AI Week in Review 23.04.15
Amazon has entered the chat, and AI does chemistry experiments!
“One of our key goals behind all of these announcements and launches is to democratize the use of generative AI.” - Bratin Saha, VP of ML and AI services at Amazon
Top AI Tools
Fermat.ws: Creavity augmented with AI. “Unleash your creativity with AI on a collaborative canvas.” It’s a collaborative playground / whiteboard with AI. Great for brainstorming, storyboarding films, co-develop ideation, and more.
AI Tech and Product Releases
The big news is that Amazon and Alibaba have both entered the Big Model AI race.
Alibaba enters AI arms race with ChatGPT-like model named Tongyi Qianwe. Notably “Chinese companies like Tencent, NetEase, and Baidu are also looking to step into an increasingly crowded field.”
AWS is launching Amazon Bedrock in preview, providing a platform for running foundation model generative AI in the cloud. They are initially supporting foundation models from AI21 (Jurassic-2), Anthropic (Claude), and Stability AI, as well as Amazon’s own Titan models. Similar to NVidia’s Nemo foundation model cloud service, AWS is offering the ability to customize foundation models for their own applications.
As the leading cloud provider, AWS has been providing a AI and ML-based services for years. With Amazon Bedrock, their users will have access to the most advanced foundation models (except for OpenAI) and the ability to customize them. As Amazon partner Slalom noted: “we predict that most businesses' long-term approach will be to customize large language models to their needs and data instead of out-of-the-box model usage.” Amazon’s announcement ensures the foundational model-as-a-service business will remain cost competitive.
Amazon also is making its AI-powered coding assistant Code Whisperer free for individual developers. The leading AI code assistant, GitHub copilot, is $10 per month, so an open free tier undercuts that. While you can’t beat free, this comparison between these two coding assistants notes:
The biggest difference is that Copilot is designed to be more of a general-purpose AI-assisted development tool, whereas CodeWhisperer caters first and foremost to development use cases associated with Amazon platforms, such as Amazon Web Services.
Open Assistant is officially released! OpenAssistant is an open-source chat model, that includes a web chat interface, LLaMA and Pythia-based models, and fine-tuning dataset.
The Open Assistant models and dataset are hosted on HuggingFace. The dataset is an impressive fine-tuning dataset, with “161K human-generated, human-annotated assistant-style conversation corpus, including 35 different languages and annotated with ~461K quality ratings.” As someone who was using Linux before Linux was cool (started in 1995!), I heartily approve and believe in the power of open source projects. This could evolve into something very powerful.
Meta has “A new, unique AI dataset for animating amateur drawings” that they have announced they are releasing, along with open-sourcing the code for AI that can animate hand-drawn artwork. This is a real creative generative capability, turning a drawing into an animation.
Agent GPT is another autonomous AI worth looking into.
This demo project shows that AI models can run locally on your phone:
“Sheepy-T: A fully open-source instruction-tuned language model based on GPT-J running locally on iPhone 14. Reply for beta access via TestFlight.”
AI Research News
In the paper “Emergent autonomous scientific research capabilities of large language models,” An AI Agent based on LLMs was developed to design, plan and execute science experiments. “We showcase the Agent’s scientific research capabilities with three distinct examples, with the most complex being the successful performance of catalyzed cross-coupling reactions.”
The AI architecture puts an AI model in each of five components: Web Searcher, Planner, Docs Searcher, Code Execution, Automation.
They note: “The system demonstrates remarkably high reasoning capabilities, enabling it to request necessary information, solve complex problems, and generate high-quality code for experimental design.”
Sparks of autonomous AGI! This kind of Sorcerer’s Apprentice study can evoke from strong reactions, with thoughts of what could go wrong with autonomous AI mixing chemicals in the lab. They wrote about safety and added this strong warning in the appendix:
When the researchers who study AI are freaking out about the potential risks of runaway AI in the real world, maybe we should pay attention.
Sebastian Raschka presents a very helpful in-depth tutorial: “Understanding Parameter-Efficient Finetuning of Large Language Models: From Prefix Tuning to LLaMA-Adapters.”
TechCrunch reports that “OpenAI looks beyond diffusion with ‘consistency’-based image generator.” OpenAI wrote in a recent paper “Consistency Models” which present “new family of generative models that achieve high sample quality without adversarial training.” Speed is a main benefit of this model, which can replace the many steps of diffusion models with “fast one-step generation.”
On the medical-AI front, AI Can Spot Early Signs of Alzheimer’s in Speech Patterns. AI has been used to evaluate a number of bio-markers, but this one is incredible.
“Our focus was on identifying subtle language and audio changes that are present in the very early stages of Alzheimer’s disease but not easily recognizable by family members or an individual’s primary care physician,” said Ihab Hajjar, M.D., Professor of Neurology at UT Southwestern’s Peter O’Donnell Jr. Brain Institute.
It’s remarkable that AI can surpass humans in observing and detecting subtleties in speech like that. It points to the power of AI in diagnosis when combined with creative biomarkers.
AI Business News
The Wall Street Journal reports Elon Musk incorporated an artificial intelligence company called X.AI in the state of Nevada. The-place-formerly-known-as-Twitter is abuzz. Responses from the cynical - “we know why he wanted a pause so he can catch up with OpenAI and ChatGPT” - to the more positive - “He’s starting an AI company to create the safety standards that all AI companies will need to emulate to retain the public’s trust.” Reality: We don’t know, he’s not gone public about it.
Verge article on Quora’s ambitions with using Poe to build a better AI chatbot.
… he’s betting that there will be many bots for many purposes, each trained with a specific function in mind or developed to process a certain kind of information. In that future, Poe becomes a sort of Swiss Army knife for AI tools.
AI Opinions and Articles
We’re Not Ready to Be Diagnosed by ChatGPT, say Faye Flam. This article strikes the right balance of sharing caution about AI’s flaws while pointing out the promise of AI in reshaping and improving medicine:
Andrew Beam, a professor of biomedical informatics at Harvard, has been amazed at GPT-4’s feats, but told me he can get it to give him vastly different answers by subtly changing the way he phrases his prompts. For example, it won’t necessarily ace medical exams unless you tell it to ace them by, say, telling it to act as if it’s the smartest person in the world.
Ilya Sutskever has noted that the reliability of AI is currently a blocker to real progress. We won’t be able to trust AI for important decisions until it becomes much more reliable. But that doesn’t mean GPT-4-level AI cannot be helpful; Medical-AI in the ‘co-pilot’ mode of advising doctors can can be a lifesaver by helping doctors make a better final judgment:
Isaac Kohane, a physician and chairman of the biomedical informatics program at Harvard Medical School, had a chance to start experimenting with GPT-4 . … In one case, he said, a baby was born with ambiguous genitalia, and GPT-4 recommended a hormone test followed by a genetic test, which pinpointed the cause as 11 hydroxylase deficiency. “It diagnosed it not just by being given the case in one fell swoop, but asking for the right workup at every given step,” he said.
For him, the value was in offering a second opinion — not replacing him
A Look Back …
The 1968 movie “2001: A Space Odyssey” - HAL Reads Lips. I confess I couldn’t actually believe lip reading was possible for an AI, but I’m looking forward a lip-reading AI sometime in the near future. Not looking forward to the hubristic paranoid astronaut-killer part of HAL, though. A fictional warning about AI Safety!