Day 20 - Some Light Reading

Thank you for joining me on this homework journey! I hope you’ve found it interesting and thought-provoking. Maybe even enjoyable in parts! 🙂

My goal was to present a series of exercises that would help you spend some time getting used to today’s AI, especially LLMs, with the idea that you need at least, say, 6⅔ hours (4 weeks, 5 days a week, times 20 minutes) to get your head around what AI is and isn’t today.

I’m looking forward to the conversation on tomorrow’s Living Between Worlds call (April 16, 2025)!

I’ve also started a community on Discord, if you’d like to continue chatting: peterkaminski.ai Discord. (It’s free, don’t pay money to use it. Ignore “Nitro” ads. If you’re new to Discord, search YouTube for tutorials. There are some other great AI Discord servers, so if you’re into AI, it’s worth getting started here.)

I’m going to recommend a few posts and articles to read. This is not a “best of list,” and it’s certainly not meant to be comprehensive, but these are pieces that I’ve found to be well-written and thought-provoking and useful.

They aren’t in any particular order. Read my notes and choose the ones that sound most interesting first!

by Nan Yu, 2025-04

This is a very short piece by a UX expert. I love this line: “A tool with infinite flexibility, no matter how powerful, is inaccessible to a general audience.”

A thing to remember today: even though ChatGPT is easy to use, it doesn’t really expose a lot of ways to usefully work for most people. It’s really early days for AI, and you still have to know what to ask for, how to ask for it, and do strange things to integrate AI with the rest of your life. (Remember Day 8 - Computer Skills? 😜)

AI will get easier to use, more useful for most people, and more integrated into our lives—all for better or worse.

And my theory: pushing yourself now to learn a little of what’s under the hood before it’s been made easy and disappears into the background is really worth it. Hence our AI homework!

interview by Michiaki Matsushima, on Wired, 2025-04-01 (paywall)

I find this to be a very thoughtful and sensible way to think about how we can move into the future along with AI. It’s a good use of your free monthly article if you don’t have a Wired subscription.

I asked GPT-4.1 to look at the article via PDF, and asked it how my response was. It said something quite reasonable, “Your response is a good start: it's positive, and it gives a quick recommendation. However, it could be made even better.” Without being prompted, offered an expansion, notionally written from my point of view, which I share here, with me taking responsibility for sharing it:

I found this interview with Yuval Noah Harari both thoughtful and sensible in how it frames our uncertain future with AI. Harari’s point that “we are now in a situation where trust between human beings is at stake. This is the greatest danger” really struck me. He makes a compelling case that as AI becomes more capable of making its own decisions, the problem isn’t just the technology itself, but whether or not humans can maintain enough trust and cooperation to steer it wisely.

The article avoids the usual extremes and instead encourages a realistic “middle path,” which feels refreshing given how polarized debates around AI can be. I also appreciated Harari’s idea that AI isn’t just a tool like the printing press, but a new kind of agent able to generate its own stories.

It’s definitely worth using a free article credit, especially if you’re interested in AI and how it’s shaping society.

by Dario Amodei, 2024-10

Exactly what it says on the tin.

Read about the poem from which Amodei borrows his title: All Watched Over by Machines of Loving Grace (Wikipedia).

by Tobi Lutke on X.com, 2025-04-07

Lutke, Shopify CEO, says the internal part out loud (publicly), because the internal memo was being leaked anyway. It's a really well-written piece about the power and opportunity available in using AI at work, and how he expects everybody within Shopify to get on board.

I think similar arguments apply to your personal life as well as your work life. If you use a computer, you should be using AI a lot. Whatever productivity and life quality boost (or loss?) you got from integrating “computers” into your life, that's sort of what to expect from integrating “AI” into your life.

