It’s becoming cliché to say the AI landscape continues to evolve at a breakneck pace. One development this week put an exclamation mark on the claim: the release of Claude 3, a chatbot that appears to rival, if not surpass, the capabilities of OpenAI’s GPT-4. As publishers grapple with the challenges of building and maintaining online communities, Reddit’s AI-powered harassment filter offers a glimmer of hope for healthier discourse. And IBM embracing Adobe Firefly highlights the growing importance of “safe” generative imagery in the corporate world. Jump into the news of the week with us with those stories and more:
Claude 3 Becomes a Go-To Chatbot Overnight
On a practical level, the past week may end up being more relevant to how AI affects the news media than most. And that’s because of the release of Claude 3, which appears to be as good as, if not better than, the standard-bearing GPT-4 — at least for most writing.
We’ve seen challengers to GPT-4 before, of course. Perplexity is better at research, Groq is faster, and there are a host of open-source models that appear to perform better for specific tasks or data sets. But none are quite as convenient and consistent as Claude, which gives access to its top-of-the-line model, Opus, in the public-facing chatbot that anyone can use. Sure, Google Gemini 1.5 also supposedly gives GPT-4 a run for its money in terms of performance, but good luck accessing its full abilities in Google’s labyrinthine naming conventions and price tiers.
Here at The Media Copilot, after initial testing, we were so excited by its performance that we rushed to include it as one of the top recommended tools in our AI Bootstrapping class earlier this week. Anthropic, the company that builds Claude, appears to have also addressed issues of over-emphasizing safety, and now the chatbot won’t necessarily shut down queries that hint at sensitive issues or matters of copyright. Anecdotally, the output also sounds a little less “AI” than ChatGPT.
We’re not alone in our impressions. Claude 3 has garnered near-universal praise from AI enthusiasts, including previous Claude skeptics. For use cases surrounding content creation that are suited to a general-purpose chatbot — such as writing first drafts, generating story ideas, analyzing or rewriting articles — Claude is now an excellent first stop, whether you’re using the free or paid version. We’ll see if Anthropic can turn this technical success into something commercial (Claude Store, anyone?).
Reddit Is Policing Harassment With AI
As the media industry continues to contract, conventional wisdom is pushing both big and small publishers to think more about the precise audience they’re serving, and how to cultivate it into a community. But communities need places to congregate online, and comment sections almost inevitably devolve into race-to-the-bottom behavior, which led many media companies to ditch them over the past decade.
Could AI help foster healthier communities? Reddit seems to think so. Now that it’s gotten a bit friendlier with AI overall via a partnership with Google, it’s implementing an AI-powered harassment filter. The filter is designed to help volunteer moderators spot harassing posts before they’re published. Moderators can enable the filter in their community’s settings and review the “potential harassment” tag for accuracy.
This is one cog in the machine of moderation, of course, but a potentially important one. If AI can sniff out not just bad language in posts (which are easy for unscrupulous users to game) but overall patterns of behavior that are toxic, that would go a long way toward media reconsidering its general “no way” stance on comment sections and perhaps reclaim a little more ownership of the communities they create.
Adobe Firefly Soars at IBM
IBM is reportedly using Firefly, Adobe’s generative image-creation tool, quite a bit internally. The company says it has 20,000 people using Adobe software, and it’s been ramping up integration into team workflows and even consulting teams. There’s a lot of Kool-Aid being drunk in this “news,” certainly, but it’s an indicator that litigation-wary corporations are responding to Adobe’s marketing that Firefly is the “safe” option for generative imagery since it’s trained only on licensed and public-domain photography. Still, you wonder if some Adobe designers long to be using Midjourney, still regarded as best-in-class by the AI community.
Australian Media Ponders Its Next Move vs. Meta
Some people in Australia’s media industry are suggesting Meta could still be designated as “news media” despite the company’s much-publicized expulsion of news from its platforms. That’s because Meta trained its AI model, Llama, partially on content produced by news websites, informing whatever output it produces.
This is just the latest (theoretical) move in the chess game between Meta and the Australian news media, which began when the government moved ahead with legislation that forces tech platforms to pay up for any content they serve up. Meta thinks it can live without the news, and — spoiler alert — it can. Copyright in AI training data is really a separate question, but there’s no doubt Meta can live without news content for training, too. It’s definitely starting to feel like there are very few pieces left on the media side of the board in this game.
More news about AI and media:
AI is like the VCR, apparently (The Verge): In its defense against The New York Times’ copyright lawsuit, Microsoft pointed to a previous technology that was almost sued out of existence due to copyright concerns.
RIP prompt engineering (IEEE Spectrum): Apparently, AI can do it better.
Deepfakes, Deepfakes everywhere (TechCrunch): Political deepfakes have spiked recently, which has a lot of people worried about the November election.
How to make photorealistic images of real subjects (X): Speaking of ?
The BBC spells out editorial guidelines for AI (BBC): Senior editors are where the buck stops for bad AI content on their watch.
Leave a Reply