Zuck’s $1 billion AI timebomb

Nobody ever said Mark Zuckerberg wasn’t smart. A jerk, maybe, um, Not nice, sociopathic, robotic, “on the spectrum” - these are some names he has been called, but never dumb. Last week he laid his cards on the table. He plans to make sure nobody (else) makes any money from AI. Meta/Facebook just released its next-generation LLM, Llama 3, as open source (i.e., free to use), and it’s a banger. 

Two versions were released, a small (8-billion parameter), and a medium (70-billion parameter) model. A large model is still training, expected to be over 400 billion parameters. The performance measures of the 2 smaller models are off the charts. If the larger one follows suit, it will be at least comparable to the best models out there, none of which are free.

Here’s the evil genius part: Facebook doesn’t need to make money from AI. Facebook makes ungodly amounts of money from its soul-sucking highly popular social media platforms. So venture-backed “foundation model” AI companies like OpenAI and Anthropic may be toast if Zuck makes top-tier models free. Google is also in the crosshairs, because AI is highly likely to disrupt Google’s ad-fueled search.

And here’s the real killer: Zuck is dropping $1 billion on training his new chatbots. Nobody but Microsoft is likely to be able to keep up with that rate of spending for long.

Meanwhile, after lighting the fuse of this timebomb, Zuck is nonchalantly stuffing little AI enhancements into his social media platforms, making them more fun, more useful - and if he is right, even more addicting.

Zuck’s $1 billion AI timebomb

Microsoft releases powerful, tiny model that can run on a phone

Yesterday, Microsoft announced Phi-3, a new family of small but powerful AI models. The smallest, Phi-3-mini is only 3.8 billion parameters, but performs almost as well as ChatGPT (which has 175 billion). This feat of engineering raises the possibility that you could have your very own AI chatbot running on your smartphone. Mark this down as a trend: the huge, clanking AI models that run in the cloud will be progressively streamlined to run on the user’s own equipment. Some day soon, your smartwatch may be a whole lot smarter.

Microsoft releases powerful, tiny model that can run on a smartphone

Boston Dynamics announces Atlas, new electric humanoid robot

Look ma, no hydraulics! Boston Dynamics is an industry-leading robotics company, whose clever videos of 2- and 4-legged robots dancing the Funky Chicken have made them social media darlings as well. Now Atlas, a fan-favorite humanoid robot, has undergone a complete redesign, losing the bulky hydraulics for a sleek new all-electric look.

Once again, Boston Dynamics made a jaw-dropping video of its capabilities, so click the link below for a quick demo.

Boston Dynamic’s new electric humanoid robot

AI can edit your genes

AI-bio startup Profluent has announced they have successfully edited the DNA in human cells using gene-editing proteins that were designed by an AI model they developed. They have open-sourced the resulting molecules under the company’s OpenCRISPR initiative.

AI can edit your genes

AI gives 30-minute early warning of AFib on your smartwatch

Researchers at the University of Luxembourg have developed a simple AI system that can predict an impending episode of atrial fibrillation 30 minutes ahead of time, just by monitoring the variability of the timing between heartbeats, known as the R-R interval. The goal is to make the model compact enough that it can fit into a smartwatch, so that users can be warned in time to take anti-arrhythmic medication to prevent the episode.

AI early warning of AFib on your smartwatch

AI speeds up Parkinson research by 10x, cuts cost 1000x

University of Cambridge researchers are using an AI model for drug discovery of anti-Parkinson’s medications. Until now, we have had medications that reduce symptoms, but none that slow or reverse the course of neural degeneration. These researchers have developed an AI model that has already proposed several promising molecules to block clumping of alpha-synuclein, the protein thought to cause the neural destruction that results in Parkinsonism. The AI screening process was 10 times faster, and 1000 times less costly, than traditional methods.

AI speeds up Parkinson research by 10x, cuts cost 1000x

GPT-4 passes 4 out of 5 of Israeli medical board exams

Researchers from the Tel Aviv University School of Medicine compared the performance of GPT-4 on the 2022 Israeli medical board exams with the scores of practicing physicians. GPT-4 achieved a passing score on 4 of the 5 specialties included in the evaluation, and scored higher than the majority of physicians in Psychiatry.

GPT-4 passes 4 of 5 Israeli medical board exams

How to get started learning about AI

The best way to learn about AI is to use it. There are several highly capable models that are free and publicly available on the internet.

Choose a model and start using it

Major free and easily accessible AI models include:

1. ChatGPT, the original - at openai.com.

2. Claude - at claude.ai

3. Gemini - at gemini.google.com

All 3 are about equally capable. ChatGPT is a good place to start, but Claude allows analysis of longer inputs, and Gemini uses Google Search to check its answers, and provides links to its sources.

All 3 share the same weaknesses. They all “hallucinate” at times, which means that they will give plausible answers which might not be correct. So do not use them as a substitute for Wikipedia. They also all struggle with math, complex reasoning, and planning.

They all share the same jaw-dropping ability to instantly generate grammatical and well-written text on almost any topic and in almost any style.

Easy, fun, and useful first projects

Try one or more of the following projects to get a feel for the capabilities of the model you choose:

1. Write a funny poem about hiking

2. Give me 5 reasons why dogs are better than cats

3. Plan a 5-day trip to Rome, Italy for me and 3 family members

4. Here is an email I want to send to my ____. Give me suggestions on how to make it better.

5. Here is a transcript of a meeting. <Note - all models allow you to upload text files.> Summarize the main points discussed, what decisions were made, what the follow-up action items were and who is responsible for each.

6. I have a difficult decision to make. <Describe the decision, and what makes it difficult for you.> Give me several options for handling this situation, and the pros and cons of each option.

Level up as you gain confidence

Once you get comfortable with using an AI chatbot, it becomes a quick assistant that can stimulate your own thinking, and make your writing and your decisions better.

A fun AI website for composing music

Udio.com is a single purpose AI website that allows you to create a song just by typing in what you want - like “a rap song that celebrates my son/grandson’s win in a soccer game”, or “a Country song that laments finding that someone ate the last cookie in the cookie jar.” You can write your own lyrics (or get ChatGPT to write them), or Udio will supply them, along with appropriate music and vocals.