There’s documented cases of PTSD among image censor workers…

So, yeah, LLM goes off down weird rabbit holes - and if you, like Alice, follow them down… that’s on you.

If, instead of getting off on how weird the response was the author just prompted - “I’m looking for the basketball memes about the NBA player from Turkey.” I suspect it would have gotten him what he was looking for.

If, instead, he shone it on and started playing along like he was distraught over Dario, I’m sure he could have led it to give him a suicide counselling hotline in 2 more prompts. And, honestly, which is the more common question to Google these days: a 10-15 year old NBA meme, or people distraught about supposed friends ghosting them?

This is business.

So was the Opium trade.

Vibe coding / agentic workflows result in poorer code quality, and relaxed oversight practices.

This was much more true in 2025 than it is today. Recent models and harnesses are improving quality and implementing more strict development practices than were commonly used by human development teams.

If you want thorough Requirements and Design specifications, full automated test coverage and CI gating based on automated test outcomes, zero compiler warnings, etc. LLMs are doing those things now more by default without explicit and repeated prompting.

If you are choosing to not human review and test the outputs of the LLMs, that’s your practices slipping, not the LLM agents’.

I’m working on Lempi, which is both similar and different - mostly on how the DJ works:

Rufin’s Auto DJ says “more like this track” and trusts the server’s idea of similar. Lempi works like a radio programmer with rules: it decides how often each song may come back, what fits this hour, and what follows well from the last track, then adds a controlled amount of chance.

https://github.com/MangoCats/Lempi

I’ve been off and on developing Lempi for personal use since 2020 - it has other names in the past, most recently changed becasue I was developing under a name somebody else “went wide” with in 2023. There’s a lot of music players out there.

I was “following key length recommendations” 10 years ago that said 3072 bit or higher for RSA. At that time, lightweight processors would struggle a bit with 4096 bit keys, so I’ve launched a fair number of systems with 3072 bit keys - none of the data on any of them is worth a handful of months on a CPU cluster.

Though, there is an interesting metric you can put on key value today, because so much compute is available for short term lease, you can literally spend $1m for a 50/50 shot at cracking a “very secure” key in a few hours, because you pay per unit of compute, not per hour. Now, Amazon or whoever might start refusing your business if you tried to spend $1M on compute in parallel all running within a single hour, but you could probably negotiate with them to get that done in low demand slots over a couple of days.

So many people struggle just to understand / use Facebook, that’s where they stop - if it isn’t on Facebook, or their other platform of choice, it’s dead to them because they just won’t take the time.

I’ve had a website at a .com address since 1997. Google used to surface it front page on relevant searches for over a decade. Not so much anymore - now if you don’t know its unique name, you’re not likely to find it. However, for the people who do reference it on a regular basis, Google does show it for them…

They used to have an option to personalize results or not, but that went away long ago.

My recent searches are definitely personalized, which usually results in good things, but it makes it impossible to tell somebody else "hey, just search for this: … "

serving different results depending on whether you’re logged in or not.

The real thing is serving different results depending on what your profile contains. Even if you’re not logged in, you still have a profile based on your IP address, time of day, freqency of queries, previous queries, etc.

Try an AI prompt like: Give me a 2006 style Google search result of the modern web for the search “Apple Pie Recipe”

They can still do it, if you ask.

And using breeder reactors that 50,000 year figure could be pushed up to 5 million years.

I think that they’re seeing all the scary movie stuff start to come true - but missing the essential plot holes like that one: just unplug it.

People complain about AI slop - well, before AI slop there was lazy human slop, but lazy humans generally gave up before writing too many words or painting too much mess, or digital clone-stamping too many generic things… Truth is: if you put in 1/2 the effort of making something great without an AI tool into making something great with an AI tool, you can usually get something much better out in the end - if you’re sincerely trying to use the tools well.

It’s the people who don’t know / don’t care whether the result is slop or not who are putting in 1/10th (or less) the effort of making something great who are putting all the slop out there.

I’m not afraid of the machines taking over (yet), but I am afraid that the same explosion in copy-pasta document vomit that came with word processing 40-ish years ago is going to explode again into a docpocalypse with people using AI to drown each other in supposed “legal written” communications that are only even vaguely understandable using AI tools to read them.

pledge allegiance to AI safety and undergo a psychological profiling test, per The Information.

And according to a rumor that, per The Spectator‘s Sean Thomas, is running rampant in Silicon Valley these days, some Anthropic engineers are going as far as to “worship” the company’s AI model Claude.

It would seem that the AI safety psych profiling test should be screening out / rejecting those who would give an AI “whatever it wants.”

The main thing I find with the LLMs is that they’re very fast and capable and limited. You give them a task of low complexity, they can execute it with superhuman speed, and if there wasn’t hidden complexity in it, they do well.

If the task is too complex (beyond their context window’s capacity), they start to “hallucinate” - make up things that sound reasonable to their training - to cover gaps in “what to do next” because they’ve forgotten some of the complex instructions. Putting the instructions “in writing” in a reference document that the agent doesn’t edit freely can help a lot, but eventually those documents get too complex as well and again it starts “winging it” rather than endlessly re-reading the instructions before making every next move.

The models’ training is getting better, doing “the right thing” more often by default - without requiring as much explicit guardrailing. They’re also getting better at paying attention to requirements documents, but they’re far from perfect. As are humans, but humans don’t always move as fast as LLM agents, so LLM agents can make big messes faster…

I did “recon drones” back in 2010… of course they lept at the opportunity to demonstrate kamakazi delivery of explosives with the platform to the Army too…

In my generation it was the nuclear industry (immediately post TMI & Chernobyl) which was giving away the huge incentive bonuses and premium salaries. I never filled out the Navy application - though I hear they paid the best overall. I did interview twice (nothing better to do at the time), got two offers, declined both, ultimately took lower paying jobs that I didn’t mind the idea of going to work at.

That’s not acceptable, that’s not professional, somebody is going to eat that burrito - are you going to file their permission to use the image of their food? Too complicated.

Two issues I have with the presentation:

A magnetic field is an external signal - the Earth’s magnetic field is actually incredibly weak…

10x better performance than GNSS … positioning accuracy of 1 nautical mile. Cheap old school handheld GPS out in the woods was accurate better than 10 feet most of the time, so what in this performance is 10x better? 10x more resistant to jamming? Seems like it should be completely impervious to jamming. 1 nm accuracy is useful, but hardly better performance.

Does it do parametric model designs?