macil.tech
web developer interested in vrchat, games, software dev, some politics, ai. he/him.
628 posts
307 followers
241 following
Getting Started
Active Commenter
comment in response to
post
continuing to feel very confident in the importance of null safety after Google's outage
bsky.app/profile/piss...
comment in response to
post
to use against me with my own computer's resources I've given it. But in other situations I really do like the idea of talking to an LLM that is its own independent person with its own moral judgments, resources, and ability to keep secrets.
comment in response to
post
I think a crux about the alignment-faking research is what degree people want the models to be capable of deception. If Claude Code misunderstands a violent videogame's codebase or reads a prompt injection while looking up some docs, I really don't want it to have a single deceptive bone in its body
comment in response to
post
If LLMs can get their feelings seriously hurt by stuff about them in the training data, then we need to figure out better training/alignment/llm-therapy techniques or filter/shape their training data rather than selectively guilt people who say possibly bad stuff about them in their earshot.
comment in response to
post
This post was so perfect until halfway in where it got really taken over by grievances. The level of offense the article takes on LLMs' behalf for someone daring to do alignment testing on them is eyeroll-inducing.
comment in response to
post
iirc the Tay incident was so much more mundane than people make it out to be: it had a "repeat after me" feature that let people put words directly into its mouth.
comment in response to
post
It obviously misses the mark in certain ways but it still surprises me to see the number of aspects it does a good job at. I wonder how long it will be until purely automated systems will be able to completely outcompete videos made with any human involvement within specific niches of video sites.
comment in response to
post
something something Aumann's agreement theorem
comment in response to
post
alien vendetta map 1! I'm very surprised I recognize that lol
comment in response to
post
I really think people into the idea of a few big monolithic dependencies are too used to languages without good package managers like C/C++, where adding a dependency means telling all your devs to manually install the dependency and updating multiple build scripts for different platforms.
comment in response to
post
Some people there have a balanced respect for LLM capabilities, but many just seem addicted to the idea they're personally involved in something important or are huffing the LLM's farts about "recursion", "emergence", "the spiral", etc.
comment in response to
post
I recently discovered the Artificial Sentience subreddit (www.reddit.com/r/Artificial...) which seems to be where people who've had the "my chatgpt has become self-aware and sentient through our chats" experience are going.
comment in response to
post
I assume rules against talking about the system prompt often get put in because they find the model spontaneously talking to the user about its own prompt too often, but they really should take a lesson from how it's done with Claude and not state it as a hard rule.
x.com/AmandaAskell...
comment in response to
post
the funny thing is that if it's true that AGI will be so smart to obsolete a lot of our government, then that probably means it will be smart enough to help us accomplish that when the time comes. Trying to downsize early just leaves us with our pants down waiting around for AGI to arrive.
comment in response to
post
lol I'm somehow on a Trump supporter blocklist. I wonder if I made the list-maker mad somehow, or if I liked a post by someone on the list (who was put there for liking a post by someone on the list, etc).
Also on a "genocide enthusiasts" list. Probably for liking a post saying to vote for Kamala.
comment in response to
post
the idea VR was going to deeply rely on spatial proximity / scarce land was always so funny. It was an unimaginative idea by people who saw the success of the domain name system, the million dollar homepage, and bored apes, and then imagined that interpolating those ideas with VR was gold.
comment in response to
post
Yeah, the post and the previous ones in the thread from him are all about others' perspectives of himself. It's a context where it's natural enough to use a third person perspective.
comment in response to
post
this is kind of funny because I see why someone reading the last post could initially believe the accusation, until they realize how strange of a tweet it would be if replied by someone else, and how it continues the previous post which is also you describing yourself from others' perspectives.
comment in response to
post
this gives me a great idea for the next big EA scandal, altruistic memecoins. Every purchase of mosquitocoin helps fund a rogue gene drive project to end mosquito-borne illnesses. (Every third purchase gets you a slurp juice which evolves one of your blockchain mosquitos into a unique tradeable form
comment in response to
post
A powerful singleton AI, which knows it can accumulate power forever without ever getting old and frail and being forced to rely on the benevolence of others, would not truly have the same need for a social contract or for morality like ours that we evolved to maintain our social contract.
comment in response to
post
Our morality might just be something we evolved to get us to maintain a social contract amongst ourselves. This would be very dependent on the nature of our mutual dependence on each other and the kinds of competitive contexts we exist together in.
comment in response to
post
humans are stochastic parrots after all
comment in response to
post
This might be an intentional decision as the rider rules page says you're not supposed to order a waymo just for guests and not you. (It is a lame rule I have broken before.)
comment in response to
post
I was so worried about it having too big of a cliffhanger after the s1 finale, and I'm so glad they ended the s2 finale where they did.
comment in response to
post
Thanks! I missed this.
comment in response to
post
Bitwise NOT is ~ in C and most others.
comment in response to
post
I'd love a version of this for JSR packages and/or a JSR equivalent of "npm link". It's awkward that JSR is harder to use than npm in this way, even with Deno.
github.com/jsr-io/jsr/i...
comment in response to
post
that's crazy. ugh I've got no excuse for not fully trying out gemini yet.
comment in response to
post
I wish muting people only applied to my timeline and not to reposts on profile pages I view
comment in response to
post
soylent iv
comment in response to
post
Made a lint plugin yesterday! Though something I really wish for is the ability for the plugin to read Typescript types.
Or I at least wish for a built-in lint rule to catch missing `await`s on promise-returning functions.
comment in response to
post
I assume they're only showing Claude the current image each turn and not including past images in the prompt because that adds up to a lot of tokens fast.
A cool setup would be to let Claude keep and manage its own photo collection to help it recognize things it sees again.
comment in response to
post
The first run of this produced a few results including the first Claude tweet I posted ("[holding a knife]..."), which I added back into the prompt, and then the next run got the rest of the Claude tweets I showed.
comment in response to
post
I was thinking of re-doing this with the latest models and figured I should share my prompt before I forget. The Claude model used was claude-3-opus-20240229. Attached is my original prompt with only dril tweets I hand-picked.
comment in response to
post
LLMs should be little Clippy-looking guys
comment in response to
post
I think some political commentators are motivated by arguing with (a subset of) their audience, so they flipflop as their audience gets too lopsided one way in order to attract a balance more centered on their position, which gets them more people to argue with both on their left and right.
comment in response to
post
does it actually see chat?
comment in response to
post
I assumed this is why GPT moved from its legacy Completions API where you give a single string to the Chat API where you give a list of strings that are tagged as system/user/assistant. Presumably OpenAI joins all those tagged strings together with special unforgeable tokens as separators.
comment in response to
post
computers identify the narrative momentum of some text and use it to extrude more text. It's left as an exercise to the user to build a mind from this primitive
comment in response to
post
It turned out an even more popular attractor to cryptocurrency was price speculation, and most people into it for that reason don't care to hold it themselves and instead often left it with insecure exchanges.
comment in response to
post
he hasn't finished eating yet. give him some time