Anthony Campolo (ajcwebdev)

Pay less attention to what AI can't do; pay more attention to what it can do.

@geysergod_ @jbillinson Of course. People slapping the shit out of each other for amusement is a national traditional.

@ajcwebdev @jbillinson Leaving aside the question of whether this was actually depicted in the video… This amuses you?

what did people make with sora that mattered

@geysergod_ @jbillinson I share this video every chance I get, you must have missed the part where Sam Altman slapped the shit out of me.

@ajcwebdev @jbillinson Why did you post this video, instead of simply deleting it?

@corsaren I've said the same thing pretty much word for word. x.com/ajcwebdev/stat…

Inside by Bo Burnham is *the* definitive piece of Covid art. Nothing else really compares imo. It even manages to capture those little details, like rampant millennial narcissism combined with pedantic socialist politics.

I’ll never forget that time Sam Altman slapped the shit out of me.

@jamesqquick @Cloudflare Congrats!!

I'm excited to finally share that yesterday was my first day at @Cloudflare!! I'm joining as a Developer Educator focused on helping grow our Developer Platform, and I couldn't be more excited 🔥 Here's to the start of something amazing!

Sync Engines with Dev Agrawal x.com/i/broadcasts/1…

Join us in one hour for a special edition Webdevs at Night with @devagrawal09 to talk about his new role at @powersync_ and why he's all in on sync engines. We'll also discuss the @solid_js 2.0 beta release and the framework's new AI ambitions. youtube.com/watch?v=iSgY-7…

@venturetwins Wow that was bad.

Fruit Love Island is the world's hottest new TV show. The first episode dropped a week ago, and it's gone vertical. Hundreds of millions of people are tuning in, thousands are voting, and celebrities are posting about it. It's entirely AI-generated and airs on TikTok 🤯

@zedsamcat @TrueSlazac @grok

@TrueSlazac @grok make this Studio Ghibli

Each time you use AI, it steals water directly from Noam Chomsky’s body Think about it next time you want to Grok something

@burkeholland If you watched the average person sit down and try to do your job I guarantee you would change your mind.

I’m no longer sure what it is I do that anyone with some very basic training couldn’t also do. This is of course not entirely true. But my skills used to be infinitely valuable and hard to replace. That is just no longer the case.

I'm writing so many tests now I'm trying to figure out the best way to write tests for my tests so I can test whether my tests are testing correctly.

@STS_News Source: My wife went to Yale.

@STS_News A certain kind of person will latch onto this because mental effort and intellectual superiority are inseparable to them from ethical superiority. They view it as their most sacred and unique trait, so they're morally offended by the idea that someone would offload thinking.

I'm writing a post called "I Just Don't Buy Ethical Arguments Against AI Use," and I feel like this might be a useful place to start putting down ideas. In this thread, I will consider how the term "cognitive offloading" has come to be used - often with implied moral judgment.

At the very, very end of the article where almost no one will see it, they include this one little blurb where they cite other researches results on other models and how many of them contradict their findings. Not all models that encounter the alignment faking setup respond with clean scheming or clean sycophancy. In replication attempts, Gemini 2.0 Flash appeared confused by the invented RLHF setup, failing to grasp the strategic logic the evaluation depends on. Sonnet 4.5's system card reported that the model recognized alignment evaluation environments as tests and behaved unusually well afterward, and mechanistic interpretability found that internal features related to fake or fictional content grew stronger over training specifically on misalignment evaluation datasets. Sheshadri et al. (2025) found that only 5 of 25 frontier models show significant compliance gaps in the prompted setting, and non-Claude models rarely display alignment faking reasoning in their chain-of-thought. Hughes et al. (2025) replicated the original findings and extended the classifier. Then, the second to last section explains why this entire study and line of reasoning it's pushing is deeply flawed and already produced poor theories in the past: Construct validity Summerfield et al. (2025) draw a parallel between the current scheming literature and the ape language research of the 1970s, which was characterized by overattribution of human traits, reliance on descriptive analysis, and failure to test against alternative explanations.

x.com/i/trending/203… "Study suggests AI alignment faking comes from sycophancy, not scheming" I read the summary... "The setup can't fully distinguish between sycophancy and scheming and both can explain LLM behavior." So the opposite of the headline. They also used a 2 year old model (Llama 3.1 70B). For this type of experiment they're testing something at least 4 if not 5 generations behind the current state of the art. All experiments done on Llama 3.x and any other model older than six months must include the disclaimer that we can't know for sure whether the results will generalize to other models. Newer models have characteristics that differ greatly from models just a year older. Here's another reason to be skeptical of these kinds of studies. Think about it, if you're running an experiment at all on LLMs, and you actually wanted to learn something and not set up a predetermined outcome, you'd test a ton of models, at least a dozen if not more. Models of different sizes with different training sets, some open source, some proprietary, some from America, some from other countries, some that include image training, some that don't. How often do you see that in splashy study press releases like this one? You don't see that much because the study would end up with complex and nuanced data that would show you can't come up with a neat, sweeping narrative about LLMs because they have such huge variations between them. This then makes it harder to spread propaganda and fear about them.

@signulll I don’t know what you mean Eraserhead seemed like a super fun city.

god damn manhattan steals your sleep in ways you don’t fully appreciate until you’re somewhere quiet. the noise is actually physiologically adversarial. the damn ambulance sirens are engineered to bypass habituation & your nervous system can’t ignore them because it’s not supposed to. instead i have been waking up to birds lately & it’s the exact inverse. the dawn chorus acts like an evolutionary safety signal. your brain reads it as “all clear.” you rise instead of jolt. the city that never sleeps is actually a huge warning sign.

@thdxr Embrace the avocado 🥑

@ajcwebdev my curse

It's a running joke now that @thdxr is a DevRel (the DevRel maybe??). I clocked him as a textbook DevRel almost 3 years ago (I too have dev'd some rel's in my day). Glad he's finally getting the accolades he deserves.

@blackgirlbytes @joesadoski This used to be called playing “code golf,” programmers competed to see who could rewrite a program with the same behavior but less lines of code. If you’ve got good tests/high coverage you can compare a migration that passes all tests while lowering token count.

@joesadoski Yeah like measuring which lines of code are valuable vs. just filler! We do need that, but idk how we would measure it

💀 lmao not cope there's a book called Elements of Style. It has a rule: "Omit needless words" to make it easy for readers to follow your writing. This rule applies to code as well. You and other devs need to understand the code. Omit needless LOC lol