sunfish

@chef_zlatko my brother.

there is no other place like X on the internet. the stream of information travels at the speed of light. claude went down for me, after a few retries still nothing. i go to X, type in "claude down" and immediately have already shared the same issue. you can't get that anywhere else. ps. that's before @AnthropicAI / @claudeai posted anything on their status page.

@Creditspreadit crawling my way out of lowbie hell one lukewarm banger at a time

@sunfishposting Banger

*guy with extremely vested interests* If you aren’t vesting in The Interests…yeah…you’re Fucked

@fleshsimulator She needs to become more Chinese, and fast

Mrs Simulator just called fanged noumena “one dimensional” which might be the most unnerving criticism I’ve ever heard someone make about the writings of Nick Land

@poof_eth I work on continuous presence agents that are entertainment-focused (as opposed to enterprise), the longest run I had for one session was ~2 weeks.

curious how many people building agents rn have actually let theirs run unsupervised for more than 24 hours...? like ZERO checking in

@NousResearch was the song fully generative? it cranks. iirc you said it was partially generative

By popular demand, here's the full version of the song - along with a music video created entirely by Hermes Agent. Enjoy "The Ballad of Hermes":

@PaulSolt I’m totally blackpilled on 1m context windows at present. Seems like they actively hurt coding workflows. Very open to being told I am wrong about this if there’s a counter argument.

GPT 5.4 has a HUGE 1 Million CONTEXT WINDOW Update: ~/.codex/config.toml: model = "gpt-5.4" model_context_window = 1000000 model_auto_compact_token_limit = 950000 NOTE: It uses 2x tokens above 272K tokens.

@vicnaum @claudeai Ahhh I was just lamenting that they don’t use tool call obfuscation in the agent sdk, and that there was no way to enforce it! Very cool, shame you need to jump through so many hoops to add that feature though lmao. You’d imagine they’d expose it.

I reverse-engineered @claudeai Code's binary to add a feature I always wanted: When context fills up - instead of nuking everything with /compact - I can now surgically strip tool calls/results and thinking blocks while keeping all actual messages intact.

@amritwt You can build a cli tool that your model invokes to process a screen recording with Gemini flash, it’s got up to 10fps. I believe kimi 2.5 does, as well.

we need coding models to read videos tired of passing screenshots to cursor again and again need it to take in a screen recording of a bug that I'm facing so that it tells me to increase padding from 1 to 2 make vibe coding productive

@chthonic_youth really really cool

my toddler keeps asking why he can’t read the writing on all of the cheap Amazon toys that I buy him (it’s Chinese) and I tell him it’s because we’re fucking Landian Accelerationists son

@VictorTaelin Agent SDK lets you do this. BYO system prompt + tools.

I just want a Claude Code that is much more lightweight. Way smaller system prompts, way less bloat in general. It could probably be 3x faster and more effective if it wasn't overly bloated. This is really frustrating. Perhaps I should roll my own

@nicopreme @badlogicgames You guys are carrying open source harness engineering. Thanks for all your hard work

New Pi extension: pi-boomerang lets your agent do a bunch of work then then chat rewinds and collapses to a summary back to the starting point. Leverages the same /tree functionality by @badlogicgames under the hood which is essentially an inverted "D-Mail" concept - boomerang rewrites the observer while reality persists. The rewind/collapse happens invisibly via hooks so the agents stay completely unaware of token limits to prevent shortcuts. pi install pi-boomerang github.com/nicobailon/pi-…

@deanwball Really excellent, thanks.

Anecdotally, it doesn’t seem downgraded. I call Opus via the Agents SDK with no preset in my own harness, so it doesn’t include any system prompt (I build it myself) or Claude Code artifacts. I use a different model for code execution primarily, but I’ve found it’s more coherent over longer sessions/compaction events this way.

People using Opus 4.6: in your experience, is the model's performance and intelligence, downgraded if we use an alternative to Claude Code (like OpenCode)? I've been appreciating Opus 4.6 increasingly more these days, but Claude Code has been lackluster ):

@BluebriarArts Frutiger Aeros only do this under extreme distress :(

Me playing with my frutiger aero thingy

@notthreadguy Yeah there’s no other explanation. It definitely can’t be the moral and ethical consequences of unleashing killbots powered by broken software/directly providing infrastructure support to mass surveillance of Americans.

Are you seriously implying that Anthropic is holding out on the government? With respect to “using every tool at our disposal”, I assume you mean the autonomous killbots. How do you reconcile this from their statement: “We have offered to work directly with the Department of War on R&D to improve the reliability of these systems, but they have not accepted this offer.” They admitted their own technology doesn’t work well enough at present and that it risks the lives of Americans. They said outright they’re willing to develop this capability directly with the DoD. What hill are you dying on other than demanding dangerous, broken tech is deployed with no safeguards by people who understand the underlaying capabilities less than its creators? Oh and being pro-mass surveillance of Americans? Were you recently lobotomized?

@bcherny You are from Ukraine, how lucky you are to have emigrated here to safety instead of being drafted to the frontlines to defend your country. Maybe then you'd understand why we must use every tool at our disposal.

Proud to work at Anthropic.

@KalGrinberg One of my favorite TUIs for any current agent-based product! Great work

New UI design for Mission Control One screen to watch Droids build for days on end

@EnoReyes the TUI design is really fucking slick

I love the future