Adobe's tasty new AI, Gingerbread

PLUS: Is ChatGPT getting dumber?


Good morning, human brains. Welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

  • Apple’s AI chatbot 🍎

    It lives on Google Cloud, and it’s called “Apple GPT.“

  • Adobe New AI, Gingerbread 🍪

    Create 3D models from text prompts, easily adjust the angle, and more.

  • Is ChatGPT getting dumber? 🤤

    A study on ChatGPT’s degradation and the AI community’s response.


Apple’s AI chatbot 🍎

Apple’s stealthily developing their very own LLM. It’s built on a custom LLM framework, dubbed “Ajax.“

The namesake: Ajax was built using Google JAX and runs on Google Cloud.

Internal engineers call the chatbot “Apple GPT.”

Not iGPT… really?

Apple employees are already using it to assist with their internal product development. Just like other LLMs, it’s a lot of text summarization and query answering.

Apple hasn’t disclosed much information about how exactly they’ll use the chatbot in the future.

Our take: The other tech giants are launching AI tools as fast as their keyboard fingers allow them. Apple? Crickets. CEO Tim Cook stated Apple would be “deliberate and thoughtful“ with AI in their Q2 earnings report. The question remains: will we see an AI masterstroke, or are they already behind?



A type of code interface that software developers leverage to build apps quickly. Most software uses the same basic building blocks, or “boilerplate”. Using frameworks allow developers to reuse these code blocks, like Legos, instead of reinventing the wheel every time.

In AI, frameworks serve as the foundation to build, train, and implement machine learning models.

Example: PyTorch.


When’s the last time your newsletter wrote itself?

Probably never, until today.

It’s the newsletter writing machine that lets you go from topic, to outline, to full draft — in mere minutes.

With an AI trained specifically for long-form copy, it’s able to write your newsletter in the exact tone you choose. And Hoppy even incorporates images, that dovetail, right into that copy.

It does all the lifting you want, without tying your hands. Tweak the outline, edit the copy on a sentence level, change the images — your piece is completely steerable, every step of the way.

Join 20,000 creators who don’t know what to do with all their new free time.


Adobe Gingerbread 🍪

Adobe teased a new AI tool at Figma Config 2023. What is it? It’s a new AI image generator / 3D modeler called Gingerbread.

Merry Christmas, ya filthy animals.

Apparently, it’s a beast. Particularly cool: it allows you to import 3D models and adjust the perspective of a text prompt-generated image.

The text prompt? “Whiskey glass on table.”

This one’s the same prompt, just a different angle.

Mind. Blown.

So, what it do?

1/ Text prompts. Standard, except for the fact that it generates 3D models instead of static 2D images.

2/ User-generated 3D templates. It allows users to create their own 3D scenes instead of relying on AI’s random composition. That lets you get selective and actually compose scenes, putting objects and details exactly where you want them.

3/ Custom 3D object support. Users can upload their own 3D objects and images to use in their designs.

4/ Adobe Creative Suite. Adobe intends to slap Gingerbread right into Photoshop, Premier Pro, Illustrator, etc.

Our take: Gingerbread looks incredibly collaborative — closer to painting and composing images with the power of AI, as opposed to shaking the magic 8-ball with prompt engineering. We’ve got our fingers crossed, hoping Adobe releases it for realsies.


Is ChatGPT getting dumber? 🤤

The ChatGPT hype train has made a brief pit stop.

Researchers fed ChatGPT the same 500 questions in March and June.

The result? On some questions, ChatGPT’s success rate plummeted from 98% to just 2.4%.

OpenAI’s Peter Welinder tweeted this response:

“No, you.”

It’s causing a stir in the AI community. Many claim they’ve experienced ChatGPT’s degradation, while others defend OpenAI.

Princeton University Researchers responded to this. They claim the study was a “vast oversimplification“ and that it used “questionable“ methods.

AI soap opera, anyone?

Our take: Jury’s still out 🍿😛🥤.



Think Pieces

Meta open-sourced Llama 2. Some speculation as to why they might have done this.

Together AI could pose a real threat to OpenAI. The startup already supports 50 of the top open-source models out there (including Llama 2).

How to jailbreak a Google Nest Mini. After this, you can run custom LLMS, voice models, etc.

Startup News

OpenAI increases ChatGPT-4’s hourly message limit. It’s doubled now to 50 messages every 3 hours.

Google Cloud beefs up its enterprise-ready AI. General availability for Vertex AI, PaLM 2 for Chat, Codey, and more improvements.

British espionage AI — in use to cut off Russia’s shipments of war supplies.


A GitHub repo on LongLLaMA. A way to scale the context window of LLMs to 265k tokens.

Paper on SHARCS (SHARed Concept Space). A new method for explainable multimodal learning.

GitHub repo on ARCTIC. A machine learning dataset for “Hand-Object Manipulation.”


Hoppy Copy [Sponsored] — write your newsletter effortlessly with the power of AI.

Texts From My Ex — brutally honest AI that gives relationship advice.

Castmagic — takes podcast audio (or any MP3) and turns it into scalable content. — AI-powered lead generation and sales tool.


Nat Friedman, former CEO of GitHub, commented on Assembly AI’s latest AI model.

Tag us on Twitter @BotEatBrain for a chance to be featured here tomorrow.


If you like Bot Eat Brain, you might like these other newsletters too:

🖊️ The Pen Pivot — AI tools and prompts curated for curators.

☀️ The Boonly — Brighten your inbox with inspirational insights that make self-growth enjoyable, not stressful. Every Sunday.

👓️ The Copilot — Weekly AI startup funding insights, and AI business opportunities.


Until next time 🤖😋🧠

What'd you think of today's newsletter?

Login or Subscribe to participate in polls.