Did AI reach AGI this week???
History rhymes... or the more things change, the more they stay the same. But we can make a bunch of money and save a bunch of time by recognizing these patterns and riding these waves. That's one of my biggest points on investing in MegaWealth: Investing.
Buy MegaWealth: Investing on Amazon
Today I will share:
Updated brief history of AI
Whether an AI winter is coming
What just happened (AGI?), and what's next
HISTORY OF AI
AI has been around since the advent of code-breaking in World War II.
Here is a timeline of major milestones in AI development from World War II to the present (don't worry about understanding each individual item... just scan for the ideas (and where I bolded words) and I'll summarize at the end):
1940s-1950s: Foundations of AI
1943: McCulloch and Pitts Neural Model - Warren McCulloch and Walter Pitts propose the first mathematical model of artificial neurons.
1945: Turing's Early Ideas - Alan Turing introduces the concept of a "universal machine," laying the groundwork for computational thinking.
1950: Turing Test for evaluating machine intelligence is introduced - Alan Turing publishes "Computing Machinery and Intelligence."
1960s-1970s: Symbolic AI and Expert Systems
1965: ELIZA Chatbot - Joseph Weizenbaum developed ELIZA, one of the first natural language processing programs.
1966-1973: First AI Winter - Missed expectations reduce AI interest.
1980s: Expert Systems Boom
1980: Rise of Expert Systems - AI applications like XCON, which configures computer orders, see commercial success.
1987-1993: Second AI Winter - Missed expectations and high costs lead to a large decline in AI funding.
1990s: Machine Learning and Practical Applications
1997: Deep Blue Defeats Kasparov - IBM’s chess-playing computer, Deep Blue, beats world chess champion Garry Kasparov.
2000s: Data-Driven AI
2011: Watson Wins Jeopardy! - IBM's Watson beats human champions in the quiz show Jeopardy!
2012: ImageNet Breakthrough - Alex Krizhevsky, Geoffrey Hinton, and others win the ImageNet competition, providing a large, standardized dataset for training and evaluating image recognition algorithms.
2010s: AI Goes Mainstream
2014: GANs Introduced - Ian Goodfellow proposes Generative Adversarial Networks (GANs) - today's LLMs are a type of GAN.
2016: AlphaGo Defeats Lee Sedol - DeepMind’s AlphaGo beats a world champion Go player. Go is far more complex than chess.
2017: Transformer Model Introduced - Google introduces Transformer architecture, revolutionizing NLP (language processing).
2018: GPT-2 Released - OpenAI releases GPT-2, a powerful language model.
2020s: AI Explosion
2020: OpenAI launches GPT-3, showcasing advanced text generation capabilities. Only available via API or specialized links that OpenAI was giving approval to one-by-one.
2021: AlphaFold Solves Protein Folding - DeepMind's AlphaFold achieves groundbreaking accuracy in predicting protein structures.
2022: ChatGPT Launched - OpenAI’s GPT-3.5-based ChatGPT democratizes conversational AI.
2023: GPT-4 Released - OpenAI enhances multimodal capabilities with GPT-4, improving text and image understanding.
2024: AI arms race between OpenAI, Google's Gemini, and Anthropic's Claude, with Xai not far behind.
Emmy's AI History Cheat Code:
1. AI started by building skills that are natural for computers: math.
2. AI made strides in other areas like vision and language, working for decades, then seemingly out of nowhere finding breakthroughs.
3. Areas of progress included:
a. Image recognition (culminating in cars recognizing animals vs people, helping us stay within lanes, and stopping before we crash.)
b. Playing games such as Chess and Go, which is an extension of their math capabilities.
c. Language processing culminating in today's LLMs (large language models). Notice how I didn't say "understanding" as these language processing algorithms may seem like they understand, but that is because what we say next follows patterns that are sadly predictable.
Is an AI Winter Coming?
History rhymes ... but rarely with predictable timing we ascribe it to after the fact. That's why I wrote last week about Warren Buffett warning of the perils of debt. He says your thesis may be right, but what really matters is lasting long enough to arrive at the point where you are proven correct.
AI Winter - Debate Style:
Yes, an AI winter is coming!
Things can't stay this high for this long. We have been in a fury of innovations, investments, and crazy stock appreciations like NVDA.
AI usage in the past quarter (3Q24) has declined, with users not seeing how it will give them the benefits of saved time or better quality output, where it is taking up their current time to learn how to use it.
I can attest to this experience whereby it took me longer to test multiple new tools, learn how to use them, decide on one, and start using it vs just doing it the old way! The time savings are yet to come.
Winters arrive when you least expect them to... They arrive when people finally decide that this time is different. If not now, when?
No, AI has plenty of legs!
What IS different this time is that average people are using AI. That wasn't the case before. And technologies that worked, like backup cameras for cars, that use visioning software, didn't go through a winter.
ChatGPT was the fastest-adopted technology EVER.
AI capabilities are increasing at a geometric rate, so fast that we can barely recognize the technologies we were using mere months ago.
More than 80% of corporations and 90% of call centers have integrated AI into their processes. It is here, and it is being used.
