LLaVA vs ChatGPT: the surprising results

PLUS: Meta marks your AI-generated images


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

Here’s what’s on the menu today:

  • Samsung’s new processor is an AI beefcake 🥩 💿

    The Exynos 2400 gets 14.7 better AI performance than its predecessor.

  • LLaVA vs ChatGPT: the dirty business 🥊 💥

    LLaVA 1.5 achieves state-of-the-art accuracy with GPT-4

  • Meta unveils Stable Signature 🤖 👀

    It’s a new, more effective AI watermarking method for images.


Samsung’s new processor = beefy 🥩 💿

In September, Samsung released “Samsung Food” — a food app with recipes, meal plans, and more.

Nice name, bruh.

Last Friday, it revealed the Exynos 2400 — a shiny, new mobile processor with the latest graphics and AI tech.

I care...? I think... Right…?

Well… Samsung claims it has 1.7 times faster data processing speed and 14.7 times better AI performance than its predecessor, Exynos 2200.

It enhances gaming experiences through improved ray tracing capabilities which allow for realistic optical effects like global illumination, reflection, and shadow rendering.

Cool, how do I get it…? You can’t yet, and Samsung didn’t specify when it was coming out. Rude.


Get a website designed by AI that doesn’t look like it was designed by AI. In 30 seconds or less.

Durable is the AI website builder and platform that makes running a business easier than having a job.

Build a fully functioning, beautiful website in 30 seconds.

Then, start attracting and booking more clients with integrated CRM, SEO, invoicing, and marketing tools.

Try it for free, no credit card required.


QA (Question Answering)

AI Systems designed to respond to user queries with precise, contectually relevant answers.

These system use techniques like Natural Language Processing (NLP) to understand and process human language.

Evaluation is done through various methods:

1. Precision and Recall: Metrics that assess the system’s correctness and completeness in providing accurate answers.

2. Mean Reciprocal Rank (MRR): Evaluates the ranking of the correct answer amoung the list of answers provided.

3. BLEU Score: Compares the system-generated text to a reference text, assessing the relevance and quality of responses.


LLaVA and ChatGPT duke it out 🥊 💥

I know you are, but what am I?

In September, we reported on ChatGPT’s new multimodal capabilities. This gave the chatbot the ability to essentially see, hear, and speak.

Last week, Stanford, UW-Madison, and Columbia researchers introduced LLaVA 1.5 — a new open-source AI that achieves favorable results against GPT-4 at visual and language processing tasks.

I’m a ChatGPT fanboy, why would I care?

GPT-4 isn’t open-source, and it costs $20 per month. LLaVA is open-source and free to use.

What’s different about it?

LLaVA was fine-tuned for two different scenarios - Visual Chat for daily applications and Science QA for scientific multimodal reasoning tasks.

It integrates a vision encoder with an LLM named Vicuna to boost both visual and language comprehension capabilities.

Here’s us playing with it:

So, I should stop using GPT-4?

No. If anything, you should consider using both. In a science-related task, the researchers combined LLaVA and GPT-4 and achieved a new state-of-the-art accuracy of 92.53%.

It’s always better when we’re together.


Meta’s new AI watermarking tool 🤖 👀

In August, we covered Google DeepMind’s SynthID — an AI watermarking tool.

On Friday, Meta unveiled Stable Signature — an invisible watermarking method that AI can detect even if the images are edited.

Know thyself, bot.

How does it work?

It’s a two-step process. First, two neural networks are trained to encode and extract messages from watermarked images. Then, the AI model’s decoder is fine-tuned to include a fixed signature in generated images.

Why is this significant?

It reduces false positive rates in distinguishing AI-generated images from human-made ones. It’s designed to withstand image alterations like cropping, compression, color changes, and more.


Meta’s Stable Signature claims to prevent this from happening, but people will continue to find workarounds.


Think Pieces

A closer look at DALL-E 3’s prompt transformations. Was it done for convenience, better outputs, or safety?

Are the EU and Japan playing nice with each other? Apparently, there are a lot of similarities in the way they look at AI regulations.

Why did NVIDIA cancel its Israel AI Summit? It’s due to safety concerns fro the escalating violence in the area.

Startup News

Spotify is developing a $20 per month plan with AI-curated playlists. It also comes with lossless audio, playlist creation tools, and more.

NVIDIA is reportedly working on graphics cards specifically for AI. It could potentially be the RTX 4080 Super or the RTX 4080 Ti.

Reka announced its new multimodal AI assistant. The AI startup was founded by researchers from Meta, Google, DeepMind, and more.


UniAudio — an audio generation model that produces various audio types like speech, music, and sounds based on diverse input conditions.

ITIT — a training paradigm that enables generative vision-language training on unpaired image and text data.

HEAP — a framework that enables LLMs to generalize across different tasks with fewer demonstrations.


Hummingbird — a free MacOS AI assistant that’s not hard on your system.

LLM Boxing — compare Llama 2 and Mistral in a fun, interactive way.

Kino AI — an AI assistant that automates video editing tasks and more.

FL0 AI — an AI tool that helps you debug, deploy, and scale your applications.


💡 Read Alts — Unique investment ideas for alpha-seeking investors, traders, and finance workers. Learn about exotic markets and alternative asset classes.


3D Gaussian Splatting is a way of representing 3D scenes with millions of particles. Stability’s CEO retweets a very realistic output that uses 25 photos as input.

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


Until next time 🤖😋🧠

What'd you think of today's newsletter?

Login or Subscribe to participate in polls.