Meta releases Llama 3

Howdy! Welcome to the 204 new wizards who joined the last week. And to all the OG subscribers reading my emails every week - I appreciate every one of you.

‍

What’s brewing in AI #36

• Dario’s Picks:
  1. Meta releases Llama 3
  2. Microsoft’s hyper-realistic talking face videos
  3. Multi-bot chat on Poe
  4. AI features coming to Adobe Premiere Pro in May
  5. OpenAI’s updated API dashboard and assistants API
• GPTs: top arrivals in the GPT store and on whatplugin.ai
• Bytes: the other important AI news this week

‍

Dario’s Picks

1. Meta releases Llama 3

‍

‍

Meta just dropped Llama 3, claiming that it’s the best open large language model publicly available. The two variants released are 8B and 70B parameters.

The new models are integrated into meta.ai, Meta’s answer to ChatGPT (available on web, insta, FB and WhatsApp). If you can’t access it yet due to geo restrictions, you can try it on Perplexity Labs, You.com or Poe.

Meta also has a massive 400B model in the works, which from early tests appears to be on the level of Claude Opus (3rd ranked of all models on the LMSys leaderboard).

Ps Mark Zuckerberg also appeared on two podcasts last week (Roberto Nickson and Dwarkesh Patel) covering Llama 3, Metaverse and more.

‍

Why it matters

Meta is continuing as a leader in the open-source models space, and are now getting on par with the best proprietary models. Llama 3 beats the free versions of ChatGPT, Gemini and Claude on important benchmarks, and is giving the paid versions a run for their money, too.

Meta has also done a great job at packing a lot of juice into a small model with this release, which allows developers to build apps that run locally on machines rather than in the cloud (here’s an example of someone running Llama 3 locally on an iPhone).

‍

2. Microsoft’s new model: hyper-realistic talking face videos

‍

‍

Microsoft released a research demo of VASA-1, a model that can create lifelike talking videos of people based on a still image and an audio clip of the subject’s voice. Micosoft’s goal here is real-time interaction with lifelike avatars.

We’re talking holistic facial nuances, natural motions, convincing lip sync (even singing!). I could on but a video says more than my words do anyway, so go ahead and have a look.

The new technique could improve access to education, accessibility for people with communication challenges and companionship.

Currently it’s just a research demo that’s released and there’s not any planned product or API.

‍

Why it matters

Imagine how strange it will be once this new tech meets the metaverse, something which I don’t think is far ahead at all. I’m equal parts excited and creeped out.

‍

3. Multi-bot chat on Poe

Quora’s chatbot platform Poe launched the ability to chat with multiple bots in a single thread. Some of Poe’s bots are similar to GPTs (task-specific agents) but also ecompass the most popular LLMs.

For example, you can switch bots and models to utilise each one’s strengths whenever you need it, even mid-conversation:

You can do in-depth analysis with Gemini 1.5 Pro’s 1M token context window, @-mention Web Search bot into the conversation to pull in up-to-date information about the topic, and then bring a specialized writing bot in to complete a writing task, using the context from all of the previous bots

- Adam D'Angelo in The Quora Blog

‍

Why it matters

This is a great new feature for easily comparing different models for your use case. Since this also works with Poe’s other bots, not just the foundation models, you can combine models and bots to do complex tasks.

‍

4. AI features coming to Adobe Premiere Pro in May

Adobe is bringing new AI features to their video editing platform, Adobe Premiere. Rolling out from May, the new features will let users use AI to add/remove objects within a scene and also extend existing clips.

The new features will be powered by Adobe’s Firefly AI models natively, but through plugins, the company will also bring on third-party models from OpenAI, Runway and Pika.

‍

Why it matters

Gen AI for video is definitely having a moment, and Adobe is not sleeping on it. These features are far from the most reality-bending innovations in AI for video I think we’ll see this year, but they’ll likely be big time savers and potentially a part of many video editors’ workflows.

‍

5. OpenAI’s updated API dashboard and assistants API

OpenAI announced two smaller but highly welcome updates for developers last week:

• New features to OpenAI’s API dashboard: create and manage individual projects.
• Expanded knowledge retrieval on OpenAI’s assistant API: upload up to 10,000 files per assistant, using vector store integration. It’s now also possible to configure a token limit to control costs.

‍

Why it matters

These updates are long-awaited by many developers. They’ll help them create and maintain advanced AI apps, build products that require accessing a large amount of information, as well as manage and control costs.

‍

GPTs

Top new arrivals in the GPT store

Highly rated, new GPTs featured in OpenAI’s official GPT store (from the last week)

• Instant Website [Multipage]  Featured

‍

Top newcomers on whatplugin.ai

Highly-rated GPTs that made it into whatplugin.ai from the last week. How rankings work.

‍

‍

Submit your GPT to whatplugin and get featured in the newsletter.

‍

Bytes

More notable AI news from the last week

• Reka Core - a new powerful model that understands video. Demo here.
• Google released 6 courses on how to use Gemini inside Google workspace.
• Stanford released AI index report 2024. Takeaways: Gen AI investment has skyrocketed, workers become more productive with AI, scientific progress accelerates thanks to AI, ++.

‍