I remember in 1996, just after the Mosaic web browser came public via IPO, the Fed Chair Alan Greenspan, gave his famous "Irrational exuberance speech: unfounded market optimism that lacks a real foundation of fundamental valuation, but instead rests on psychological factors.
The market, especially the internet stocks he criticized for being in a bubble, went straight UP for 4 YEARS following his speech.
He was eventually correct, but he would have lost his job and wealth if he was a money manager betting against those stocks in 1996.
As I write about in MegaWealth: Investing, my smartest bosses had long-term theses, but let each day's evidence guide their investments. It's like driving a car. Know the weather forecast, but the most important evidence you'll react to comes from looking up at the sky and ahead on the road.
Right now, the road of AI has a LOT of innovation and some dissapointment.
Where are we now?
The common refrain on AI today is: who is using it? I was at dinner the other night and a husband was shocked and curious to find out that his wife was using AI at work fairly extensively. People are using it, and some don't readily admit is.
The question is, are we just trying it out, or is it becoming part of our workflow, where it saves us more time than it took to learn the tech?
Later we can talk about whether AI can make our work better, but for right now, I have a low bar of saving me more time than it took me to learn and use the technology. Not all the apps I used passed this test.
Over the next few weeks, I will categorize and go over the best AI apps I have found: video, virtual assistant (agents), and more. Next week, we will cover the macro view.
Before and after OpenAI o3.
When I started writing this newsletter, OpenAI hadn't released o3.
If I had written this email a month ago, I would have told you that Anthropic's Claude is outperforming OpenAI and especially Google on many tasks.
If I had written this email a week ago, I would have told you that Google's AI leapfrogged both Claude and ChatGPT/OpenAI, especially being able to read my books and then create content, like podcasts based on my book content. These AI-generated MegaWealth Explainer Podcasts are currently on my website.
But I'm not writing you a month ago, or even a week ago. I am writing you on December 21, 2024, one day after OpenAI announced o3.
OpenAI has been running a fun marketing gimmick online.
They announce a new product each day for the 12 days of Christmas. Like perfect bookmarks, Day 1 they announced o1, and Day 12, they announced o3. They also said there is not going to be o2 because that's the name of the British telecommunications company.
ChatGPT4 vs o1
I asked ChatGPT to compare itself with o1. Here's the summary:
ChatGPT-4 excels at a broad spectrum of tasks requiring general language understanding and interaction, while O1 is likely tailored to address specific challenges in business or technology, excelling at precision and efficiency in those areas. o1 works within the context of what you are meaning to accomplish.
o1 truly opens the door for specialized vertical applications. And if you want to experience it first hand, you can call it and have a conversation:
1-800-ChatGPT
You can also download it on your phone and turn on video access. Then you can show it what you are seeing. Are you walking down a street? Want it to tell you what it sees? Draw a diagram and it will guide you. Show it your computer screen or pages in a book and ask it questions.
o1 coached Anderson Cooper of 60 minutes on anatomy by instructing him to diagram on a blackboard where the heart, liver, and brain were and then making corrections to his work.
That's the magic of o1.
Keep in mind: we don't know when these models were "ready" vs when they were announced. o1 was announced and released only 12 days ago. And it's impressive.
Yesterday, December 20, 2024, OpenAI announced o3. They didn't release the model due to wanting to do more extensive safety testing. Many online are calling this AGI (Artificial General Intelligence).
Here's why:
Sam Altman, OpenAI's CEO is not willing to call o3 AGI.
But it seems to me that it's significantly close enough that they want to test it for longer than previous models. He brought the creator of this ARC-AGI scoring system on the announcement video with him to say they are collaborating to make a new measuring system.
How smart is o3? It received 97% on PhD math problems. It received 88% in PhD science problems that expert PhD's get right 70% of the time. It scored higher than every software engineer at OpenAI except one.
It's one thing to be super smart in one area. But this model is smarter than the smartest people in multiple areas at once.
What's Next?
If o3 isn't AGI, we are a lot closer than we thought we were at the beginning of this week...
There is so much we still don't understand about the human brain. So how can we create a technology to match or exceed what we don't understand?
In the meantime, think back to the 1950s and early 1960s when we dreamt of sending men and women to the moon. We did. But perhaps even more importantly, that goal then inspired us to develop many technologies along the way...
We are building some exciting new capabilities in AI, whether or not we reach the holy grail of AGI, and indeed whether or not we even think AGI is the goal!
It seems to me that o3 is smarter than the smartest PhD's, and when that AI starts teaching itself and improving upon itself, we may be mere months away from AI being smarter than the smartest humans.
In this historic moment, I have paused to reflect, learn, and share with you what I'm learning.
Here's what I'll share over the next few weeks:
Macro1: What do the most experienced minds (Gates, Bezos, Schmidt, etc.) in tech think about AI? Opportunities and risks...
Macro2: What type of companies will be built on AI or by AI? What will the industry structure look like? What does it look like now?
Micro1: AI tools - video. What can AI tools do to save you time?
Micro2: AI tools - productivity. Which ones can save you time or more?
Micro3: AI Assistants. Are we there yet? Can you build and use an AI assistant or agent today? (This is called Agentic AI)
Micro4: AI Applications. Can AI help you make personalized applications?