Nvidia is in advanced talks to acquire AI startup Inception at a valuation of at least $500 million, potentially with performance-based earnouts. Founded in 2024 by Stanford, UCLA and Cornell professors, Inception builds diffusion-based large language models; its Mercury model claims sub-300ms time-to-first-token and 5-7x throughput via parallel generation. If completed, the team could join Nvidia's Nemotron open model efforts and extend diffusion techniques to robotics and physical AI.
Autoregressive Generation
自回归生成
A generation method that predicts the next element or action step-by-step based on preceding sequences.
Curated Intelligence
A research team from Nvidia, NYU, UC Berkeley and others introduces TacitVLA, showing that reasoning text in vision-language-action models mainly matters during training rather than at inference. The method matches explicit-reasoning success rates on four manipulation and autonomous-driving benchmarks while cutting latency by up to 3.4x, and sets new records on long-horizon memory tasks.
Yann LeCun's new paper mathematically proves the inefficiency of autoregressive next-token prediction and challenges the scaling hypothesis. By formalizing the Platonic Representation Hypothesis into a theorem, the study demonstrates that a 109-million-parameter world model can outperform a 72-billion-parameter LLM, providing strong theoretical backing for the JEPA architecture.
Inception Labs released Mercury 3, claiming it outperforms GPT-6 Luna and Claude Haiku 5.5 on several benchmarks while running 4-8x faster. The model uses a diffusion architecture for parallel text generation, showing strong gains in instruction following, though the overall lead is narrower than advertised.
TypeSafe AI released Jev in September 2026, a novel model based on the Transformer architecture that abandons text generation. Specializing in classification, scoring, and judgment tasks, Jev outputs probability distributions directly to achieve millisecond-level inference, significantly faster than traditional LLMs. The article analyzes the architectural evolution from BERT/GPT to Jev, highlighting its core innovation: retaining the Encoder's understanding capabilities while removing the Decoder's generation mechanism, offering an efficient alternative for specific discriminative scenarios.
This 30,000-word comprehensive primer explains Large Language Models (LLMs) from first principles for readers without machine learning backgrounds. It traces the logical thread from information theory and next-token prediction to the Transformer architecture, clarifying how simple probabilistic modeling leads to emergent intelligence. The text bridges theoretical foundations with product implementation, offering a structured mental model of how LLMs function.
The vLLM team conducted systematic benchmarks of speculative decoding on AMD GPUs, validating the mechanism where multiple drafted tokens are verified in a single target-model pass. Results indicate that throughput gains vary significantly depending on drafting methods, proposal lengths, model families, and acceptance behaviors, rather than offering uniform acceleration. This study provides granular engineering reference data for optimizing LLM inference on heterogeneous compute infrastructure.
This tutorial demonstrates implementing EAGLE3 speculative decoding on AMD Instinct GPUs using vLLM and AMD Quark to address autoregressive latency bottlenecks. By employing a lightweight draft model for parallel token prediction verified in a single target pass, the method significantly boosts inference throughput for MoE models like Kimi-K2.5 while preserving lossless output distribution.
Google introduces a method to retrofit Multi-Token Prediction (MTP) onto existing 'frozen' Gemini Nano models for Pixel devices. This approach addresses mobile autoregressive bottlenecks by improving inference speed and energy efficiency without retraining the base model.
Hugging Face released LeRobot v0.5.0, its largest update to date, introducing first-time support for humanoid robots and reintegrating autoregressive Vision-Language-Action (VLA) policies. The release also features significantly faster dataset processing and includes over 200 merged PRs, aiming to comprehensively expand the capabilities of the robot learning ecosystem.