CONVO OVER COFFEE ☕

Big topics to discuss with friends and colleagues

GPT Store from OpenAI is coming next week 🟠
OpenAI is gearing up to open its much-anticipated GPT Store, a platform where users can sell and share their custom AI agents based on the GPT-4 large language model. Initially announced at a developers conference and slated for a November release, the launch was delayed but is now confirmed for next week.

My take: This feels like OpenAI is setting up a digital AI farmers market where everyone's invited to set up a stall and sell their homegrown chatbots. It's a pretty neat idea, making the power of GPT-4 more accessible and letting all sorts of creative (and probably quirky) AIs bloom. I'm curious about the weird and wonderful bots we'll see and how they'll be used. It's also a smart move for OpenAI, pushing for new iteration of their dominance over LLM realm. Whether I like this monopolistic play or not, let the AI bot bazaar begin!

Google's is thinking about automating Ad Sales Roles 🚩🚩
Google is reportedly exploring the use of generative AI in restructuring its workforce, particularly within its large customer sales unit. According to a report by The Information, the tech giant is considering staff consolidation, including potential layoffs, as it automates functions traditionally performed by employees overseeing relationships with major advertisers.

My take: You may be surprised to see this so high here, but there are many gossips circulating around about massive layoffs at Google. While these are rumours, a Black Swan event for the industry could be Google’s bet on its own AI capabilities exemplified by massive staff reduction. A message of ‘we can help your enterprise with AI, and to prove it, look as we layoff 30k people’.On one hand, you've got to admire the efficiency and innovation of AI – these AI tools are cranking out ads that would take teams of humans ages to put together. But then there's the real-talk part of it: jobs are changing, maybe disappearing. The big question is, what happens to all that "natural intelligence" once the AI's done its thing? Especially that, as we hear from everywhere, the current AI is the stupidest we will ever see, and innovation will make it way more efficient over the next decade. Before you go further, let me underline: this is, as of now, only a rumour.

Perplexity AI Secures $73 Million to Revolutionize Search 🟠
Perplexity AI, a San Francisco-based startup, has successfully raised $73.6 million in a Series B funding round, with significant contributions from Institutional Venture Partners (IVP) and prominent figures including Jeff Bezos and NVIDIA. The company aims to transform online information access through its AI-powered search engine. In less than two years, Perplexity has garnered 10 million monthly active users and processed over half a billion queries in 2023.

My take: Unlike traditional search engines, Perplexity provides direct answers to queries using natural language processing, integrating its search index with leading AI models like Claude 2 and GPT-4 for summarized responses and citations. If they can keep the momentum and truly crack that better, smarter search code, we might all be in for a new way to satisfy our digital curiosity.

DocLLM: JPMorgan's and Multimodal Document Understanding 🟠
JPMorgan has introduced a new AI model, DocLLM, designed to revolutionize how financial institutions and other businesses understand and process documents. This "lightweight" model, enhances the comprehension of complex documents by incorporating both text and spatial layout, a significant leap from traditional language-based or image-encoded methods.

My take: In practice, DocLLM employs a text infilling pre-training objective, conditioning the model to manage disjointed text segments and irregular arrangements common in business documents. The fine-tuned model has demonstrated remarkable performance, surpassing leading models like GPT-4 in many cases, particularly in understanding forms and generalizing to new document types. This is cool this comes from non-tech institutions! The question is: what data was used to train it?