Alpaca, GPT-4 and Google AI goodies

While Facebook fires another 10.000 employees, Google and Open AI make big announcements and launch new products

OpenAI manages to keep up the crazy release flow of updates to its product line. Today they have released GPT-4, the successor to the much-acclaimed GPT-3.5 large language model.

GPT-4 is a multi-modal model which means that it can also take images as input, so among its long list of tricks, it can also analyze images, write captions for them or interpret what “it sees” and deliver the output in natural language text.

In the live demo, they showed a really cool example where they turned a napkin sketch of a website idea into HTML Javascript code. That was pretty impressive!

Of course, it can pass the bar exam with a score of 90%, can win the latest Biology Olympiad with a score of 99%, and is considerably more capable than the old version on many other tests. I am expecting an explosion of awe and novelty in the next few days as people start using it.

The first test also showed that it’s 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5, but the announcement had a clear disclaimer: GPT-4 is still flawed and has limitations… You can read the full update here.

The new model is available via API (waitlist) and for ChatGPT+ subscribers. ChatGPT free users will have to wait for a while.

One other interesting thing about this launch is that it comes with some examples of live products that already integrate GPT-4, straight out of Apple's playbook.

Some of them:

- ChatGPT for customer care from Intercom
- Duolingo enhanced by GPT-4
- Khan Academy offers a customized tutor based on GPT-4 for every student
- BingChat from Microsoft already uses GPT-4 (of course it does)

What about the other Guys?

Well, just a few hours before the OpenAi announcement, Google underwhelmingly updated the industry about their plan for Generative AI with two blog posts. The change log is pretty hefty, not gonna lie.

For once they integrated Generative AI-based tools for Google Workspace users. In Gmail and Google Docs, people will be able to type a topic they want to write about and will and AI instantly generate a draft. Users will be able to adjust the pone or length of the generated text.

It’s still some sort of closed beta, but they’ll make it available for more testers in the coming weeks.

Google also announced updates for PaLM and Google Cloud.

PaLM is a large language model (LLM) rival to the GPT models of OpenAI. To encourage developers to train PaLM, Google has also released an app alongside the API for this model. the app is called MakerSuite and it helps with prototyping ideas.

Google Cloud users will have support for Generative AI in Google’sVertex AI platform, which is a product launched to assist businesses in training and deploying machine learning models. Also, they will be able to use the new Generative AI App Builder to build their AI-powered chat interfaces and digital assistants. More details are here and here.

A very busy day for Generative AI, but what is an Alpaca?

Well, Alpaca i is a language model that can follow natural language instructions developed by Stanford researchers. It is fine-tuned from another model called LLaMA 7B on 52,000 instruction-following demonstration and is comparable to OpenAI’s text-DaVinci-003, but it is smaller and cheaper to reproduce.

How much cheaper? It cost less than $500 to fine-tune this model and it can be used on a single GPU or CPU, such as Apple Silicone or Raspberry PI.

It’s for research purposes only, but I still find it really cool! You can read the paper here.