The Hemingway Effect and Generative AI Coding Revolution
Sundays, The Sequence Scope brings a summary of the most important research papers, technology releases and VC funding deals in the artificial intelligence space.
Next Week in The Sequence:
Edge 289: Our series about new techniques in foundation models continues with an overview chain of thought prompting techniques, Google’s original paper in that area and the OpenChatKit framework. Dont miss that series. It’s super cool!
Edge 290: We discuss Koala, Berkeley University open source LLM that was fine tuned using ChatGPT conversations causing quite a bit of debate about the correctness of that approach.
📝 Editorial: The Hemingway Effect and Generative AI Coding Revolution
In his 1926 breakthrough novel, "The Sun Also Rises," Ernest Hemingway coined one of his many famous lines that have transcended to pop culture:
"How did you go bankrupt?" Bill asked.
"Two ways," Mike said. "Gradually, then suddenly."
I feel the same principles can be applied to the evolution of AI programming automation methods.
Programming is surprisingly becoming one of the forefronts of the generative AI race. I say surprisingly because, as a problem domain, programming is not only magnitudes more complex than, let's say, email generation, given that it involves complex reasoning mechanics, but programming is also the type of AI method that can face strong barriers in terms of adoption. After all, programming is at the foundation of the world's digital economies. Should we incorporate machines into that? Additionally, and yet, it is happening at an astonishing speed.
GitHub CoPilot has become the gold standard for the application of AI to programming, but it's certainly not the only one. Amazon recently entered the race with Code Whisperer. Salesforce has been super active in the space with solutions such as CodeGen. Startups such as Replit have been able to raise nine figures in funding to bring AI to programming. Just this week, Hugging Face and ServiceNow open-sourced StarCoder, a new code generation model. Microsoft Research also published a paper describing a self-play fine-tuning method for code generation LLMs. Stanford University published research around Parsel, a model that can break down programming problems into smaller tasks.
Certainly, AI-based programming automation doesn't receive the same level of attention as other LLM domains. It is also likely to experience longer market adoption cycles. But make no mistake. The AI coding revolution is already happening. Gradually, then suddenly.
🔎 ML Research
Using Generative AI to Imitate Human Behavior
Microsoft Research published a paper detailing a technique that uses diffusion models to imitate human behavior in interactive environments. The model uses an imitation learning technique that can extrapolate human behavior from images —> Read more.
MaMMUT
Google Research published a paper introducing MaMMUT, an encoder-decoder model for vision-language tasks. MaMMUT is a “small” 2B parameters that is very efficient on training in competing objectives which is often the foundation of multimodal architectures —> Read more.
Parsel
Stanford University published a paper detailing Parsel, a new code generation method. Parsel’s main contribution is to break down complex coding problems into smaller tasks —> Read more.
Self-Play for Programming
Microsoft Research published a paper detailing a LLM and self-play method for solving programming puzzles. The technique uses self-play as a way to fine-tune the LLM in a large variety of synthetic programming problems —> Read more.
🤖 Cool AI Tech Releases
StarCoder
The anticipated collaboration between Hugging Face and ServiceNow produced a new open source code generation model —> Read more.
Lamini
Lamini, a new framework for fine tuning language models with RLHF was just released —> Read more.
DeepFloyd IF
Stability AI released DeepFloyd IF, a diffusion model that can integrate text into image —> Read more.
StableVicuna
Stability AI also open sourced StableVicuna, an open source LLM with RLHF capabilities —> Read more.
🛠 Real World ML
Generative AI at Salesforce
Salesforce Research published some guidelines for the responsible adoption of generative AI —> Read more.
📡AI Radar
AI legend Geoffrey Hinton resigned from Google and spoke out about the dangers of AI.
Salesforce announced Slack GPT based on Einstein GPT integration with Slack.
Microsoft announced new AI features for Bing and Edge.
AI supply chain startup Pando announced that it has raised $30 million.
Nova announced the beta release of BrandGPT, a model designed to protect brand integrity.
NewRelic announced Grok, a new system observability assistant.
AI-based employee experience platform Simpplr announced a new $70 million funding round.
AI advertising platform Alison raised $5.1 million.
Box announced the release of Box AI, a new set of generative AI capabilities for content management.
Databricks announced that is acquiring AI governance Okera.
Salesforce and Accenture announced an extension of their partnership focused on bringing generative AI capabilities to enterprises.
DataStax launched Luna ML to streamline real time AI deployments.