OpenAI’s Sora breaks the internet

Welcome to the 158 new subscribers who joined in the last week.

Sit back, grab a coconut (or coffee), and check out some pretty mind-blowing advancements in AI.

What’s brewing in AI this week:

• OpenAI’s Sora breaks the internet
• Gemini 1.5 dwarfs ChatGPT’s context window
• GPTs: Editor’s choice and new arrivals in the GPT store
• The other top stories in AI this week

‍

Dario’s Picks:

I. OpenAI introduces Sora

OpenAI introduced Sora last Thursday – a next generation text-to-video model which, by the look of it, vastly outperforms anything we’ve seen so far.

An AI generated video says more than a thousand words:

‍

‍

Sora has currently only been made available to early testers without a set release date to the public.

‍

Why it matters:

Sora understands and simulates the physical world in motion with unprecedented realism – all from some simple text instructions. If it delivers what it says it does we’re officially at the point that, without a lot of scrutiny, you can no longer tell if a video has been AI generated or not.

Our perception of reality and “what’s real” is going to be challenged. The opportunities for productive use and solving real-life problems is massive, as is the potential for abuse. The dynamics and costs in media production, with video now being as easy as typing on a keyboard, will change too.

Paraphrasing the philosopher Marshall McLuhan: technology is a means to extend our senses and – as soon as we’ve adopted it – societal norms, values, practices, and structures are bound to change. I’d say we’ve just taken a groundbreaking leap as far as extending and enlarging the visual human imagination goes. But don’t take it from me. Take it from Will Smith eating spaghetti.

‍

II. Google launches Gemini Pro 1.5

Google has announced a new model (again): Gemini 1.5.

‍

‍

The most notable thing is the massive context window of 10m tokens (GPT-4 has 128k). To get a glimpse of the opportunities it brings, I recommend Google’s own demos:

• Reasoning across a 402-page audio transcript
• Multimodal prompting with a 44-minute silent video
• Problem-solving across 100,000 lines of code

Source: Google

Key features:

• It’s Google’s first Gemini 1.5 model, and the first model available for testing is Gemini Pro 1.5.
• It’s mid-sized, but performs in benchmark tests at a similar level to the larger Ultra 1.0.
• Right now, the model has a 128k context window for most users – same as GPT-4 Turbo. However, a limited group of developers have gotten access to the model with a massive context window of 10M tokens. This feature seems like it will be rolled out eventually, but first needs to be optimised in terms of latency and computational requirements to enhance the user experience (and, I’m presuming, control Google’s costs).  
• It’s based on a new Mixture-of-Experts (MoE) architecture.
• It’s multimodal – understands images, video and audio. In other words, it’s not just big PDF files you’ll be able to feed the model, but also lengthy audio transcripts, image databases, entire movies and more.

Why it matters: The context window jump is huge compared to all other existing models. According to Google, the performance persists at a very high level even as the context window increases. Judging by Google’s demo videos, we could soon be able to feed AI all our data and get high-quality responses back.

‍

III. AI sound effects coming soon Eleven Lab

Eleven Labs, a leader in text-to-speech, is getting a notable feature soon: AI generated sound effects. That’s right, you will soon (no dates give yet) be able to simply describe a sound with a prompt and generate the audio for it. For their demo, the company cleverly chose to overlay AI generated sounds on some of OpenAI’s Sora clips.

Why it matters: Not as revolutionary compared to the other news above, but nevertheless a cool new feature that seems inevitable and pairs well with the advances in AI video. Our AI generated worlds will need sound, won’t they?

‍

GPTs

Editor’s Choice

Weekly picks from me to you

Enterprise AI Use Case Advisor

💬100+

This GPT will help you develop AI use cases for your company.

‍

‍

❞ This GPT asks the right questions, and suggests tailored solution on how to best leverage AI in your organisation. It’s made by one of the biggest influencers in the AI sphere, Allie K. Miller.

‍

A couple of useful SEO tools

‍

Search Intent Optimization Tool

💬900+  ↑100

Enhances content relevance based on Search Quality Evaluator Guidelines, utilizing methodologies from recent academic research.

❞ Reviews the relevance of your site’s content based on Google’s official guidelines. Tested it on a page on my site and found a good optimisation opportunity.

‍

Blog Post Title Generator

💬600+  ↑100

Generates 25 SEO-focused blog titles.

❞ Simple tool to create optimised titles for articles (or get ideas for new articles). Enter your target keyword and it gives you title suggestions categorised into lists, stories, opinions, questions and frameworks.

‍

New arrivals in the GPT store

New GPTs featured in OpenAI’s official GPT store in the last 7 days

1. Diagrams ⚡PRO BUILDER⚡  Rank 7 in   Programming
2. Website Generator  Rank 9 in   Programming
3. AI Humanizer Pro  Rank 8 in   Writing
4. Physics Oracle  Rank 7 in  Education
5. Math Solver  Rank 12 in  Education
6. math  Rank 7 in  Lifestyle

‍

Bytes

• Groq (not to be confused by Elon’s Grok) is a new AI model with really fast response time. It uses LPU (language processing units) instead of GPU, unlocking faster speeds. Here’s a demo by Matt Shumer. BTW, the naming similarity to Elon Musk’s chatbot is incredible. As if this space isn’t already confusing enough when it comes to names 😄 Groq was actually first though, founded in 2016.
• Fresh of the rumour mill: OpenAI is developing a Web Search product to challenge Google.
• Reddit signs AI content licensing deal ahead of its IPO.
• Deutsche Telekom showcased an AI phone at MWC 2024. The phone is launched in collaboration with Qualcomm and Brain.ai, and uses AI assistants to replace the apps.
• ChatGPT’s web traffic is down 11% since it’s peak in May last year.
• Anthropic is testing Prompt Shield to avoid misinformation in the upcoming elections. It redirects questions on politics and voting to “authoritative” sources of voting information.

‍