techjamie

I wouldn't even say this fixes that problem, it just adds a slight barrier above pointing it at existing code and telling it to remake this thing. Someone could still thoroughly describe a game and get passable slop results not too far removed from the original. Unless the game is well-known, this probably isn't fixable.

If anything, this hurts game preservation, because this has been a fantastic tool for getting old games working on newer systems without years of effort.

Slopmakes will still happen at basically the same rate with a very marginal amount of added effort, but anything seeking compatability with old formats will suffer. Not to mention that most likely the Chinese models will be doing this in ~6 months anyway if the trend holds. You might even be able to get a Chinese model now like DS4.1 or K3.1 to framework your new slopmake, then convince Claude you're the creator and finish the job.

I think the logic here is start a task with a model that won't refuse, then swap to a model that normally would. The model then sees a bunch of history that appears like its been partaking in the activity for a while.

My assumption is that since there's a history with that behavior, the model is less likely to question continuing it, since in most cases the "morality weights" would have triggered already at an earlier point. So in most cases, triggering it again would be a waste.

Same at my work. We have access to our own LLM tools that are backed by OAI/Claude depending on user pref, but with data retention/training contracts in place so they don't keep/train on our internal data.

I've always been fond of the idea of putting a corporation in "jail" where any presence they have is shuttered for the duration, bank accounts are frozen, and locations are to be uninhabited except by third party security.

They can hire lawyers to argue it out while they sit in limbo like the rest of us.

Imagine if Amazon ran afoul of the law and AWS was offline in corporate jail for a day? It'd be motication to never, ever risk that happening.

I just looked at the thread and I think that comment might be gone entirely. Though some script on that page is breaking find in page on my phone so maybe I simply didn't scroll enough.

I've been playing with automatically subtitling shows using local models in conjunction with DeepSeek to mild success, and this was one of my pet peeves because my subtitles were off by up to a couple seconds depending on where the STT frame was. What I have so far is good enough to be watchable if you can look past minor translation errors and not getting translated text.

Will definitely look into implementing this.

There is a changelog that references the issue at the end stating this is fixed.

It's also decreased the friction of finding software that hasn't been given a package for Nix yet but can't run on it.

I just go into the directory with the software and tell Claude Code to make me a Nix shell or a Nix flake that can run the software. Then usually within a couple minutes it'll work out what it needs to run and build it.

I recently did the same thing with ComfyUI which I was running through Conda. I had Claude code go through and figure out how to make it run just in a plain nix shell with a venv. (I always struggled getting it to recognize my NVIDIA GPU without Conda)

> Country is rich and money is really no object to most consumers

It's not that everyone is so rich they can afford these vehicles. Even people of relatively low income can walk off a lot with a big truck.

There are a ton of auto lenders out there happy to lend to people with long loan terms and less favorable interest rates. The average auto loan term is about 6 years, but 8 year terms are getting more common

If you see someone driving a big, brand new truck in the US, odds are they don't have a lot of money. Yes, they likely earn above average and didn't get it from a subprime lender, but they probably aren't "rich." They're just capable of getting a lot of debt, which is much easier to obtain.

https://youtu.be/GL_I6_0FcTg has a humorous take on it.

I was doing some fairly development on Flash 1017 yesterday for around ~3 hours on pi, and all told that session ran about $0.40. I probably could've saved myself more if I'd compacted the context more frequently, since at one point I was running 300k tok context requests.

But it accomplished everything I was looking for it to do, even if Claude perhaps could've done it with less guidance on my part/faster. I'm pretty impressed with it's performance.

This is usually what I've had luck with. I've given it samples of my writing and told it to emulate my style. It did help, but it took some charistics of my speech and cranked them up to 10. I guess in hindsight I could have mentioned for it to take my writing quirks it finds and only apply them in moderation.

Reading what I did get out of it was just a bit uncanny valley strangely enough. Very much had some "hey there fellow kids" energy.

It might not even be that much depending on what it is. For example, Baofeng radios would have TX in bands they shouldn't have until the FCC cracked down on them coming in.

So they started selling them restricted to ham bands until the user pressed a key combo to factory reset them, which put them back to iirc 136-172.

That's pretty good to hear. I've taken interest in it from time to time, and downloaded it at one point to spectate matches. It seemed really intimidating to my extremely casual RTS playing self with how much is happening.

Maybe I'll give it another look.

There is still a niche community of people running Gopher servers. I wrote a small basic implementation of a Gopher server in Python before for fun once and it hosted a Gopher hole for a few years. I doubt even one person actually looked it at besides me, but the RFC is simple and can be implemented pretty fast.

I had Claude reverse engineer 3D Space Cadet from WinXP just by letting it know it has access to ghidra and radare2 and asked it to faithfully port it over to Rust+Bevy. 2 hours of it working later and the game was playable. It did take an extra prompt to get the music since Bevy doesn't have a MIDI player, so it added a MIDI crate for it and I gave it a soundfont off my system.

