building something agi wants. prev: ml lead @tryskylink, @microsoft. nlp & infosec research @columbia @iitkgp. alien of extraordinary tism. e/advaita. COYG!
Palo Alto
Joined July 2022
- Tweets1.6K
- Following898
- Followers1.1K
- Likes16.7K
me!
chasing money/status though as i realized that they are useful assets to have in the arsenal.
my "poet" era is definitely from another lifetime (~2016-2021).
reminds me i should add make my scribblings public as they are currently locked behind a private insta account.
amusing how my angsty seeker years (2017-21) spent trying to grok the universe/self/consciousness felt completely orthogonal to my professional life back then
but turns out they were good post-training for dealing with the era of machine intelligence
codex has native support to rename thread titles. a scheduled task made it easy to configure custom auto-update behavior.
i just checked if claude code has it so that you can use a similar routine :(it doesn't).
codex got more of the r in rsi.
xitter.cf/curious_queue/status/2…
yep!! i have a scheduled task that runs hourly and does something like this for me for all active threads
using this rename-active-codex-threads skill: github.com/ma08/botfiles/blo…
that's why it's one of my most used skills recently
xitter.cf/curious_queue/status/2…
lol this makes me feel a bit less guilty about letting codex handle the xfinity chat
(i wasn't sure what percentage of human/ai was on the other end)
overall codex helped me save $25/month by cancelling xfi complete and optimizing the modem/extender replacement wifi setup
welcome to the club!
it's also great for hackernews and any random website without a good dark mode :)
also fromthe foundation series: the use of nuclear tech's lock-in factor to conquer the territories (that use them like an API) serves as a great metaphor for the proprietary LLM-powered agentic stacks that limit choice
xitter.cf/curious_queue/status/2…
discovering that exponential with codex usage
(it's a cumulative usage graph)
while it's a fair argument that models just got more token hungry, i am pretty sure that the increase in reliable autonomous work i am able to delegate to agents is a big factor
~60% of 13B tokens in lifetime codex usage was during last two weeks lol.
codex app (now chatgpt lol) is quickly becoming a superapp/OS.
while the sol-terra-luna family models' performance and the liberal usage quotas/resets are ofc a driver of this increase, the app features/harness itself feel very underrated.
what i love about it:
- phone + remote machines support
- first class inter-thread messaging
- goal mode (use natural language to just let codex set it up using set_goal tool)
- scheduled tasks (great composability with inter-thread messaging)
- in-app browser, terminal, PDF, sheet viewer etc. with great annotation support to share feedback
- browser-user, computer-use
- recursive use (use codex chats to improve codex setup or invoke features)
- pets
it's already just too good not to use even there are some obvious improvements to be made:
- gpt-live voice model support
- better handling of threads from multiple providers like azure)
- questionnaires (via request_user_input tool) should have "add notes" option when selecting one of the given option like we have in cli
- all input modalities should have the native dictation option. some of them do not right now (like the freeform text input option in questionaires by request_user_input tool).
- smoother chatgpt work <> codex : right now some things like adding/submitting annotations via return key are flipped in interfaces based on the form factor i believe.
gg @ChatGPTapp @OpenAI
in light of @OpenAI signing the open weight letter:
always believed in you!!
microsoft.com/en-us/corporat…
(pls don't rugpull lol)
tbh codex is fully open-source and is taking on much more of a community-driven building approach (including their codex app server architecture, allowing use of chatgpt subscriptions in other products etc.)
while claude code is closed source (their DCMA notice response to people hosting the leaked source code on github showed how closed they want to be)
also openai actually released a few open-weights models
whatever might be the motivations, that's enough reason to think of them as more open than anthropic