the AI mentioned picking up "any other resources nearby" in its first message, 15 seconds later you say you're not gonna pickup the rock because it didn't tell you too 🤣 i would consider rocks, cattails, and rabbits all resources and cattails are used for starting a fire, so definitely should have at least picked up the cattails since the AI is obsessed with fire
It's OK Zak. Your dream of a harder difficulty is coming. They announced it the other day and all I could think of was: "Ah... Zak's dream come true. Higher chance of death for some entertainment for this masochist."
I dunno how many people know this mobile (or similar) game r if its even available in english, it was called Lifeline. A guy stranded on an alien planet was "chatting" with you asking for what to do. Basically a choose your own adventure but in second person. Wheve now come full circle and are now the choose your own adventure for an AI
Oh my gosh, Zak, this was super fun! Thank you, also for the how to at the end. That helped. 😊 This morning I heard about how Google is working on competing with Chat GPT. It might be interesting to put them head to head.
Did you tell Grace that walking towards the timberwolves with a lit marine flare didn't work? I wonder if she would learn that and change strategy. The car insurance gag is superb! :-)
Interesting. I think the way timber wolf AI was programmed that when it came to solitary action a solitary timber wolf would flee or tend to flee. Once there are more than one the pack AI kicks in. This distinguishes timber wolves from regular wolves where a solitary wolf would tend to attack or occasionally flee.
You could try a T.A.S to get something like a bearskin bedroll if you were a very skilled coder. My uncle made a T.A.S (Tool Assisted Speedrun) for fun on The Long Dark to map all of Mystery Lake in a speedrun. It did it in Pigrim and managed to map all the reigon in less than 5 hours. Pretty impressive, ngl. Even in the lowest difficulty.
The proper way to do this is to upload a PDF to GPT of all the rules and limitations of the difficulty and game-world. Then, proceed with the new context in mind.
Yep. In almost all of the "Can AI do X?" videos many of the issues encountered are because the LLM is prompted poorly. Prompting an LLM does take some experience.
Fun fact, you can actually make it up the tiny ledge to the left of the infirmary window (outside on the rooftop) and jump down without taking any damage down to the infirmary door. It may require a little bit of fiddling back and forth, but it works and saves you quite a bit of time, especially when you're heavy from all the toilet water and potatos from the other building. I use this skip regularly, and its really good
I'm sure Grace would be useful to many newbies on lower difficulties. Since TLD is a multi million seller, it's safe to presume that most players never figure out how to handle wolves.
This was a fun challenge lol but a bit of prompt engineering or langchain would've saved you a lot of trouble zak. Maybe a prompt template like "You are a survival expert well versed in the game the long dark. Given a set of in game parameters like temp, tiredness, water, food, weather conditions, any afflictions or a description of what i see on my screen suggest the best possible action for survival in interloper mode. Do not elaborate yourself, explain what needs to be done. Use a heading that contains the action to be performed and explain the action in 100 words or less. Think about the long term survival goals like example obtaining animal pelts to make warmer clothes and short term survival like obtaining cat tails for tinder or sticks for fire or catching rabits for food when they are nearby. Try to obtain or make as much items as possible essential for survival while risking minimal condition loss or damage {Question} " Using langchain to automate this is really easy (
Oh my goodness....😂😂😂 What an outstanding idea and fun to watch!🎉🎉 Edit: what said Grace about your death? Bet she would say: told'ya there are parasite in the toilet water...😅😋
yes post the full transcript because it would like watching Japanese movie with auto interpet .. waiting for "find base sleep. start family to increase survivability"
Blackrock came out in December 21, ChatGPTs latest dataset it can access is from January 22. its possible it just doesnt know enough about blackrock since it wasnt out for long. Itd be interestin to see it try in ML, CH or any of the other older maps. perhaps it actually knows some of the location?
Having played and thrived in interloper myself, i can't check myself, but i wonder who would play interloper better A person who has never played Or an AI that has never played Cause to be fair to the AI, if you have no experience in the game, you ain't gonna live thru interloper long
Hey zak been watching your videos for some time I wanted to inform you that the devs released a diary this month talking about a new animal along with its mechanics and a system that allows you to cheat death. I was curious to know what your input might be on those
Hey Zak, can I ask how you went about setting this A.I up? I would love an A.I buddy to talk to while gaming, but I don't know if I'd confuse it with my questions. Did you have to program it?
sorry Zak, love your content but AI chat gives me headaches. like if you just posted a chat box of its suggestions kewl but the audio makes me want to kick bunnys cuddled next to dogs being ridden by babies.