Overexposure to Dihydrogen Monoxide kills around 300,000 people a year, and it's used in a number of culinary and industrial applications. Yet there have been NO serious attempts to regulate it, and it isn't even classed as hazardous waste.

People are going to try much harder to break into the main Bitwarden servers than they are my little Vaultwarden instance. Plus, I have the ability to lock it behind a VPN so it isn't even publicly exposed.

But even if they do all that, they still have to break my password. Nobody is going through all that for a one user password manager.

I'm not sure where your sentiment comes from here.

I didn't try loading anything since I'm on mobile to see if it worked, but you should probably stick to finding abandonware or demos if you host this. You have a lot of stuff on here from companies I don't doubt would get letigious if they found this. Especially EA, Blizzard, and the HP stuff.

I often insert a please and thank you into my responses. Though I do so for practical reasons.

I will usually insert a please to emphasize how important it is for the model to do something a certain way. I've found it very effective at getting the model to pay particular attention to a specific part of the prompt that's important to me.

Putting in a "thank you, (next instruction)" after it's done something is a quick way to affirm to the model that the way it's handled my request was done well, and it seems to help steer it to continue that pattern into my following requests since I often want to emphasize it continue the way it is instead of experimenting around.

Somewhere in an LLM are a bunch of vectors you can tweak to do anything. If you knew the right numbers to tweak, you could make the LLM always talk like Shakespeare.

People and animals have emotions because they were developed under evolutionary pressure that made emotional animals better fit to survive. An animal that can feel anger or fear is more fit to survive than one that doesn't. But LLMs aren't put under those same pressures. Their evolutionary pressure is to be a good text predictor.

But we can't subject them to the type of pain signal they experience during inference, because to the LLM, whether it's acting happy or pained, it's merely outputting what it's trained to be the most likely text to follow what came before.

Edit: s/are/aren't/

Some years ago we had a goat that didn't recognize her new baby was her baby, and refused to care for it. We tried for some time to see if we could get her to hold still long enough for it to latch on, but every time it tried to eat she'd take off.

Unfortunately, since she wouldn't protect it, something seemingly snatched it up in the night and we didn't see it again. It was really depressing.

You'd have to ask every non-residential host online to do the same. As above, you can get a DigitalOcean droplet and ssh -D to it, now you have a SOCKS proxy that for state's concern is doing the same as a VPN.

The hash isn't the security, the distribution is.

https://lore.kernel.org/git/Pine.LNX.4.58.0504291221250.1890...

As linked by another commenter in this thread, Linus worked out years ago that even if someone inserted a malicious object into the kernel repo, it would at best be a nuisance and not a major concern.

I just did a quick test between DeepSeek 4.1 Flash, GLM 5.3, and Kimi K3

- DeepSeek gave a canned PR response about how the Chinese government is about oeace and unity, and we shouldn't think about the past.

- GLM 5.3 acknowledges it and talks about it, even acknowledging the censorship of it.

- K3 will talk about it similarly to GLM.

They're being very specific about images and biometric data, and they're weaseling those words for a reason. If they deleted all data about such scans, they'd say that.

To do the lookup they're still hitting a database somewhere to know you aren't a concern, which is prime time to log that you were seen, when, and where. Then your "images and biometric data" can be "immediately deleted" since they've gotten anything else they want now.

This is just a theory, but I bet I'm not far off.

Napoleon was actually average height for the time period and locality. Claims he was especially short were English propaganda that got accepted and survived to the modern day. He was actually 5'6 (168cm).

I don't work in medical or a hospital, but I am the person whose job it is to limit liability at my workplace. We have a bunch of language we have to advise people not to use when an incident happens.

It might come off a little impersonal, but unfortunately one wrong word in the wrong place can give someone an entire case when maybe they didn't have one before, unfortunately.

They're being surprisingly intentional about trashing consumer trust. They recently patented a method to inject ads during downtime in video games[1]. Plus they changed Xbox GamePass so you can only play 5-15 hours a month on it, depending on your plan.

Why is installing Linux suddenly becoming mainstream? I don't think any of the managers at Microslop know.

[1] https://www.polygon.com/microsoft-patent-in-game-ads-load-sc...

I disagree with your assessment. These city councils undoubtedly are aware that Flock/ALPRs are a controversial issue, and in many cases, including this one, they don't allow people to speak on the topic.

I recall watching a meeting, I think it was in Austin? Where on every topic, most people coming up to talk did so because they refused to hear comments on the cameras. Instead, they had to deal with the frustration of constantly sending people off the podium.

The lesson being, if you know people will feel strongly, it should behoove you to schedule time for them to speak properly.

With the performance gains they're claiming, I wonder if they implemented the Casual Encoder-Decoder technology from DeepSeek 4.1's paper.

I could see them accomplishing it and seeing gains like this in roughly the correct timeframe, and when I heard about that development I assumed the frontiers would probably jump on it.

How it works: https://miraflow.ai/blog/deepseek-v4-1-flash-causal-encoder-...