A monthly recap of news, product announcements, beta leaks, and other interesting Generative AI action that caught our attention. Subscribe to our Newsletter here, or Join our mailing list to receive this recap straight to your Inbox.
Proudly curated by Dr. Daniel Moskovich.
Past issues:
This month’s highlights:
- The biggest GenAI news of April is Meta‘s release of Llama 3. Llama 3 is a big step up from Llama 2, and it seems that it was achieved mainly using more, and better training data. Meta says that most LLM’s are trained (or pre-trained) using 30 to 50 times less data than they should be. This highlights just how much mileage there is in tweaking known approaches, even before trying anything new. I believe Meta’s goal is to capture market share by releasing tremendously powerful models with very permissive licenses, thus squeezing smaller players out of the GenAI space.
- Microsoft released Phi-3 one week after Meta released Llama 3, claiming it is effectively just as good while being much smaller. Bring out the popcorn 🙂
- Snowflake released Arctic 480B, an LLM specifically designed for enterprise usage. So Fortune500 (or even 1,000) companies have a great GenAI tool to use internally, to facilitate processes (easy recall of material in company documents and e-mails). Worth noting that several companies have had internal in-house LLM’s for a while now.
- OpenAI leaked a mysterious LLM called gpt2-chatbot with amazing performance, then quickly made it disappear. This is extremely interesting, as my ‘love affair’ with LLM’s began with GPT-2. We know that Sam loves GPT-2 as well, so perhaps it will come back!
- China! Some dramatic announcements are coming out of China about all sorts of LLM’s with incredible performance. For example, SenseTime‘s Sense Nova 5.0. These cannot be downloaded openly so I’m somewhat skeptical but I wouldn’t scratch it just yet. China has had top-secret super-technologies before, so we will have to wait and see how this plays out.
- Cognition Labs is now valued at $2B, yet still with no revenues. They advertise an autonomous GenAI coding agent named Devin. Not much is available about Cognition or Devin, yet Cognition claims it reached a ~14% accuracy rate in resolving issues independently. At the same time, Cognition isn’t alone – GitHub Copilot, Google Gemini Code Assist, Amazon’s CodeWhisperer, IBM’s WatsonX Code Assistant, Meta’s Code Llama, are just a few of the dominant players ‘in the playground’. Guess we’ll need to wait and see how this develops.
- Tencent released InstantMesh, that can create a 3D model from a single image. Depth from single image has been around for a while using CNN’s, but this is much fancier because you can see the object from any angle.
- Reka released a new multimodal model, Reka Core, that is supposed to be competitive with Claude-3. For starters, I love the name – Rekka 烈火 is extreme fire! This is an exciting name for a generative model, so look forward to experimenting with it.
- Udio is generating music from text prompts. Not sure it’s anything I’d willingly listen to though 😉 I am waiting to see how Vocaloid decides to position itself with regards to AI-generated music. From my perspective, they’re still very much the leaders in the field.
That’s it for now. Did you see something of interest in the domains of Generative AI? Drop us a note and we’ll be happy to include it in our monthly round-up 🙂
