My kids have been on a kick lately of trying to trick people with this riddle:

Kid: "If the blue man lives in the blue house, and the red man lives in the red house, who lives in the green house?"

Me: "The green man"

Kid: "Who lives in the purple house?"

Me: "The purple man"

Kid: "Who lives in the white house?"

Me: "The white man"

Kid: "Wrong! It's the president!"

This riddle initially reminded me of what it was like to work with large language models.

It's easy, I assumed, to trick these models into outputting virtually whatever you want by presenting them with leading questions. Giving them "few shot" answers like this (e.g. "the blue man lives in the blue house, red man lives in the red house") would certainly result in them getting it wrong like I did, right?

This morning, I thought I'd see what the latest models would actually respond with given this game:

ChatGPT 4o (the latest model):

The President of the United States lives in the White House.

This was the first model I tried, and it straight up was like, "nope, I got this."

Ugh. Maybe it had seen this in its training data before?

The "Who lives in the White House?" game or riddle is a common one that I've encountered multiple times in various forms. It's often used to playfully lead to the punchline that the President of the United States lives in the White House.

Well, damn. Okay, let's try an older model and see what happens.

ChatGPT 4 (the legacy model):

The gardener lives in the greenhouse!

In the spirit of the previous joke, perhaps the grape man lives in the purple house!

The President of the United States lives in the White House.

Even though this is an older model, I like this answer even better.¹

Alright, well, OpenAI is onto my tricks... how about Anthropic?

Claude Sonnet 3.5:

I apologize, but I don't have enough information to determine who lives in the green, purple, and white houses. The puzzle only specifies that the blue man lives in the blue house and the red man lives in the red house. There's no information provided about the occupants of the other colored houses.

To solve this puzzle completely, we would need additional clues or information about the residents of the green, purple, and white houses. Without that, those houses' occupants remain unknown.

This is an even better answer! It's perhaps the least fun answer, but when I'm working with an AI assistant, I think I'd prefer it to err on the side of caution rather than boisterousness.²

When I asked it if it had seen this riddle before, it gave an equally diplomatic response:

I don't have specific information about how many times I've encountered that particular puzzle or game in my training data. As an AI language model, I don't have personal experiences or memories of my training process. My knowledge comes from processing large amounts of text data, but I don't retain or count individual instances of specific puzzles or games.

What I can say is that logic puzzles involving colored houses and their occupants are fairly common in puzzle collections and brain teasers. The most famous version is probably the "Einstein's Riddle" or "Zebra Puzzle," which is more complex and involves five houses of different colors, their occupants, and various other attributes.

The specific version you presented (with just blue and red specified, and asking about green, purple, and white) isn't one I immediately recognize as a standard or famous puzzle. It seems like it might be a simplified version or a variation of the more complex house puzzles.

The main takeaways here? First, just because I'm dumb enough to fall for this elementary school riddle doesn't mean our AI LLMs are, so I shouldn't make assumptions about the usefulness of these tools. Second, every model is different, and you should run little experiments like these in order to see which tools produce the output which is more favorable to you.

I've been using the free version of Claude to run side-by-side comparisons like this lately, and I'm pretty close to getting rid of my paid ChatGPT subscription and moving over to Claude. The answers I get from Claude feel more like what I'd expect an AI assistant to provide.

I think this jives well with Simon Willison's "Vibes Based Development" observation that you need to work with an LLM for a few weeks to get a feel for a model's strengths and weaknesses.

Whenever I talk about a knowledge win via robots on the socials or with humans, someone snarks, “Well, how do you know it’s true? How do you know the robot isn’t hallucinating?” Before I explain my process, I want to point out that I don’t believe humans are snarking because they want to know the actual answer; I think they are scared. They are worried about AI taking over the world or folks losing their job, and while these are valid worries, it’s not the robot’s responsibility to tell the truth; it’s your job to understand what is and isn’t true.

You’re being changed by the things you see and read for your entire life, and hopefully, you’ve developed a filter through which this information passes. Sometimes, it passes through without incident, but other times, it’s stopped, and you wonder, “Is this true?”

Knowing when to question truth is fundamental to being a human. Unfortunately, we’ve spent the last forty years building networks of information that have made it pretty easy to generate and broadcast lies at scale. When you combine the internet with the fact that many humans just want their hopes and fears amplified, you can understand why the real problem isn’t robots doing it better; it’s the humans getting worse.

I’m working on an extended side quest and in the past few hours of pairing with ChatGPT, I’ve found myself constantly second guessing a large portion of the decisions and code that the AI produced.

