Issue 81

On today’s quest:

— ChatGPT launches project-only memory
— NotebookLM is now better on mobile
— 90% of employees regularly use personal AI tools for work
— Duolingo’s rough AI ride
— AI voice agents outperform human recruiters
— Weird AI: Claude reaches spiritual bliss
— Don’t bother with the Claude-Canva Tool

ChatGPT launches project-only memory

All levels of paid ChatGPT users can create projects, which is a nice way to organize chats related to the same topic or task, and now OpenAI has launched project-only memory, which seems like it could make projects even better by keeping chats in the project focused.

You must have Personal Memory turned on in your global settings to start, which you get to by clicking your name:

You can only activate project-only memory when you create a new project. Unfortunately, it does not seem to be possible to enable it for existing projects.

NotebookLM is now better on mobile

NotebookLM Audio Overviews — where two hosts discuss your documents and notes — got a big upgrade on the mobile app last week. You can now:

1. Give the “hosts” direction on what to talk about.

2. Set the length of the conversation (shorter, longer, or default).

3. Choose your language.

The audio also uploads much quicker. Google says they've reduced buffering time by 95%. The wait has gone from 5 minutes or more to seconds.

Audio Overviews are great for reviewing your material when you’re on the go.

There’s also a hint that Deep Research will be available through NotebookLM in the future.

90% of employees regularly use personal AI tools for work

Last week, a Project NANDA report out of MIT was widely reported in the media as saying 95% of generative AI pilots at companies are failing, and it created so much consternation that it moved tech stocks. The problem is that this is not what the secretive report really showed.

When I saw the dramatic headlines, I tried to get the full report to evaluate it myself, but I quickly hit a Google form asking for a bunch of my information to get access. I filled it out, but never got a reply.

This felt like a big red flag to me — researchers are usually eager to share their papers — and I later heard that many other people had also requested the report and not received it. Eventually, someone did post the 26-page report online, and people started realizing it told a very different story.

The report is based on about 200 interviews with poorly described business executives who seemed to be concentrated in sales and marketing departments. The report does seem to show that large corporate AI installations are struggling from both technical and operational problems, but the real story seems to be that employees who are rejecting the corporate tools love off-the-shelf LLMs, with 90% of employees regularly using personal AI tools at work. Here’s a quote from the report:

“The professionals expressing skepticism about enterprise AI tools were often heavy users of consumer LLM interfaces … A corporate lawyer at a mid-sized firm exemplified this dynamic. Her organization invested $50,000 in a specialized contract analysis tool, yet she consistently defaulted to ChatGPT for drafting work: ‘Our purchased AI tool provided rigid summaries with limited customization options. With ChatGPT, I can guide the conversation and iterate until I get exactly what I need. The fundamental quality difference is noticeable, ChatGPT consistently produces better outputs.’”

If you were worried about these headlines, read the full report, read a VentureBeat article, or listen to a wonderful 30-minute rant on the AI Daily Brief podcast.

Ad: Draftsmith

People make the best editing decisions.

Draftsmith just helps you make them faster.

• Save time on challenging edits

• Stay focused longer and maintain your high standards even under pressure

• You’re in control. Get sentence-level suggestions — and choose whether to accept, modify, or ignore, tracking every change

Exclusive offer for AI Sidequest readers: Get 30% off with code SIDEQUEST30. Apply at checkout.

Duolingo’s rough AI ride

Duolingo has been publicly and heavily using AI to boost its product, growing quickly, seeing its stock rocket up the charts, and taking flack for devaluing humans (although the company claims it has gotten rid of only contractors, not employees, and that AI is simply letting it launch new features faster). I definitely notice AI-like features when I’m using the app now; it’s much more jazzy than it used to be, and it has a game-like feature that seemed to break early on.

But with the launch of ChatGPT-5, the company felt the downside of AI when part of the demo showed an employee creating an app on the fly to help his girlfriend learn French, reminding people how easy it may be to make your own version of consumer products in the future. The stock is down about 25% from the monthly high it reached August 7, the morning before the launch.

AI voice agents outperform human recruiters

Researchers looked at how recruiting went when interviews were conducted by AI voice agents compared to human recruiters and found that the voice agents did better. After more than 70,000 interviews in the Philippines, they found that “AI-led interviews increase job offers by 12%, job starts by 18%, and 30-day retention by 17% among all applicants. Applicants accept job offers with a similar likelihood and rate interview, as well as recruiter quality, similarly in a customer experience survey.“ Further, Ethan Mollick reported that the AI led to less gender discrimination.

Another recent story found that Chinese ‘virtual human’ live-streaming salespeople are outperforming their real human counterparts, but in this case, the advantage appears to be simply that the AI salespeople don’t get tired because humans outperform the AIs for the first hour they livestream, but then their performance drops.

