Why is DALLE3 so good?

PLUS: OpenAI started selling shares at $80B valuation

Hi folks!👋🏻 This is The Prompt!

Here’s what you need to know about AI today:

  1. We have a better clue why DALLE3 is so good;

  2. StackOverflow laid off 25% of workers, yet maintains good traffic;

  3. OpenAI started selling shares at $80B valuation;

  4. Chain-of-thought prompt technique helps LLM to reason.

Let's read on

FEATURED

Why is DALLE3 so good?

OpenAI shipped DALLE3 for all Plus users. You can now generate images with simple prompts directly in ChatGPT. The model also makes it very easy to add text on the photos.

So, why is DALLE3 so good?

For GPT4 we didn’t get any info on how OpenAI made the model so smart and capable. The technical report didn’t include relevant info on the model/infrastructure.

For DALLE3, we have more info, specifically around the prompts and the text-on-image capability.

How DALLE3 prompts work👇🏻

To make it possible for DALLE3 to work with short and simple prompts they did two things.

First they fine-tuned an image captioner with short descriptions of the main object of the image. Then fine-tuned it with longer descriptions that describe all details in the image. At the end 95% of DALLE’s dataset was synthetically generated by this captioner.

But, then they did something interesting.

They used GPT-4 to “upsample” short descriptions into longer ones. Meaning that you can use short descriptions, and GPT-4 will be able to expand those and generate accurate images.

Wild!🤯 

This clearly illustrates the impact of multi-modality on image generation.

How DALLE3 generates text on images👇🏻

When testing their image captioner, they made sure to include important words (that are usually found in a particular image) directly in the caption. As a result DALLE3 learned to create some text on images.

That’s it - so simple and yet it works very nice (see image below).

But, they said it's not good enough. They want to work on character-level models, which can train a model on individual letters, not whole words.

Generated with DALLE3. From the OpenAI website

🚨 What else is going on

  • StackOverflow is laying off 28% of their workspace, and it might be the first large layoff due to AI; mostly if developers are asking ChatGPT for answers than resorting to StackOverflow forum answers.

    But is traffic going down?

    StackOverflow says they only have 5% drop compared to 2022.. but we think this is just the start. Other companies might face layoffs, think Quora, e-learning platforms like Udemy or niche platforms like TripAdvisor.

  • OpenAI is considering whether or not to release an AI detector for generated images, but the team has not made a decision yet.

  • ThriveCapital is leading the purchase of OpenAI employees shares at $80B valuation. The deal would boost OpenAI's valuation by three times and it will make it one of the world’s most highly valued private companies.

📕 Resources

🧰 Tools

  • AdEditor: Repurpose your influencer content for ads

  • ZipChat: Sales and E-commerce support chatbot

  • Warmly: AI powered sales orchestration (just raised $11M Series A)

  • Hexofy: Capture data from every page

  • TileDesk: Design, test and capture different chatbot flows

  • Impaction: Analytics for conversational AI chatbots

✍🏼 Prompt of the Day

TOOL

Midjourney

PROMPT

a woman walking a basket and a cat on the beach, in the style of tracie grimwood, lucy grossmith, dee nickerson, daguerreian, expansive, naturecore, meticulously crafted scenes --ar 13:16 

RESULT

Subscribe to keep reading

This content is free, but you must be subscribed to ThePrompt to continue reading.

Already a subscriber?Sign In.Not now