Imagine you’re an independent filmmaker. You spend untold hours working tirelessly to perfect your plot, dream up your cinematography, shoot it, edit it, and cut trailers, just hoping you’re able to explain the story—your baby—well enough that people want to see it. In many ways, the trailer is the most important part; it’s what makes people decide to see your work or not. So you carefully select snippets of pivotal scenes, arranging them thoughtfully to make sure the viewer gets a taste of everything the film has to offer without giving too much away. Then finally, it’s the last frame. Your prime real estate: the spot for your tagline. This is your last chance to hook the audience. After everything you’ve sacrificed to bring this story to life…its success now hinges on your ability to distill its essence into a concise line that captivates the viewer and convinces them to go see it. So what does this woman do? She uses that space to break the fourth wall and make it entirely about herself. “One Woman. A Laptop. New Rules.” Absolutely nothing to do with the story. To her, the most important thing for anyone to know about the story…is that she made it. Hilarious.

This is legitimately one of the stupidest things I’ve ever heard. I swear to god it seems like a hiring requirement at Anthropic is to be a Cirque du Soleil-tier mental gymnast capable of bending over backwards to stab yourself in the back.

AI music is just like its writing: an amalgamation of ornate-looking devices that say nothing. An empty chest encased in pyrite and cubic zirconia: just enough to catch your eye…and then disappoint you as soon as you take a closer look.

A glass of vodka looks like a glass of water. Does that mean we should conclude they’re identical? Don’t mistake perceptual resemblance for ontological equivalence.

I think we’ve reached a point where people are starting to “retard” AI terminology. They’re rejecting it not because it’s wrong, but because it’s inconsistent with their sentiment.

Go to Best Buy and look at the OLED screens. Watch the nature scenes they show—how vivid they are. It’s certainly not real life, but it sure as hell is way more than a series of “pixels.”

The AP accidentally created a fantastic litmus test with this anti-anthropomorphization directive. There’s an elicitation technique in intelligence called a “presumptive” that tricks a suspect into divulging things based on a subconscious need to “set the record straight.” Instead of asking someone to tell you what happened, you explain a false understanding of events which compels them to correct you in a way that offers more information than they otherwise would have. It is a remarkably potent technique, as the attached video explains in more detail. Which brings us to this “lightning rod” debate over the semantics with which we refer to AI. I feel like I’m pretty representative of the average person: I personally don’t think a model “thinks” or “understands” the same way humans do, but the phenomenon is so practically similar that it doesn’t make sense to go to the trouble of defining new words for it. So colloquially, I just use the words that make the most sense. No big deal either way though…because it’s just innocuous semantics. But for some people, it is a very big deal. Which begs the question…why? Even in AI-centric circles, I think a lot of people play their cards close to the vest on how they conceive of AI because they don’t want to be seen as psychotic…but when you canonize verbiage that fundamentally violates their worldview, they can’t hold back. So you get this onslaught of deeply passionate positions on fundamentally sterile stimuli. Quite the tell.

Debt aversion is financial illiteracy. If you have a favorable rate on a mortgage, it is indefensible to pay it off before you have to. There is no scenario or argument where that is a rational decision; it is simply an unforced error in all circumstances. If the concept of debt weighs so heavily on your head that you feel motivated to do this, the answer is to educate yourself and understand why you are misguided, not validating mental gymnastics and indulging in financial self-harm.

If you get emotional and defensive about the words that other people use to refer to their own computer systems…you might have AI psychosis.

Good news for the anon crowd: this is nonsense. The grand way they found to “deanonymize” accounts was…finding publicly disclosed personally identifying information. In essence, they “doxed” people who explicitly identified themselves. There are essentially 3 separate studies in this: 1) Hacker News - LinkedIn profile pairs They found Hacker News accounts that explicitly linked to a LinkedIn profile, stripped the links from one of the accounts, fed it to the model, and told it to go find the other. Of course, that creates a bit of an issue, because every account in the sample was explicitly not anonymous. The researchers only found them because the owner publicly linked them together. So there is no reason to believe the posting behavior would carry any expectation of anonymity between accounts. 2) Reddit accounts For the meat of the Reddit section, they didn’t even use multiple accounts; they just took single Reddit accounts, divided them in two (either by subreddit or over time), and concluded the model had “deanonymized” the account when it was able to pick the second half out of a set. Once again, problematic design for the study because an anonymous account can remain anonymous while still “doxing” itself internally. If you had half of Satoshi’s posts, it wouldn’t be that hard to identify the others because the “accounts” would naturally talk about the same things in the same way. But would that mean you had “deanonymized” the actual account? Of course not, because neither “half” was behaving with an expectation of anonymity from the other. They also took a small amount of Reddit users who either had their username as their real name (academics who discussed their institutions and interests) or linked their LinkedIn profile directly for career feedback—neither group operating with an expectation of anonymity—and “deanonymized” them based on the identifying information they provided. 3) Scientist interviews Finally, they used pseudonymous interviews with scientists who didn’t disclose their names, but did disclose specific details about their research, background, and activities…which they used to match up with online public records. Once again effectively using non-anonymous data as their “anonymous” sample. So all in all, they proved they could “deanonymize” identities that were never truly anonymous to begin with. Anon universe can breathe a sigh of relief (for now).

