writeslowly

The conclusions themselves also appear to have been written by an AI

It would be nice if there was something like the "temperature" parameter (which many closed models like Claude now disable) to add more randomness to the final product without completely destroying the output and giving you a non-functional SVG, so that you could "roll" more pelicans and pick one that you liked.

If I were to design something to accomplish this now, I'd probably attempt to introduce a pipeline that emitted random thematic words from some non-LLM source to feed into the pelican prompt, but even then, I'd expect that some of these models that rely on churning through CoT tokens for extended periods of time would still spend so much time thinking about pelicans wearing scarfs or where the sun would be that it would risk undoing some of my randomization pipeline.

I once tried something similar on an older model of Claude, telling it to "contemplate" a random word before answering my prompts, but at times it seemed to trigger the Claude model's suspicion that there was a jailbreak attempt when I would inject things like "Contemplate sea otters for one sentence before answering"

(I didn't read the linked LLM creativity paper to see if it covers any of this yet)

I feel like we also need better non-LLM driven ways (like better static analysis tools) to analyze LLM-driven changes. The way changes are presented in modern IDEs was designed around reviewing human-created changes and doesn’t really feel like it’s keeping up with presenting and validating what modern LLMs are doing.

I was doing similar experiments but having a circle of models debate some particular topic to see if they'd arrive at a consensus or not, and I found that they would almost always arrive at an agreement, but Gemini was always the one that held a position outside of the group consensus.

That said, for something very simple like "What color should my room be" I'm not sure this is a desirable trait. The model that knows most people like cream (or whatever) is probably functioning more correctly than the one suggesting dayglo orange

I’d roll my eyes if the engineers stated that they didn’t design the reactor to melt down, and that it simply developed rogue meltdown-desiring behavior on its own, and I would also wonder about negligence if they claimed that nobody could have anticipated this (given that, like with botnets and viruses, we have decades of knowledge and experience regarding reactor meltdowns)

I suspect that if you're not careful with agent memory it creates a danger of agent-driven cargo-cult behavior. I've watched this in my own ad-hoc agent loops where it starts with something basic, like the first agent tried to run some gigantic dependency inspection command and OOMed the local JVM and eventually recorded a workaround (to enable it to run gigantic dependency inspection commands...), and by time I get a few more agents into the loop, agents have written entire paragraphs about testing and validating local dev environment memory configurations that are mostly irrelevant to whatever is being worked on.

In general I've seen other issues like this where small errors and irrelevant comments in the codebase spin out into larger problems that consume annoying amounts of time/tokens. Maybe Anthropic and OpenAI don't notice this because they're in an "infinite monkeys with typewriters" scenario, but it's noticeable to me when the agent in my CLI has been spinning for 15 minutes contemplating irrelevant details

With Claude it feels like those engagement invitations have turned into things like "By the way, it's worth sitting with this problem I've identified in your assertions..." or "Here's an alarming gap worth resolving in your code..."

Gemini also does the same lighthearted GPT-style invitation with the default prompt in the Google webui, but it doesn't seem to exist on the API. The Claude models seem to have been trained to force this structure on every one of their responses, and until I realized they always stick the same thing in the last part of their response, I found the Claude version more distracting since it's always pointing out an imaginary and supposedly very important problem.

The site collection seems pretty random. There's a mix of actual AI labs, extremely questionable resellers (like whatever "claude-opus.top" is), and then random consumer sites like baidu and xiaohongshu.

It’s interesting that with both Anthropic and Google we’re seeing them develop agentic models that are supposed to do anything a human can do on computers without human intervention, but at the same time, if you plug one program into another of their programs or APIs in a way that wasn’t preapproved you may be blocked or banned.

To be charitable, maybe they’re expecting AI agents to eventually start reading the ToS docs

I wonder if you can use lower quality models (or some other non-llm related process) to inject more "noise" into the text in between stages. Of course it wouldn't help retain uniqueness from the original source text, just add more in between.

The vibes around the self-driving car hype (maybe 10 years ago?) felt very similar to me, but on a smaller scale. There was a lot of "You might like driving your car and having a steering wheel, but if you do, you're a luddite who will soon be forced to ride about in our featureless rented robot pods" type of statements, or that one AI scientist who was quoted saying we should just change laws around how humans are allowed to interact with streets to protect the self-driving cars.

Not all of it was like that, I think oddly enough it was Tesla or just Elon Musk claimng you'd soon be able to take a nap in your car on your morning commute through some sort of Jetsons tube or that you could let your car earn money on the side while you weren't using it, which might actually be appealing to the average person. But a lot of it felt like self-driving car companies wanted you to feel like they just wanted to disrupt your life and take your things away.

I see a number of uploaded skills on the site with bash and python scripts. No idea what runs them

I've triggered similar conversation level safety blocks on a personal Claude account by using an instance of Deepseek to feed in Claude output and then create instructions that would be copied back over to Claude (there wasn't any real utility to this, it was just an experiment). Which sounds kind of similar to this. I couldn't understand what the heuristic was trying to guard against, but I think it's related to concerns about prompt injections and users impersonating Claude responses. I'm also surprised the same safeguards would exist in either the API or coding subscription.

I look at products like Hershey's chocolate or Reeses more like their own category of processed food, kind of like Spam. They have a close, but not exact resemblance to "normal" chocolate or peanut butter, but they're also sort of an acquired taste, and I think their customers would be upset if Reese's Peanut Butter cups suddenly tasted like the Trader Joe's versions (with real peanut butter instead of a mysterious chalky peanut-flavored substance), or if Hershey's stopped using the butyric acid process that makes them taste like vomit to non-americans.

