Azalia Mirhoseini

@azaliamirh

Founder @RicursiveAI, Asst. Prof. of CS at Stanford. Prev: DeepMind, Anthropic, Brain. Co-Creator of MoEs, AlphaChip, Test Time Scaling.

following
followers

Ricursive Intelligence aims to automate chip design with AI

TechCrunch reports its founders are developing AI to learn from each chip-design iteration and improve later designs.

1 Source, 6d ago, first seen 6d ago

Next-Concept Modeling is claimed to be up to 31× faster than similar-sized diffusion models on summarization and translation

Its creators also report 10–21× faster unconditional generation on three datasets, with comparable or better generative perplexity.

23 Sources, 1d ago, updated 1d ago

ThunderSync is claimed to make reinforcement learning up to 1.9x faster without stale policies

A post introducing the method reports that it matches the performance of synchronous updates on long-horizon agent tasks.

1 Source, 7d ago, first seen 7d ago

AI-as-compiler approach claims speedups over Triton baselines on B200

A post describes a model that translates source code such as Triton directly into low-level assembly such as PTX, with a verifier checking functional correctness, race conditions and deadlocks.

2 Sources, 20d ago, updated 20d ago

Turbo-dLLM debuts as an open-source library for training diffusion language models

The Turbo-dLLM team reports DFlash2 speculative-decoder training speedups of 2.48x at 512K context length and 7.59x at 1M context length on eight H100 GPUs.

4 Sources, 32d ago, first seen 32d ago

Local AI’s efficiency gains came more from hardware than models, a user says

Discussing a Stanford University and Together AI paper, the user reports an 18-fold rise in accuracy per joule over 16 months: 5.9-fold from better accelerators versus 3-fold from better models.

2 Sources, 40d ago, first seen 40d ago

Technology

Azalia Mirhoseini

Ricursive Intelligence aims to automate chip design with AI

Next-Concept Modeling is claimed to be up to 31× faster than similar-sized diffusion models on summarization and translation

ThunderSync is claimed to make reinforcement learning up to 1.9x faster without stale policies

AI-as-compiler approach claims speedups over Triton baselines on B200

Turbo-dLLM debuts as an open-source library for training diffusion language models

Local AI’s efficiency gains came more from hardware than models, a user says

AlphaChip Pioneers Recognized in TIME100 AI

Sunny Madra and Elon Musk Flag Intelligence Per Joule

Maithra Raghu Shares DiscoLoopAI Swag With Jeff Dean

Azalia Mirhoseini

Ricursive Intelligence aims to automate chip design with AI

Next-Concept Modeling is claimed to be up to 31× faster than similar-sized diffusion models on summarization and translation

ThunderSync is claimed to make reinforcement learning up to 1.9x faster without stale policies

AI-as-compiler approach claims speedups over Triton baselines on B200

Turbo-dLLM debuts as an open-source library for training diffusion language models

Local AI’s efficiency gains came more from hardware than models, a user says

AlphaChip Pioneers Recognized in TIME100 AI

Sunny Madra and Elon Musk Flag Intelligence Per Joule

Maithra Raghu Shares DiscoLoopAI Swag With Jeff Dean

Featured in

Ricursive Intelligence aims to automate chip design with AI

Next-Concept Modeling is claimed to be up to 31× faster than similar-sized diffusion models on summarization and translation

ThunderSync is claimed to make reinforcement learning up to 1.9x faster without stale policies

AI-as-compiler approach claims speedups over Triton baselines on B200

Turbo-dLLM debuts as an open-source library for training diffusion language models

Local AI’s efficiency gains came more from hardware than models, a user says

AlphaChip Pioneers Recognized in TIME100 AI

Sunny Madra and Elon Musk Flag Intelligence Per Joule

Maithra Raghu Shares DiscoLoopAI Swag With Jeff Dean

Featured in

Ricursive Intelligence aims to automate chip design with AI

Next-Concept Modeling is claimed to be up to 31× faster than similar-sized diffusion models on summarization and translation

ThunderSync is claimed to make reinforcement learning up to 1.9x faster without stale policies

AI-as-compiler approach claims speedups over Triton baselines on B200

Turbo-dLLM debuts as an open-source library for training diffusion language models

Local AI’s efficiency gains came more from hardware than models, a user says

AlphaChip Pioneers Recognized in TIME100 AI

Sunny Madra and Elon Musk Flag Intelligence Per Joule

Maithra Raghu Shares DiscoLoopAI Swag With Jeff Dean