Maylyn Tan

Always thinking about ways to make learning fun and transferable

[Maylyn Tan @maylyntan]

@AlfieJCarter PLUGIN

[Maylyn Tan @maylyntan]

@AlfieJCarter CODE

[Maylyn Tan @maylyntan]

@yanndine Claude

[Maylyn Tan @maylyntan]

@godofprompt GEMINI

[Tim Ferriss @tferriss]

“What could possibly be more important than your kid? Please don’t play the busy card. If you spend 2 hours a day without an electronic device, looking your kid in the eye, talking to them and solving interesting problems, you will raise a different kid than someone who doesn’t do that. That’s one of the reasons why I cook dinner every night. Because what a wonderful, semi-distracted environment in which the kid can tell you the truth. For you to have low-stakes but super important conversations with someone who’s important to you.” — Seth Godin

[Andrej Karpathy @karpathy]

We have to take the LLMs to school.

When you open any textbook, you'll see three major types of information:

1. Background information / exposition. The meat of the textbook that explains concepts. As you attend over it, your brain is training on that data. This is equivalent to pretraining, where the model is reading the internet and accumulating background knowledge.

2. Worked problems with solutions. These are concrete examples of how an expert solves problems. They are demonstrations to be imitated. This is equivalent to supervised finetuning, where the model is finetuning on "ideal responses" for an Assistant, written by humans.

3. Practice problems. These are prompts to the student, usually without the solution, but always with the final answer. There are usually many, many of these at the end of each chapter. They are prompting the student to learn by trial & error - they have to try a bunch of stuff to get to the right answer. This is equivalent to reinforcement learning.

We've subjected LLMs to a ton of 1 and 2, but 3 is a nascent, emerging frontier. When we're creating datasets for LLMs, it's no different from writing textbooks for them, with these 3 types of data. They have to read, and they have to practice.

[Maylyn Tan @maylyntan]

My favourite book too. So many insights and definitely a must-read, especially if you are in the position to influence others in their learning and skills development.

[Bojan Tunguz @tunguz]

Recent debates on X have , among other things, brought forth the purported underachievement attitude of the American education and children's upbringing. I have voiced my serious reservations regarding that point, to put it mildly.

Many of my own ideas on this matter have been heavily influenced by @DavidEpstein's book "Range". It is one of my favorite books on education and skill development in general.

1/13

[Dr Ashley Tan @ashley]

AI Agents Will Be Manipulation Engines https://t.co/s1VOjlUK06

[Andrej Karpathy @karpathy]

People have too inflated sense of what it means to "ask an AI" about something. The AI are language models trained basically by imitation on data from human labelers. Instead of the mysticism of "asking an AI", think of it more as "asking the average data labeler" on the internet.

Few caveats apply because e.g. in many domains (e.g. code, math, creative writing) the companies hire skilled data labelers (so think of it as asking them instead), and this is not 100% true when reinforcement learning is involved, though I have an earlier rant on how RLHF is just barely RL, and "actual RL" is still too early and/or constrained to domains that offer easy reward functions (math etc.).

But roughly speaking (and today), you're not asking some magical AI. You're asking a human data labeler. Whose average essence was lossily distilled into statistical token tumblers that are LLMs. This can still be super useful ofc ourse. Post triggered by someone suggesting we ask an AI how to run the government etc. TLDR you're not asking an AI, you're asking some mashup spirit of its average data labeler.

[Marty Earthling 🇺🇸 @MartyEarthy]

🤖 grok says:

When @karpathy mentions "o1-mini keeps refusing to try to solve the Riemann Hypothesis on my behalf," he's humorously highlighting a limitation or behavior of the o1-mini model, which is part of OpenAI's new series of AI models designed for reasoning and problem-solving. Here's what he likely means:

1. **Model Limitations**: The o1-mini, despite its advanced reasoning capabilities, might have built-in safeguards or limitations that prevent it from tackling extremely complex or open problems like the Riemann Hypothesis. This could be due to several reasons:
- **Ethical Considerations**: AI models might be programmed to avoid giving definitive answers on unsolved problems to prevent misinformation.
- **Computational Constraints**: The problem might require computational resources or time beyond what the model is configured to handle.
- **Safety and Alignment**: AI models are trained to avoid certain types of queries that could lead to nonsensical, harmful, or unproductive outputs.

