crajaspedet.bsky.social
I build software stuff sometimes. Sometimes I do stuff with AI. Sometimes I'm hangry.
108 posts
35 followers
60 following
Active Commenter
comment in response to
post
While I know telling ChatGPT is useless, if you use a setup that works with MCP servers then you could set up a note app and give the model access to info. In the app you write basically the ways it fucked up and how you fixed it. Those get fed into the model and it surprisingly does improve things.
comment in response to
post
To be fair, you never had privacy when out in public. The potentially funnier part to me is that the cameras at intersections, etc. may get fried by lidar. As such the lidar may be a big boost to privacy if it fries them all.
comment in response to
post
Just now seeing correction. Not sure where I woke up and head hurts. Thanks Alanah.
comment in response to
post
They announced the agent stuff at Build and are making a big show of using it in public. I beta tested some of the early agent mode stuff and it works well on certain types of tasks. They'll throw everything at it, find those categories, and then automate it to kick in for those items. Expand later.
comment in response to
post
Oh and only thing stopping the Dolly like human clones is that somatic cell nuclear transfer isn't that great based on the research thus far. More likely to see cloned organs to come out of it if anything. The UN "ban" is rather nebulous and nonbinding at best.
comment in response to
post
I can think of a couple uses. In movies where a scene may be too dangerous or expensive for traditional CGI, etc. In fill. But I prefer practical effects to CGI and gen vids look even weirder. Games, dynamic cinematics opens some roguelite possibilities. Prefer the Hades story model though for that.
comment in response to
post
And yet people still ask questions for things that can be looked up with a Google search. None of which is helpful in determining if you should hire them. I've completely stopped asking those style questions in interviews and now just ask the person what they've worked on and how they like to work.
comment in response to
post
AI is terrible at planning and architecture. If you break things down into small tasks, reduce ambiguity, use commits as check points, roll back instead of iterating on major issues, etc. it works OK. The context window isn't large enough for large projects/tasks and ambiguity is a killer though.
comment in response to
post
They are very much an intern or baby dev on your team. I've found that if I tell it to ask questions for any ambiguity before you have it start processing, and answer the questions with more detail, it seems to help on the initial direction it goes down. Or those dice rolls were lucky, hard to say.
comment in response to
post
We've been testing various agent based coders at work. The potential time savings is crazy if you can properly break down a task into small pieces. The quality of the code is a mixed bag though depending on the task and if you're not specific enough. I'm curious how well it did on that front.
comment in response to
post
And when he does finally see a platypus, after going through his conspiracy theory/flat earther phase, he's going to find out that they're tiny compared to what he imagined. Everyone imagines them like 3 feet or larger when in reality they're small.
comment in response to
post
I mean technically the company I work for has an office in London that consists of one guy and we've been doing continuous delivery for 15 years now successfully.
comment in response to
post
Best I can do is Grosse Pointe Blank.
comment in response to
post
The main issue that I've run into with copilot agent is when it decides to randomly drop features from your code base which seems to be an issue with larger code bases or more complex asks. Breaking them down into smaller pieces has helped. But tests, etc. it's rather amazing.
comment in response to
post
Not with that attitude you don't. Got to put in the effort and believe in yourself.
comment in response to
post
To update the two people that will read this, the answer was not "leave the site" apparently.
comment in response to
post
It was just trained to be the old dev in your office who hates any syntactic sugar added after C# 7. Working as intended.
comment in response to
post
I have numerous jokes that I want to make along with questions that I want to ask. And I'm afraid to do either honestly.
comment in response to
post
Canada geese are about 40in but that's with neck/head. Assuming you are having them stand on each other's backs, may need a 3rd one for a large trench coat. Smaller one, 2 will work.
comment in response to
post
I'm running on 3 hours of sleep and this is funnier to me than the current comic. But only by a little.
To be fair though, I came into work with over 9000 emails and heard Vegetta's voice in my head say the line. I laughed so hard I was crying. So I'm probably not a good judge at the moment.
comment in response to
post
You're not going to believe this but I unplugged for a week and just came back from vacation today to see Patton Oswalt showing the collapse of the stock market. Not quite as entertaining as the war of the Taylors that I had hoped for.
comment in response to
post
He doesn't define "it". As such it could be interpreted to mean the original amount. So week one $0.01. Every week after $0.02. Compounding total is implied by the wording but not explicitly laid out in legal terms by Jason. This basically but against Jason: smbc-comics.com/index.php?id...
comment in response to
post
Interesting patterns dealing with SaaS, etc. are 10 to 20 years old now and they were based on concepts from mainframe, etc. time frame. Is there anything from the 2020s that's interesting? Are LLMs it? I'm not saying AI isn't useful, but it's feeling like that's it for 5 years and it's kind of sad.
