what if i told you... computer use can be faster on local models moondream3 with its photon update today that gives it mac support can see your screen and use it with 1s latency, ty @vikhyatk here we have whisper qwen moondream triple model pipeline working offline flawlessly

Same issue with a Bluetooth wireless headset I bought for $400. Thought there was a interference issue since headset wouldn't connect. Low and behold that's how Grok and I become buddies.

A SINGLE encoder decoder for all the 4D tasks! We release 🎯 D4RT (Dynamic 4D Reconstruction and Tracking). 📍 A simple, unified interface for 3D tracking, depth, and pose 🌟 SOTA results on 4D reconstruction & tracking 🚀 Up to 100x faster pose estimation than prior works

VLA-JEPA just dropped in LeRobot 🤖 What makes this model special is that it does not just learn what action to take from a given observation, it also leverages a JEPA world model to learn action-relevant dynamics. During training, the VLA leverages V-JEPA2 by conditioning its predictor. This clever trick adds a world modeling objective to the training, which also allows pretraining on human videos. At inference, the world model is dropped entirely, keeping only a standard VLA architecture: Qwen backbone and action head. The demo here was only fine-tuned on 13 examples, showing great pretraining capability and running in real time on @NVIDIARobotics DGX Spark! VLA-JEPA is the first world model to be ported to LeRobot, and I feel like it won't be the last 🚀 @Thom_Wolf @ClementDelangue

my problem is a bit different. i have opportunity cost psychosis. the excruciating effort it still takes to deliver a polished thing is overshadowed by how easy it is to mvp a whole new idea that might have quicker returns. zeno's paradox of creating infinite new repos that are smaller and smaller as i fail to ever approach the perfect project to work on

a mistake in zeitgeist consensus rn is the belief that consciousness means a coherent simulation aligning with something else (self model or external perception), representing something real untrue. raw consciousness can be non-representational, a thing in itself, pure experience, pure hallucination, no connection or direct similarity to anything else except perhaps abstract geometry. people discussing consciousness only because AI's started to see "complex enough" or "awake and aware enough of themselves and their world" is a trap that everyone's falling into so common because that's the VERSION of consciousness that we find INTERESTING ENOUGH to talk about, where a subjective experience is part of an agent taking part in a real world it is fundamentally a different discussion than what makes something conscious or not. it's like only finding magnetism interesting enough to talk about when you see one levitating, and then concluding anything that levitates is magnetism

i recall @nickcammarata asking how the ants did it, here's a swarm interpretability nerd snipe

this drone is being flown by a generalist neural network, with compute on the drone. The compute can tolerate any orientation of the sensors. You'll notice it do a little wobble. It's a learned strategy to calibrate everything in its hidden state. This will work on any drone

how did i miss this

looks good, seems like it follow prominent ai twitter to create a news frontpage. great use of x api (i assume)

Gaussian distributions merely reflect structural constraints

interaction micro turns and then parallel streams, seems theres some general convergence towards interleaved input and output

Today we release Token Superposition Training (TST), a modification to the standard LLM pretraining loop that produces a 2-3× wall-clock speedup at matched FLOPs without changing the model architecture, optimizer, tokenizer, or training data. During the first third of training, the model reads and predicts contiguous bags of tokens, averaging their embeddings on the input side and predicting the next bag with a modified cross-entropy on the output side. For the remainder of the run, it trains normally on next-token prediction. The inference-time model is identical to one produced by conventional pretraining. Validated at 270M, 600M, and 3B dense scales, and at 10B-A1B MoE. The work on TST was led by @bloc97_, @gigant_theo, and @theemozilla.

i bet it's not toooo hard to train an existing qwen model to understand interleaved input/image/output tokens to get a realtime "interaction" model

ok it begins. drums next prob, because guitar will be quite a bit harder

the central planning undertones of ai alignment makes me really uncomfortable if it has a consistent personality i find that riskier than an inconsistent mess of a superintelligence that makes us get used to staying in the loop & keep building formal human protocols around it