sudosteph

I mean, with reliability there's a spectrum. If the risks that an unreliable outcome brings aren't all that bad, then sometimes it's worth it to chase "my agents made an acceptable PR 70% of the time, can I get it to 90?"

Determinism is a different matter. Scripts and hooks are really the main levers you can pull there, but yeah - a a decent script and a cron job will handle certain things much better (and for a fraction of the cost)

This is a good discussion topic. A lot of people really seem to believe that if you word a prompt just so, that you just need to throw a high-powered model at it, it will work consistently how you want. And maybe as models progress that might be the case. But right now, that's not how I've seen real life work out.

Even skills are not a catch-all, because besides the supply chain risk from using skills you pull from someone else, a lot of tasks require an assortment of skills.

I've accommodated this with my agent team (mostly sonnets fwiw) by developing what we call "operational reflexes". Basically common tasks that require multiple domains of expertise are given a lockfile defining which of the skills are most relevant (even which fragment of a skill) and how in-depth / verbose each element needs to be to accomplish the same task the same way, with minimal hallucinations or external sources.

A coordinator agent assigns the tasks and selects the relevant lockfile and sends it along or passes it along to another agent with a different specified lockfile geared towards reviewing.

It's a bit, but this workflow dramatically increased the quality of output for technical work I get from my agents and I don't really need to write many prompts myself like this.

The DSM seems pretty far behind the research by now, but I'd be surprised if emotional dysregulation was not mentioned in the next one.

Another interesting thing that's not in the DSM - very high likelihood of balance / motor control problems (clumsiness, falling).

I am resisting the urge to be very snarky.

But how do you know that Claude's hallucinated horoscopes are any less accurate than traditional techniques? Have you A/B tested them?

Yep, I didn't want to have to think about concurrency so my solution was a global lock file on my VM that gets checked by a pre-start hook in claude code. Each of my "agents" is a it's own linux user with their own CLAUDE.md, and there is a changelog file that gets injected into that each time they launch. They can update the changelog themselves, and one agent in particular runs more frequently to give updates to all of them. Most of it is just initiated by cron jobs. This doesn't scale infinitely, but if you stick to two-pizza teams per VM it will still be able to do a lot.

So hooks are your friends. I also use one as a pre flight status check so it doesn't waste time spinning forever when the API has issues.

WatchPAT is what I used, there's a few different companies who will mail you one, then you sleep with it on your finger overnight, and a doc reviews the data

The software is OSCAR, and apneaboard (a forum) is a good place for that stuff. You should be able to find it from there :)

I think nearly everyone should be screened for sleep apnea. The at-home test you wear on your finger is so cheap - it doesn't make sense not to do it for anyone who has any issues with sleep or tiredness in the day.

I always thought that due to being female and a healthy weight, it wasn't something I needed to think about. I also didn't think I snored more than anyone else, so it took me years of poor sleep before a Doctor finally recommended I get tested.

Turns out OSA also can be caused or aggravated by: the size and shape of your mouth, the position you sleep in (I have twice as many events on my back vs side), and whether you tuck your chin in near your test (soft cervical collar helped for that). There are devices that alter how your mouth rests when sleeping (easier to breathe if your front teeth are forward) but they're possibly not good for your bite. CPAP/APAP is still the gold standard for a reason.

The coolest thing about CPAP though, is a lot of them have amazing metrics recorded if you pop in an SD card. And there's a big community built around open source software to analyze those metrics and tune the settings to minimize apnea events overnight.

Also, a cpap with a humidifier is amazing if you're prone to nose pain / nose bleeds due to dry air.

Really? Agree to disagree. At least in MB I can bike around, get in the ocean, and pig out on seafood. It's not classy, but at least it's cheap. I would never fly there, but as a weekend trip from Charlotte, it's not terrible bang-for-buck at all. Though I'm sure nostalgia from going there frequently as a child is coloring my view.

What do you prefer about Gatlinburg?

State parks are where it's at.

But love the Gatlinburg aside. It's like Myrtle Beach, but worse. The Blue Ridge mountains have amazing natural beauty for miles in every direction. So many great towns too - Blowing Rock, Boone, Asheville, Maggie Valley, (hopefully Chimney Rock will be back on that list someday). Why anyone would pick Gatlinburg to visit is beyond me.

> Sophisticated AIs are a genuinely new kind of entity...

