Sabitlenmiş Tweet
ㅤㅤㅤ retweetledi
ㅤㅤㅤ
9.8K posts
@pvncher @0xblacklight @aisdk assuming single step. if you have so much as even 1 more step the caching already pays for itself. no idea who is doing a single query without any tool loops or follow ups. also applies if you reuse the same system prompt between differing queries
English
@0xblacklight @aisdk For agentic coding - sure, but if you want to just do simple queries, the 25% upcharge for cache writes can be a real pain.
English
dang have y'all ever tried using opus for agentic coding without cache-control
like lighting money on fire
found the @aisdk option for it for anthropic, seems like it should be enabled by default since it costs u (almost?) nothing
GIF
English
ㅤㅤㅤ retweetledi
ㅤㅤㅤ retweetledi
“oracle-zoned” is such apt phrasing for the realization I had last year that most people actively resent anything resembling clairvoyance, and at best view you as little more than a human hot-take aggregator kept around solely as some sort of spectacle
Vivian@suchnerve
Autistic people who specced into manually-learned social skills get oracle-zoned rather than passing for allistic. You generally get seen as a useful and interesting font of mysterious wisdom, not as a fellow normal person who should be socially included.
English
ㅤㅤㅤ retweetledi
I did the math a couple weeks ago and it turns out a vegan prompting a frontier LLM *every second, 24/7* consumes less water than the average omnivore who never uses AI.
mary@theoceanblooms
people are trying to argue with this and it’s literally the truth. a kilogram of beef requires over 15,000 litres of water to produce. a vegan who uses chatgpt every day is living a more sustainable lifestyle than someone who regularly eats beef while boycotting AI.
English
@KentonVarda except it compacts way sooner than 1M on the subscription plan
English
People pointed out that under the subscription plan, input token cache reads are *free* (don't count against quota), as opposed to API pricing where they cost 1/10 of uncached tokens. I guess that means if you use Claude Code you can actually use 1M context without getting screwed over on cost.
Us poor Opencode users paying API pricing, though, basically can't.
Seems like if you stay under 100k tokens, API pricing is basically negligible and I'd rather pay it than lock myself in, but a single 1M conversation would be expected to be 100x the cost of a 100k conversation, which is quite substantial -- hundreds, maybe thousands of dollars for one convo. Yikes.
English
Claude has a 1M context window now, but realistically costs skyrocket around the 200k mark since input token costs are O(n^2) (due to resubmitting the chat history on every turn/step -- even cache hits can get pricey).
English
@zebassembly your colleagues should read my application so that i can join in on the vagueposting
English
ㅤㅤㅤ retweetledi
ㅤㅤㅤ retweetledi
ㅤㅤㅤ retweetledi
This fucking rat was the guy who ran the Seattle Supersonics into the ground and when he didn't get his way he sold it to a group knowing they had every intention to move the team out.
If I knew that's what it took to get rid of him I'd be campaigning for that tax 18 years ago
The Needling@TheNeedling
Millionaire Tax Already Working: tinyurl.com/4sm3azwb
English
@noahzweben how is it even this slow? no other harness has startup times this egregious, and it behaves like this even spinning one up from the desktop app
English
Remote Control - Session Spawning:
Run claude remote-control and then spawn a NEW local session in the mobile app.
* Out to Max, Team, and Enterprise (>=2.1.74)
*Have GH set up on mobile (relaxing soon)
* Working on speeding up session start-time
English
OB1 could not beat @droid the fair way. Shame, but clever way of cheating Terminal Bench
GIF
Monk Zero@NoCommas
English
Maria is speaking straight facts here.
Please don’t spam us with giant PRs adding new features. Just tell us what you want and we’ll triage accordingly.
Appreciate the hype around T3 Code but please chill a LITTLE bit guys
maria@maria_rcks
to the people contributing to t3.codes could you please... stop? - do not file a 9k loc pr that adds 7 features on top of what you advertised in the title. - ANY ui change should have before / after screenshots, idk how this isn't common knowledge. - and for the ones adding providers, i'd just wait for @jullerino / @theo to add them. if you want to make things right: - first open an issue with something you'd like changed / added with clear examples of how it'll work / look, also attach examples from other open source apps that do that specific thing well - something as simple as changing the wording "open pr" to "view pr" - removing a stray dot in the ui for the working animation and most important of all... make it easy to review
English
@SeriesInfinityy because, ultimately, a human can always review the resulting code, modify it, or prompt further iteration. the contribution guidelines are drawing a fake line in the sand.
English
@SeriesInfinityy using AI for “codebase understanding” is effectively the same thing as asking it to perform the implementation for competent developers. the only difference is imposing yourself the additional steps of manually altering the individual files based on the guidance the LLM spat out
English
@transpiracy im gonna have to run my own version that just fixes the build system without any of this silliness going on, i wanna keep it as close to the original code as possible
English
ㅤㅤㅤ retweetledi
I remember writing prompts as comments inside an empty function body to get those sweet vscode co-pilot completions a hundred years ago
English