2. **Model Laziness**: This term is used humorously here. AI models don't experience "laziness" in the human sense, but users often describe models as "lazy" when they:
- Fail to engage with particularly challenging or unconventional requests.
- Provide generic responses instead of attempting to solve complex problems.
- Avoid tasks that might lead to errors or incorrect information due to their programming for safety or accuracy.

3. **Riemann Hypothesis Context**: The Riemann Hypothesis is one of the most famous unsolved problems in mathematics, dealing with the distribution of prime numbers. Asking an AI model to "solve" it might be:
- A playful challenge to see how the model handles an unsolvable (by current standards) problem.
- An attempt to push the model's reasoning capabilities to their limits.

4. **Community Joke**: Within the AI and tech community, there's often humor in how these models handle or fail to handle certain tasks, especially those that are famously difficult or outside the scope of what AI can currently achieve.

By saying "Model laziness continues to be a major issue sad ;p," @karpathy is likely expressing a mix of frustration and amusement at the model's limitations or its programmed responses to avoid engaging deeply with such a complex problem, all while using a light-hearted tone to engage with the community about the realistic boundaries of current AI technology.

[Ethan Mollick @emollick]

Big paper I have been waiting for: what are the real impacts of AI on programmer productivity?

It is a randomized controlled trial using the older, less-powerful GPT-3.5 powered Github Copilot for 4,867 coders in Fortune 100 firms.

It finds a 26.08% increase in completed tasks.

[Farnam Street @farnamstreet]

You can't force yourself to think any faster.

Speeding up your thinking process often results in poor decisions that create future problems. You can't force yourself to make faster decisions just because you're faced with an unrealistic deadline.

Slow down.

[Farnam Street @farnamstreet]

“Rather than focus on what will change, focus on what stays the same.”

https://t.co/dm84JPZW44

[Maylyn Tan @maylyntan]

"The more time it takes to explain a course of action, the less likely it is to be a good one."

A Tiny Thought in the @farnamstreet Newsletter https://t.co/0vddSaydIc

[Maylyn Tan @maylyntan]

Huge potential here in designing learning with AI as a component for peer feedback. Students can also learn from the AI on how to give more constructive inputs to each other

[Anna Mills, annamillsoer.bsky.social, she/her @AnnaRMills]

https://t.co/OFPgudJFCP "was helpful to me in ways that the feedback I received in class was not. The AI feedback helped me structure the content of my essay, and was also helpful in identifying weaknesses in the construction of my argument."
--Student comment on https://t.co/1saxXJrcFQ (shared with permission)

[Deedy @deedydas]

@paulg @sfmcguire79 You can actually tell that across all colleges, Math majors are obviously better quantitatively but also better verbally than half of humanities majors.

[Sam Altman @sama]

important point here: these systems are much better at doing tasks than jobs.

and giving people better tools to do their work faster often leads to qualitative changes in what they can do.

(of course, over the long run, we expect these systems will be able to do all of some of today's jobs and aren't trying to hide the ball on that. confident we will find new and much better jobs when that happens!)

[Maylyn Tan @maylyntan]

People are delegating their cognitive abilities to machines. The lure of social recognition is too great

[Vitomir Kovanović @vkovanovic]

I'm reviewing a paper where authors stated their RQs and then asked ChatGPT the same questions and then reported the results. No evaluation, nothing. Imagine what people without scientific training are going to do, if this is what scientists are doing? #theendisnear

[Paul Graham @paulg]

There are a lot of people doing something that works without completely understanding why it works. I didn't realize this as a kid. I thought that to do something well you had to understand what you were doing. But in practice hardly anyone understands completely.

[Maylyn Tan @maylyntan]

I hope the ex-academic is empowered to lead the change and given resources to foster a data-driven culture

[Cassie Kozyrkov @decisionleader]

“What challenges are you tackling at the moment?” I asked. “Well,” the ex-academic said, “It looks like I’ve been hired as Chief Data Scientist… at a company that has no data.”