openrouter UX for the key security page is quite poor, almost impossible to horizontally scroll and see the multi-select ui but codex using cdp found it and did my config change easily, やっぱり another victory for DOM-based navigation. and it will be viable for yet a while since llm proficiency with the web stack is perpetuating its use

things have not been looking good for the portfolio since near left and stopped posting investment advice

andrej didn't mention this in his 2025 YC talk but lately I've been merging his software 1.0 and 3.0 in the form of fallback and automatic self-repair - you give input to program X - X tries processing it with its deterministic logic - if X detects a failure, it spawns an llm agent to examine the log (including the input) and path to X's src code - agent diagnoses the failure, implements a fix, while (!) preserving the running X process. such that the agent can then manually push the corrected post-fix output to the original user so from the pov of X's user (itself an llm agent usually), there was never any issue with X in the first place besides a longer running time. this is how I've set up all of my mcp servers to work

iterating on some reverse engineering with sol 6.1 for over a week and opus 5.5 more or less one-shot it to a better state than the final with sol lmao I need to improve my proxy benchmarks

my number of subagents is now bottlenecked not by the optimal structure for solving the problem but by available RAM, since the iteration each one is doing is heavy I have like 8 at once right now, with my custom orchestrator, not the built in cc/codex subagent systems

still scared of accidentally getting banned for some technical misuse of claude SDK so we went with cc TUI tmux automation for my t3-equivalent easier than I thought; you can automate nvim/similar as EDITOR and then use Ctrl-g for composer input, much less janky than pasting

just accidentally did my first >1h voice prompt looks like I've been improving on speech stamina, this would have winded me and made my throat ache like a month ago

I tried glm 5.3 flash for some ~cybercrime and it seemed to start heavily hallucinating when the context hit ~300k tk, despite it advertising 1m as its supported size otherwise pretty good and inexpensive despite no subscription-level subsidization

enabled the codex auto review bc of the day 2 free usage but I'm seeing sol complain about blocks constantly the best convenience tradeoff might still be yolo mode + good backups unless you're really expecting prompt injections or something

I made a small luna voice assistant as like a router on top of my codex manager so with one physical button press, I can tell it to start a new codex thread with some X specific model on my Y chatgpt account, or forward a message to some old existing thread, etc

to what extent does a micro todo list increase subjective experience of working in the abstract maybe you'll have infinite dopamine if an llm could monitor your work and show a +5 XP animation on every productive action like submitting a codex prompt

anyone else surprised that the $500 plan didn't change its usage/cost ratio against the existing 20-100-200 doesn't this decrease wealth inequality

is it considered autistic to be voice prompting llm agents outside at a public park

here's my ms paint masterpiece of an entry for this meme format couldnt fit the 8-9th but it'd be ds2 and nioh maybe deliberately avoided including names otherwise