There are lots of pull quotes in Lutke's piece; this is the one I want to share:

What we have learned so far is that using AI well is a skill that needs to be carefully learned by… using it a lot. It’s just too unlike everything else. The call to tinker with it was the right one, but it was too much of a suggestion. This is what I want to change here today. We also learned that, as opposed to most tools, AI acts as a multiplier. We are all lucky to work with some amazing colleagues, the kind who contribute 10X of what was previously thought possible. It’s my favorite thing about this company. And what’s even more amazing is that, for the first time, we see the tools become 10X themselves. I’ve seen many of these people approach implausible tasks, ones we wouldn’t even have chosen to tackle before, with reflexive and brilliant usage of AI to get 100X the work done.

It is similar to what I've been saying; if you've tried AI a little, you haven't done enough. You need to use it a lot, even if it feels awkward and dumb at first. Love it, or even hate on AI if you want, but you gotta get familiar with it first.

by Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean

I told you that you could pick whichever, but I think you should make this one of your must-reads.

AI 2027 is a sobering, informed deep dive on AI takeoff scenarios. It reads like science fiction, but it’s not meant to; it’s a straightforward extrapolation.

AI takeoff scenarios are a way of thinking about how fast and how dramatically things could change once AI reaches a certain level of capability. We are really really close now.

AI takeoff is one of the main reasons I'm trying to teach people to use and get informed about AI. The AI we have now is powerful, but not world-ending.

Humans need to get together to understand and make good decisions; in five years it could be all over.

by Alexandra Samuel, 2025-04-15

I love Samuel’s Canadian perspective in this post—hey, this is our sovereign country! Sometimes people in the US forget.

But especially, she’s really eloquent and correct here:

We are now firmly in the era of AI, and the use and demands of AI are only going to accelerate. We might dream of resisting by refusing to use AI in our own personal or professional lives, but the impact of individual AI abstainers will be dwarfed by the power- and water-hungry footprint of government, industrial and business AI usage. We need conscientious citizens, smart NGOs and democratic governments to harness the power of AI so that we can effectively challenge the ever-growing power of AI-enabled companies and autocracies.

by Stephen Wolfram, 2023-02-14

And lastly, this article is now long in the tooth, but it is a nostalgic favorite of mine even still.

Wolfram is both really smart and a good writer, and reading this article helps me feel like I understand how ChatGPT 3.5 worked.

IMPORTANT NOTE: We are now a generation and a half or two past ChatGPT 3.5, so this isn’t the only way LLMs work now. For more information, ask your state-of-the-art LLM:

• What major advances have there been in LLM technology since ChatGPT 3.5?

And for more detail, ask your LLM to dive into some of these (this list prepared by GPT-4.1):

Major Advances in LLM Reasoning Since ChatGPT 3.5

• Mixture of Experts (MoE)

  • Models use specialized sub-networks ("experts"); only a few activate per input.

  • Enables larger, more efficient models (e.g., Mixtral 8×7B, GPT-4).

• Chain-of-Thought (CoT) Prompting

  • Encourages step-by-step reasoning.

  • Boosts performance in logic, math, and complex tasks.

  • Variants: Tree-of-Thoughts, ReAct (reasoning + tool use).

• Self-Consistency & Reflection

  • Multiple reasoning paths → best answer selected or aggregated.

  • Models can reflect and self-correct.

• Tool Use & Retrieval-Augmented Generation (RAG)

  • Integration with tools (e.g., search, calculators).

  • Supports real-time, grounded answers.

• Scratchpad / External Memory

  • Stores intermediate steps for multi-part reasoning.

  • Improves coherence over long tasks.

• Agentic Planning

  • Models decompose goals, plan steps, and execute workflows (e.g., Auto-GPT).

  • Enhances autonomous problem-solving.

• Generalization & Robustness

  • Better at novel, abstract tasks.

  • Curriculum learning improves reasoning over time.

In summary: LLMs now reason more like humans—step by step, with specialization, memory, tool use, self-correction, and goal-directed planning.

[Note by Pete: “more like humans” there means in outcome, not in the underlying artificial vs. biological neural architecture.]

And that’s a wrap! Thanks for joining me for Pete’s AI Homework!

When I get a round tuit, I’ll prep an overview of all the lessons and post that, so look for that in your inbox at some point.

In the words of Sergeant Esterhaus, Let’s roll. And hey... let’s be careful out there.

~ Pete