Morph 2K pre-orders are now open!
A $199 analog video scaler for retro gaming with composite, S-Video, SCART, component, and VGA support. Incredible video quality, CRT simulation, auto-sampling, Wi-Fi updates, and more.
To celebrate launch day, we're giving away one Morph 2K.
To enter:
✓ Follow @PixelFXco
✓ Like & repost this post
Winner selected June 20th.
https://t.co/xQqM20tdnt
@Maxnotcreative@PubWanghaf Is that what you think is happening here? Is this more likely a real video of a six fingered woman sat next to a fat man doing wild tongue gymnastics?
Moonshot just told you the agentic era requires swarms, not single-agent chains, and nobody’s repricing what that means.
The benchmark headline says “50.2% on HLE, beats Opus 4.5.” Look closer at how they got there: Agent Swarm. K2.5 spawns up to 100 sub-agents executing 1,500 tool calls in parallel. That’s 4.5x faster than sequential execution.
This tells you something critical about where frontier AI is headed. The models themselves are becoming orchestrators. K2.5 trained a “Parallel-Agent Reinforcement Learning” system that teaches the model when to parallelize vs when to reason deeply. Without predefined roles. Without hand-crafted workflows. The model decides.
Why does this matter? Because everyone building agents right now is doing it wrong. They’re writing deterministic workflows in LangChain or CrewAI, manually defining which agent does what. Moonshot trained the coordination layer into the model itself.
The irony: Opus 4.5 still wins on pure coding benchmarks (76.8% SWE-bench vs K2.5’s 76.8%). The gap is in agentic tasks where execution time matters. BrowseComp, DeepSearchQA, HLE. Tasks where “thinking longer” loses to “thinking in parallel.”
OpenAI, Anthropic, and Google are all racing to build agentic capabilities. But they’re optimizing reasoning chains. Moonshot is optimizing swarm coordination. One of these approaches scales compute linearly with task complexity. The other scales exponentially.
K2.5 is also open source under modified MIT. Anyone building AI products can now access swarm orchestration at $0.15 per million input tokens.
The real question: How long until the frontier labs ship native swarm capabilities? Because right now, the best orchestration model in the world isn’t from San Francisco. It’s from Beijing.
@lydiahallie I fed it a 2.5gb file of a concert, along with a time stamped song list I pulled from the comments, and it made me an audio album complete with metadata. Neat
@jasonfried@TPLINKUS Look at this, man: https://t.co/jxZBZwsUdf. You’re telling me this is an acceptable tradeoff? How many people use double tap compared to just turning the damn lightning and EXPECTING IT TO WORK QUICKLY
@jasonfried The geniuses as @TPLINKUS decided it was worth implementing a 200ms delay on a lighswitch press to ensure the user wasn’t going to use the double tap feature — even if that feature was disabled. 200ms delay. On all smart dimmers. I campaigned so hard they gave me custom firmware