musings

is everyone getting cooked on battery? like my mac just drains battery so fast do i need to move to ghostty + claude? or like claude is cooked so no shot

lol you get not usage limited but now rate limited by claude when using ultracode gg

TIL the prompt cache is at an effort level? maybe just for `xhigh` on anthropic? but doesn't this mean `xhigh` is a different model? or at least on different hardware? either intusions about how prompt caches work are wrong or it is a clue about something janky anthropic is doing and not telling

ultracode has the adverserial review thing built in

built a lib

6 days, a 1M line PR claude code is fully built on bun, so when this goes upstream will affect them

very cool that you can just have so many sub agents

TIL `gh search code` for codesearch across org

playing with codex browser-use and one thing that is obvious to me that will happen is UI state being stored in the nav is super useful but also will be abused the hell out of all and you almost want to support and encourage this abuse resend by default shows 150 items in a page, they don't virtualize but the agent can set load 10k items and its loads but slowly no infinite query and no virtualization i think this is both very cool the state is in the search so the agent can fuck with it but UI should also support that agents will fuck with this

anyone using `pi`?

^ sent by codex computer use

codex computer use onboarding was one of the best things i've tried

such a fun document, it came back up in convo today and fun read when you have a few hours to kill

he found the way to skip the getRect by getting agents to just look inside the browser quirks and end run it

very cool

so within this you can apparently you can reverse engineer api

<chrome://inspect#remote-debugging> ^^ feels like a superpower especially with openclaw

very cool product and and like the human/agent landing page

anthropic jobs data

harness engineering

why europe lacks tesla

recent reading

recent reading.

saaspocalypse

the one in which sarah guo and elad gil talk through the saaspocalypse.

stripe minions

ramp ai index

very interesting.

claude in excel

codex agent loop

cowork + claude chrome is more useful than atlas it is single threaded vs atlas able to open many threads but atlas just goes on for too little time

very interesting, like this feels like revenue pressure seems uncharacteristic for linear maybe the agent is doing very well and expansion of that

book marking

very interesting

> The `vm_bundles` (8.2GB) is the Linux VM images for Cowork mode - that’s what I’m running in right now. *Don’t delete that* or Cowork won’t work. cowork is using a linux vm? that is interesting

energy of tastemakers

i love this description, so much of what i do and the pain with it.

todo read:

also an interesting thing friend from ramp shared on structuring for agents

physical intelligence

insane. would also recommend reading their original paper, super interesting.

chain of thought monitorability

bookmarking for myself for later.

very curious to see how this will go spotify is the app i use on this platform, i understand why they made the chociies they did and i like that it exists but still sorta sucks i am curious why openai is so bad at building this kinda of ecosystem ontop of them

factorio learning env

forgot to play with this, but it was my favorite paper from neurips.

langwatch scenario

works in progress run

i've loved works in progress the last few months. they are publishing great work.

—- sent this a few days ago but very good article around hiring from delian/rabois

the new codex app is really good, the steer feature is great (it is the update thinking that the pro models have but in more usable models) 5.3-codex is also great i could see myself moving over to this from claude

on dairo's writing

dairo is such a good writer, the articulation and choice of words is so good.

this podcast is very interesting

nat and daniels talk at sessions was very good

as more saas vendors cross the innovators dilemma and start building for agents first, might be interesting to think about where value acrews the permissioning as just humans vs other permissioning structures discussion is interesting too

spacex flywheels and process

davos has been interesting (not thinking about the geopolitics) to calibrate timelines, dario had two talks. would watch both, one alone and one with demis. haven't seen elon but satya brought this up and i’ve seen him talk about it and would want to expand that framework internally token factories. these are fundamentally different business than software, infact microsoft always thought of themselves as a software factory, to build any software and not thinking themselves as a token factory. in long term thinking and strategy thinking, we should mechanically think about token factories as different than software companies, they might be better (grow faster) or worse (lower margin) but they are different token factories are not software business, not in the traditional sense. maybe there is a software business on top of a token factory, this is to be determined i think this framing of seeing them as different businesses is important when benchmarking. now the next question is should we be a token factory, and this is a maybe

gpt-5.2 browser demo

wow. this is crazy.

brie wolfson on craft

i think brie articulates company culture, craft, and taste things very well.

red bull founder

i knew very little about the founder of red bull and this was just very interesting.