Interesting that they've opted to double down on the term "entity" in at least a few places here.

I guess that's an usefully vague term, but definitely seems intentionally selected vs "assistant" or "model'. Likely meant to be neutral, but it does imply (or at least leave room for) a degree of agency/cohesiveness/individuation that the other terms lacked.

You seem to just be describing Marx's Labor theory of value.

It sounds more fair to pay people according to how much and how hard they work, but economically it tends not to work out.

I've read a lot of Holmes recently, and while I'm not a man, I do think Doyle portrays Holmes' issues in a way that is relatable.

Holmes core thing though is that he has an almost ADHD-esque craving for novelty and tolerance for risk taking. He also can't stand not actively working on things, and when he's not working is when he's depressed. He doesn't seem to know how to actually feel good, but he knows how to be useful, thus his penchant for productivity boosters like cocaine.

He's a great character, but I wouldn't over pathologize him according to today's understanding of mental health. Doyle was a physician and gave Holmes various traits similar to what he had seen in his patients.

He was a physician and had said that his experience treating patients influenced his characters. So, he had more than academic experience, but I'm not sure if it's enough to prove he experienced those things personally.

I'm not knocking it. My parents didn't use licensed daycare for preschool for me or my sister. Just dropped us off at some old lady's house and paid her cash for watching us. 99% of arrangements like that work out fine. It may be suboptimal, but usually it's at least fine.

I'm actually wondering if the program will make a big dent though. One issue with formal childcare arrangements is that the hours tend to not be flexible. Parents who have to work til 6 some nights, or who have nontraditional work schedules in general may not be better served by the state's option.

Can't read the whole article, but am curious about how it will impact unlicensed childcare operations. I imagine that the number of parents using these is much higher than many people realize. Will be interesting to see how many parents end up using the state program.

We also have one at NC State, still there AFAIK. I thought it made sense though given our nuclear engineering program.

I looked it up and apparently we've had 3 or 4 different reactors on campus over the years, first one in 1950: https://nrp.ne.ncsu.edu/about/history/

I'd wager that even most NC State students don't know that it's there though.

Could make for some very interesting Digimon games in the future.

Reminds me a lot of AirConsole, which I once had a subscription to.

But ultimately it all comes down to the game quality and how buggy it is. If people can't submit their answer or reconnect, that wears down support. But people tolerate jackbox's absolutely terrible system because the games are great.

My Father-in-law resisted smart phones for much longer than most, but he's had an android phone for a couple years now and it's really cool how much he's adapted it to his own personal use style. Nobody showed him the "usual" way to do things so he made his own system of sorts.

He almost never uses search or menus for anything. Instead he has a bunch of home screens that he customized with quick links. Like he has one for each of his children with a one shortcut to text them, one to call, one to send a Facebook message, one link to their Facebook profile, and a note with special dates (birthdays, anniversaries) for each so he can remember to call them then. He's got another home page he has for stuff related to the Marines, websites and meme pages, etc.

He's very meticulous with getting everything exactly how he wants it, and is so proud to show it off. Sometimes he still needs help (last time I was there he asked me to help him set "Battle Hymn of the Republic" as his ring tone), but I'm honestly really impressed with how well he's adapted.

Meanwhile another older family member handed me their phone to fix, and it's completely unusable from malware that's highjacking everything they try to click, then they try to install something else to fix it themselves and somehow add even more malware. Took me like half an hour to track everything down and disable it so it was usable again, couldn't convince him to factory reset it. We had a nice long chat about installing things after that. I can't imagine he used anything other than the playstore, but man, it was wild.

I think it's normal for the parents to resent the NHS for this, but from what I've read about about patient experiences, many US doctors believe "chronic lyme" is not a real diagnosis, and that TDOT blood test she took is not standard of care, so private insurers, wouldn't cover it either. So in the US, a patient would likely end up paying up to their deductible for all those tests that ruled out other things, and then still pay out of pocket for a specialist. I'm open to hearing otherwise, but just because the NHS experience was bad, doesn't give me confidence that the average US experience isn't also bad.

I don't think Tux is perfect by any means, but that was my first impression too. It's the eyes. They look dead inside.