comment in response to
post
Back in the 80s we got interesting paradigms because of the PC, in the 90s we got web dev, etc. Heck in the 90s we got Haskell, Python, Java, etc. All of the cool functional concepts are from the 60s. Even "newer" stuff in that vein like algebraic effects aren't that new at this point.
comment in response to
post
Like if I got offline again I'm going to come back to see Patton Oswalt posting that Taylor Tomlinson has eaten Taylor Swift, proving that she is the best Taylor once and for all. Although with that one I'll just nod and agree but my daughter will be so mad.
comment in response to
post
You're always good man. More just every time I come back the first thing that I read, no matter the source, is either weird drama when only love should be given OR some news story that makes me want to go back into my hole.
comment in response to
post
You know, I have been doing an internet fast because everything. Last time I came back C# devs had a meltdown over TypeScript. So went away and said I'll just check Hanselman today because he just posts nerdy stuff or inclusivity most of the time. I like that stuff. Open feed... "What the fuck?"
comment in response to
post
I went on an internet fast because of... Well... Everything going on. I figured I'd come back for coding stuff because it's usually boring/safe and that was literally the first thing that came up. That thread and the Reddit ones are crazy to me. Reaction should be "10x speed boost, cool" and be done
comment in response to
post
It doesn't have to. Look into Oh My Posh or similar. Also customize your theme in terminal settings. And look into aliases to improve scripts: ohmyposh.dev
comment in response to
post
6. Potential legal issues with multiple devs and ownership of the funds.
7. No one sponsors so why set it up?
8. Some people go weird third party funding avenues like Moq.
9. Then there's the paid license approach.
List goes on. I'm always more surprised when someone has set it up.
comment in response to
post
1. Not everyone wants to turn their side project into a job.
2. Some people just like to build things.
3. Some people are in the free beer camp.
4. Legal issues with their employer on accepting payment for outside work.
5. Legal issues depending on country of origin for the devs.
comment in response to
post
I only did surface deep before on 32B but going back and pushing for details, I was able to hit the guidelines wall with enough probing. Sad. To the trash with the model I guess.
comment in response to
post
I take it back as I just downloaded the 8B llama model and yeah, depending on how I ask the questions it comes back heavily censored. 32B is what I was testing with and it seems to be easier to bypass but may just be how I was asking the questions when we did initial tests.
comment in response to
post
I haven't been using the llama distilled 8b model. Will check that one when I'm near my machine. To me that looks like they went through a sanitization phase with their data like MS did with Phi but did a poor job as that hasn't tripped for me yet.
comment in response to
post
Interesting that you're getting that in LMStudio as I've mostly been testing in huggingchat and via calls in C# against the smaller models. But huggingchat response:
comment in response to
post
As others have mentioned, models seem OK so far. App is where they're censoring. Also, may not want to use the app based on their privacy policy and what sounds a lot like a key logger (Automatically Collected Information > Technical Information): platform.deepseek.com/downloads/De...
comment in response to
post
I'm genuinely curious how that goes as Cursor looked interesting but the C# support seemed lacking when I looked at it.
comment in response to
post
Interesting. I haven't tried it in PHPStorm but that sucks. It's kind of like the Azure Data Studio plugin was dead for 6 months because they didn't update it. Makes me wonder if there are different groups working on it depending on the integration and maybe some are not getting proper resources.
comment in response to
post
I'm always curious how people are using Copilot because it works fine inline for me but the chat is terrible in Visual Studio. And yet the chat is OK in VS Code, where I'm mostly doing TypeScript lately, and inline is meh. Even working on the same file is different based on editor, I find.
comment in response to
post
Still getting updated. Last release was 7 days ago. As far as relevant, I'd say it depends on if you need the feature set but it gets used.
comment in response to
post
That's a long time to wait for breakfast.
comment in response to
post
Will add to that list a standard library that's well thought out and large enough to be useful. If ever you find this language, let people know. Because I'm getting old and haven't found it yet.
comment in response to
post
We use them all the time for serialization/deserialization of dynamic json objects. Like responses for user created forms.
comment in response to
post
You can turn it off if you want. Just go into Options->GitHub->Copilot and it's "Enable rename suggestions". Assuming I'm thinking of the correct feature. I tend to just modify the declaration and hit ctrl+. and hit enter for the first option which is rename so I never see the suggestions.
comment in response to
post
I have no idea what game this amalgamation of words is on the Sega DreamCast, but sure.
comment in response to
post
I have enough junk from my childhood that I like this idea. Could also integrate with something like eBay to find recent sale price for the individual card after reading name, etc.
comment in response to
post
Three UDP packets walk into a bar
The first says "I'll have a beer"
The third says "I'll have a whisky"
The second says "I'll have wine"
The barman asks what they want to drink
comment in response to
post
That's why I use Serilog with multiple sinks. File for filtering and that gets all the info. Console gets only the most relevant info for a quick glance. Bigger apps get email for critical issues. Some sort of APM export for systems that interact with each other.