Dan Becker

Cursor chat lets you include screenshots. I have given it screenshots and says "update the CSS to make this look better"

In theory we should think critically about our code. But if you are going to skip that step, ui code is easier than most code to judge by locking at the results

Books 🤝 AI

Instead of learning from a static text, I'd rather learn ideas/insights through interactive conversation. I'll pay for a long prompt that an LLM can talk through with me.

I really want this!

What are they trying to accomplish with fake job posts?

I watched a 10-year build a video game in 15 minutes on bolt.new just by describing what she wants

Making a custom video game was faster than picking up a birthday gift from Target.

2025 will be the year where we fundamentally rethink how custom apps fit in our lives

What do you know about people's motives when creating these bots

Beyond normalizing cruelty or causing reactive hardening, why do they want those things?

Is it foreign actors weakening a country's cohesion for political reasons?
Competitor hurting platform to steal market share?
Something else?

Well, as long as no one ever writes a popular movie script where AI starts a Global Thermonuclear War, we should be just fine

I would expect stocks to be hard because the current price reflects the views/analysis of so many professional full-time analysts and investors.

I've spent a couple weeks using LLMs for prediction-market predictions (e.g. on Metaculus & Polymarket)

I'm using conventional ML ideas (iterating based on validation scores). But it's >10X more broadly useful than conventional ML prediction.

Who else is trying LLMs for pred markets?

I don't really see a clear path where we keep an open internet that is not mostly full of AIs talking to each other. We can't reliably detect AI content, it is cheap and easy to generate, and there are lots of incentives to do so, even besides scams.

You can see the problem on all the social sites.

I'm glad to see LLM innovation evolve from train-time scaling to test-time scaling

Train-time scaling favors a winner-take-all outcome. Whoever spends the most upfront will have the best model.

Test-time scaling means more people can experiment and innovate... lower barriers to entry.

You can audit LLMs better than human processes

Asking a person a question can influence how they'd subsequently respond to variations on that question

So you can't experiment with any precision

With LLMs, set the temp to 0 and you can experiment very precisely

I should do more to lobby them to prioritize it, though they know that the AI features don't work with notebooks. Here's the error message if you hit apply

Nothing says government efficiency quite like hiring two executives into one job

I frequently do that in general (including with Python scripts & React apps). But the problems with notebook support are deeper:

- Apply button doesn't work
- Chat can't see cell output
- Code responses combine code from all cells, so you can't even copy/paste suggested code easily
- etc

I've assumed schoolteachers don't know how to integrate AI yet. But my 4th grade kid has a creative writing assignment where she iteratively steers AI to write a story for her using magicschool.ai

It looks very fun

Do you have any examples of what you are doing to skate to where the puck is going, given that explanation about what you think will change?

Most people want to skate where the puck is going. Being correct about where it's going is the hard part :)

Otherwise you skated to nowhere in particular

"Be in strategic spaces" is right too

Knowing what's a good strategy is the hard part. What do you think is a smart strategic move right now?

Is there something better than Cursor for getting AI help in Jupyter notebooks?

Or a trick to improve cursor for notebooks?

Where do you think the puck is going?

What's your prediction for how AI changes in the next 12 months (or whatever timeline you like)?