Happy Birthday, ChatGPT!

A little more than a year ago, neural networks were the Rodney Dangerfield of the artificial intelligence field - they got (almost) no respect. Then OpenAI released what they thought was an amusing little side project, a chatbot they called ChatGPT. To their very great surprise, it quickly became a monster hit. They had thought that if it did really well, it could eventually attract a million users. It reached a million users in 5 days. It eventually reached 100 million users in 2 months, the fastest takeoff of any web app in history.

In the year since this meteoric debut, OpenAI has never taken its foot off the gas, shipping feature after upgrade after feature. Venture capitalists are pouring billions into AI startups. Tech giants are revamping their entire legacy product lines to make them more AI-powered. Governments are scrambling to figure out how to regulate AI (or not.) It is more than possible that this almost-accidental launch will mark a transition point in human history. Not bad for a silly little chatbot.

Infographic: The 50 most visited AI tools in the past year

The folks at Writerbuddy produced this nifty little graphic that shows the relative popularity of the 50 most visited AI tools on the web. A few points stand out. OpenAI accounted for 60% of all the 24 billion visits in that time, a stunning domination. Even more surprising (at least to me) is that Character AI, the “AI friend” company, was second, at 3.8 billion visits. As we have reported in the past, Character AI has insane user engagement figures, with the average user spending up to 2 hours a day chatting with their personalized AI chatbot friend. Click the link for more analysis.

Top 50 Most Visited AI Tools

Google Deepmind announces Gemini, its next-gen chatbot

Google invented modern AI. The seminal paper that laid out the foundational architecture of all the most powerful AI systems today (“Attention is All You Need”) was written by 8 Google AI researchers. But Google slept on its opportunity, and all 8 of these mega-talents eventually left for greener pastures (much greener, with $billions in VC funding.)

Ever since ChatGPT burst on the public consciousness, Google has been scrambling to retake the initiative. Mostly with pronouncements about how really, truly, super super good their next AI product, codenamed Gemini, is going to be. This week they did it again, with more detail. You can use a midlevel version of it in Bard, Google’s chatbot, similar to the way Microsoft’s Bing search engine is powered by GPT-4. Google released some performance metrics against benchmarks that looked impressive, but there is as yet no independent verification. Full release of Gemini is not expected until sometime next year.

Google Deepmind tells us again how good it’s going to be

Stable Diffusion positions itself for acquisition

How the (high and) mighty have fallen. Stable Diffusion, one of the early breakout successes in the text-to-graphics sector of AI, and once valued in the multiple $billions, had 2 big problems. First, its 2 top competitors, Midjourney and OpenAI’s DALL-E, were really, really good. Second, the company CEO, Emad Mostaque, was an egomaniac whose terrible management style caused all of his best talent to quit. Investors revolted, and now the company is shopping itself around for acquisition. Sic transit egomania.

Stable Diffusion seeks acquisition

“Visual Anagrams”: AI-generated optical illusions

AI-generated graphics are getting cooler all the time. Recently, an open source project has developed a method for creating “visual anagrams” - optical illusions that switch their appearance from one object to another when rotated. Definitely worth clicking the link to see some of the creations in action.

New open source graphic AI generates optical illusions

Google Deepmind uses AI to design 380,000 new materials.

Researchers at Google have developed an AI system that can design entirely new materials - and they have already designed 380,000 candidates that are predicted to be stable. Each new compound has to be made and tested for its detailed properties, so Google teamed up with Lawrence Berkely National Laboratory to have a robot methodically make and test each one. Researchers are hoping that better materials for batteries, solar cells, and other critical needs are hiding in the AI-genrated designs.

Google Deepmind uses AI to design 380,00 entirely new materials

Study shows that GPT-4 works as a math tutor for SATs

Sure, students can use ChatGPT to cheat on their homework. But it turns out that chatbots make awesome tutors. A new, well-designed study found that students who received explanations of math concepts from GPT-4 performed significantly better on SAT math problems.

Tutoring by GPT-4 improves student performance on SAT math problems

Llamafile is the best way to run an AI on your own computer

ChatGPT, Bard, Claude, and all the other mainstream Large Language Models run on superclusters of AI chips in the cloud. But sometimes a girlgeek (or guy) wants to curl up with a nice bite-sized AI on their own Macbook (or, in my case, a Corsair Vengeance GamerPC with an Nvidia 4090 GPU, but that’s just moi.) The Open Source movement is made of, and for, folks like us, and now we have a simple way to get the AI onto our machines. It’s called Llamafile, and it uses Llama 2, Meta’s open source LLM, and greatly simplifies the usually fussy installation of such models. 10 out of 10, recommend.

Llamafile simplifies installing AI on your own computer

AI agents learn from humans by imitating them

In a recent paper in Nature, researchers from Google Deepmind report on a method to allow AI agents to learn from humans in real time using imitation. Human experts would use various techniques to interact with a virtual 3-dimensional world, and the agents would learn to perform the same operations simply by imitating what the humans had done. Easy for humans to do, but very, very hard for robots. This cuts out all that tedious programming - just show the bot what to do, and it does it.

Nature article on AI agents learning by imitating humans

Ex-Apple employees want to reimagine desktop computers

Three ex-Apple employees have founded a startup that aims to make a computer operating system for the AI era. Their intuition is that modern computers are optimized to run pre-packaged software applications, which are efficient but inflexible. The future, they believe, is that everyone will create their own personalized software, just by talking to their friendly local chatbot, which will do all the coding for them.

Ex-Apple employees are bringing AI to your desktop

Robots made of human cells repair human neural tissue

Researchers at Tufts University have developed a method for growing tiny mobile robots from human tracheal epithelial cells, which they call “anthrobots.” They then were able to use the anthrobots to repair a sheet of neurons that had been scratched. Within 3 days, the damage was completely healed. The team foresee that in the future, robots grown from a patient’s own cells could roam the body to “clear arteries, break up mucus, or deliver drugs.”

Robots made of human cells repair neural tissue

AstraZeneca backs startup for AI-designed oncology drug

Pharmaceutical giant AstraZeneca has invested $247 million into a collaboration with startup Absci to develop an AI-designed antibody to fight cancer.

AstraZeneca bets on AI to develop cancer drug