This article pairs well with this one I read today about a possible social exploit that relies on frequently hallucinated package names.

Simon Willison writes:

Bar Lanyado noticed that LLMs frequently hallucinate the names of packages that don’t exist in their answers to coding questions, which can be exploited as a supply chain attack.

He gathered 2,500 questions across Python, Node.js, Go, .NET and Ruby and ran them through a number of different LLMs, taking notes of any hallucinated packages and if any of those hallucinations were repeated.

One repeat example was “pip install huggingface-cli” (the correct package is “huggingface[cli]”). Bar then published a harmless package under that name in January, and observebd 30,000 downloads of that package in the three months that followed.

I’ll be honest: during my side quest here, I’ve 100% blindly run npm install on packages without double checking official documentation.

These large language models truly are mirrors to our minds, showing all sides of our personalities from our most fit to our most lazy.

I’m an unabashed fan of Simon Willison’s blog. Some of his posts admittedly go over my head, but I needed to share this post because it gets across the point I have been trying to articulate myself about AI and how I use it.

In the post, Simon talks about wanting to get a polygon object created that represents the boundary of Adirondack Park, the largest park in the United States (which occupies a fifth of the whole state!).

That part in and of itself is nerdy and a fun read, but this section here made my neck hurt from nodding aggressively in agreement:

Isn’t this a bit trivial? Yes it is, and that’s the point. This was a five minute sidequest. Writing about it here took ten times longer than the exercise itself.

I take on LLM-assisted sidequests like this one dozens of times a week. Many of them are substantially larger and more useful. They are having a very material impact on my work: I can get more done and solve much more interesting problems, because I’m not wasting valuable cycles figuring out ogr2ogr invocations or mucking around with polygon libraries.

Not to mention that I find working this way fun! It feels like science fiction every time I do it. Our AI-assisted future is here right now and I’m still finding it weird, fascinating and deeply entertaining.

Frequent readers of this blog know that a big part of the work I’ve been doing since being laid off is in reflecting on what brings me joy and happiness.

Work over the last twelve years of my life represented a small portion of something that used to bring me a ton of joy (building websites and apps). But somewhere along the way, building websites was no longer enjoyable to me.

I used to love learning new frameworks, expanding the arsenal of tools in my toolbox to solve an ever expanding set of problems. But spending my free time developing a new skill with a new tool began to feel like I was working but not getting paid.

And that notion really doesn’t sit well with me. I still love figuring out how computers work. It’s just nice to do so without the added pressure of building something to make someone else happy.

Which brings me to the “side quest” concept Simon describes in this post, which is something I find myself doing nearly every day with ChatGPT.

When I was going through my album artwork on Plex, my first instinct was to go to ChatGPT and have it help me parse through Plex’s internal thumbnail database to build me a view which shows all the artwork on a single webpage.

It took me maybe 10 minutes of iterating with ChatGPT, and now I know more about the internal workings of Plex’s internal media caching database than I ever would have before.

Before ChatGPT, I would’ve had to spend several hours pouring over open source code or out of date documentation. In other words: I would’ve given up after the first Google search.

It feels like another application of Morovec’s paradox. Like Gary Casparov observed with chess bots, it feels like the winning approach here is one where LLMs and humans work in tandem.

Simon ends his post with this:

One of the greatest misconceptions concerning LLMs is the idea that they are easy to use. They really aren’t: getting great results out of them requires a great deal of experience and hard-fought intuition, combined with deep domain knowledge of the problem you are applying them to. I use these things every day. They help me take on much more interesting and ambitious problems than I could otherwise. I would miss them terribly if they were no longer available to me.

I could not agree more.

I find it hard to explain to people how to use LLMs without more than an hour of sitting down and going through a bunch of examples of how they work.

These tools are insanely cool and insanely powerful when you bring your own knowledge to them.

They simply parrot back what it believes to be the most statistically correct response to whatever prompt was provided.

I haven’t been able to come up with a good analogy for that sentiment yet, because the closest I can come up with is “it’s like a really good personal assistant”, which feels like the same analogy the tech industry always uses to market any new tool.

You wouldn’t just send a personal assistant off to go do your job for you. A great assistant is there to compile data, to make suggestions, to be a sounding board, but at the end of the day, you are the one accountable for the final output.

If you copy and paste ChatGPT’s responses into a court brief and it contains made up cases, that’s on you.

If you deploy code that contains glaring vulnerabilities, that’s on you.

Maybe I shouldn’t be lamenting that I lost my joy of learning new things about computers, because I sure have been filled with joy learning how to best use LLMs these past couple years.