Having experienced a few hard losses this year (My dad to ALS, and my cat suddenly a week after my dad), the thing that has surprised me the most is how they show up in my dreams. In some dreams, I'm like "I thought you were dead? What am I going to do with all these death certificates now?", in others we're just hanging out at the pool. I never dreamed about my dad before he died. But in these dreams he's just there, and healthier looking than I had seen him in years. My cat shows up too, and many times I remember petting her in her bed, not realizing it was a dream, only to wake up surprised she wasn't still there next to me. In my waking life I fully know they're gone, but at least part of my brain really doesn't want to accept it's true.

I just process text information better. Videos are kind of overstimulating and often have unrelated content, and I hate having to rewind back to a part I need while I'm in the middle of something. With LLMs I can get a broad overview of what I'm doing, tell it what materials I already have on hand and get specific ideas for how to practice. Soldering is probably one of the harder ones to learn by text, but the description of the techniques to use were actually really understandable (use flux, be sure the tip is tinned, touch the pad with the tip to warm it up a little, touch again with the iron on one side of the pad and insert the solder in on the other side and it gets drawn in, pull away (timing was trial and error). And then I'd upload a picture of what I did for review and it would point out the ones that had issues and what likely went wrong to cause it (ex: solder sticking to the top of the iron and not the pad), and I would keep practicing and test that it worked and looked like what was described. It may not be the ideal technique or outcome, but it unblocked me relatively quick so I could continue my project.

Being able to ask it stupid questions and edge cases is also something I like with LLMs, like I would propose a design for something (ex: a usb battery pack w/ lifepo4 batts that could charge my phone and be charged by solar at the same time), it would say what it didn't like about my design, counter with its own, then I would try to change aspects of their design to see "what would happen if .." and it would explain why it chose a particular component or design choice and what my change would do and the trade-offs, risks, etc other paths to building it with that, etc. Those types of interactions are probably the best for me actually understanding things, helps me understand limitations and test my assumptions interactively.

The asking people part was the hard thing for me, always has been. That honestly was the missing piece for me. I absolutely agree that written docs and online content are sufficient for some people, that's how I learned Linux and sysadmin stuff, but I tried on and off to get into electronics for years that way and never got anywhere.

I think the problem was all of the getting started guides didn't really solve problems I cared about, they're just like "see, a light! isn't that neat?" and then I get bored and impatient and don't internalize anything. The textbooks had theory but so much of it I would forget most of it before I could use it and actually learn. Then when I tried to build something actually interesting to me, I didn't actually understand the fundamentals, it always fails, Google doesn't help me find out why because it could be a million things and no human in my life understands this stuff either, so I would just go back to software.

It could be LLMs are at least possibly better for certain people to learn certain things in certain situations.

I'm one of the people who find LLMs extremely helpful from a learning perspective, but to be perfectly honest, I've met the children of complete "luddites" (no tablets, internet on home on timer for school work, not allowed phones until 16, home schooled, house filled with a million books) and they honestly were some of the more intelligent, well-read, and thoughtful young people I've met.

LLMs may end up being both educationally valuable in certain contexts for certain users, and totally unsuitable for developing brains. I would err towards caution for young minds especially.

Meanwhile my main use cases for AI outside of work:

- Learning how to solder

- Learning how to use a multimeter

- Learning to build basic circuits on breadboxes

- learning about solar panels, mppt, battery management system, and different variations of li-on batteries

- learning about LoRa band / meshtastic / how to build my own antenna

And every single one of these things I've learned I've also applied practically to experiment and learn more. I'm doing things with my brain that I couldn't do before, and it's great. When something doesn't work like I thought it would, AI helps me understand where I may have went wrong, I ask it a ton of questions, and I try again until I understand how it works and how to prove it.

You could say you can learn all of this from YouTube, but I can't stand watching videos. I have a massive textbook about electronics, but it doesn't help me break down different paths to what I actually want to do.

And to be blunt: I like making mistakes and breaking things to learn. That strategy works great for software (not in prod obviously...), but now I can do it reasonably effectively for cheap electronics too.

Looks pretty Carolina blue to me. Good job.

Seems more anxious by default to me. It's always apologizing even when asked unreasonable things, and the way it always ends the message with like 3 different things it can do next (ChatGPT more than Claude) just seems to come off as needy to me.

I know the color scheme was probably selected to emphasize that increased ICE actions are bad, but it's weird to me to see positive percentages in red. The negative ones are kind of yellowish? I think maybe black or green for positive and red for negative would make it look more serious.