ThePrompt
Posts
This AI outperforms GPT3.5 and humans in most tests🧪

This AI outperforms GPT3.5 and humans in most tests🧪

Anita Kirkovska
February 24, 2023

Hi folks! 👋🏻 This is the Prompt, we deliver the juiciest AI news straight to your inbox.

We're like DoorDash for AI.

Here's what the dasher brought today:

This AI outperforms GPT3.5 and humans in most tests🧪

The key challenge with LLMs is that models under 100 billion parameters tend to generate hallucinations that generate wrong answers.

To mitigate this challenge, Multimodal-COT, combines both text and images to improve performance.

The model has only 738M parameters and outperforms GPT3.5 by 16%.

For comparison: GPT3.5 has 175B parameters. This kind of model could be run on a home computer, with about 4GB of RAM.

Aside from being better than GPT3.5, the model beats humans in every test except social sciences. Which is incredible!

This is where it gets interesting folks. Get ready for much smarter AIs.

💥 The Latest

OpenAI is launching a developer platform 🤩

OpenAI is quietly launching a new developer platform that lets customers run the company’s newer machine learning models, like GPT-3.5, on dedicated capacity. Foundry will provide some level of version control as well. Running a lightweight version of GPT-3.5 will cost $78,000 for a three-month commitment. Really, lightweight?

$43m Series B for a pre-revenue AI storytelling startup Tome

Tome is a storytelling app, that lets you create powerful visual presentations with the help of AI. They have massive adoption — 500m+ MAU, 30M+ daily decks across many products - but they still haven't made any money with it. Is this the AI bubble?

Spotify launches AI-powered DJ

DJ is a personalized AI that picks music for users based on taste and past listening history. Or, as Spotify says, it’s putting an “AI DJ in your pocket.” Some Steve Jobs vibes right there.