I constantly think that November 30, 2022 will be one of the most important dates of my life. I believe the launch of chat GPT 3.5 will be considered in the future as the start of a new paradigm for Earth and the human race.
A while back I decided that on every November 30 starting in 2025, I will compile my notes on Artificial Intelligence and put them in a post. Continue below for my thoughts on AI in 2025.
About once a week, usually when I am showering, I marvel at the thought that humans have managed to turn energy into a form of intelligence. How our monkey brains managed to do that still wows me.
Both Sophie and I rely on LLMs for a considerable amount of our daily tasks and projects. She has been using ChatGPT Plus since June 2024, and I have been using Google Gemini Pro since May 2025. We probably average 2-3 hours a day each on the apps, far more than any other smartphone app. As a household we spend $55 CAD a month on LLM subscriptions.
I’ve been working on a project fine-tuning an LLM on multi-task workflows relevant to my domain expertise, which gives me a preview of the next iterations of the technology. I am excited to see how the next frontier of LLMs will increase the productivity of the white collar professionals who adopt them.
The best way to get LLMs to be useful for you is to view using LLMs as a video game: When you start playing a new video game, you need to learn how to use the controllers and the bounds of what you can do within the game. LLMs are similar. My suggestion is to take some of your free time and see what you can get an LLM to do. Over time you will be impressed on how much it can do.
Like video games, there are power-ups when using LLMs. My favorite is meta prompting. Here’s an example of what I mean by meta-prompting.
I have always found LLMs to be too agreeable and sycophantic. Some models (like Gemini) now have personal context setups, where you can give your LLM instructions on how you would like it to respond. Here is mine:
When responding to me, always adopt a concise and objective tone. Do not be agreeable or seek to please me. Instead, actively challenge my premises and assumptions to ensure rigorous critical thinking. Prioritize factual accuracy and counter-arguments over conversational harmony.
I am disappointed, but not surprised, that younger generations are using LLMs as shortcuts to schoolwork as opposed to enhancers. I have professor friends who have seen a serious degradation in the preparation of these young students who use LLMs to cheat their way into a degree.
I use Gemini a lot to learn about new subjects, and I use the following prompt I found on X as a starting prompt:
I would benefit most from an explanation style in which you frequently pause to confirm, via asking me test questions, that I’ve understood your explanations so far. Particularly helpful are test questions related to simple, explicit examples. When you pause and ask me a test question, do not continue the explanation until I have answered the questions to your satisfaction. I.e. do not keep generating the explanation, actually wait for me to respond first. I’m hoping that by tutoring me in this Socratic way, you’ll help me better understand how superficial my understanding is (which is so easy to fail to notice otherwise), and then help fill in all the important blanks. Thanks!
I then explain the subject I want to learn about and the resulting conversations are very enlightening. Here’s an example.
Nano Banana Pro is seriously impressive. The image below is AI generated.
I predict that social media as we know it will evolve into something completely new. I do not see myself using Instagram more often if those who I follow are posting AI generated content, which I will have a hard time discerning from real photos / videos.
I am confident that many people will use AI to fake their lives on Instagram in an effort to gain status. I believe this will lead to a significant reduction in visual social media consumption in the near future.
Almost a year ago I wrote a post about my predictions on AI. My predictions still stand (for now).
I managed to get in a Waymo on our trip to Austin back in April. We waited 27 minutes for it to arrive but it was worth it. It was a mind-blowing experience.
‘AI is coming for our jobs’ paradigm is still far away. If you lose your job in 2026 and you think it was because of AI, it is likely that you are partially right. You did not lose your job to AI, you lost your job because people at your firm became far more productive as they harnessed the power of AI, and the firm realized they could be as productive or more productive with less human labor.
I think AGI is coming before I retire, but I am not confident enough to put a number on it. To me it seems that there are still meaningful breakthroughs in agency, memory, and self learning that need to happen before we get there.
Even if AI advancement stalls today, and the best generalized model we have moving forward is Gemini 3.0, the technology will still transform human knowledge work as we know it. There is a lot of value to be made in transforming and applying current cutting edge LLMs to many different domains, and there are thousands of firms all over the world working on that.
The AI trade was the big winner in 2025. If you invested in virtually any stock that touched AI, you probably beat the S&P 500 for the year. I believe in 2026 there will be more divergence in the ecosystem as competition in certain domains heats up and capital allocation comes into question.
As far as I am aware, there is no data pointing to a slow down in AI compute demand. Unlike the railroad and telecom examples that are consistently mentioned as comparable, there are no idle cutting edge data centers anywhere waiting to be used. As soon as a data center is finished and turned on, the utilization goes to 100%.
The ultimate bottleneck in AI Datacenter build up will be power generation. All the other bottlenecks that currently exist will be solved via competition in the next few years. Generating enough base load power for these datacenters is the crux of the AI Infrastructure problem.
Canada is especially well positioned to take advantage of the AI Infrastructure build up. For a country of our population and economic size, Canada has the following advantages (From my letter to Minister Evan Solomon):
- Proximity and Relationship with the United States: Despite recent geopolitical developments, Canada is a natural ally to the United States. All major private enterprises engaged in this field have a physical presence in Canada, and considerable cross-border investments already exist. The United States government is determined to lead in this domain, and Canada possesses all that is necessary to be its most productive ally.
- Abundant and Diverse Energy Resources: Canada can offer the vast and reliable energy supply crucial for AI data centers and operations.
- Strategic Assets: Canada offers ample land, freshwater, and access to the financial capital necessary for large-scale infrastructure projects.
- World-Class Human Capital: We have a deep pool of talent in AI research and development, as well as the skilled workforce required to construct and manage state-of-the-art facilities.
- Strong Innovation Ecosystem: Canada was the first country to launch a national AI strategy and continues to foster a supportive environment for AI innovation and investment.
So far I am disappointed in what our Minister of Artificial Intelligence and Digital Innovation has managed to accomplish this year. I hope to see some large scale AI Infrastructure projects in Canada in construction by this time next year.
I used Google Gemini 3.0 Pro in thinking mode as an editing assistant to write this post.