OpenAI unveils DALL-E 3 with support for text and typography

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each weekday, we scour the web to aggregate the many stories related to artificial intelligence; we pass along the news, useful resources, tools or services, technical analysis and exciting developments in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on under the hood in AI.

OpenAI announces DALL-E 3

The story: OpenAI has unveiled the third version of their AI image generation model called DALL-E, which can generate high-quality images from text prompts. DALL-E 3 looks like a huge improvement over the lackluster (compared to other state of the art offerings) DALL-E 2 model, with enhanced abilities to render readable text and understand spatial relationships described in prompts.

More details:

• DALL-E 3 can now generate images with readable text baked directly into them. This allows for more natural integration of text instead of overlaying it, including font control.

• The model has a better grasp of spatial relationships described in prompts, allowing prompts to be rendered more accurately in the generated image. A prompt describing separate objects in relation to each other can be translated visually.

• OpenAI says the model understands context much better than DALL-E 2, and that they have taken steps to limit harmful or biased content present in the older models.

Takeaways: DALL-E 2 came out nearly a year and a half ago, so these improvements aren’t totally unexpected. But this offering might put OpenAI ahead of new entrants like Ideogram and continuing innovators like Midjourney. OpenAI plans to roll out DALL-E 3 to premium ChatGPT users in October.

“Building AGI is a deeply political move. Why aren’t we treating it that way?”

Amazon will also allow developers to integrate their own custom LLMs with Alexa. Some of the new capabilities enabled by Alexa's upgrades include more expressive and contextual text-to-speech, improved automatic speech recognition using a massive transformer model, and visual ID so Alexa recognizes users just by looking at them on Echo Show devices for personalized responses.

Obligatory note: Amazon shares your personal data with social media and marketing companies. Amazon employs hundreds of people to read and annotate the transcripts of voice recordings to ‘improve the Alexa service’, but claims Alexa does not record unless it is activated.

• The gruesome story of how Neuralink’s monkeys actually died

• Naturally, Neuralink is now accepting human patients for trials of its brain implant

• More Amazon: AWS joins the UN summit on generative AI

• George R.R. Martin, John Grisham and other major authors sue OpenAI, alleging ‘systematic theft’

• OpenAI announces open call for Red Teaming Network applications

• First Google search result for Tiananmen Square “Tank Man” is now an AI generated selfie

• Google DeepMind AI tool assesses DNA mutations for harm potential

• What to expect from Microsoft’s ‘special’ Surface and AI event

"Query understanding" - using language models to parse the user's input - can help plan a more sophisticated search strategy, and dispatch queries across multiple backends like databases, search engines, email, calendars, etc. Rather than just returning text search results, these systems can reformat outputs for each backend and combine results into a unified response - a powerful paradigm for building language model APIs that can interoperate cleanly with existing infrastructure.

• Fine-tuning GPT with OpenAI, Next.js and Vercel AI SDK

• (Open source) CulturaX: a cleaned, enormous, and multilingual dataset for Large Language Models in 167 languages

• (Couchbase) Capella iQ enables developers to write SQL++ and application-level code more quickly by delivering recommended sample code