When you read about AI, you've got to approach it first and foremost with this mindset: most people have no idea what the fuck they're talking about.
One of my main hobbies is reading short-form internet articles, in greater quantity than you'd actually think is healthy, with an intentionally broad exposure spectrum. I do read the "big names" in AI, and a fair number of research papers - but I also want to know what everyday people think about it! Labs build the tech but everyone builds the culture.
And let me tell you, for as much as we self-criticize, the on-Bluesky discussion culture is actually some of the better AI chat that's happening out there. You would not believe how godawful it's gotten in the deepest cesspools of X, LinkedIn, Reddit, and even Facebook.
This is especially true among less-technical users. There are some "end user" communities I won't link to as I don't want to draw attention to users who have more or less auto-gangstalked themselves via use of AI tools. It's tragic stuff that goes deeper than mainstream media is reporting on.
I started writing a longer post on AI cost effectiveness and budgeting, but I wanted to talk about why real world outcomes are so hard to understand. I blame incentives from many directions to spread bad content about AI, such that it drowns out almost any high quality "ground truth" info about real world AI adoption. And that got long enough to pull out on its own!
So, here's my typology of why AI posts are especially bad, ranked from the honest to the craven:
- 1.
Most people giving opinions about anything aren't really doing so in a rational or measured way. But it's easy to get some kind of numbers out of AI adoption efforts, and we have a tendency to believe those with numbers must be more informed than those without. (Even when there isn't really justification for why those particular numbers, or how they were collected, or what they even mean.)
- 2.
Everyone using modern LLMs started doing it this decade, and most people, far more recently than that.1 There's very rarely any acknowledgement of humility around this! Some of these awestruck authors have never even written code before, so their articles about the amazing power of AI are just as much about them realizing the benefit of any automation for the first time.
- 3.
Agents, their interactions on particular codebases, and your interactions with a provider can all have quirks. If you've only used one agent in a particular context you might be incorrect about what they can do2 or what they cost in a general sense.
- 4.
The most agentpilled and Yeggeified among us really do believe they've built something That Changes Everything. So do the people who think their agents have solved Navier-Stokes or quantum gravity or built a new virtual society. I'm not sure I love the phrase "AI psychosis" or "bliss attractor" but there are people producing vast amounts of self-referential work-shaped churn with no real-world impact other than, well, blog posts. I have even seen many cases of creators taking out paid advertisements on platforms to showcase non-plausible solo-author LLM-based "new math" and "new physics" results.
- 5.
There's a culture of openness around AI "discoveries"... but a lot of them are actually trivial extensions, outpaced by ongoing harness development, or just errors in benchmarking. This is often promoted by folks with a DIY hacker or tinkerer mindset who don't realize that their improvements aren't generalizable or cost-effective, or are long since implemented by inference providers. People who have discovered legitimately non-obvious and interesting secrets have a strong incentive to not talk about them publicly!
- 6.
If you're a director or executive, you might not have ever used AI for a substantive coding or knowledge work task. Even if you got with "AI transformation" early, even if you were previously technical: you're now responsible for managing hundreds of early adopters who know the nuts and bolts daily uses of this technology better than you do. Like any "upwards" feedback, it's going to get filtered and warped into something that they think you want to hear.
- 7.
Many AI critics are grounded in how models worked several years ago, because they haven't actually used them since then. That was before vast improvements in inference cost, context windows, tool use, safeguards, and many other adjacent areas - plus a lot of user experimentation about what works best. Taken together these have dramatically changed the quantitative performance, qualitative working experience, and cultural expectations around these models.
- 8.
If you're a newly launched startup looking for funding, you have to at least credibly lie about how relevant you are to modern AI. There are enormous numbers of companies seeding social media with comments, blog posts, and studies (occasionally in covert ways) for viral traction. These are like the prior failure cases, but with even more motivated reasoning.
- 9.
The broader community of braggarts, blowhards, content creators, sweaty grinders, and scam artists have followed the money and moved on from crypto into AI. Not everything they post is an outright lie or scam, mind you - but it is all in service to a "get rich quick" mindset, which inevitably trends towards scamming people. You need to develop a near-allergic distaste for this category of people.3
Of course, much of this boils down to "you can't trust everything you read". Always true, but there's many different reasons to not trust people! Understanding why something is bullshit in specific detail means that you can better explain it to your friends without just saying that "the vibes are off".
Perhaps you can even address it:
Your CEO needs some education from someone who isn't a McKinseyite
The "afflicted" need help if possible and to be politely ignored if not
The scam guys need to be driven out of the industry the way we drove crypto posters off of Bluesky
I don't know if any of these flawed discussion modes are really solvable at a structural level, but with a little discernment and effort you really can improve your local information environment for you and your friends.