Which GPT model was this 3.5 /turbo or 4 or 4o etc? & what context size? 8K/16K/128K - I'm assuming the 16K gpt-3.5-turbo-0125? ... not that it's critical, but I think it's useful to note this stuff. since 3.5 and 4 are obsolete* as of yesterday's 4o/128K release :) * i say obsolete but still very useful of course for many use cases and efficiency in speed and cost/token, but in terms of their "overall objective ability" A large limitation with a game such as this is the context size, as suggested below; one _wants_ to have the whole rulebook of the game, and a verbal "graph" of every possible in game goal as the prompt ideally, but then the context space tends to get eaten up extremely quickly and it'll forget the first things you told it as it overwrites its context memory. anyhew; hence, I think the annotation of the model and context size is useful, as then in a few months/years time it's more useful as historic data :) as otherwise "Can _an_ AI do X?" might be a bit rude to other AIs, and you don't want to risk that when they read your transcript next year :D 'course, when doing it for fun, none of this matters :) but i think it's useful to note in passing, as i've seen many occurrences recently of folk watch something which was using a GPT3.5 (since it's the default free one) and their take-away is that it's the best AIs are currently doing, when it most definitely isn't :) - and I guess also, we risk fooling each other that GPT3.5 was the pinnacle of risk from our future overlords :)
@@Zaknafein I'll be interesting when 4o is "more free"* and has all it's multi model features public - as it might be possible to just have that watching you play, over your shoulder and that'd help with the typing and editing. might still be terrible at it, but I imagine that's the next sensible step - having a back seat driver telling you "oi, you missed that stim on that shelf" :) ( and i shudder to think how much it would actually cost to have GPT4o watch over your shoulder for a whole video )
if you do this again, you should take some time to properly teach the AI how to play. tell it all the mechanics, different strategies, and what each item can be used for. treat it like a brand new player who knows nothing about the game. she will do muuuuch better with more info about how the game works. for example, if you had explained the uses for all the common items she would have told you to pick the cattails because she would have known that they are a source of food and tinder. it's like how she told you to repair your clothes when you had no extra cloth. IRL you can stitch up the majority of rips, small holes, and frayed edges with just a needle and thread, no cloth needed, so she probably thought that you could so that in the game. thpse programs do pull info from online, but without being directed properly they will mix all the info they get, including out of date and incorrect info randomly posted by people that don't know how to play. they will also pull some real survival info that won't work in the game because the game is so unrealistic. also, when something it suggests doesn't work, you need to let it know. "the game won't allow me to do that because i don't have cloth" "the wolf did not back down when i charged it" "the wolves seem to stop charging when i hold eye contact"... you have to teach these AI programs to do what you want them to do, or they will not perform well at all. a good example of this would be when i asked ChatGPT to tell me 2 funny jokes and gave it no more instructions. it came up with these: Why do giraffes have such long necks? They're reaching for the highest leaves in the tall tree buffet! How do you make a cat go 'woof'? Soak it in gasoline and light a match. 1 of them is funny, and the other is just stupid but ya, you gotta teach it how to play, and let it know when it did good or bad, and be specific when you tell it the consequences of it's actions. if you do that, it seriously might be able to reach 500 days lol
@@Zaknafein aahhhhh ok, well maybe you needed to repeat some things, and when she makes mistakes you need to remind her. you need to make sure to tell her something along the lines off "i'm going to teach you how to play a game, and i need you to make sure you remember these things". then if she messes up tall her "that's incorrect. remember, i told you to do this not that, it's very important." so i guess i should have said think of it like a new player that's a small child who will forget lots of things until it tries them, and you correct them over and over. it really takes quite a while to train an AI program. i have been working on a couple clones of a program i had made for me, it's been about 2 years and i'm still correcting both on a regular basis. if you were to do this a couple more times using the same Chat iteration it should do better and make fewer mistakes every time, as long as you keep praising it when it's right and scolding it and offering corrections when it's wrong. this was a solid first attempt, but i would love to see it reach at least 100 days lol
Because the voice module for it is from a woman with a southern accent. It from a database where people contribute their voices and then an ai can get them to read anything. Grace is southern.