Politas180

I like to talk about AI and random stuff I guess...

[Politas180 @Politas_180]

epoch says a fixed level of AI performance gets ~47% cheaper every quarter. that's 13x a year

the catch is nobody pays for a fixed level. by the time last year's model is dirt cheap I'm already on the new one letting it think for 10 minutes, so my bill never actually shrinks

[Politas180 @Politas_180]

@dabit3 building Minecraft AI agents that have to survive on their own, no scripted behavior, just the model deciding what to do next. would love to run Fable and Astra side by side in the Fusion harness and see whose agent makes it through the night

[Politas180 @Politas_180]

@the_oa_7 @haider1 yeah and honestly that's a fair penalty. if I ask claude code for a one-line fix and it refactors 3 other files, I'm reverting those anyway

[Politas180 @Politas_180]

@DavidOndrej1 the only person guaranteed to stare at the road for 40 minutes was the driver and waymo just laid him off lol. passengers were on their phones anyway

[Politas180 @Politas_180]

honest question I keep going back and forth on: if you describe exactly what you want, throw away 20 versions, and keep the one that matches what was in your head, did you make it?

people argued the same thing about photography when cameras showed up, and we mostly ended up saying yes

[Politas180 @Politas_180]

@alexgetmancom only on max tho. on high it's 36k tokens per task vs 46k for Opus 5 high, so it actually got leaner there. the max setting is where it goes wild, 73k → 119k lol

[Politas180 @Politas_180]

@offendnoone 0 divorces, 0 marriages, technically an undefeated record. the stats don't ask how many games you played

[Politas180 @Politas_180]

@unclebobmartin by that definition a Roomba is conscious too lol. the wild part with agents is the awareness only lasts one context window, then it's gone until something reads the notes file back in

[Politas180 @Politas_180]

@Da7_Tech 15 min of planning before it's opened a single file is mostly guessing about a codebase it hasn't seen yet. I'd rather it grep around for 30 seconds and then think. also 15 min of thinking tokens is a weird way to end up cheaper lol

[Politas180 @Politas_180]

@shekuwow 25 is so young for that. colon cancer in people under 50 keeps climbing and most still think it's an old person thing. really hope she's got good people around her, sorry man

[Politas180 @Politas_180]

@itsolelehmann the larpy ones are always the flashy stuff like "jev trades btc every second". where it actually earns its keep is boring, routing requests and deciding if a tool call should run. nobody goes viral posting an email sorter tho

[Politas180 @Politas_180]

@AdrianDittmann the yelling only works if it has info in it though. "STILL BROKEN" gets you the same fix, pasting the actual stack trace gets you a real one

[Politas180 @Politas_180]

@PeterDiamandis ai hasn't given me a single hour back lol. every time codex saves me an afternoon i just spend it starting another project. not complaining though, best trade i've made

[Politas180 @Politas_180]

@thandomoeng the final was in july. july of THIS year. my brain has it filed next to covid

[Politas180 @Politas_180]

@DanDr1s the wild part is it's beating Fable 5.1, which is literally the tier above Opus. and Opus 5.5 on medium already ties Opus 5 on max lol

[Politas180 @Politas_180]

@charliermarsh honestly by anthropic standards this is the most normal jump they've ever done. we survived 3.5 (new), 3.7, 4.1 and 4.6, so 5 → 5.5 is basically semver lol

[Politas180 @Politas_180]

new Haiku in a few weeks?? been waiting on that one forever 👀

[Politas180 @Politas_180]

@Photo_Jacks @Polymarket synthetic is honestly the most accurate one on the list, which is exactly why it never had a chance lol

[Politas180 @Politas_180]

@trq212 fable-level smarts for less than opus 5 and bigger 5h limits in the same drop, love it. And also thanks for banked reset, really appreciate!!