🇮🇹 Italy bans ChatGPT + LLMs for Finance

PLUS: Open source GPT-4?

Author

Sudharshan
April 04, 2023

Sup folks,

We’ve had a good week in AI. I’ve been tweeting less and focusing on two things.

(1) Building Frontier AI Resources (Curated AI products and Learning resources - launching soon)

(2) Scaling Frontier AI Consulting. We started 2 weeks back and are helping a few clients build custom AI solutions. Will keep you all updated on our progress and the cool stuff we build!

Alright, let's get to it. What's cooking today? 🧪

  • Italy bans ChatGPT 🇮🇹

  • OpenFlamingo - An open-source multimodal model (Like GPT-4) 🦩

  • ChatGPT for finance - BloombergGPT! 📈

👋 Psst! New here? Subscribe to receive the top AI news every Friday + exclusive content. Follow us on Twitter!

🇮🇹 ITALY BANS CHATGPT

Italy just because the first major country to ban ChatGPT.

Why the ban?

Europe has always been icky about privacy laws, and this is no different. They banned ChatGPT because OpenAI "improperly collected and stored user information"

"There appears to be no legal basis underpinning the massive collection and processing of personal data in order to ‘train’ the algorithms on which the platform relies"

Italy's privacy regulator

What now?

OpenAI has 20 days to find solutions or will be fined $21.8 million. I'm not too concerned though, OpenAI will find a solution - like every other major software provider that navigates EU laws. They’ll be quick too before other European nations start taking action.

It's funny how the news came out after Sama announced his OpenAI tour.

Read more about the ban here

🦩 OPENFLAMINGO - OPEN-SOURCE MULTIMODAL FTW!

I don’t get why this isn’t BIG. This is an open-source multimodal model i.e a LLM capable of processing and reasoning about images, videos, and text.

It's an open source version’s of deepmind’s flamingo. 

OpenFlamingo is a bunch of things:

  1. A 9B multimodal model trained on top of LLaMA-7B.

  2. A multimodal Dataset! (for others to train more models)

  3. A framework to train multimodal models

  4. A framework to evaluate multimodal models

Is this the stable diffusion moment for multimodal models? Perhaps the MVP. But this is an important step in accelerating open source research in this direction.

📈 BloombergGPT - CHATGPT FOR FINANCE

Bloomberg launched a 50B LLM trained on a LOT of financial data ~ 40 years of data.

“BloombergGPT outperforms similarly-sized open models on financial NLP tasks by significant margins — without sacrificing performance on general LLM benchmarks”

I read the paper so you don’t have to.

  • The model is based on BLOOM, the largest open source LLM (176B params)

  • Dataset - 700B tokens dataset. 50% is private bloomberg data and the rest is public financiald data (GPT-3’s training data was around 500B tokens)

  • Performance ~ fairly good on financial data!

Why did blomberg train a custom LLM?

Bloomberg has relied on AI and NLP for trading since 2016 and is infact, quite AI savy. They have a 200 member large ML division in NY. They trained a custom LLM for specific financial NLP tasks 👇

  1. Trading heavy tasks - Sentiment Analysis, news classification, and question answering.

  2. Make bloomberg terminal MUCH better. (Bloomberg’s $25K / year trading terminal)

This is a significant development fellas. This is a company with access to extensive sensitive data train their own LLM. We’re going to see similar data rich companies launch their own LLMs. (I smell a business idea here)

Betting you 50 bucks we're seeing 3 massive custom LLMs this month.

Their paper is pretty dope too; going into the weeds of how they trained their model and everything involved (except the cost ~ assuming a few million $)

PS: If you want to train your own custom LLM, hit me up! I started an AI consulting agency (with a small team) and we're helping awesome folks do exactly this.

🤯 SNIPPETS 

  • Andrew Ng responds to Gary Marcus’s call for a ban on AI progress (link)

  • Cerebras Systems open-source a suite of GPT models (this is big) (link)

  • Zoom unveils the ideal video companion with a partnership with OpenAI (link)

  • Elon Musk and other industry pioneers urge a halt to the "dangerous race" to develop AI that is on par with humans (link)

  • Character.AI is now valued at $1 Billion after recent funding (link)

  • OpenAI Plugin Protocol gets integrated into the LangChain library (link)

  • Google denies Bard was trained with ChatGPT data. Ouch! (link)

  • Summarised debate on AI Progress and Potential Regulation (link)

🛠 TRENDING AI TOOLS

  • HarvyAI - Write replies to your emails in your preferred tone and style

  • SplineAI - AI-powered 3D design and collaboration.

  • Decktopus AI - Create presentations in the shortest time without any design skills.

  • Motion AI - Build Chatbots to do anything on any platform.

  • MapsGPT - Use AI to generate map content

  • Kittl AI- Create stunning graphics with simple tools.

  • DomainsGPT - Generate brandable & memorable domain names using AI.

  • Roam Around - Plan your trips through AI travel planner.

  • MyGPT by SamurAI - Access to ChatGPT plugins without ChatGPT Plus.

😂 AI MEME

"will smith slapping will smith at the Oscars"

That's a wrap for today. Stay curious & see ya tomorrow! If you want more, be sure to follow on Twitter (@sudu_cb)

🧠 CHATGPT PROMPT OF THE DAY