Claude and I set up some new functionality, which we call Office Hours, for the online seminar. It allows you to have interactions one-on-one with the professor, and to have a conversation that lasts up to five interactions.
The underlined phrases are clickable, so that you can get a short definition of the concept.
The user interface for the seminar is pretty straightforward. You go to the seminar page, type in a name (AKling in the example above), and when you get into the seminar choose any chapter from the dropdown menu (it defaults to the first chapter).
As you read through the dialogue, you will come to the end of a section. At the end of the section, there is a CallOnMe button, which allows you to have a single interaction with the professor and the other students. After a brief interlude of side conversations, your question or comment will get their reactions. These are not pre-canned responses. They are generated from Claude, based on my desired atmosphere, a list of key concepts that I have provided, and Claude’s general knowledge.
Keep reading the sections, and when you get to the last section in the chapter, you will see an “Office Hours” button. If you click on that, you can then type in a question or a comment. Submit it, and after a while a response will come back from the professor. Then you can submit a follow-up. You can have as many as five interactions per chapter.
I should emphasize that when you try CallOnMe or Office Hours, the answers you get back are not pre-canned. They are created on the spot by Claude, based on prompts that get sent to the AI by a programming interface.
Here is another screen shot where I joined the seminar using a different name (Arnold) and asked a question.
The professor goes on to say "You seem interested in how statistics can mislead. You might find John Ioannides' paper 'Why Most Published Studies are False' particularly enlightening. It's become quite influential in discussions about research methodology across all sciences. "I hope that helps explain some of the key challenges in scientific research."
My Achievements
First, from a software development point of view, this is now a very complex application. There are 30 separate modules of code, along with the scripts for the current seven chapters.
Working with Claude, I would say that my productivity gain was on the order of 100X. If it had been up to me to code the project without Claude, it would have taken more than a year, which means I never would have even attempted it. Instead, the site with the various chapters and the CallonMe function took about a week, and adding the Office Hours function took only half a day.1
Second, I am able to convey my insights during both the chapter scripts and the real-time interactions. This is an achievement because Claude and other large language models are biased toward conformity with mainstream, left-leaning, politically correct positions in social science. The Social Code Seminar will provide an educational experience that (a) comes from an independent point of view that is informed by libertarian and conservative thought and (b) allows for you to engage with an AI that also is informed by the sources that have influenced me.
I would describe using the seminar as like reading a textbook in the format of a play, with various characters conversing. In addition, you are able to obtain reactions to your own questions or comments.
This will only appeal to a thin sliver of the population. The content is over most people’s heads. A lot of you will not have the time to devote to it. But I think that there will be people who find it a rewarding experience, once I have it extended and revised. As usual, comments and suggestions welcome at the end of this post.
Shaping the Content
For the non-interactive dialogue in the chapters, I gave Claude a lot of context, but then let Claude come up with the scripts. Rather than edit the scripts, when I was not satisfied with a script I revised the prompt. I can always think of improvements for the scripts, but right now I place a higher priority on adding chapters and adding functionality.
This is not to say that Claude is polished as a software engineer. When there is a bug, Claude does not think logically about what the problem must be. Instead, it seems to be trying to fit the bug into a pattern of bugs it has seen in the past. This leads to a very circuitous process of debugging, and I find that I am better off intervening by telling Claude “It can’t be X, because Y” or “Here is code that worked before.”
I love this! In my view AI is simply a new computer interface that allows you to get your ideas into the world more easily. I've had many "I never would have built this without AI," moments recently. This is the power. Old jobs will go like pony express riders did with the telegraph but a thousand new ways of sharing ideas and making money from them will bloom.