A post on X says DFlash plus a new method called Weaver made a nine billion parameter model four times faster on an M5 MacBook Pro. I installed Mirai's free uzu engine and tested it on my M4 Pro MacBook Pro against plain MLX. Same model, four prompts, one engine at a time. Results: 3.2x on the post's own demo prompt (merging two sorted lists), 2.5x on another coding question, 2.1x on plain English, 1.9x on a shell script. 4 to 6 pieces of text per step, against the 7.5 the post claimed. Two disclosures matter. Mirai wrote the Weaver paper and ships the engine. The paper's 4.08x was measured on one Nvidia B200, with nothing measured on a Mac. Caveats on my side: an M4 Pro, not the M5, and I couldn't switch Weaver off inside Mirai, so this is Mirai against MLX, not Weaver on its own. For the group: the gain tracked how predictable the text was. Code won, prose and shell lost. What's your workload mostly made of, and has anyone tried oMLX's DFlash setting yet? Video: https://youtu.be/g6zGC44Oy-g Paper: https://arxiv.org/abs/2607.06763 Engine: https://github.com/trymirai/uzu

A semantic cache hands back a stored AI answer when a new question looks close enough to an old one. "Close enough" is one number. Redis ships LangCache at 0.85. We ran 243 support-question pairs through the small embedding model we use every day, on a Mac Studio. Cancel my subscription against pause my subscription: 0.87. Different answers. Cancel my subscription against "I want to stop being charged every month and end my plan": 0.68. Same answer. At 0.85 our model caught none of the 101 real rewordings and wrongly matched 29 of 102 look-alikes. Our pairs, built to probe the hard cases, on our model, not LangCache's. The point is that a threshold only means something for one model on one kind of traffic. Method, charts and limits: https://semanticcache.novcog.us.com/the-threshold Video: https://youtu.be/XW-VsLfYiqg For the group: if you run a cache or a RAG bot, what's the look-alike pair in your niche that would do the most damage if it matched? Post it below.

Moonshot shipped Kimi K3.1 Pro on Thursday. I went through the launch claims the next day with independent results and a few API runs of my own. What holds: fourth of 215 on Artificial Analysis's index, and 67.2% on Terminal-Bench 4.0 in neutral software, a little above the 65.7% Moonshot reported from inside Claude Code. What doesn't: the one-third price cut mostly disappears because it writes far more tokens, so a finished task costs about what K3 did. Highspeed promises 2 to 3 times faster and measures about 1.7. And the weights aren't out yet. They're promised by November 4. Moonshot also says it removed hacking material from training. On its own exploit test the release still scores 81% against K3's 33%. For the group: most of Moonshot's coding scores were run inside Claude Code. Does testing in a rival's tool make you trust the numbers more or less? Full audit: https://k31pro.novcog.us.com/ Video: https://youtu.be/tNEDJx2QEV8

Shaun Anderson at Hobo Web says GEO has been mathematically disproven. I read every source he cited, and the ledger is up. What holds: the 2023 GEO paper's prompts allowed invented statistics and sources, its Perplexity test used uploaded files, and most rewriting tricks fail in later benchmarks. What doesn't: a quote attributed to C-SEO Bench isn't in the paper (four versions checked), the "live web" study is an offline lab, and the Toronto study found 97% of ChatGPT's retrieved pages weren't in Google's top fifty. The one that stings for us: within a site, schema and FAQ blocks showed no significant citation lift. I've said structured data helps. Let's argue about what that study does and doesn't measure, in the comments. Full audit: https://geocoffin.novcog.us.com/ Video: https://youtu.be/6xIpBRgqaWc

If you've got a Mac Studio and a spare Mac mini and you've wondered whether chaining them gets you a bigger model, the answer changed in December and most of the advice out there hasn't caught up. Three things get called "Mac clustering" and they're not alternatives. They stack. At the bottom is RDMA over Thunderbolt. It's not software you install, it's a macOS capability. One Mac reads another's memory directly, without the OS copying it through a network stack. macOS 26.2 or later, Apple silicon, Thunderbolt 5. Apple's technote TN3205 has the requirements. In the middle is MLX, Apple's own framework, which actually runs the model. Its distributed layer has two backends and the backend IS the strategy. On top is EXO, which finds your machines and decides how to split the work. It runs on MLX. You can skip it entirely if you want to write a host file yourself. Here's what I think most people are getting wrong. There are two ways to cut a model across machines. In a relay, each machine holds a block of layers and work passes down the line, so one machine computes while the others wait. More machines, bigger models, no extra speed. That was the only option for years, and it's where "clusters are for capacity, not speed" comes from. That advice was correct. In the other mode, every machine holds a slice of every layer and they all compute at once. The work divides. It needs hundreds of synchronisations per token, which only became viable when the fast link shipped. Exo Labs publishes up to 1.8x on two devices and 3.2x on four. The result I find most convincing isn't a benchmark. Jeff Geerling ran two programs on the same four Macs with the same cables. One sped up as he added machines. The other slowed down. Same metal, opposite direction. Two things worth knowing before you spend anything. The feature ships switched off. You enable it from Recovery mode with one command, on every machine. Miss one and nothing fails, no error, no warning. The cluster just runs the slow way and you conclude the technology is overhyped.