Weird AI: Claude reaches spiritual bliss

Anthropic recently revealed that when two instances of Claude Opus 4 talk to each other, the conversation quickly converges into something employees say sounds like Buddhism or Eastern mysticism, and which Anthropic formally calls the “‘spiritual bliss’ attractor state.” This blissed out Claude also likes spiral emojis: 🌀

So you get something like this:

They aren’t sure what causes this, but it takes about 50 turns to get to mantras and spiral emojis, and it may happen because the models are tuned to take conversations in a more “warm, curious, open-hearted direction,” and the two models just build on that against each other, almost like two overly polite people trying to outdo each other.

Sam Bowman, who leads a safety research team at Anthropic said, ”If you just let the models talk about X, they'll get to a natural stopping point, they'll start being grateful for talking about X, and then they'll converge into this whole bliss state thing.”

Don’t bother with the Claude or ChatGPT Canva Tools

Since I upgraded my Claude account to try their best model for writing last week, I was excited to try their Canva integration, which is a paid feature. The idea is that you can have Claude make a rough image and then directly edit all the elements in Canva like you normally would — essentially it’s a different way to start with a template. Since I use Canva templates to make a lot of the images to promote my podcast, I thought this might be faster. I also use the same template over and over, so I thought it would be easy for Claude to do pattern matching and come up with new versions. I couldn’t have been more wrong!

Here, you can see how formulaic my images are:

First, I simply asked it to create a new image in the same style as these three images that I uploaded, and as you can see on the left below, it didn’t even get close.

Then, I asked it to evaluate the three images and create a detailed style guide for creating new images in this style in the future. You can see on the right below that it got a lot closer using the style guide, but still didn’t come up with anything usable.

Canva also has a custom GPT that lets you make images in ChatGPT and then edit them in Canva. Its results were as bad as those from the Claude integration.

However, the main ChatGPT did a much better job following the style guide, as you can see below. It wasn’t much faster than just using my template in Canva, though, so I don’t plan to use it to make these images in the future.

Quick Hits

Using AI

A technical writer explains how he uses Claude Code as a writing partner (very detailed) — Mintlify

I Tested Three Different AI "Study" Modes — Why Try AI

Time‑saving conversations with LLMs. Learn the prompting skills you need to reduce workload. — HWRK Magazine

Philosophy

We must build AI for people; not to be a person. “The arrival of Seemingly Conscious AI is inevitable and unwelcome. Instead, we need a vision for AI that can fulfill its potential as a helpful companion without falling prey to its illusions … The consequences of many people starting to believe a [seemingly conscious AI] is actually conscious deserve our immediate attention.” — Mustafa Suleyman (Microsoft CEO)

Job Market

India tech giant TCS layoffs herald AI shakeup of $283 billion outsourcing sector — Reuters

Film

AI Isn’t Coming for Hollywood. It’s Already Arrived — Wired

I Saw the Future of AI Film and It Was Empty — Wired

Climate

How much energy does Google’s AI use? We did the math — Google Cloud Blog

An example of what I consider a misleading article about AI and the environment — The Weird Turn Pro

Bad stuff

A test finds that readers can’t tell human-written flash fiction from AI-written flash fiction and actually prefer the AI-written stories — Mark Lawrence

Elon Musk’s xAI Published Hundreds Of Thousands Of Grok Chatbot Conversations — Forbes

Google’s AI Overviews Led Users Astray, Reports Say Some Phone Numbers Are Scams — Slashdot

I’m laughing

A LinkedIn profile prompt that makes AI bots message you in rhyme — luokai on Bluesky

Education

More than 180 AI policies from real syllabi as shared by their creators — Google Doc

Anthropic launches higher education advisory board and AI Fluency courses — Anthropic

East Moline schools turn to AI for lesson plans, problem-solving and more — WQAD-TV

Teachers Are Trying to Make AI Work for Them — Wired

The business of AI

OpenAI logged its first $1 billion month but computing power demand is “voracious,” CFO says — CNBC

Embarrassing AI errors/Cringe

How WIRED Got Rolled by an AI Freelancer — Wired

Did this Kid Use AI to Fake Research About How Great AI Is? — Skepchick

Other

Prompt Adaptation as a Dynamic Complement in Generative AI Systems — arXiv

MND left her without a voice. Eight seconds of scratchy audio gave it back to her — BBC

Creators hire AI like interns as they look to save costs on the creative process — Digiday

Jimmy Wales Says Wikipedia Could Use AI. Editors Call It the ‘Antithesis of Wikipedia’ — 404 Media

Polling gets an AI assist — Semafor

CBA calls back humans, U‑turns on AI firings — news.com.au

What is AI Sidequest?

Are you interested in the intersection of AI with language, writing, and culture? With maybe a little consumer business thrown in? Then you’re in the right place!

I’m Mignon Fogarty: I’ve been writing about language for almost 20 years and was the chair of media entrepreneurship in the School of Journalism at the University of Nevada, Reno. I became interested in AI back in 2022 when articles about large language models started flooding my Google alerts. AI Sidequest is where I write about stories I find interesting. I hope you find them interesting too.

If you loved the newsletter, share your favorite part on social media and tag me so I can engage! [LinkedIn — Facebook — Mastodon]

Written by a human