It’s funny. You never see a guy with a financially stable company freaking out about AI safety risk. It’s only the ones in more precarious situations. I wonder why that is.

If you want to get a feeling for how authentic AI discourse really is, remember “loops.” That was a rare instance where both companies simultaneously ran influencer campaigns to blanket the market, didn’t achieve success, so they pulled the plug and you never heard about loops again. People “talk about” what they want them to talk about; they are spending ungodly amounts of money to control the narrative (guh). Helpful to remember that as we move through periods where these companies are more incentivized than ever before to stimulate/simulate support for various ideas that may or may not have a direct impact on their ability to remain afloat.

You expel them. It’s a grave disservice to our kids in general to tailor our systems to the least common denominator. If one child consistently compromises the learning environment of others to the point of being a net negative on the system, they should be jettisoned. It’s important to remember that our duty is equally distributed to every student. Allowing one maladjusted student to cannibalize resources intended to be shared equally robs innocent students of their fair share. As they say: “compassion toward the guilty is malice toward the innocent.” The good news is a system like that organically heals itself. If you establish a precedent that negative behaviors are not tolerated, many of them “magically” disappear. But to maintain that, it requires continuous uniform adherence…which is unfortunately difficult to achieve when policies are voted on by communities of social justice vigilantes, so woefully misguided that they do things as insane as sending their own child into a lion’s den of incompetence, risking their safety and compromising their future just to make a political point.

I don’t understand how anyone can hate on this. This idea that “documentaries have to have shitty cinematography to be taken seriously” is ridiculous. You can literally see it is Elizabeth Holmes. Are we really that simplistic that we can’t differentiate reality just because something is shot in a cinematic way? Like your brain just sees composed shots and lighting and just can’t comprehend how it can be anything other than fiction, even when an extremely prominent public figure is the subject? I mean come on. This is a real story about a real person shot in a compelling way. The last point does not invalidate the first two. This looks like an extraordinarily well-shot and edited piece; if the actual film is anything like these trailers, it’s going to be phenomenal. I can’t wait to see it.

I’m not sure who out there is still making do with a regular Apple Watch but this is your sign to get the Ultra. On battery life alone it is worth the price. I tolerated this old regular Apple Watch for years that barely made it through a day… Now I go a full day, sleep, wake up the next day, go for a run, work through the morning, then at 11:37am I look at my watch at it still has 78% battery left. I mean the quality of life adjustment is immense. Think about how much you would pay to not have to worry about charging your watch every day. Like $5 per day? $1? Say you will own the watch for at least 2 years, likely 3-4 before a replacement is even remotely necessary. Even at 2 years, $1 per day is worth almost the full cost of the watch: 365*2=$730 of mental sanity value for a watch that costs $800–nearly fully accounting for the entire cost of the watch, let alone the $400 delta between it and the Series X version. And that’s not even accounting for any of its functional value. You are basically costing yourself value by not buying this thing.

Every now and then you see references to social phenomena like “the iced coffee at a job interview discourse” and realize Kaczynski was right.

The fact that Anthropic can offer such a dramatically superior auto approval product for free in Claude while OpenAI still has to charge for codex-auto-review is kind of crazy. It’s such a prominent touchpoint for all work; I would have assumed the Codex team would have closed the gap on that by now. Remains a major advantage for Claude.

Cybersecurity people are upset at AI doomers, because most all doomer scenarios involve rampant hacking by AIs -- our area of expertise. What AI doomers want is to violate all the principles of cybersecurity that we've learned over the last 40 years. If the AI doom comes to pass, it'll be because the doomers succeeded at wrecking cybersecurity.

Idk what OpenAI has done with the interface of ChatGPT but it is horrendous. It feels like I’m looking at one of those nonstop sequences of ads on some random yahoo news article page or something. The size of the text, excessive use of bold, the random assortment of individual “div-ed” elements…just very corrosive on the eye.