This is exactly what I’ve been waiting for 🔥 Meta just announced its new VR Glasses. Basically an Apple Vision Pro-like experience, but MUCH smaller and lighter. • Only 100g • Micro-OLED displays • Snapdragon Reality Elite • Eye + hand tracking • External compute/battery puck • $1,299 This is the direction I want VR to go. Smaller, lighter, more comfortable, while keeping the full spatial computing experience. Launching Spring 2027. I’m getting these 100% 🙌

Benchmark suite:

Rick Rubin understands what many artists still refuse to accept. Creativity isn't in moving the brush. It's knowing what to paint. AI doesn't change that.

Phosphor glitch with Qwen 3.8 Flash Next before bed js / canvas / 0 shot / 1 turn you know the drill It follows pretty complex directions very well.

Made an interesting discovery when re-aligning Timewave Zero to INTERNET DENSITY! Revealing the potential TRUE TIMEWAVE ZERO on April 15 2048! According to the alignment this year's novelty is just about to start ramping up! Try it out on timewave-zero.com there is cool customizable merch as well!

🚨Qwen4家族首次曝光！！ 刚刚，在2026年云栖大会的开幕式上，新任@Alibaba_Qwen LLM负责人刘大一恒官宣了即将到来的Qwen4家族！ 包含Qwen4-Max Qwen4-Flash&Qwen4-Plus 还有Qwen4-27B！！！ 未来Qwen会训5-10T的模型

Latest recipe for the Qwen3.8-27B Dynamic DFlash2 n=10 for c1-c2 n=8 for c3-c4 n=7 for c5-c8 n=6 for c9-c10 … and so on since DFlash acceptance is inversely correlated to concurrency You get the best of whatever concurrency you are choosing to use. github.com/AEON-7/Qwen3.8-27…

And just to top it off Introducting: A Piece of Human Cortex, Remembering HumanBrain-Pollard-H01 — 13,473 human neurons as a language model's memory, 100% exact recall In 2024 a collaboration between Harvard and Google published H01: a cubic millimetre of human temporal cortex, reconstructed from electron microscopy at nanoscale resolution. Roughly 57,000 cells. About 150 million synapses. A petabyte of imaging, for a volume smaller than a grain of rice. It is the finest-grained map of human neural tissue that exists. And the obvious question, once you have been using a fly's connectome as a language model's memory, is whether the human one works too. It does. But the interesting part is what had to be got right first, and what the result does not mean. Getting the graph right The naive pipeline gives you a graph that is confidently wrong. H01's soma table covers all ~57,000 cells, and most of them are not neurons. Roughly 32,000 are glia — astrocytes, oligodendrocytes, microglia. Run the synapse extraction without filtering and the single largest edge class in the entire dataset is astrocyte → pyramidal, at 52,177 connections, comfortably ahead of pyramidal → pyramidal at 38,713. Astrocytes do not form chemical synapses onto pyramidal cells. What is happening is that astrocyte processes physically wrap real synapses, and the automated detector reports the wrapper. A graph whose commonest connection is biologically impossible is not a connectome; it is a map of a detection artifact. Neurons only. The second trap is sign. Excitatory and inhibitory matter enormously — they are the difference between a network that settles and one that runs away. H01's detector makes its own E/I call per synapse, and it is tempting to use it. We checked it against Dale's law — the principle that a neuron releases the same neurotransmitter at all its terminals, so a pyramidal cell is excitatory everywhere and an interneuron inhibitory everywhere. They agreed 57.5% of the time. Barely above a coin flip. We take the sign from cell type, and report the disagreement rather than quietly picking the one that flatters us. The third fact is scale. Of H01's ~166 million detected synapses, only about 0.3% connect two cells whose soma is inside the volume. The rest land on neurites cut off at the block boundary — a cortical neuron's arbour extends far beyond a cubic millimetre. So the extractor streams all 166 shards, keeps that 0.3%, and discards each shard as it goes: peak disk of 200 MB rather than 33 GB. What survives is 13,473 neurons and 75,452 connections, carrying 114,227 synapses, with cortical layer recorded for every cell. The result Same trainer as the fly brain. Same recipe. One argument changed — the graph. No brain — fact outside the window0.0% With the human brain100.0% Control — a word the document never contained0.0% Control — the brain read a different document0.0% Live state17.7 MB, constant at any length 100% at step 44 from a cold start, held flat thereafter. A six-letter string stated once, buried under filler, asked about far beyond a 128-token attention window, with words generated fresh every sample so there is nothing to memorise. Human against fly cellssynapsesstate100% at Fly MaleCNS (pruned core)8,552300,88011.2 MBstep 26 H01 human cortex13,473114,22717.7 MBstep 44 The human graph has 1.6× the cells and far fewer synapses per connected pair — 1.51 against the fly's dense wiring — because the fly connectome is a complete brain and H01 is a fragment of an enormous one. It converges a little slower. It arrives at the same place. Experiment, It would be easy, and wrong, to read this as "human connectome enables byte-exact memory." huggingface.co/PollardWeight…

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here: darioamodei.com/post/we-must…

This benchmark just came in and it blew my expectations out of the water! This is for the NVFP4 Mixed quant running on a SINGLE DGX SPARK! Best score of all models of all time on Aeon Bench, beating DeepSeek v4 and GLM5.2 and of course even stock Qwen3.8-27B! aeon-bench.com/share/aeon-7_…