Good morning! Stability AI has released its powerful Stable Diffusion 3 text-to-image AI model for developers via an API. PyTorch has launched a new library called torchtune to simplify the process of fine-tuning large language models on consumer hardware. A recent paper by cryptographer Yilei Chen claims to have developed a polynomial-time quantum algorithm for solving a class of lattice problems that underlie many leading post-quantum cryptography schemes.

— Forrest Knight & Meghanadh Vasireddy

Stability AI Releases Stable Diffusion 3 API for Developers

Stability AI has released its powerful Stable Diffusion 3 text-to-image AI model to developers via an API. The API provides access to two variants:

• Stable Diffusion 3

• The faster Stable Diffusion 3 Turbo

What makes Stable Diffusion 3 unique is its new Multimodal Diffusion Transformer (MMDiT) architecture. This improves the model's text understanding capabilities as well as its ability to render text and spelling accurately in generated images. It achieves this by using separate trained components for image data and language data.

Stability AI claims that Stable Diffusion 3 outperforms other state-of-the-art systems like DALL-E 3 and Midjourney v6 when it comes to rendering typography and adhering to the input prompts. To ensure a reliable, high-performance delivery of these models via the API with 99.9% uptime, Stability AI has partnered with Fireworks AI, a provider of enterprise-grade API platforms.

While currently the models are API-only, Stability AI has stated that they plan to make the model weights available for self-hosted use via a paid membership in the near future.

In addition to the API, Stability AI is launching a limited beta of something called "Stable Assistant." This is an interactive chatbot that combines the image generation capabilities of Stable Diffusion 3 with their Stable LM language model. It enables a conversational approach where users can iteratively create and refine images through a back-and-forth dialogue.

Read More Here

PyTorch Launches torchtune to Easily Fine-Tune Large AI Models

PyTorch has released a new library called torchtune that makes it easier to fine-tune large language models (LLMs) like GPT-3 and LLaMA. The rise of these powerful AI models has been impressive, but customizing them for specific use cases is often needed. However, fine-tuning such massive models poses significant challenges, especially on consumer hardware with limited GPU memory.

torchtune fixes this by providing:

• Modular building blocks and training recipes

• Optimization of popular LLMs across different GPU types, including consumer-grade hardware

• Support for the full fine-tuning workflow - from data/model preparation to enabling efficient inference

A key strength of torchtune is its flexibility and extensibility. It stays true to PyTorch's design principles by using composable components and hackable training loops under 600 lines of code. This easy extensibility is combined with features that democratize fine-tuning:

• Works on gaming GPUs like the NVIDIA RTX 3090

• Integrates seamlessly with open-source tools such as Hugging Face Hub, Weights & Biases, and ExecuTorch

Additionally, torchtune provides interoperability with the LLM ecosystem through integrations like EleutherAI's LM Evaluation Harness for benchmarking and torchao for quantization. It also supports distributed training using PyTorch FSDP.

Read More Here

Quantum Algorithms for Lattice Problems

A new paper by cryptographer Yilei Chen on the IACR preprint server has been talked about a lot recently in the quantum computing and cryptography fields. Chen claims to have developed a polynomial-time quantum algorithm for solving a class of lattice problems known as GapSVP.

Some context: The GapSVP problem asks to approximate the length of the shortest non-zero vector in a high-dimensional lattice. While this may sound fairly niche, lattice problems of this kind underlie many of the leading candidates for post-quantum cryptography, including schemes like:

• NTRU

• Kyber

• Dilithium

If valid, Chen's result would be extremely significant. It doesn't outright break these post-quantum crypto schemes yet, but it puts them in a precarious position, vulnerable to further improvements that could undermine their long-term security against quantum attacks.

The proposed algorithm utilizes some novel and slightly perplexing techniques, including:

• Complex Gaussian functions

• Windowed quantum Fourier transforms

• Mysterious objects called "Karst waves"

A lot of experts agree Chen's claim is a serious one deserving of careful study, despite the convolutedness of the algorithm making it difficult to evaluate its validity at this stage. If it stands up to scrutiny, it would greatly expand the scope of exponential quantum speedups beyond what was previously known.

Read More Here

🔥 More Notes

• Meta releases Llama 3, claims it’s among the best open models available

• Netflix to hide its subscriber numbers starting in 2025

• Hardest Problem in Computer Science: Centering Things

Youtube Spotlight

All New Atlas | Boston Dynamics

“We are unveiling the next generation of humanoid robots—a fully electric Atlas robot designed for real-world applications. The new Atlas builds on decades of research and furthers our commitment to delivering the most capable, useful mobile robots solving the toughest challenges in industry today: with Spot, with Stretch, and now with Atlas. Stay tuned to see what the world’s most dynamic humanoid robot can really do—in the lab, in the factory, and in our lives.”

Was this forwarded to you? Sign Up Here