1. Google builds a “Chat with data” app

Google has announced the release of NotebookLM, previously known as Project Tailwind. NotebookLM is a notebook designed to enhance learning by leveraging the power of language models.

Why is it important? As per Google: “The product aims to address the challenge of synthesizing information from various sources. It will automatically generate a document guide and facilitate a better understanding of the material.

Unlike traditional AI chatbots, NotebookLM allows users to "ground" the language model in their own notes and sources, creating a personalized AI assistant. Users can upload Google Docs and interact with NotebookLM by obtaining summaries, asking questions, and generating creative ideas. While the product reduces the risk of misinformation, fact-checking against original sources is still recommended. “

In a summary this is a “Chat with data” app using Google Docs.

What’s next? You can join the waitlist here.

I have never been bullish on the “Chat with data” apps. Most of the documents are stored in Google Docs and Microsoft Word. It is trivial for Google and Microsoft to build this integration.

That doesn’t mean all such apps are going to die. But the numbers would go down quite a bit. The ones which remain will need to differentiate themselves.

2. Llama’s commercial release

As per FT, Meta is preparing to launch a commercial version of its artificial intelligence (AI) model, allowing start-ups and businesses to build custom software using the technology. This move positions Meta to compete with Microsoft-backed OpenAI and Google in the race to develop generative AI.

While Meta has explored the possibility of charging enterprise customers for fine-tuning the model using proprietary data, there are currently no plans to charge for the upcoming release.

Why is it important? The current offering of models is mostly paid and there is no scope for fine tuning on private data. The best approach you have is to create embedding pairs and then rely on “chat with data” type apps.

Open-source models have the advantage of broader adoption, enhanced data input, bug identification, and improvement. You can always finetune your models using a variety of techniques like LORA.

What’s next? OpenAI is also looking to release a GPT-3ish model as open source. Google on the other hand has not announced any free PaLM models.

It is going to be interesting to see Llama and the OpenAI model go toe to toe.

3. Stability AI’s Troubles

Stability legal troubles don’t seem to end. Previously it was Tayab Waseem who claimed to be a co-founder of the startup. He had sued for 10% of the startup.

There is another suit filed by another co-founder - Cyrus Hodes. His suit accuses Emad and Stability AI of fraud and embezzlement. Hodes alleges that Mostaque misled him into selling his 15% stake in the company for $100. This sale came just months before Stability AI secured over $100 million in funding at a $1 billion valuation.

He claims that Mostaque withheld information about the fundraising attempts and a planned pivot into AI text-to-image generation. This resulted in Hodes missing out on a stake that would have been worth $150 million. The lawsuit also accuses Mostaque of destroying evidence and embezzling funds.

Emad responded on twitter denying all the allegations. He claims that Stability’s first focus was on COVID related projects which didn’t take off. Hodes wanted a clean break and wanted his share repurchased at the original price. That is what happened and Hodes left the company willingly.

Musk weighed in with his response as well.

🗞️10 AI news highlights and interesting reads

1. Speaking of Musk, his AI startup xAI was launched last week. You have to wonder what is his obsession with “X”, his son being named X Æ A-12. The company will collaborate with Musk’s companies - Tesla and Twitter.

2. After Europe, China has come up with AI regulation. The major change is that previously models needed to register within 10 days. But now models need pre-approval to launch. I wonder if Musk’s prediction of super-intelligence taking over the country spooked them. China wants to control the rate of AI growth but this kind of regulation is untenable in long term.

3. Jeremy Howard at fast.ai echoes this thought by pointing out that model licensing and surveillance will lead to concentration in power. He proposes open source, regulating high risk applications and disclosures as the focal point of AI regulation.

4. Semi-analysis wrote a piece on overall cost and setup of GPT-4. It is behind a paywall. All the attempts to share the details from the post have met with a copyright strike. You can read some details here.

5. First, there was the Writers Guild of America strike. Now Actors are also on strike. This effectively brings Hollywood to a standstill. The issue has everything to do with AI. Hollywood wants actor AI replica for free and forever.

6. Hallucinations are coming to bite OpenAI. FTC has received complaints about ChatGPT making misleading claims about people. It is now suing OpenAI to determine if this was intentional.

7. Dominos keep falling for data copyright. Now AP and OpenAI agree to share data and technology.

8. Train AI model once and deploy anywhere using Nvidia and Run:AI

9. LLMs might be better at classification compared to traditional ML. Though I’d take this claim with a grain of salt. Things are often context dependent.

10. How to Use AI to Do Stuff: An Opinionated Guide

1. A close look into AI detectors and why it often fails.

2. Beginner’s guide to unstructured data analysis in Langchain

3. Save hours at your tech job using Generative AI