OpenAI’s latest small model, GPT-4o mini, made a splash last week. It was one of several recent unveilings of small language models (SLMs), which, taken together, are a positive sign of progress — cheaper, more performant models — something that will become necessary as cost-sensitive newsrooms deploy AI tools.
Before I get to that, I have a personal story for you:
I spent this super uneventful weekend in Halifax, Nova Scotia, reuniting with friends and faculty from the journalism school I attended, the University of King’s College. In the 25 years since my class graduated, journalism has been decimated by the rise of tech platforms that took ownership of the advertising market, something my old journalism ethics professor, Bruce Wark, observed in his remarks at the reunion.
Now the industry stands on a new precipice with AI. In contrast to how we felt about the internet at the turn of the century, the media is taking a much more skeptical attitude toward what AI might represent to journalism. While that’s healthy, it’s important to remember that skepticism isn’t an end state — it’s a tool journalists use to get closer to the truth.
We’re long past the point where journalists could simply dismiss AI or stubbornly refuse to acknowledge its obvious influence and utility. It’s right to be suspicious of tech companies and platforms claiming to change the world for the better, but change is still inevitable, and finding a viable path forward for the media in an AI-saturated world is exactly what I try to do in this newsletter. Returning to my journalistic roots was an excellent reminder of that (and you thought this was just a self-indulgent digression!).
And now a little bit of housekeeping: The Media Copilot is moving to a twice-weekly schedule for the rest of the summer. On Tuesdays, you’ll get my weekly column that dives into a specific aspect of how AI is transforming our information ecosystem, and on Thursdays I’ll have something a little shorter, but with bonus thoughts on recent news stories from that area of the Venn diagram where media and AI overlap. The Media Copilot podcast — where I interview journalists, builders, and executives architecting the future of media — continues. New episodes land every other Friday.
Finally, John Biggs is moving on from The Media Copilot to focus on his wildly successful podcast, Keep Going. We’ll miss him, but we admire and respect him following his passion and wish him well.
Now, on with the show…
Keep Your SSN Off The Dark Web
Every day, data brokers profit from your sensitive info — phone number, DOB, SSN — selling it to the highest bidder. And who’s buying it? Best case: companies target you with ads. Worst case: scammers and identity thieves.
It’s time you check out Incogni. It scrubs your personal data from the web, confronting the world’s data brokers on your behalf. And unlike other services, Incogni helps remove your sensitive information from all broker types, including those tricky People Search Sites.
Help protect yourself from identity theft, spam calls, and health insurers raising your rates. Plus, just for The Media Copilot readers: Get 55% off Incogni using code COPILOT.
Smaller Models, Better Tooling
“Small is the new big,” goes the old joke. It’s usually mocking fashion or food or apartment square footage. But it’s an apt summary of the progress in smaller AI models over the past few months, including recent announcements from Hugging Face and Mistral (in partnership with Nvidia). The latest — and arguably most impactful — is OpenAI’s unveiling of GPT-4o mini.
The two ingredients these smaller models bring to the table are value and speed — they’re considerably faster than “regular-size” models while also being more affordable (in terms of compute cost). However, those advantages generally come at the expense of quality: the answers just aren’t quite as good.
For many real-world applications of AI, “not quite as good” is still good enough. In media specifically, you may be able to get away with using a smaller model for, say, a custom tool that writes social media copy to promote an article. Since that kind of captioning tends to be short and checked by humans anyway, you probably don’t need to apply the more expensive power of a full model to the task. Plus it’s part of a workflow you’re doing several times a day, so using a cheaper model will let you save on compute costs.
The danger in a cash-strapped newsroom is that, in a business where it’s important to keep costs in control, small models will likely become the default for most applications. That may lead to less productivity gains, and may even bias the staff against AI. After all, if most of your experience is that AI output is mediocre, why would you use it for more challenging projects?
Applying AI systemically, however, means organizations need to make choices about what’s “good enough” for editorial applications. That’s why it’s essential product teams consult with informed editorial leaders: senior managers who don’t just understand the work their teams do but also the nuances of AI: its inherent strengths and weaknesses, the differences between models, and where humans need to be in the loop.
This kind of leadership is essential for AI to graduate in newsrooms from the simple, low-hanging fruit of AI-written headlines and social copy to more capable assistive tools. For example, a reporter following a court case would greatly benefit from an AI with access to all the court documents and reporting relevant to that specific case — something relatively straightforward to set up with a custom tool leveraging retrieval-augmented generation (RAG).
With a research use case like this, inference and reasoning are key. For many (if not most) queries, you’ll want the power of a larger model that is better equipped to find “nuggets” — those parts of court documents that are written succinctly and innocuously in legalese but actually contain vital, new, or unusual information. Legal reporters often spend hours finding these, and for an AI tool to come close to replicating their probing abilities, you need performance.
That’s what GPT-4o mini brings — without breaking the bank. OpenAI’s latest small-fry model is said to be a big step up over its predecessor, GPT-3.5, which OpenAI was offering as “the small model” up until now. OpenAI has a bunch of fancy stats that make it look good, and you don’t have to look very far on X to find developers praising it, but its best feature is by far the cost, coming in significantly cheaper than even other small models like Claude Haiku. Here’s how it ranks in MindStudio (a no-code AI platform and Media Copilot partner):
To be clear, there are still some open-source models cheaper to run than GPT-4o, but not by much, and almost certainly with much less robust performance.
As newsrooms and content operations go beyond simply handing out subscriptions to ChatGPT or Gemini, the cost to deploy AI weighs heavier, and product design becomes a negotiation. With performant small models GPT-4o mini becoming the norm, those negotiations just got easier.
Leave a Reply