Why does business news feel like it's written for people who already get it?

Morning Brew changes that.

It's a free newsletter that breaks down what's going on in business, finance, and tech — clearly, quickly, and with enough personality to keep things interesting. The result? You don't just skim headlines. You actually understand what's going on.

Try it yourself and join over 4 million professionals reading daily.

Check it out

We all know AI makes a lot of shit up, but recently OpenAI published a blog post noting that the LLM has been obsessed with gremlins and goblins for a while.

The model has become so obsessed that there are at least two instances of instructions in Codex that note that the model is not to refer to everything from the mythical troublemakers to raccoons, pigeons, trolls, or other “animals or creatures unless it is absolutely and unambiguously relevant to the user's query,” according to reporting at Business Insider.

“Where the Goblins and Gremlins Came From,” OpenAI’s title on the post, is a post-mortem deep dive on why the LLM has those instructions: The upshot being that the model has a “Nerdy” personality mode, which is obsessed with weird and fictional creatures and the trained behavior did not stay contained within the “Nerdy” personality and began to show up in other queries.

Beyond giving me a chuckle, this weird AI-tick gives me some pause. 

First, OpenAI only caught this because someone took a closer look at why the model was producing goblins. Engineers only traced it back to the feedback loop after a safety researcher outside the company flagged it. 

OpenAI frames it as harmless, whimsical, and fixable, but it's important to recognize that the mechanism that decided to replicate these figments of our very human imaginations is a pattern-recognition machine. It runs on reward feedback.  AI cannot tell the difference between a creature metaphor and, say, a political claim, a medical assertion, or a completely made-up conspiracy theory, and that’s where we can get into trouble. 

The bigger question that this seemingly harmless blog post raised for me is: What happens when the goblins and gremlins are a lot less obviously fictitious?

I think of it as similar to what has happened to America’s trust in journalism, thanks in large part to the advent and rise of cable news. I spent more than 10 years booking and producing guests for CNN television, and I worked with everyone from Anderson Cooper and Lou Dobbs to Gerri Willis and Poppy Harlow. In that era, it was a core pillar of the company to book guests who were actual experts with factual, real data to share with viewers. Guests we discovered were pushing falsehoods, and snake oil undermined the company's core mission as a news network, and they were not invited back. After all, our job as producers and reporters was to report the news, not to platform conspiracy theorists.

It’s been more than two decades since I sat in the newsroom in New York, and I can tell you that the industry has changed dramatically since I left. Fox News and its “news is entertainment” approach, with its unscrupulous, rage-baiting, and platforming of right-wing conspiracy theorists, did a lot of unquestionable damage to journalistic integrity and our democracy as a whole. That network’s approach to “news” helped usher in the era of distrust and fake news we currently live in today (along with the rise of social media and a multitude of other shifts in media, of course). 

It’s no wonder that everyday people don’t trust “journalists,” least of all television journalists, after all the damage that the network did. In reality, what really happened was that the business saw an opportunity to make a shit ton of money on outrage and fake facts, and it ran with it. Viewers rewarded Fox News for making up conspiracy theories about stolen elections and Lizard people at the Denver Airport, and kept watching.

Just like rewarding ChatGPT for talking about gremlins and goblins. 

This all serves as a useful reminder that these systems are reward-based, and those reward signals don’t always yield factual, accurate, or even true information. The model’s goblin-heavy responses have been used to train future versions, creating a feedback loop that has compounded the system drift over model generations, the same way that conspiracy theories and the platforming of white supremacists and pedophiles have created a drift in our democracy. 

While it's easy to laugh about fantastical creatures appearing in LLM answers, it serves as a good reminder that harmless is only harmless until it isn't. The goblins are the canary, not the coal mine.

Chess is a surprisingly good training ground for critical thinking.

If you play chess online (on Chess.com, for example), there’s a tool you should be using to get better at the long game. 

The app is called Chessalyze, and it is a new Android app that imports your games from Chess.com and Lichess, then gives you move-by-move Stockfish analysis with accuracy ratings, move classifications (brilliant, blunder, and everything in between), and an interactive evaluation graph, plus player insights including strengths and weaknesses, all running locally on your device. You can rerun old games to learn where you went wrong, and you own all the data.

Sign up via this Google Doc link to try it out and give feedback before he releases it into the wild. Once you sign up, you’ll get a link to download the app for free. Right now, it's only available to Android users, but he plans to open it to Apple users once he completes this round of testing. If you signed up last week, be sure to find the automated email response and download the app to start learning from your games.

The jobs list is huge today and includes more than 30 new opportunities for journalists to make as much as $295k per year and $100 per hour.

All of the roles are hybrid or remote, and I include all email addresses and pitching guides so you can chase down the right editors to pitch.

On the freelance front, there are opportunities to pitch and work for:

There are tons of brand new opportunities to go full-time, too, at publications like:

There are lots of Comms jobs on the list, and Dishonorables are solid, as is the Passion Project.

All jobs pay at least $1 per word and up to $295,000 per year, and have been posted in the last 7 days. If you’re tired of searching for brand new roles yourself this is the newsletter for you.