About that viral content

This article about clipping and “narrative campaigns” is a wild read. Clipping is a thing in part because dozens of specialized agencies are creating thousands of sock puppet accounts that use clips to generate millions of posts across Instagram, X, TikTok, and YouTube. The agencies are used by record labels, celebrities, politicians, film studios, anyone who has $25,000 to spare. The article is packed with revelations about the machinery and the consequences.

The kerfuffle over Bad Bunny’s Super Bowl performance? Fake, according to Gudea, a firm that analyzes viral behavior. 4% of the accounts venting about it generated > 25% of the content. And it looks like one party was responsible for seeding both pro- and anti-Bunny posts. : )

The American Eagle Sydney Sweeney “good jeans” shitstorm? Also fake. The bot-detection firm Cyabra found that “15% of TikTok accounts commenting on it were fake and created a disproportionately large percentage of the uproar.”

Fun fact: Not labeling marketing content as such is illegal. The reporter contacted the FTC for comment, and the FTC said they don’t care. So much for the rule of law.

Theoretically, this would be a reputational hit for the social media platforms. Nothing seems to really matter any more, so it’s hard to say if this will do any damage.

The piece is definitely worth reading in its entirety. It’s paywalled, but worth a $5 sub to NYM.

Remember last week when LLMs seemed poised to gobble up Google’s search business? It turns out Google has some of the cards, namely very personalized search results. Personalized search used to be limited to Gemini, but it will filter into all of its search results. Improved AI abilities mean your results will now be tailored for you based on everything you do in Google’s ecosystem: your search history, calendar, map searches, and email. Also, eventually AI mode will be the default in Google search, not a toggle. You and your next door neighbor could enter the same query and get totally different results. This kind of kills the idea of key words. User-context will drive results now, not key words.

It also gives Google a massive leg up over LLM-based searches. OpenAI and Anthropic may mimic the personalization, but Google has a big advantage because everyone uses Google’s products.

The post speculates, however, that at some point users might be able to export their Google activity into a different LLM.

Personally, I’m not sure hyper-personalized results are always a win for consumers. You can imagine some of these search results generating legitimate concerns about privacy.

AI slop can make using the web a crappier experience and degrade online communities. It can also rain on IRL parades too. The piece also sees a future where, instead of getting better at identifying and avoiding AI slop, AI degrades our ability to write because of its ubiquity.

I think some of the scenarios discussed in this post veer into hyperventilation, but it’s still interesting to consider some of the possibilities. What would prevent OpenAI or Anthropic from deciding to USE rather than SELL their super-powered models?

Through a combination of using a better model (Mythos), and better harnessing of the model, Mozilla saw significantly better results hunting for security issues with Firefox. They decided to publish some of the bugs. As someone who is falling in love with arcane coding terminology, I really dug this sentence: “A raw NaN crossing an IPC boundary can masquerade as a tagged JS object pointer, turning double deserialization into a parent-process fake-object primitive for a sandbox escape.” Those damn tagged JS object pointers!

The downside of course is that Mythos can be used by the bad guys to find bugs too.

The Spanish AEO search firm Omnia plugged 20,000 search queries into Google, ChatGPT, and Perplexity. 98% of the time, URLs that came back would appear on only one engine. Only 2.37% of the URLs appeared in all 3. This was true for informational and commercial searches. Caveats: EU- and Spain-centric data.

A month ago, OpenAI rolled out a feature in their Codex app that gives you a browser window in the app. You can not only see what you’re building, but you can also comment directly in that in-app browser window.

On May 7 they announced a Chrome plug-in that lets you use Codex directly in Chrome. It can run projects in background tabs; it doesn’t need to take over the browser from the user. I personally would be leery of letting it work in a browser.

Last week they announced a new feature that lets you use Codex via the ChatGPT app on your phone.

OpenAI also claims Codex has over 4 million weekly active users today, an 8X increase from the beginning of the year.

Paywalled, but the gist is that women are 20% less likely than men to use generative AI, and that the AI boom is having gender-based effects at home, and these effects are not benefitting women. (I recommend subscribing to Wired, it’s cheap and they do great journalism.)

One of the main layers of crazy today is that companies boast about increased use of AI as an end, not as a means to an end. Who cares about consumer experiences, workforce morale, product quality, or profitability(??!!) when you can instead focus on tokenmaxxing? Developers report that AI output is often flawed. Using AI to code is often more time-consuming, harder, and more frustrating, because they have to check its work and fix its mistakes. Added bonus: they also say that using AI is degrading their coding skills.

The robot isn’t doing much, just making sure the mailing labels are always face down, but it’s still kind of nutz.