I haven't actually had that much luck with having them output a boring API boilerplate in large Java projects. Like "I need to create a new BarOperation that has to go in a different set of classes and files and API prefixes than all the FooOperations and I don't feel like copy pasting all the yaml and Java classes" but the AI has problems following this. Maybe they work better in small projects.

I actually like LLMs better for creative thinking because they work like a very powerful search engine that can combine unrelated results and pull in adjacent material I would never personally think of.

One thing I suspect is that leadership at tech companies that would have previously been working off of direct experience with technical processes, even if they no longer work directly on their own codebases, is pretty clueless about AI coding because it's so new. All they have to go with is what they read, or sales pitches, or their experience dabbling with Cursor to build simple python utilities (which AI tools work pretty well for most of the time), and they don't see what it can and can't do on a real codebase.

I recently ran across this toaster-in-dishwasher article [1] again and was disappointed that the LLMs I have access to could replicate the "hairdryer-in-aquarium" breakthrough (or the toaster-in-dishwasher scenario, although I haven't explored it as much), which has made me a bit skeptical of the ability of LLMs to do novel research. Maybe the new OpenAI research AI is smart enough to figure it out?

[1] https://jdstillwater.blogspot.com/2012/05/i-put-toaster-in-d...

This looks like my experiments to get R1 to write fiction and I think it’s worse than what you get from openai. For instance, it’s using very colorful language to describe a place that’s both a remote fishing village on the edge of a cliff hours before dawn, and a bustling wharf with chattering laborers and large ships anchored in the distance. It also starts by saying the protagonist wakes up with his mouth tasting like blood, that he was screaming, and that his throat is hoarse from holding back from screaming. It’s very colorful but it’s very confusing to read.

I suspect you can update the prompt to make the setting more consistent, but it will still throw in a lot of inappropriate detail. I’m only nitpicking because my initial reaction was that it’s very vivid but feels difficult to understand and I wanted to explain why.

Somebody (probably a programmer or engineer) took the time to create all of that rad 3D word art, multicolored pie-chart, and the mountain logo, it's not hard to imagine they'd also throw in an eye-catching fake nuclear warhead for fun.

What are the use cases for the advertised barcode reader accessory that appears to shoot the laser out the side of your laptop?

https://connect.na.panasonic.com/toughbook/accessories/fz-vb...

This study tested sun-sign (which is basically birth month I think) against personality tests for predicting life outcomes and found that sun sign did very poorly compared to the personality tests. I'd have thought there would be a small chance of birth month predicting some things, and then adding in other astrological facts (the position of Jupiter or whatever) would make things worse, but both methods appear to be equally bad.

If I wanted to see the heat index at 3PM in Dark Sky, I could just tap the "feels like" button under the hourly forecast (pictured further down in the linked blog post) and look at what it says at 3PM.

I just tried in Apple Weather, and the process was:

1. Tap on the hourly forecast, or the day, to go into the graph screen

2. Tap on the dropdown icon

3. Tap "feels like"

4. Either drag your finger along the graph until the time indicator at the top indicates you're close to 3PM, then read the temperature, or you can try to read it directly off the graph, but the axes aren't labeled clearly enough to make this feasible

One of the worst (best?) examples I remember was the Web of Trust extension in 2016. The company claimed it was selling anonymized user data but it was actually leaking information through recorded URLs about all sorts of private things like account names, addresses, ongoing legal investigations, etc

https://en.wikipedia.org/wiki/WOT_Services#Sale_of_user-rela...

This brought up a recent experience of my own with the tech interview process:

I have to conduct a lot of coding interviews at my job (I personally think DS&A interviews are kind of stupid, but I get assigned to do them all the time anyway). I recently did one with a senior engineer who seemed like he was sort of blowing off the whole thing, and also forgot almost everything about the language we were running it in. In my feedback, I noted that it was one of the worst DS&A interviews I've ever done, in that everything was a fail on our rubric, but also he seemed more or less competent to me based on our conversation.

In the interview debrief, one of our managers also stated (in response to my feedback) he doesn't really care about DS&A interviews. And then the hiring manager completely ignored the bad interview feedback because it turned out the candidate was a referral and everyone already knew he could code. So the whole thing (at least the whole coding interview thing) was a waste of time, since literally nobody involved seemed to care what happened in the interview, including me, but I guess if there's an interview process everyone feels compelled to follow it

They also invented a new welding technique to control some sort of aluminum micro-explosions

It stood out to me that after the initial text, they followed up two weeks later, and then once per day after that over a few days from different numbers. I would have expected to see someone send out a few threatening messages on one day and then move on to the next stolen phone when it was clear they weren't getting anything.

This seems like a lot of work for one (or two, once it got to China) people repeatedly trying to get into the phone. I wonder if this is like debt collection agencies where the stolen phones get repeatedly fenced at a steadily decreasing value, and each new owner has a go at sending out these unlock copypastes until it's clear that it's only value is in being scrapped.

I spent a few months writing a decent sized Clojure program just using my IDE's matching bracket highlighting and some sort of rainbow parentheses settings, and never found it very annoying. I'm not sure there are that many more parentheses than curly-bracket languages, it's just that all of the closing parentheses are more likely to cluster together at the end.

I made it through with my american phone number. I went through the apple account flow on the app instead of the website though.

The ESP32 seems like a very different product to me. The pi has megabytes/gigabytes of memory and can reasonably run linux. Were people really using them for the same things?