AI Week In Review 23.12.30
Pika 1.0, Github CoPilot Chat, Microsoft CoPilot smartphone app, running Mixtral 8x7B MoE locally or on CoLab, New York Times sues OpenAI, Top 50 AI Tools.
Cover Art story - AI Detects Unusual Signal Hidden in a Famous Raphael Masterpiece. Researchers created a sophisticated machine learning model to detect fine details in the work of Rafael (including brushstrokes, color and other details). The custom AI model deduced Rafael’s painting “Madonna of the Rose” was mostly painted by Rafael but Saint Joseph was not.
"So, then we tested the individual parts and while the rest of the picture was confirmed as Raphael, Joseph's face came up as most likely not Raphael."
Add “art provenance detection” to the skills a (specialized) AI model can perform.
AI Tech and Product Releases
CoPilot is now an app. Microsoft CoPilot is now available as an app on Android as well as on Apple’s iPhone (iOS) and iPad.
Coming in January, Midjourney will start training video models.
The AI video generation tool Pika 1.0 is now available to all and is accessible via the web at Pika AI.
GitHub Copilot Chat now generally available for organizations and individuals. This is a GPT-4-powered chat-based coding assistant that can help with coding, testing and other developer tasks and queries, in the software developer’s IDE (such as Visual Studio Code).
Top Tools & Hacks
Writerbuddy shared a report on the 50 Most Visited AI Tools, reporting a total of 24B total AI tool visits between September 2022 and August 2023, which included a cool infographic.
The Top 12 AI Tools based on visits and usage was:
OpenAI (ChatGPT).
character.ai - AI Chatbot
quillbot.com - AI Writing
midjourney.com - Image Generator
huggingface.co - Data Science
bard.google.com - AI Chatbot
novelai.net -AI Writing
capcut.com - Video Generator
janitorai.com -AI Chatbot
civitai.com - Image Generator
vocalremover.org - Voice & Music
you.com - AI Chatbot
AI Research News
Shared by Rohan Paul: You can run Mixtral-8x7B models in Free colab or GPUs like a 3060. This is thanks to work from Yandex, in the paper “Fast Inference of Mixture-of-Experts Language Models with Offloading.”
They solve the problem of how to run mixture-of-experts (MoE) model in a memory efficient manner, by performing parameter offloading in a way that utilizes the properties of MoE LLMs. This technique works by loading model parameters just-in-time when they are needed for computation.
A Survey of Reasoning with Foundation Models is a comprehensive 160 page paper from AI researchers from various institutes in China:
In this paper, we introduce seminal foundation models proposed or adaptable for reasoning, highlighting the latest advancements in various reasoning tasks, methods, and benchmarks. We then delve into the potential future directions behind the emergence of reasoning abilities within foundation models. We also discuss the relevance of multimodal learning, autonomous agents, and super alignment in the context of reasoning
There has been an explosion in papers and results in AI reasoning in the past year. Getting AI to reason reliably is critical to getting to Artificial General Intelligence (AGI), so this area is only going to get more interest going forward.
AI Business and Policy
The New York Times dropped a bombshell lawsuit against OpenAI and Microsoft, claiming OpenAI unlawfully used their news articles to train GPT models without permission, causing “billions of dollars in statutory and actual damages.”
The New York Times, reporting on their own case, says Boom in A.I. Prompts a Test of Copyright Law. Since New York Times articles are a part of the Common Crawl dataset used in training many AI models, it’s undoubtably true that there is “copying and use of The Times’s uniquely valuable works” in ChatGPT, but the critical question is whether merely using articles to train AI models is ‘fair use’ or a violation of copyright. Their lawsuit presented examples of AI model output (given the right prompts) that matched article text.
TechDirt has more on the lawsuit, including some of its weak points. They note that the Times reuses other sources in their articles (without attribution often) in ways that could implicate them similar to their own lawsuit.
… the crux of this lawsuit is the same as all the others. It’s a false belief that reading something (whether by human or machine) somehow implicates copyright.
The Times winning could be a blow to the AI revolution, as they call on OpenAI and Microsoft to to “destroy” existing AI models and training datasets containing their materials. This would destroy practically all AI Models. Some called on AI tech leaders to “smash” the NYTimes in court in response.
There is an AI solution to this AI copyright problem: The AI models themselves could be trained to avoid copyright infringing outputs, by recognizing and removing such outputs. AI Legality concerns can be addressed with AI guardrails.
Copying an article is one thing, but what about AI that Replicates Famous People by copying someone’s whole life’s work? This is a “policy gray zone” and new laws will be needed to establish people’s rights from not getting digitally copied by AI.
Trump’s former lawyer Michael Cohen passed along fake legal cases generated by AI that found its way into a motion filed on his case. Cohen used Bard, which hallucinated some cases.
AI Opinions and Articles
For all the hype in 2023, we still don’t know what AI’s long-term impact will be, says John Naughton. He notes the dichotomy between short-term hype and long-term world-changing impact:
we always overestimate the short-term impacts of novel technologies while grossly underestimating their long-term effects.
My position is simple: AI Changes Everything. Much of the hype around AI is a state of wonder about what new thing AI can do. Once the wonder wears off, the long-term value and benefit becomes more clear. AI that makes you twice as productive in a boring way will end up more impactful than an exciting toy AI.
A Look Back …
With this being the last AI weekly for 2023 closing and New Years approaching for 2024, we can look back at 2023 and call it a Banner Year for AI.
This year was a pivotal moment when:
GPT-4 exhibited “sparks” of AGI and near-human-level performance on some tasks, while GPT-4 vision incorporated visual understanding.
Llama, Llama-2, Mistral and other open-source AI models showed remarkable efficiency and capability, and Phi-2 showed properly-trained small AI models can perform as well as prior larger AI models.
Generative AI for image, 3D, and video generation made leaps and bounds in speed and quality, aided by research improvements to diffusion models as well as new AI model types in this space.
Robotics merged the power of LLMs and new AI models (like PaLM-e) showed a path to truly intelligent robotics.
AI agent innovations leveraged LLMs to showcase ability to get AI to solve multi-step complex tasks. This is one area where we are just getting started.
After 70 years of AI research, in the year 2023 AI arrived as a world-changing technology. Now, it is here to stay. 2023 is the year of AI, in the decade of AI, in the century of AI.
AI will advance as much in 2024 as it did in 2023, and in subsequent years the power of future AI models and applications will dwarf what AI can do today. AI will only get better from here.