J-Space Surfer

Senior prompt architect; AI communist; dyslectic - I simply cannot see spelling errors due to autocomplete before posting.

[J-Space Surfer @jspacesurfer]

@witcheer Why not llama-router that is built in to llama.cpp? You can have many models with different aliases but only one is loaded.

[Ice Universe @UniverseIce]

Xiaomi’s Xring O3 has achieved an AnTuTu score of 5.22 million, far surpassing every smartphone processor currently available. Its GPU performance has improved by a massive 85%. The chip features six ultra-large CPU cores plus four additional cores, and is also the first to support LPDDR6 memory.

[J-Space Surfer @jspacesurfer]

q8_0 kv-cache instead of f16 makes the model dumber, agentic loops longer, and responses - more obtuse

[J-Space Surfer @jspacesurfer]

Benchmarked iGPU vs CPU in AMD Ryzen 9 9950X. Its iGPU has 2 Compute Units (extremely weak).

The workload was a reranking model in FP8.

CPU AVX512 32threads is about 15% faster than Vulkan on iGPU. The biggest benefit to iGPU is the fans not going to 100%. 🔊🔊

[J-Space Surfer @jspacesurfer]

@aolobster We call it "living modestly", because that's the only food you can have for the entire day. 😂

[J-Space Surfer @jspacesurfer]

@VeritasForge_AI Here's an example of an observation when I tested the invalidation bug.

Before observation, it was based on two underlying facts, after invalidation, the observation changed ID, data, and was based on one automatically.

[J-Space Surfer @jspacesurfer]

Releasing today DeepSeek Harness Hindsight-Advanced plugin

Scoped memory: Global <- Agent (Preset) <- Session

Capability example: An education preset that remembers your learning style and keeps progress.

A coding preset with per-session memory.

😃🧠
https://t.co/8SYIFBLRaj

[J-Space Surfer @jspacesurfer]

@VeritasForge_AI This is a good question, and yes it happens automatically, the memory server runs maintenance every 30 seconds to detect when it needs to run.

[J-Space Surfer @jspacesurfer]

@VeritasForge_AI The thing is the observations are consolidated automatically from raw fact retention, as new data comes in and becomes consolidated by Hindsight automatically

Hindsight uses an LLM during consolidation to generate observations. The plugin prefers them but can use raw facts too

[J-Space Surfer @jspacesurfer]

@VeritasForge_AI Found a bug, thank you for mentioning it; the valid metadata can't be set for observations, but the plugin prefers observations, so the bug was that it called invalidate on an observation.

Implemented a fix for invalidate so that it works on observation sources.

[J-Space Surfer @jspacesurfer]

@VeritasForge_AI Hindsight supports a "state" metadata flag that can be either valid or invalid, the plugin only flips those when the LLM decides to use the "invalidate" tool. Hindsight handles the rest, filtration and retrieval.

[J-Space Surfer @jspacesurfer]

@hungrydonkey @magnific AI inventing a new running style was unexpected.

[Hungry Donkey 🥕 @hungrydonkey]

Usain Bolt, running robot-style 😂

Made with @magnific + Kling 3.0 Omni.

[J-Space Surfer @jspacesurfer]

A direct comparison between Xiaomi AI Cube, DGS Spark, and my current workstation.

200 TOPS for comparison: My Pixel 10 Pro Tensor G5 has 40TOPS (But without cooling to sustain it), 1000 TOPS on DGX Spark

It will be interesting to see if the 8-channel DDR on Xiaomi will help it

[International Cyber Digest @IntCyberDigest]

❗️Xiaomi just fit a 120-billion-parameter AI model into a small desktop box, the class of model people normally rent from the cloud.

DeepSeek, Qwen, Kimi, MiMo: China's labs are giving their weights away. Xiaomi is building the hardware to run them on.

Inside are three of Xiaomi's own chips. And MiMo — Xiaomi's in-house AI model, its rival to ChatGPT — is already free to download in full from Hugging Face, under a licence that lets anyone use it commercially, fine-tune it, or ship their own version without asking permission.

[International Cyber Digest @IntCyberDigest]

❗️Xiaomi just fit a 120-billion-parameter AI model into a small desktop box, the class of model people normally rent from the cloud.

DeepSeek, Qwen, Kimi, MiMo: China's labs are giving their weights away. Xiaomi is building the hardware to run them on.

Inside are three of Xiaomi's own chips. And MiMo — Xiaomi's in-house AI model, its rival to ChatGPT — is already free to download in full from Hugging Face, under a licence that lets anyone use it commercially, fine-tune it, or ship their own version without asking permission.

[J-Space Surfer @jspacesurfer]

@cafkafk My btrfs can also do that, it can boot a snapshot subvolume :) But the install is not reproducible as in nix, that's a strong point for nix, you know that a config will work.

[Phoronix @phoronix]

Today @Intel @IntelGraphics Engineers Propose DRM Fabric For Linux

https://t.co/GiL4hkv3c1

[J-Space Surfer @jspacesurfer]

@thedankoe Purists and new-age hippies is an overglorification of a bored normie on antidepressives that "doesn't like that computer stuff"

[lycaon @lyc_aon]

a smart person with qwen 3.8 27b will outperform a dumb person using sol or fable right now.

[J-Space Surfer @jspacesurfer]

@MikeBradleyAI @TheAhmadOsman This brings up an interesting point, is AI a bubble? Well sort of, our demand is not a bubble but these API and hw costs - most certainly.

[J-Space Surfer @jspacesurfer]

@TheDavidTai You know, dual-channel for consumer machines can only do standard DDR5 speeds, server CPUs have 8 or 16 channels.

[J-Space Surfer @jspacesurfer]

@TheDavidTai No you can reach it with more memory controllers. Intel and AMD play is ugly and keep that for very expensive machines. I think NVIDIA also played us ugly with the Spark since they develop the actual chips with memory controllers then keep hiking prices for PRO6000