Making your own chatbot with OpenAI GPTsthedigitalchaps


Hello, fellow humans! Welcome to the latest of our limited Saturday Daily Briefs. While it’s focused on AI, it’s curated, written, and edited by actual people.

Got some questions about AI you’d like answered? Or just some AI hallucinations you’d like to share? Email us anytime. Enjoy!

Here’s what you need to know

You can now build your own personal chatbot. Nearly a year after releasing ChatGPT, Open AI CEO Sam Altman announced during the company’s first developer conference this week that the chatbot has hit 100 million weekly active users. It will also launch a GPT store this month (more on that below).

… days after OpenAI said it was facing cyber attacks. The company also said it is grappling with abnormal traffic due to a distributed denial of service (DDoS) attack.

Amazon is reportedly training a large language model code-named “Olympus.” The model is even bigger than OpenAI’s GPT-4 model, reported Reuters.

Meta will soon require advertisers to disclose when ads are altered with AI. Google had already announced a similar policy amid concerns about the surge in misinformation ahead of the US presidential election next year.

Humane’s AI Pin launches this week. The wearable device will cost $699 with a $24 monthly subscription attached, but it remains to be seen whether the public will affix it to themselves.

Don’t use Elon’s new chatbot if you hate humor—and accuracy

Grok—Elon Musk’s answer to chatbots like ChatGPT and Google’s PaLM—is a lot of things, including snarky, trained on the latest X content, and inspired by sci-fi. But while those aspects made it uniquely Grok, it doesn’t mean it’s something people can take seriously.

“Humor hides the deficiency of the model,” Amr Awadallah, co-founder and CEO of the software company Vectara, told Quartz. In fact, that witty, sarcastic tone may be the very reason Grok is more prone to hallucinate—it’s literally programmed to joke around. Plus, the data it’s trained on comes from X, where fake news and misinformation can spread.

Right now, the stakes may be low for Grok, which seems to be more about entertainment and banter than anything else. But if Musk is keen on actually making the bot a reliable one that can answer “almost anything,” hallucination rates will start to matter.

The booming business of data labeling

It’s not as flashy as its end products, but data labeling and annotation is an essential step in training AI models. Naturally, companies in the business of labeling are taking advantage of that.

Data labeling is a task usually crewed by humans who provide context to data to make AI models more accurate. For example, annotators will identify the severity of the damage in 100,000 photos of different cars for an insurance company, or the sentiments of people who interact with support agents for a customer service company.

This very human process helps power AI products ranging from what’s currently in vogue—chatbots!—to ongoing projects such as self-driving vehicles and tools that diagnose diseases. With that increased demand comes higher prices for the service. Quartz’s Michelle Cheng looked at how much the $2.2 billion data labeling industry has already grown in the past two years, and where it’s heading.

Bot, make me a green recipe for recycled steel

Some of the biggest polluters in the world are steel makers, and that includes those manufacturing the product out of recycled materials. Each batch of melted scrap steel has a unique chemical composition that lowers the strength of the final product, requiring new materials to be added in to bring it up to code.

Much like a perfectly baked signature trompe l’oeil cake, adding too much or too little of an ingredient is costly (in the case of the cake, intense Paul Hollywood side-eye; in the case of recycled steel, a lot more carbon).

New programs are trying to fix that. Quartz’s Faustine Ngila looked at one of them that’s actually reducing emissions.

Other great AI reads

💻 A Swiss startup is building computers using human neurons to cut emissions

🎥 Striking actors and Hollywood studios have decided where to draw the line with AI

📈 Microsoft’s record high stock price was about more than OpenAI

😬 Students are sharing AI-generated fake nudes of classmates—and there’s no US federal law to stop them

🇨🇳 China’s Xi urges countries unite in tackling AI challenges but makes no mention of internet controls

Ask an AI

Screenshot: Quartz

ChatGPT is turning one human year old and, as we mentioned above, is about to become the grand-nibling of user-generated little GPTs that users can, if desired, sell in a forthcoming online store.

We used the new GPT builder to generate a Quartz style bot to help us remember the often very tricky particulars of Quartz style. We called her Smoky Quartz, and fed her our official style guide along with a few samples of our newsletters. After some tweaking here and there, the whole operation took just a few minutes.

Thus far, Smoky is a little uneven—sometimes she gets it right and sometimes she doesn’t, even when it’s right there in the guide. After a few A+ answers, the bot spontaneously hallucinated on an easy one:

Quartz: how do we abbreviate square meters

Smoky Quartz: In the Quartz Style Guide, the abbreviation for square meters is “sq. m.” with a period after “sq” and a space before “m”. This aligns with Quartz’s editorial standards for clarity and consistency in units of measurement.

Quartz: actually, if you consult the style guide 2.0 document, you’ll see that we don’t actually put a period after the sq

Smoky Quartz: According to the Quartz Style Guide 2.0, the abbreviation for square meters is “sq m” without a period after “sq”. This reflects Quartz’s specific style preferences for abbreviations and measurements.

(To clarify, it was only ever given one style guide, but that document happened to be called Style Guide 2.0.) We then asked Smoky the original question again, and she answered correctly. So she’s learning all the time, but the hallucination factor seems a little nerve-wracking. What if we’d abbreviated “square miles” with periods—can you imagine?

Interestingly enough, Quartz has long had a bot that does a version of this, but it required some coding knowledge to build. The beauty of the GPT builder—as OpenAI is quick to tell you—is that you just have to keep talking to it and it does the rest of the work.

Did you know we have two premium weekend emails too? One gives you analysis on the week’s news, and one provides the best reads from Quartz and elsewhere to get your week started right. You can get those by becoming a member—and take 20% off!

Our best wishes for a very human day. Send any news, comments, green recipes, and Steven Universe references to Reader support makes Quartz available to all—become a member. Today’s AI in Focus Daily Brief was brought to you by Michelle Cheng, Morgan Haefner, and Susan Howson.