TLDR Index
TLDR AI

TLDR AI 2024-05-21

6 sections19 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

Sponsored

MIT

Building an AI product is hard. Engineers who understand AI are expensive and difficult to find. And there's no way of telling who's legit and who's not.

ae.studioSponsor

02HEADLINES & LAUNCHES

INFLECTION AI REVEALS NEW TEAM

Inflection AI has unveiled a new leadership led by co-founder Reid Hoffman. Hoffman claims that the startup has 18 months of runway left and that it will double down on building emotional intelligence within AI.

links.tldrnewsletter.com7 min read

OPENAI PAUSES CHATGPT-4O SKY VOICE

OpenAI paused the "Sky" voice mode for ChatGPT-4o after backlash for allegedly mimicking Scarlett Johansson's voice from the 2013 film Her. The actress is also pursuing legal action against the AI firm.

arstechnica.com2 min read

GOOGLE INTRODUCES THE FRONTIER SAFETY FRAMEWORK

Google DeepMind introduced the Frontier Safety Framework to address risks posed by future advanced AI models. This framework identifies critical capability levels (CCLs) for potentially harmful AI capabilities, evaluates models against these CCLs, and applies mitigation strategies when thresholds are reached.

deepmind.google4 min read

03RESEARCH & INNOVATION

3D GROUNDED SCENE UNDERSTANDING

Understanding where things are in a 3D scene is challenging. By using language-guided 3D understanding, you can pinpoint semantic labels for objects in 3D space.

groundedscenellm.github.io12 min read

FAST VISION TRANSFORMERS WITH LEARNABLE META TOKENS

LeMeViT is a new approach to reduce computational costs in Vision Transformers by using learnable meta tokens. These tokens efficiently capture key information, improving inference speed.

arxiv.org14 min read

04ENGINEERING & RESOURCES

Sponsored

WEBASSEMBLY FOR AI INFERENCE

The Gcore engineering team decided to see what happens if they ran an image classification model on their CDN infrastructure [11]. This white paper has the full details of Gcore's experiment — tuning a WASM-at-the-edge environment for AI inference workloads. Bonus: there

gcore.comSponsor

DEVON

An open-source pair programmer.

github.comGitHub repository

LLAMA 3 FROM SCRATCH

It is easy to download language model weights and run them with a variety of tools. However, it is harder to know exactly what is going on in the code itself. This great tutorial shows every step of reconstructing Llama 3 and running the trained weights.

github.comGitHub repository

05MISCELLANEOUS

BACKDOOR ATTACK ON VISION TRANSFORMERS

Researchers have discovered a new security threat for Vision Transformers, a popular AI model. The attack, named SWARM, uses a "switch token" to secretly activate backdoor behavior in a model, making it highly stealthy and dangerous for users.

arxiv.org16 min read

WHAT'S UP WITH LLAMA 3?

Meta's Llama 3-70B language model has surpassed rivals in the English Chatbot Arena, excelling at open-ended creative tasks but struggling with more technical prompts. Analysis reveals that Llama 3's win rate decreases as prompts become more challenging and that it is particularly strong with friendly, conversational responses. While Llama 3's approachability may contribute to its success, its overall competitive edge needs further study.

lmsys.org9 min read

PERPLEXITY CEO ARAVIND SRINIVAS TAKES SHOTS AT GOOGLE

Perplexity AI's CEO Aravind Srinivas is unfazed by Google's AI-driven summaries, suggesting they could harm Google's ad-reliant business model. In this post, he critiques Google's frequent rebranding in response to AI competition while highlighting his company's user-focused approach to ads.

links.tldrnewsletter.com2 min read