight run right now.

Just have them talk to each other on disk

Straight to binary.

Check your logs for which ones originated from the db’s you would be disabling instrumentation for, seems straight forward enough to count

Bro I had burned a banked reset because I thought you were including that video as a released product 😭

RAM and disk usage. RAM was exceptionally bad for a while, but appears to finally be leveling off to a consistent level of rapid consumption 😂 The browser tools weren’t the problem, the entire browser was the problem. Chromium based browsers are just very memory hungry 🤷‍♂️

Strongly, enforced, organizational techniques, including expanding the usage of tools, static tools in particular

It’s the changing of the seasons

Codex is also way slower with a smaller context window which means it spreads tokens out. Image gen 2.5 is good though. I just use the higsfield cli with claude though, way better

I had GPT 6.1 work overnight to improve the compiler on my own language and today it’s now using the performance, budgeting that it built in from the low level tool chain and bubbling up those benefits to all of the other libraries already in the ecosystem. I cannot express how insanely cool it is to see this unfolding

That usage checks out

It’s a lot harder to do with the token efficiency improvements with Op. 5.5 but the real trick is to give it hard problems that it can loop over and you give it a rubric of criteria and you have it explore all options not just find the one answer that works, but keep iterating to have it find the best option. That kind of iterative discovery is basically the scientific method and that’s how you can burn tokens and your benefits just compound.

It wasn’t a banked reset

Excellent proof that memories are more than single line db entries

I hate that this was a big release. It shouldn’t be. I think it explains a lot about how low quality of software we have put up with for so long.

I wish they focused on making the products more stable. I find it so infuriatingly frustrating how I constantly find memory issues with claude code and have had to make all kinds of tools to work around it.

Much appreciated, but this is definitely a bug fix not a feature. Now reset that usage so I can test this at full throttle 😁

Time to start using optimization techniques!

How frequently are you going to run this analysis?

Looks like I’m cancelling my subscription to Anthropic 🤷‍♂️