The Generative Edge Week 29
Claude 2 wants to be a contender, Meta shows off a new Llama and animations is all you need.
Welcome to week 29 of The Generative Edge. Here’s the gist in 5 bullet points:
Anthropic's Claude 2 offers a 100k context and impressive code generation, challenging GPT-4's dominance in the large language model field.
Meta's LLaMa 2, fine-tuned as a chat assistant, outperforms its predecessor and can be used commercially, with support for various cloud platforms.
We test Pika Labs, which is advancing AnimateDiff technology and produces cool temporally consistent animations.
Midjourney and AnimateDiff can be combined to create dynamic animations from generated images.
Autochain is positioning itself as a lightweight alternative to LangChain.
For the details, let’s hop right in!
Claude 2
GPT-3.5 and GPT-4 are (still) the best LLMs available. Slowly, some of the model creators are inching closer. Anthropic, a research lab founded by former OpenAI employees, is such a creator. They’re driving development of Claude 2, and the goal is to be a contender that unseats GPT-4 from its throne.
Claude 2 boasts a 100k context, compared to GPT-4’s 8k (32k is available in private beta)
This is good because it means more knowledge can be given to the model at runtime
Claude 2 is quite good at generating code
It is multi modal and will accept various file formats, not just text.
GPT-4 still better in terms of understanding complex tasks, but Claude 2 impresses
Capabilities maybe somewhere between GPT-3.5/ChatGPT and GPT-4, the additional features however are quite useful
You have to be in the US or UK to officially test this model, but you can try it via https://poe.com.
Llama 2
Meta’s kind-of sort-of open large language model, Llama 1, had a weird rollout. Released with a research-only license that you had to apply for, the model weights quickly leaked and proliferated into various open source projects. While non-commercial research license made it difficult to use it for various tasks, holding back adoption, it did drive a lot of innovation in the space!
This week, Meta released version 2 of their popular Llama large language model.
Llama 2 is better at everything, has been trained longer, on more text, and comes fine-tuned to behave like a chat assistant
Released in 3 sizes, 7B, 13B and 70B, the 2 small models can run on local hardware
Performs very well in a variety of tests
Can be used commercially (some restrictions apply)
Support announced for a lot of platforms: Azure, Google Cloud, AWS, etc.
Give it a spin here: https://www.llama2.ai/
Animations is all you need
We’ve talked about AnimateDiff last week, and it continues to be an interesting topic. Research lab Pika Labs is the first player that tries to productize this tech, which creates similar outputs to what AnimateDiff can create: a temporally consistent (so no trippy AI flickering) animation, 3 or so seconds long, generated by a motion AI model.
We tested it, and the results are surprisingly good!
We start out with a generated image out of Midjourney, it will be our input image.
Midjourney allows to zoom out of generations (technically outpainting), so we can create 3 zoom levels, near, middle and far.
Pika Labs (or alternatively AnimateDiff) then generates the animated segments:
And finally, we edit them together with a bit of music and sound effects:
This tech is pretty cool, expect a flurry of content to appear that uses this.
Will likely improve quickly (longer animations, more coherence, higher resolution, etc.)
Pika Labs has a waitlist (www.pika.art) though AnimateDiff based tools will likely appear soon.
… and what else?
Autochain wants to be a lightweight replacement for LangChain
And that’s it for this week!
Find all of our updates on our Substack at thegenerativeedge.substack.com, get in touch via our office hours if you want to talk professionally about Generative AI and visit our website at contiamo.com.
Have a wonderful week everyone!
Daniel
Generative AI engineer at Contiamo