Hardware Podcasts

88. Dall-E, But Better

The Next Byte Podcast · 20 Sept 2022 · 34m

What this episode covers

Hosts Daniel and Farbode discuss OpenAI's DALL-E 2, explaining how it uses CLIP embeddings to generate photorealistic images from text prompts. They analyze DALL-E's limitations with spatial relationships and text rendering, and explore the ethical concerns regarding training data bias and artist copyright. The episode highlights MIT research into 'composable diffusion,' a method that combines multiple AI models to improve accuracy and creativity in generated imagery.

Most relevant to

makers and hobbyists

Topics

Mentioned

Companies
OpenAI, Mouser Electronics
Products
DALL-E, CLIP, GitHub Copilot, StyleGAN 2, EBM, Glide

Related episodes from other shows

Other shows on the same ground, closest first.

  • Embedded.fm · 1h 10m

    275: Don't Do What the Computer Tells You

    Janelle Shane, creator of the AI Weirdness blog and author of the upcoming book 'You Look Like a Thing and I Love You', explains how character-level Recurrent Neural Networks (RNNs) generate surreal and often hilarious outputs like paint colors, ice cream flavors, and cocktails. She details the technical limitations of these models, such as short memory windows and lack of semantic understanding, which lead to nonsensical results like 'raspberry turd' pony names or recipes requiring a blender for 2 hours. Shane also discusses the dangers of algorithmic bias in serious applications, citing Amazon's failed resume screener, and contrasts the limited 'jellyfish-level' intelligence of current AI with the complexity of biological neural networks.

    AI & machine learningDebugging & failure analysis
    7 chapters

  • Oxide and Friends · 1h 33m

    Open Source LLMs with Simon Willison

    Simon Willison, a software engineer and creator of Django, joins Bryan Cantrill and Adam Leventhal to debunk an IEEE Spectrum article claiming open source AI is uniquely dangerous. Willison details the rapid evolution of open source Large Language Models (LLMs), highlighting how tools like llama.cpp and models from Mistral AI have democratized access to powerful AI, allowing them to run on personal laptops. He argues that open source models are essential for security research, citing the discovery of universal adversarial attacks that work across both open and closed systems, and discusses the ethical and legal complexities surrounding training data and copyright.

    AI & machine learningSecurity & privacy
    9 chapters

  • Embedded.fm · 1h 10m

    275: Don't Do What the Computer Tells You (Repeat)

    Janelle Shane, founder of AI Weirdness, explains how character-level Recurrent Neural Networks (RNNs) generate nonsensical but humorous outputs like paint colors and ice cream flavors by predicting the next character in a sequence. She discusses the limitations of these models, such as short memory windows and the inability to understand context, which leads to absurd recipes and image generation artifacts in BigGAN. Shane warns against trusting AI for critical decisions, citing biased training data in hiring algorithms, and highlights that current AI lacks the complexity of biological brains, comparing them to jellyfish in terms of neuron count.

    AI & machine learningEmbedded firmware
    10 chapters

  • Oxide and Friends · 1h 23m

    Adversarial Machine Learning

    Nicholas Carlini, a security researcher specializing in adversarial machine learning, joins Bryan Cantrill and Adam Leventhal to discuss how sequences of seemingly random tokens can cause Large Language Models (LLMs) to ignore safety restrictions. Carlini explains that these attacks, which include 'poisoning' training data by exploiting expired domain names and using adversarial suffixes to bypass jailbreaks, demonstrate that LLMs are vulnerable to inputs that look like noise but trigger specific internal features. The conversation highlights the transferability of these attacks across different models, the importance of open-source models for security research, and the practical risks of deploying LLMs without understanding their failure modes.

    Security & privacyAI & machine learning
    8 chapters

The summary, chapters and topics are generated from the episode’s audio by this directory, not written by The Next Byte Podcast. The episode itself belongs to its publisher.