One Station Plus a Sidecar Beats Two Stations at 16 Users

@alexzfunk reverse engineered the gen 1 Apple TV remote in about 45 minutes and published it at github.com/a-funk/open-appleremote: Linux over Bluetooth, full button control, tracking and microphone push-to-talk. His stack: “literally just opus 5.5 max effort ultracode in cli claude code.” The remote becomes a controller for SparkHome, his voice-native OS for a DGX Spark, where he reports 60 fps on Star Wars Galactic Racer and Control Resonant after Arm optimization. Audio records only while the mic button is held.

@alex_prompter shares a prompt built on Feynman’s advice to Caltech’s class of 1974 to report everything that might make a result invalid. It instructs Claude to mark every claim as checked, likely or guessed, and to list what could make the answer wrong. It must also name the one source to check first and refuse to fill a gap with a confident guess. @aron_brand answers that prompts are the wrong lever: “Fix the schema: no ‘unknown’ option, no nullable fields, and the model has to guess.”

@ashxhart releases TensorFold 1.0.3 with Living Weights, about three months of engineering that lets a model update its own weights as it is used and grow a persona in real time. @taofeng_field’s test came back weak: on a document holding 5 facts, only 5 of 11 recall questions were answered correctly, and 1 of 16 related questions was contaminated. @namakosan3 filed an issue that it fails to learn Japanese effectively because it treats spaces as delimiters.

@ivanfioravanti trained a decision model on a single DGX Spark in about 2 minutes using Unsloth, with Qwen3.5-2B as the base model. The demo video shows @MiaAI_lab’s sparkDash v2 monitoring the box. @danielhanchen of Unsloth says he will fix the model’s slow loading. @vermaizer notes Unsloth’s own benchmark reports about 40 minutes of training for 81% test accuracy, and asks what accuracy the two-minute run reached on a held-out test set.

@satyanadella argues we cannot trace model behavior to specific training inputs or weight configurations, yet agentic systems hold sensitive data and take mission-critical actions. His answer: separate the supply of intelligence from the authority over it, and put the controls outside the model so it cannot tamper with its own permissions. He calls CoT transparency non-negotiable, saying “Neuralese” cannot justify opaque reasoning, and wants models treated as insiders: assumed compromised, with an authorized person able to pause one mid-task.

@saxena_puru reads falling token prices against sticky-high GPU prices as Jevons paradox in real time. Efficiency, competition and Blackwell-scale gains make inference cheaper, which expands agentic workloads, coding, automation and users instead of cutting demand. Total token volume grows far faster than unit costs fall, so the scarce resources, GPUs, HBM, power and data centres, stay tight. He expects GPU makers to keep benefiting, durable demand for memory, networking and power, and a margin squeeze for hyperscalers and neoclouds if token declines outrun volume growth.