1/5 — $200 of AI is not always $200 of AI.
My local Codex logs project to ~28.3B raw tokens/month.
$200 of DigitalOcean Kimi K3 buys ~65.8M–270M.
That’s only 0.23%–0.95% of the raw token volume. 🧵
This is not a 105–430× intelligence claim. It’s subscription agent access vs. metered inference.
For skeptics: exact counters, cache assumptions, formulas, sources, and limitations:
https://t.co/vObRi2gEYE
I thought my parser was wrong.
My Codex logs project to 28.3B raw tokens/month on a $200 plan.
The same $200 on Kimi K3 buys 65.8M–270M.
That’s only 0.23%–0.95% of the raw volume.
The number stayed after I checked the logs. 🧵
5/5 — The mindful takeaway:
A token is not a unit of value.
The better question is: what did those tokens help me finish?
Full math, sources, graphics, and limitations:
https://t.co/vObRi2gEYE
1/5 — $200 of AI is not always $200 of AI.
My local Codex logs project to ~28.3B raw tokens/month.
$200 of DigitalOcean Kimi K3 buys ~65.8M–270M.
That’s only 0.23%–0.95% of the raw token volume. 🧵
4/5 — My strategy:
• Codex for tool-heavy, repo-aware work
• Kimi K2.6 for bounded routine inference
• Kimi K3 when premium quality or vision pays for itself
• Benchmark outcomes before migrating
Coming this Fall: AMBOSS Qbank sessions, launched instantly from your Anki flashcards! Sign up to try out the new add-on feature. https://t.co/hJxdoqEMo4
We didn't think it was possible... but we just launched a beta of image to SMILES conversion for chemical diagrams in Snip!
Amazing it works at all, but much more improvement to come soon. Looks like diagram parsing with deep nets IS possible! 🤯🧠
#JavaScript tip for today:
`some` & `every` are one of the most underutilized array methods.
So don't just `filter` everything. Recognize when what you are asking is a "yes/no" question and often these will be the methods you need for a more elegant solution.