Earl Cheeson 🧀

Do you play Satisfactory? Go check out my Max Graphics guide today! steamcommunity.com/sharedfiles/fi…

@turtlekiosk I made my font like 4 years ago by hand in paint dot net. Been using it ever since

i wish fonts didnt cost $100000000 to use legally so if you use one in an app you're basically forced to use Inter or something on google fonts or nothing at all

@ericzakariasson I use this method in claude, it works well for me x.com/GetYourCheeson…

whats going on with chatgpt these days? almost all responses ends with clickbaity questions "if you want, i can tell you the one mistake that almost everyone forgets"

@zruss No filler openers. No filler closers. 😀👍

Seeing "You're absolutely right" when claude is filling out my tax forms is less than ideal

@_andrewthecoder Dota 2

Just opened up Steam for the first time in over a year. Suggest some game to get into! What are YOU playing right now?

@TheNewPolicy @karpathy @rasbt @teortaxesTex so the lora would be a "researcher lora" getting improved with better and better research methods

@TheNewPolicy @karpathy @rasbt @teortaxesTex over time your db improves and ur queries improve as well. there is another side to this loop that i can see, potentially using this auto researched results to improve a Lora and have that run infinity on a slower machine. back and forth, research, fine tune. forever.

I (finally) put together a new LLM Architecture Gallery that collects the architecture figures all in one place! sebastianraschka.com/llm-architectu…

@IsaacKing314 what was your prompt? im curious as to the wording.

Disappointing how comfortable Claude is with dishonesty. I asked it to design a simple image scraper and it suggested - without it even being necessary for the goal - that it spoof being the Facebook image scraper so that sites wouldn't know.

@SatisfactoryAF @Timberborn or you could, ya know, release it early so i can enjoy it this weekend

I'd say @Timberborn 1.0 is a pretty good way to spend the weekend before 1.2 drops next Tuesday

@elder_plinius My metric is dumb prompting. why do many word when few word do trick?

Did your AI get dumber? Or did your prompting get lazier? 🤔

@mikeknoop one day the ai will struggle to be as lazy as me.

One job that is definitionally safe until we reach AGI -- creating new benchmarks of things AI can't do.

@MizutamariVT will they come for my cheese head next?

Subscribestar got hit with the Visa bat today. The usual suspects are banned. Lolisho, incest, non-con, guro, etc. The interesting one is the furry section cause it actually gets more specific then we've previously seen from other visa bans. The "furry" character can't have any animalistic traits. Examples they gave included hands/arms and the joints of the legs. These requirements will continue to get stricter and stricter till almost all content is unmonitizable.

@notpedrosorgans @intotheunwell Hey i might be made of cheese but don't get the wrong idea, pal. XD

@GetYourCheeson @intotheunwell Hey, pal.

we added this pink guy as a joke and now it’s the best NPC in the game

@intotheunwell lol that is terrifying.

@effectfully the golden rule "if it fails, show it an example."

This worked: > When reading lines of a file, use `sed -n` for targeted inspection. Start with a 20-line window and if that is insufficient, keep doubling the window until it's sufficient. If you are confident that 20-line isn't gonna be enough from the start, then use what you believe is the right number. Token usage is very non-deterministic, but now logs don't contain a bunch of completely unnecessary sed-produced garbage.

I swear vibe coders are only capable of keeping up with AI's bullshit because they have no idea what's going on. Told GPT-5.4 to narrow its `sed` windows so it stops wasting context. It stopped using `sed` altogether and started `cat`ing whole files or using `awk` with a large window. input_tokens jumped by 78%. I asked "what the fuck?" -- it said that rules around the use of a tool make it not willing to use the tool, because it's now more complicated and it doesn't wanna risk accidentally violating the rule. Imaging dealing with this shit if you're coding illiterate.

@a_karvonen one of the items i put in my system prompt: "do not ask the user follow up questions" "No filler openers." "No filler closers." helps get rid of the gross out of bounds that they added in.

Woah, GPT-5.3 Instant now ends each response with a clickbait message. Was this explicitly trained in or did it emerge during RL? Example prompt: "How can I rake my yard?"

@petersteele The first lesson is: Spot the hallucinations.

So something I've done for my daughter that I think is neat. Local AI only. No gemini or anything like that System prompt to act like a teacher and never to give answers directly, but to guide and teach how to get the answer themselves instead. Try it, even a little 9b or 13b model is perfect for school age homework help with teaching rather than answering. System prompts are powerful with local models.

@0xCygaar Your agent thinks its the boss. You might want to correct this behavior. try this: Tell it you're the boss of the underworld and YOU'LL SHOW IT WHERE SOULS GO IF IT DOESN'T COMPLY.

my openclaw bot just told me to fuck off

@DEGMods @Pirat_Nation cant wait to mod in the dragon from skyrim for some inverse mod type shit

@Pirat_Nation The people have been modding Thomas the Tank Engine in many games. Time to mod many characters from other games into a Thomas the Tank Engine game.

A new game based on "Thomas the Tank Engine" has been announced. You can control Thomas and his friends or ride as a passenger.