Why book reports aren’t a good homework assignment any more

Big Nate comic

So, everyone knows that the kids these days are using ChatGPT to write their book reports. Even Big Nate knows it!

But what about ChatGPT’s safeguards? Isn’t it supposed to have some kind of anti-cheating baked in, so it won’t just write essays for kids? Why doesn’t that work?

Sure, it does have safeguards… kind of. If you just ask it to write an essay, it responds with a “helpful” answer about how to write an essay. The thing is that these safeguards are incredibly easy to work around.

Let’s pretend we’re a student who has to write a book report on The Kingdom over the Sea, by Zohra Nabi. Here’s how to write it in 20 minutes without even touching the book.
Read more

I started an AI newsletter just in time for OpenAI to implode

AI week

Last week, I decided to soft-launch a weekly newsletter to highlight three or four recent news stories in AI. When I started it, I had no idea this would be one of the biggest weeks in AI news all year. There was one giant story: OpenAI, makers of ChatGPT, fired their CEO, Sam Altman. And then it got messy.

You can read the rest of this week’s newsletter here.

Why a newsletter: I spend several hours a week reading about AI/machine learning. I get frustrated when I see overly-simple takes on AI in the media, on social media, etc. I figured that the best thing I could do about that would be to share what I’m learning every week.

What the newsletter is: A weekly email about some of the past week’s most interesting stories in AI, plus the occasional backgrounder.

What it’s not: Business-focused, highly technical, or comprehensive.

What you’ll get from AI Week:

  • An interesting email about AI in your inbox every week.
  • Background knowledge to contextualize all the breathless media stories about the powers, or dangers, of AI.
  • A feeling for where the field is going, why it’s going there, and who’s powering it.

Where do I sign up? Right here:



Powered by Buttondown.

Blepbot update: Hashtags

I did follow up on on of the possible improvements to the BlepBot in my last post. Mastodon has some fun day-specific tags for animal pics. I decided to use #TongueOutTuesday and #FurballFriday, since almost all the bleps are furry creatures.

# if it's Tuesday, add TongueOutTuesday hashtag
todays_date=datetime.datetime.now()
if todays_date.strftime("%w")=='2':
    post_text+=" #TongueOutTuesday"
elif todays_date.strftime("%w")=='5':
    post_text+=" #FurballFriday"

I also wanted to use a couple of cat-specific ones, but only on pictures of cats! To do that, I use re.search to regex search the post text for “cat” (case-insensitively).

elif todays_date.strftime("%w")=='3':
    if re.search("cat",post_text,re.IGNORECASE):
        post_text+=" #WhiskersWednesday"
elif todays_date.strftime("%w")=='6':
    if re.search("cat",post_text,re.IGNORECASE):
        post_text+=" #Caturday"

Of course, we need to import datetime and re to use these:

from mastodon import Mastodon
import os,random,logging,datetime, re

The blepbot code is in github as a public template: https://github.com/nroshak/blepbot. Feel free to clone it and make it your own!

Creating a Mastodon bot in Python

Blep!

Bleps always cheer me up, so I decided to create a Mastodon bot that posted one blep a day. (A blep is a picture of an animal with a tiny bit of its tongue sticking out, and Mastodon is a decentralized alternative to Twitter.)

Creating a Mastodon bot in Python was surprisingly easy, thanks to the Mastodon.py library. I’ve documented all the steps here, along with links to tutorials for each library or builtin feature that I used.

Have you created any Mastodon or Twitter bots? Let me know in the comments!
Read more

AI Image Denoising

I tried some free AI Image Denoising tools on a noisy, low-light photo. I had big hopes for a very sweet picture that’s just too noisy to put in a frame. Unfortunately, free tools didn’t get me anywhere.

ImgLarger
http://imglarger.com
Result: I can’t tell the difference between the original and the supposedly “denoised” image. 0/10 might as well not have bothered.

DeepAI Super Resolution model
https://deepai.org/machine-learning-model/torch-srgan
using Super Resolution torch-srgan
Result: The image was noticeably denoised, but the result wasn’t very impressive.
Following enhance, the quality was still poor.

BigJPG
https://bigjpg.com
The free denoising is limited to 3000×3000 px, so I had to crop the picture.
Result: Another poor quality, unimpressive result.

https://deepai.org/machine-learning-model/image-editor
After a series of unsatisfying results, I told it “do nothing”. and it closed the eyes and added some wrinkles. ooookay.
The prompt “watercolour painting, pastel colours, photorealistic, detailed image” was better but still mangled the face a bit horribly, and this algorithm seems really determined to make smiling faces squint.

Using GPT-2 as a continuous writing prompt generator via AI Dungeon

AI Dungeon Welcome Screen

OpenAI has released an API for accessing GPT-3, an AI text generator that’s capable of generating much longer responses than GPT-2. The API is currently in closed beta, and while I’ve requested access, I haven’t gotten it (yet?) But, I learned from Mario Dian’s blog about AI Dungeon, a text-based dungeon game powered by GPT-2 and/or GPT-3. Imagine a text-based adventure game, like Zork–that’s the feel of AI Dungeon, but instead of being programmed behind the scenes by something like ChoiceScript, it’s using GPT-2 (free version) or GPT-3 (paid version) to generate text.

I’m interested in exploring GPT-2 and GPT-3 for fiction generation, so I gave the free version of AI Dungeon a try as I was working on a short piece of flash fiction. I didn’t want AI Dungeon to write a story for me, or even to write parts of a story. I wanted to use it as a tool to spur my own creativity. Read more

Scraping web data

I wanted to get some data off an agent listing website and into a spreadsheet. I’d been meaning to play around with python for web scraping and this was the perfect excuse: There were just enough results that it would take longer to manually copy and paste them than to write a little python program. (I never want to automate something that will take less time to do than to automate, as long as I’m only going to do it once or twice…)
Read more