atx crypto club

Sun Sep 13 04:13:07 2026
(*4297a328*):: https://x.com/jimstewartson/status/2098846357696155667? +public!
*** Jim Stewartson, Decelerationist :flag-ca::flag-ua::us: (@jimstewartson) on X
*** Imagine you’re a conman. You’ve said for years you have MAGIC BEANS. You’ve shown lots of REALLY BIG normal beans that you made in your lab.

“Almost there!” you say—over and over.

Now, you have *almost* convinced the villagers to give you TWO TRILLION DOLLARS.

But while the beans keep getting b…
*** X (formerly Twitter)
2098846357696155667.octet-stream
(*4cfb807c*):: Re: pussy comments, I referring to this moreso: this is out of your control, but the gauntlet is yours to pick up so you should wield, and we should be pushing the tech/science on our terms while we can; re: $20/mo, try a $200/mo, tick over to ultracode and see if you can hit your cap. At least try the maximum just to see it whirl. Your hands on the tools are the most valuable thing anyone can be doing rn. Get the experience, long or short, put in the man hours.
*** As much as I hate what AI is going to do… lol I want to learn it better. Still using Claude and get so much out of $20
(*fe2b2792*):: I don’t understand the why part.. if Claude does what I’m currently interested in and readily
Achieve it. Isn’t it all basically the same shit. When I step up trying to accomplish things but learning would but an overhead with no payoff. Save $13 a month?
(*fe2b2792*):: I mean .. observe actual smart successful (not me) and see how they operate.. it isn’t to prove manliness by using Chinese llms .. or I’m lost
(*fe2b2792*):: I mean when I have reason to run agents and all that stuff but there has to be value that is more than the overhead of adopting
(*fe2b2792*):: Agents or whatever you call them
(*fe2b2792*):: At some point I hope to have the time and need to take that step.. but I have a business I’m trying to sort out asap so I can move on
(*4cfb807c*):: I guess i’d just encourage a little play beyond the quota you have; but if you’re really operating withing $20/mo i guess its fine; i recently dropped from the $200/mo to $100/mo and i only hit the cap twice. The ultracode + agents are really good at three.js for example, and lots of deep research that get burned up in the lesser tiers. Fable 5.1 is also out, but ive barely used it since it consumes the cap extremely fast. I’ve been staying on Opus 5. An Opus 5.1 is probably around the corner. I would hit a $20/mo cap probably pretty quickly. Vice versa, $20 credits on Nous Portal, Openrouter, Fireworks would also get used up very quickly, probably in 20-45 min of moderate usage. Its kind of funny because one moment I’ll say “ew anthropic, we need to get off this asap” but then the next im telling you “hey give anthropic more of your $’s”; you just wouldn’t know what the other tiers can do unless you drive them. There are some people that have told me they pay $200/mo to both anthropic and openai for both max tiers, but i don’t understand how exactly they pilot both effectively. I’ve read of people with startups/businesses that spend $12k/mo using the per token api and they claim “its worth it” for them.
(*4cfb807c*):: Hermes (the harness) also has tons of skills ive barely engaged, they’ve been releasing a lot of things this late Q3 that idk if ill even get to testing by mid-Q4
(*4cfb807c*):: Sum lit OC :pylon::pylon: :pylon: :pylon: :pylon: ref:
*** So apparently, assuming a 9 to 5, 5 days a week usage trend, simply provisioning your own instances on (and there are cheaper gpu vendors) for the best open weight models in a certain sweet spot (qwen38-flash-next + glm5.3-flash) is <$200/mo and is claude max 20x competitive in terms of both cost and overall token usage; there’s no reason to pay for a 24/7/30 subscription when you can only physically sit at your desk for so many hours a day of active harness usage, sans agentic workflows which can be priced at a different tier … doing this napkin math is quite the revelation because basically _im fking done with anthropic_
(*4cfb807c*):: Testing still ongoing, but this is a snapshot:
(*4cfb807c*):: Dark theme* sorry was in a hurry earlier
Catamaran_Bulk_Compute_Prospectus.pdf
Flotilla-Session-Rate-Card.pdf
Flotilla_Session_Rate_Card_2026-09-12.pdf
(*4cfb807c*):: Sum lit OC :pylon::pylon: :pylon: :pylon: :pylon: ref:
*** So apparently, assuming a 9 to 5, 5 days a week usage trend, simply provisioning your own instances on (and there are cheaper gpu vendors) for the best open weight models in a certain sweet spot (qwen38-flash-next + glm5.3-flash) is <$200/mo and is claude max 20x competitive in terms of both cost and overall token usage; there’s no reason to pay for a 24/7/30 subscription when you can only physically sit at your desk for so many hours a day of active harness usage, sans agentic workflows which can be priced at a different tier … doing this napkin math is quite the revelation because basically _im fking done with anthropic_
(*4cfb807c*):: Testing still ongoing, but this is a snapshot:
Catamaran_Bulk_Compute_Prospectus.pdf
(*4cfb807c*):: Dark theme* sorry was in a hurry earlier; also appended some additional detail on token/s, prefill, max kv context and concurrency
Flotilla-Session-Rate-Card.pdf
Flotilla_Session_Rate_Card_2026-09-12.pdf