gen-x : lost boy : source-destination-unknown : father-husband-son : analog-2-digital sojourner : work in tech : x-talk is my own
Near OK Route 66
Joined April 2011
- Tweets11.2K
- Following370
- Followers581
- Likes14.7K
Our GLM-5.3 Flash lane (recipe 1104, 2× DGX Spark TP=2) came up at 498,688 ctx instead of 1M after a clean cold boot. System healthy, memory healthy. What we found + how we fixed it in ~2.5 min: 🧵 @MiaAI_lab
2
2
60
Side find: two identical-settings 1M launches split leftover memory differently — Nov 9: 12.5 GiB KV + 3.7 GiB prompt cache. Today: 17.1 GiB KV + zero prompt cache (0-byte budget warning). Same window, different posture, and the serve line doesn't say why.
1
14
Takeaway: if your served ctx shrinks and STAYS shrunk across relaunches, check .state/context-fit.tsv before suspecting host memory or the engine. Full evidence + repro table filed here: github.com/MiaAI-Lab/GLM-5.3… @MiaAI_lab
20
Well @mem0ai is changing their free plan from November 25 and now needs your own LLM key but no option to use a self-hosted LLM for that rather than their limited options (docs.mem0.ai/platform/pricin… ). Choice: OSS mem0 or another option (prefer 0 or low-cost). What are others using?
36
Hey Joe 💻🎸🎭🏀🏈🛣️ retweeted
You might want to make sure you have a few open-weight models backed up and some basic gear to run them. Who knows what really happened at NVIDIA, but the world just lost one of free AI's biggest proponents.
This interview by @pmarca explains what is really behind Jensen Huang's ouster as Nvidia CEO.
Hey Joe 💻🎸🎭🏀🏈🛣️ retweeted
Thank you @JensenHuang for 33 years of bold leadership. You built a small startup into America’s most valuable company, with relentless improvement. Your grit and perseverance should always be remembered. Your personal generosity with time and advice to founders like me has secured a bright future for America and AI.
Hey Joe 💻🎸🎭🏀🏈🛣️ retweeted
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
I'm such a fan of him and love his moxy but for the greater good the team needs him to be a healthy QB1. Please no more injuries to any of his 4 limbs.
Bucs: “Baker, you won’t complete a pass today, but we need you to get physical in the run game to help us win”
*Mayfield hits smelling salts*
Baker: “Say less.”
45
Good lawd, @MiaAI_lab's 2x Spark recipe for GLM 5.3 Flash EXL3 with Tensorfold v0.6.0 is fast. I regret not swapping out Qwen3.8-Flash-Next for it sooner.
41
