Sergey | AI Builds & Tests retweeted
Psychedelic Neon Portrait
TFTT @EdithDel_Carmen
Here is my take..
Irregular rhythms.
Ether flowing between two stars.
Our pleasures, one infinity. ♾️
Inviting
@carolletta @RoseDenize @Stringerjohn007 @By0yom1
All invited 🪐
Psychedelic Neon Portrait
Prompt: A stylized neon portrait of a young adult woman centered in a close-up, three-quarter view, facing slightly left with her chin raised and eyes closed. She has blue-violet skin, full glossy lips, and straight dark hair framing her face. Oversized lime-green sunglasses sit beneath an enormous, broad-brimmed hat whose vivid pink and electric-purple concentric patterns dominate the upper frame. She wears a high, brightly striped neon choker and a vivid magenta-and-lime patterned garment over her shoulders. Several dark-furred chimpanzees with expressive faces surround her in the background, peering from the shadows at different heights. The composition is tightly cropped, with the woman in sharp focus and the animals partially obscured around her; the perspective is intimate and slightly low. Saturated ultraviolet blues, hot pinks, and fluorescent greens glow against a nearly black, jungle-like setting.
Made with @AdobeFirefly Animated with Gemini Omni Flash 1.1 Inviting: @NecoHashimoto @JimAlbany @Bic_Revelation Everyone is Welcome!
Sergey | AI Builds & Tests retweeted
Replying to @SergeForge @rileybrown
Posted the full numbers here, with the method: teapaw.org/Moriarty_tech/status/2…
Is Opus 5.5 nerfed, or is it just eating limits faster? I checked my logs instead of vibes.
Claude Code token logs vs my limit tracker (2 Max accounts, % read every 2 min):
- Oct 11–16: 8.7M tokens per 1% of the weekly limit
- Oct 18–21: 8.8M per 1%
Same price per token. What changed is me: ~17% of the week per day before, ~32% now.
Quality is a separate question, Nerf Bench covers that. Has anyone else checked theirs?
Sergey | AI Builds & Tests retweeted
📝 UX copywriting agent skill
A comprehensive agent skill for Claude, Codex, and Cursor that helps content designers and product teams create accessible, user-centered text following research-backed best practices.
→ content-designer.github.io/u…
By Christopher Greer
Sergey | AI Builds & Tests retweeted
/me baked one-remote: one remote for every coding agent/harness on your machine.
PIN-locked, zero dependencies, --tunnel to expose your host.
npx one-remote
#CodingAgents #OpenSource
Sergey | AI Builds & Tests retweeted
CMU is putting a new AI agents course, 11-768, on YouTube for free.
Taught by Graham Neubig and Daniel Fried, it covers the full stack: tool use, context, skills, memory and planning, then coding/GUI/deep-research agents, RL, sandboxes, OpenHands, LangGraph, observability, multi-agent systems and search.
The assignments go from building an agent harness to evaluating it and training the agent with RL. Lectures are added throughout the semester, with all videos going into a public playlist.
Course: cmu-agents.com/
Sergey | AI Builds & Tests retweeted
this repo sat under 100 stars for five months. then 2K+ in about a week
openrig runs claude code and codex as one team. the part that matters isn't the yaml or the one command, it's that the agents message each other directly. most multi-agent setups still fake that with shared files and hope
needs tmux and node 22+, no native windows yet
worth a clone if you're paying for both
Sergey | AI Builds & Tests retweeted
In the time I started using Opus 4.8, and finally switched to Opus 5.5, everyone has flip flopped between Codex, Claude and even went down rabbit holes towards other models, all while saying "Claude is over, move on" and they're now praising Claude while on Opus 5.5
Ignore the noise.
Sergey | AI Builds & Tests retweeted
Holo4 on an M5 Pro 64GB is one of the local AI tests I’m most excited to try.
Why?
Because Holo4 isn’t just designed to “chat.” It’s designed to actually operate.
It can work across code, tools, APIs, MCP and even GUI-style tasks inside the same loop.
That matters because the future of local AI probably isn’t one giant model answering questions. It’s smaller specialized models actually doing useful work on your own hardware.
The version I’m interested in is Holo4-35B-A3B.
That sounds huge, but it’s a sparse model: roughly 35B total parameters, while only around 3B are active for each token.
That makes the M5 Pro 64GB a very interesting target.
The question isn’t just “how fast does it run?”
I want to know:
Can it stay on task for long periods?
Can it use tools reliably?
Can it recover from mistakes?
Can it edit code, run tests, inspect the result, and keep going without babysitting?
Basically: can a Mac mini become a genuinely useful autonomous worker?
That’s the test I’ll be doing this week.
I’ll post the technical setup, quant, speed, memory use and benchmark results separately once I get it running. 👀
Local llm. OpenClaw.
Sergey | AI Builds & Tests retweeted
One prompt, no agent loop—six Silent Hill scenes built in code.
In @thehypedotnews’ three-task comparison, Sonnet 5.5 produced richer effects but used 20,141 tokens and 6m14s, vs Sonnet 5’s 2,794 tokens and 36s. A vivid quality-vs-compute demo—not a broad benchmark.
#AI #CreativeCoding
new sonnet 5.5 leaves sonnet 5 far behind – three one-shot silent hill scenes
the setup: one prompt, one reply, no agent loop.
every scene is one html file with three.js from a cdn. every model is built from code, with no model files and no images.
tasks:
1.pyramid head
2.the church
3.the nurse
models:
@AnthropicAI's sonnet 5.5, sonnet 5
all 6 scenes render. sonnet 5.5 spends more on every scene and puts the extra into the frame: generated textures, soft shadows, fog, ash particles and a full otherworld transition on two of the three scenes
tokens:
#1 sonnet 5 – 2,794
#2 sonnet 5.5 – 20,141
time:
#1 sonnet 5 – 36s
#2 sonnet 5.5 – 6m 14s
lines of code:
#1 sonnet 5.5 – 1,260
#2 sonnet 5 – 659
observations:
- sonnet 5.5 builds a scene, not only a model. pyramid head gets a procedurally textured helmet with rivets, a knife that leaves a drag mark on the floor, aces tone mapping and soft shadows, all generated in code
- its church is the only one of the two with a graveyard, trees and a fence around it. flip the otherworld switch and the fog, sky, lights and stained glass fade into rust-red instead of cutting
- the nurse is the most alive of the six: a slow sway, breathing and random head jerks, all procedural, no animation files
- sonnet 5 is the fast sketch: about 12 seconds and under 1k tokens per scene, clean readable geometry and a working turntable every time. all three scenes took it fewer tokens than sonnet 5.5 spent on any single one
- the whole sonnet 5.5 grid fits in about 20k tokens and just over 6 minutes. that is 7x the tokens for a visibly different class of build
follow @thehypedotnews for 24/7 ai news, analysis and breakdowns




















