Monday morning is the perfect time to set goals for the week ahead. I like setting up progressive goals. Daily goals contributing to weekly goals contributing to monthly goals, etc. Of course not all goals need to be progressive. What are your goals for the week?

Zig on macOS is cool. 😎 mlx-serve (github.com/ddalcu/mlx-serve) provides the inference, nullclaw (github.com/nullclaw/nullclaw) provides the agent. 100% local 200% fun 🎉🥳

Which MLX engines do you actually use? I can't test and eval them all 🤷🏻‍♂️ Community poll: pick up to 5 of 19 (oMLX, LM Studio, Nativ, MLX-Serve, …). Add your chip (M1 to M6) and RAM if you like, to see what each setup prefers. No login, live results, open for 7 days: poll.devocracy.it/mlxengines Please repost to spread the word and don't cheat out there! 🧐

The two biggest Apple fumbles under John’s leadership so far: 1. Apple first said the new Photographic Styles is coming to iPhone 16 and 17 series. But later completely backtracked and made it exclusive to the 18 series. 2. They are keeping the new Readiness feature in watchOS exclusive to Apple Watch Series 12 and Ultra 4. Even tho older watches are completely capable of calculating it. Older models also support HRV, training load, vitals, sleep score, and other metrics they use to calculate readiness. They could’ve totally made it work. I hope these decisions are not specifically from John and do not reflect what he’s gonna do at Apple in the coming years.

THIS is INSANE. 🤬 Spotlight with macOS 27 Golden Gate cannot search Settings. On a brand new Mac Studio M5 Max. Come on Apple. Do better! @AppleSupport @Apple @johnternus

Cancelled my Codex plan. It burned through the tokens before the work was done. Last night: about five hours in Zcode + GLM-5.3, merging PRs, deploying, smoke testing. It held. The stack that survives for me isn’t the most famous harness and LLM; it’s the one I can afford, run, get real work done.

Had an "interview" for a blockchain project last week. Camera was on, we're chatting, and the guy tells me to clone a GitHub repo and run it locally before we go further into the technical round. I said sure, but first can you do me a favor hold up 3 fingers in front of your face for me real quick. He froze. Didn't move. Just sat there for a few seconds before the call cut off and he blocked me. That's when I knew. A real interviewer doesn't glitch out over a random ask like that. A deepfake/AI overlay does. These "run this repo" scams are getting scary common in crypto and dev hiring right now. The setup is always the same: - flattering DM - real-sounding project - rushed timeline - a "quick step" before the call that's really just remote access or a credential stealer in disguise. If someone wants you to run code or install something before you've even had a real conversation, that's the whole scam. Trust the instinct. Stay safe out there.

About 5 hours in Zcode tonight fixing issues, merging PRs, deploying and smoke testing. Never missed a beat. Zcode + GML-5.3 is rock solid. Thanks @Zai_org!

MiMo-V3 is getting a new architecture. The core of it, HySparse2, is out today. Less prefill, a smaller KV cache, better long-context retrieval—and we got all three at once. Compared with MiMo-V2.6's Hybrid SWA architecture: • 5.02× lower prefill FLOPs at 1M tokens • 4.5× smaller KV cache at 1M tokens • Better MRCRv2 and RULER-v2 scores, plus lower AgentPPL and LongPPL Why build a new architecture? Agentic inference is a very different workload. Each round, a short action can return a long observation that needs to be prefilled, while the context keeps growing. That puts prefill cost, KV-cache size, and retrieval accuracy on the critical path at the same time. HySparse2 tackles all three with two levels of KV sharing: • KV Bridging: Following YOCO, full-attention layers in the cross-decoder build their K/V from self-decoder hidden states. • KV Reuse: Within each hybrid block, sparse layers reuse the preceding full-attention layer's KV cache and selection indices. Two more changes: token-level selection replaces block-level selection, and a forced window of recent tokens replaces the separate SWA branch, so local and global tokens share one KV cache. Since all cross-decoder KV caches now come from the self-decoder, prefill can stop once the self-decoder finishes. Paper: arxiv.org/pdf/2609.26368

People are now selling “shared agent memory” as a product. Most of it is still one agent talking to itself. The failure I care about is split-brain across tools: harnesses that forget each other’s tragedies always-on agents that invent a second set of facts A memory product that can’t be read by Claude Code and Pi and a *Claw agent is just a nicer tab. Working agreement first (AGENTS.md, PLAN.md). Then one store. Then docs every agent can actually reach. Memory that never leaves the session is still archaeology.

Yes

My guess. 90%+ of users have no idea there are great open weight models available that cost less than Anthropic, OpenAI, and X. OK, maybe 90% is an overestimate since everyone in China seems to be using Doubao.

Fair review.

If you've followed me and I've not followed you back, it could be because I'm up against the following limitation.

Oh. MacOS 27.2 beta has a new icon/animation for audio adjustment. I like it.

Another example of Xiaomi MiMo in creative work. Xiaomi MiMo-V2.6-Pro generated an orchestral piece featuring around ten instruments based on creative requirements, then autonomously converted the score into MIDI. The process demonstrates its understanding of different instrumental roles and orchestration, as well as its ability to apply musical knowledge to melody creation and overall arrangement. #XiaomiMiMo #XiaomiMiMoV26 #AIGeneration #XiaomiAI