Text Diffusion — Brendon Dillon, Google DeepMind

AI Engineer · Intermediate ·🧠 Large Language Models ·1mo ago

Key Takeaways

Demonstrates Text Diffusion using Brendon Dillon's Gemini Diffusion model and compares it to GPT-4 and Gemini 2.5 Flash

Original Description

GPT-4o answered 40. Gemini 2.5 Flash answered 42 and stuck to it even after working through the reasoning incorrectly. The Gemini Diffusion model, considerably smaller than both, answered 60 on the first forward pass, then 49, then corrected itself to 39 once it finished reasoning. Bidirectional attention means it can see future tokens and go back to fix mistakes. Autoregressive models cannot do that. Brendon Dillon covers why text diffusion is fast (24 denoising steps to generate 256 tokens means roughly 10x fewer memory transfers than autoregressive generation), what the tradeoff is (lower throughput at large batch sizes makes it expensive to serve at scale today), and what gets unlocked when latency drops to 2,000 tokens per second. The demos include a fake Wikipedia generated on the fly, a Reddit clone with AI generated comments and images, an operating system where every click generates the next screen, and a todo app built in 15 seconds by voice.
Watch on YouTube ↗ (saves to browser)
Sign in to unlock AI tutor explanation · ⚡30

Related Reads

📰
The Sophistication Trap: Why the Smarter AI Technique Keeps Losing
Smaller AI models can outperform larger ones due to overfitting and complexity, and understanding this phenomenon can inform better AI development strategies
Medium · Deep Learning
📰
MoE Enabled Scale. ZEDA Makes It Practical
Learn how ZEDA makes MoE enabled scale practical by reducing friction in post-trained MoE systems, allowing for more efficient routing of expert compute
Medium · LLM
📰
How to actually test if your LLM app is working
Learn how to effectively test your LLM app with a 3-layer evaluation approach to ensure it's working as expected
Medium · Programming
📰
LangChain cost audit: what ConversationBufferMemory actually costs you at scale
Understand the true cost of ConversationBufferMemory in LangChain at scale and optimize your usage
Dev.to · wartzar-bee
Up next
5 Levels of AI Agents - From Simple LLM Calls to Multi-Agent Systems
Dave Ebbelaar (LLM Eng)
Watch →