🤝 AidfulAI Newsletter #12: Embracing the Power of Bullet Points!
Dear curious minds,
Welcome to the ultimate newsletter for those interested in Artificial Intelligence (AI) and Personal Knowledge Management (PKM).
In today's issue, I introduce a formatting change – the use of bullet points. This decision was made in an effort to make the content more digestible and user-friendly for you, my valued reader. At the same time, it will make the creation of the newsletter easier for me, as bullet points are already the style which I use to collect information for a new newsletter issue.
Major AI News
📊🐍 Powering Analytics: Unravel Data Insights with ChatGPT Plus Code Interpreter!
The Code Interpreter is now available to all ChatGPT Plus subscribers.
You have to activate the Code Interpreter in the ChatGPT settings.
You can upload one file with up to 100 MB, but you can use a zip file to circumvent the limitation to only use one file.
The Code Interpreter uses Python as a programming language to create and execute code to answers your request.
There are many nice results created and shared on Twitter, but I want to share two rather old ones out of the Alpha phase. The first visualizes crime data from a CSV file, and the second segments music markets from an Excel file and generates business strategies for them.
My take: Uploading files and executing code are two new superpowers for ChatGPT. The limit is only your imagination.
💬🚀 ChatGPT competition: Claude 2 from Anthropic Started Free Beta
The LLM Claude from Anthropic is so far mainly known for supporting huge inputs of up to 100k tokens, which corresponds to roughly 75k words.
You can now test and use their new version Claude 2 in a free Beta on their webpage.
So far, officially only for users from the US and UK, but by using a (free) VPN like Windscribe you can circumvent that limitation.
You can upload up to five files which are used at the same time to answer your question.
In my first test, the results for answering a question with content from a PDF were excellent.
The latest version (v1.29.0) of the browser extensions ChatHub supports Claude 2 by using the web-app in the background.
My take: Happy to see competition for OpenAI, but the privacy policy of Claude is rather scary. I need an AI to analyze it. 😅
🎨🚀 Pushing Artistic Boundaries: Leonardo's Canvas v2.0
Leonardo launched Canvas v2.0, which looks like a nice competitor to Photoshop’s generative fill feature.
You still get daily free credits to use Leonardo, which are enough if you only occasionally create an image.
However, there is still a waitlist and if not already done, you should consider registering yourself for accessing Leonardo.
Matt Wolfe showed in his AI news video the official demo of the Canvas v2.0 feature.
My take: Leonardo is a great service if you want to generate an AI image occasionally without a monthly subscription.
Privacy-Friendly AI
🔬🧪 Stable Diffusion XL 0.9 Model Released for Researchers
The model of Stable Diffusion XL 0.9 was released as Researcher Early Access on HuggingFace.
So far, you only could, and still can, use Stable Diffusion XL 0.9 on ClipDrop.
Official launch blog article from 22nd of June.
The official Paper includes many impressive images generated with it.
My take: Midjourney seems still ahead, but Stable Diffusion is making a big leap forward. Looking forward to the 1.0 release to the public.
🐋🧠 Unveiling OpenOrca: Revolutionary Steps in open-source LLMs
OpenOrca-Preview1-13B, a new model developed using the OpenOrca dataset, has been fine-tuned on LLaMA-13B. This dataset is a reproduction of the one generated for the Orca Research Paper by Microsoft Research.
The Orca research paper is about a new neural network model that learns from step-by-step explanations generated by GPT-4.
The model has been trained on less than 6% (200k entries) of the total OpenOrca data, after filtering out certain responses that could harm model reasoning capabilities.
Even with this small training dataset, the model has shown state-of-the-art results, with training costs under $200.
Learn more about the OpenOrca model on HuggingFace, where you also can test it in an CPU only playground.
My take: I love to see how people commit to the reproduction of published approaches in open-source implementations. Costs for training compute is often the limiting factor, but this work shows that there are approaches which allow the generation of competitive results in a cost-efficient way.
PKM and AI
🤖💡 Decoupling Prediction and Judgment: The Evolution of AI in Decision-Making and Society
The following is a short summary of the article The Rise of Prediction Factories: How AI Allows Us to Replace Rules with Decisions from Tiago Forte.
AI can decouple prediction from judgment. Judgment assigns value to outcomes, while prediction says what's likely to happen.
AI can generate predictions much cheaper than humans by analyzing large amounts of data.
AI allows us to replace rules with more customized decisions based on personalized predictions.
It will take a long time for AI to fully transform society, requiring new systems and organizations.
People will still be needed to exercise judgment, act, and provide new perspectives.
We can enhance our value by optimizing for fewer but better decisions using AI predictions.
Accessing unconventional data sources will be important to feed AI systems and improve predictions.
My take: If this sparks your interest, go and read the full article yourself, which will take about 20 minutes.
Podcasts
🤖🌐 Marc Andreessen's Vision on AI, the Internet, and Our Future!
Episode #386 of the Lex Fridman Podcast with Marc Andreessen about the Future of the Internet, Technology, and AI.
Marc Andreessen is the co-creator of the early web browser Mosaic, co-founder of Netscape, and co-founder of the venture capital firm Andreessen Horowitz (a16z).
He wrote an essay with the title "Why AI will save the world".
Small teams and individuals have often developed revolutionary software in the software world. This pattern gives hope for the future of AI and technology. Software transmutes labor into capital, creating value through human thought.
Despite the availability of powerful tools, there appears to be a lack of hyper productive people. Distraction and lack of motivation may play a role.
Striving for satisfaction and purpose rather than chasing happiness is important.
My take: Worth listening to the complete episode as my highlights only scratch the surface, and it was a fascinating conversation.
⌨🔮 George Hotz on AI, Coding, and Our Future!
George Hotz is a guest in the Latent Space podcast episode from June 20th, 2023.
George Hotz was the first person to unlock the iPhone, jailbreak the PS3. He founded the AI start-up Comma.ai which works on self-driving car software and hardware and started a new company named tiny corp in November 2022 which aims to port machine learning instruction sets to GPUs.
Dumping code on GitHub without active development and outdated issues does not encourage contributions to an open-source project.
Improving models can be done by focusing on inference instead of training, and on-device fine-tuning is desirable for Comma's operations in cars.
AI will supercharge workflows in coding, similar to how Photoshop created new opportunities for artists.
While tools may eventually replace humans in certain tasks, a complete replacement is still 20 years away.
The prevalence of chatbots and AI-generated content can be frustrating, and a payment model for emails may help combat spam.
George Hotz stated that his next company will be AI girlfriend, representing the merging of humans with technology.
My take: After the interview of George Hotz by Lex Fridman highlighted in the last issue of the newsletter, in this one he shares many more thoughts which are not less interesting.