TLDR AI 2023-09-07

Siri’s generative AI improvements 🤖, Falcon 180B model released 🦅, simulating an order book in Jax 💹

[Livestream] Come check out the LLMs and the Generative AI Revolution event on Sept 14 — stream talks in the comfort of your living room on the power of agents, fine-tuning, fighting hallucinations, security, and more (Sponsor)

AI is moving so fast that it can feel impossible to keep up. There’s so much news, so many new tools and techniques, and only so many hours a day 😩

The AI Infrastructure Alliance’s upcoming livestream event is the perfect way to catch up with twenty five fantastic talks on:

  • Building Agents the right way
  • Fighting hallucinations in LLMs
  • Security and compliance for Gen AI

We'll see you at the big show!

See the full agenda and register (free) here

Headlines & Launches

Falcon 180B model released (6 minute read)

The Falcon models from the UAE have long been the best open models available. The newest 180B parameter model slightly outperforms Llama 2 70B and has a 2k context window. The models have historically been quite tunable. However, given the resource requirements, it's not clear whether the community will adopt this newest model.

Apple Is Pouring Money Into Siri Improvements With Generative AI (2 minute read)

Apple has boosted its budget for developing artificial intelligence, emphasizing creating conversational chatbot features for Siri — allegedly spending millions of dollars daily on research and development.

OpenAI’s first developer conference (2 minute read)

OpenAI will hold its inaugural DevDay conference in San Francisco on November 6, 2023. This event will preview new tools and foster idea exchange, attracting hundreds globally. Currently, over 2 million developers use tools like GPT-4 and DALL·E through OpenAI's continuously updated API.
Research & Innovation

Reducing PPO memory usage in RLHF (25 minute read)

PPO is a pain in RLHF because you need to have 3 models floating around. However, they all stay close and only require small updates from one another. Enter LoRA. If you instead use adapters you can dramatically reduce memory costs without losing performance. Beautiful idea in its simplicity.

Enhancing Speech-Driven 3D Face Animation (14 minute read)

This study delves into the complexities of 3D face animation driven by speech, highlighting two important aspects: global factors that change how the face moves over time (composite nature) and how different parts of the face move independently based on local muscles (regional nature).

Simulating an order book in Jax (28 minute read)

Financial institutions use something called a limit order book to maintain all of the transaction information for trades on their platforms. This is useful, but it usually runs on the CPU, which makes running RL on them hard. This paper talks through creating an order book that runs on the GPU written in JAX.
Engineering & Resources

TokenFlow (GitHub Repo)

Using a pre-trained text-to-image model to edit videos produces dream-like and strange results. TokenFlow is much smoother and maintains many semantic and structural features of the original video. It almost seems to outperform Runway Gen-2.

Comgra (GitHub Repo)

A library for use with PyTorch that makes it easier to inspect the internals of your neural networks.

Making Video Avatars Adapt to Different Lights and Backgrounds (3 minute read)

ReliTalk is a cool tech breakthrough that lets us make video avatars look natural even when you change the lighting or background. It uses a single video and the sound of your voice to create a 3D face model.

A New Way to Tell if Two Similar Pictures Are Really the Same in 3D (3 minute read)

Doppelgangers is a new tech tool that helps figure out if two pictures that look almost the same are actually showing the same 3D object or not. It's smart enough to avoid mistakes that even people might make.

How The AI Revolution Will Reshape The World (4 minute read)

An argument that the impending technological wave, largely driven by AI, will herald a historic redistribution of power.

Can LLMs Learn From A Single Example? (10 minute read)

It appears that AI models were able to rapidly memorize examples from the dataset after seeing them just once. This astonishing feat contradicts most prior wisdom about neural network sample efficiency.
Quick Links

Buzzy AI Startup For Generating 3D Models Used Cheap Human Labor (5 minute read)

Kaedim, a 3D generation startup, often used human artists to make models, at one point using workers to produce the 3D design wholecloth themselves.

OpenAI Will Host Its First Developer Conference On November 6th (2 minute read)

OpenAI will host a developer conference, OpenAI DevDay, on November 6, the company announced today.

Seona (Product)

Seona is an AI assistant designed to optimize your SEO.
TLDR is a daily newsletter with links and TLDRs of the most interesting stories in startups 🚀, tech 📱, and programming 💻!
Join 500,000 readers for one daily email