Marc Seal | Aevonix Research

🗞️The write-up and exploit to jailbreak iOS 27.1 and below will be released really soon. No doubt that Dopamine and Relaxing teams will look into it📣 nitter.gq/pwnlambda/status/21176…

Introducing the Microdock! 😅 The first self charging dock prototype 🦆

🔥 @cfontes got the full 753B GLM‑5.3 generating code at 73.6 tokens/sec on four DGX Sparks at 1M CTX. 🚀 That’s the FASTEST single-stream code result I found across the four public recipes I compared. When I released my 3.38 bpw SAGE‑EXL3 quant, we had all 19,200 routed experts in 319 GB. Chris funded the compute himself, then spent four weeks working on what happens after the weights load. Here’s what his TensorFold TP4 + DFlash2 recipe delivers on one request at a time, measured at p50: 💻 69.2 tok/s on code ✍️ 55.7 tok/s on prose 🧩 73.6 tok/s on structured output For perspective, the other published single-stream code figures I found were 57.8 tok/s from @u1tra_instinct , 46.6 from @Knapcio and ~43 from @bertholomusai . Each used its own benchmark harness. The number I care about most for my own agent sessions is 19.4 seconds. By turn six, the conversation had grown to 15,930 tokens. A cold turn took 85.3 seconds. Chris’s prefix cache reused 12,288 tokens of history and brought that same turn down to 19.4 seconds. That’s over a minute saved on one turn, with a 1M-token context window configured for the session. On my independent fidelity check, the quant picked the same next token as BF16 9,521 times out of 10,240 (93%) with a perplexity increase of 2.1%! @cfontes Chris, thank you for putting your own time and money into this and publishing the exact setup so the rest of us can run it. 👏 🛠️ Chris’s four-Spark recipe huggingface.co/cfontes/glm53… 🤗 My SAGE‑EXL3 weights huggingface.co/vcruz305/GLM-…

Today we’re introducing Zyphra Hierarchical Attention (ZHA), a path to true subquadratic attention. ZHA searches a learned tree of compressed summaries, with projected speedups of 35× over dense attention and 5× over MSA at 10M tokens, increasing to ~400× and ~55× at 1B. 🧵

Following my earlier post about the M5 Ultra 512GB Mac Studio, more than two dozen people asked me to set up crowdfunding for it. I always try to pay my own way, so asking for help at this scale doesn't come naturally. But I couldn't make the numbers work this time, and my loan got rejected after placing the order. I'd love to get one for TensorFold and to see how far we can push it. If you'd like to help towards it, I've put the goal on Ko-fi below. Please only contribute if you're comfortable doing so. ko-fi.com/ashxhart/goal?g=2 Thank you to everyone who's encouraged me to do this, and to everyone already supporting TensorFold. I'm incredibly grateful and will continue to push TF. ❤️

Have Codex optimize your iPhone Home Screen with Computer Use + iPhone Mirroring Use the following prompt: Help me make my iPhone Home Screen more efficient. Using iPhone Mirroring with Computer use, review my current layout and iPhone-specific Screen Time data. Ask which apps I use most, which need one-tap access, and which actions I already access through widgets, shortcuts, or hardware buttons. Before suggesting moves, ask which app positions I know by muscle memory and which placements I want to preserve. Don’t assume my most-used apps are the ones with the most screen time. Research relevant best practices for app placement, folders, widgets, reachability, and notification badges. Explain the tradeoffs behind your suggestions. Generate 2–3 layout previews for me to review before making any changes.

Using a stick of RAM to stir my coffee this morning, fuck you, Sam Altman

🤯 I don’t know where to even start, is this even real? 🥹 I could never have dreamed that TensorFold would bring this much support and so many good people into my life. To everyone who's followed, subscribed, donated through Buy Me a Coffee, lent me hardware to test on, contributed to the project or sent me some encouragement, thank you so much. 🫶🏼🫶🏼 I've met some incredible people over these last few months. The time and kindness you've shown me mean a lot, and I'm incredibly grateful for all of it. Thank you for being here ❤️

Glycine + NAC for aging (fascinating study) A placebo-controlled trial PMID: 35975308 showed: 16 weeks of glycine + NAC (100mg/kg of each a day, about 7g each for a 70kg person) pushed almost every aging marker measured back toward young adult levels. The placebo group didn't move. oxidative stress (F2-isoprostanes) -> down 72%, back to the level of the young adults IL-6 -> down 78% TNF-alpha -> down 54% hsCRP -> down 41% insulin resistance (HOMA-IR) -> 10.8 to 3.9 walking speed -> 1.13 to 1.34 m/s, same as the young adults grip strength -> up 4kg in both hands systolic blood pressure -> 132 to 124 waist -> down 3cm Why it works: glycine and NAC are the two raw materials for glutathione, the body's main antioxidant. Older adults can't make enough of it. Give the body the parts and it makes its own. No side effects reported.

My DeepSeek v4.1 Flash recipe can now run its own code mid-response ⚡ Introducing Compute Lanes. (Early alpha.) Same engine, but when it's on, the model can fire compute tasks on the DGX Spark's CPU cores while it's still generating the response, and the result feeds straight back into the same generation. No client round trip. Basically, an internal tool call the model triggers itself so it doesn't have to finish, wait, and go again. Across 40 coding tests, same 38/40 solved: - 9.67s per solved task vs 11.24s - 1 request instead of 1.53 average - 9% fewer tokens That's 16% faster across all 40 tests. idea came after seeing what @divagr1925 was able to ship in the weights. What if we could do the computation before the model's finished responding? OHHH, and with lanes toggled off, the model itself is faster too thanks to @bertholomusai's KV mirror: - code 107.2 tok/s (+5.1%) - prose 48.6 tok/s (+3-4%) It's on by default in the recipe but not 100% stable yet. More updates once it is. pr's in the comments if you want to test it and read how it works.