HNHacker News
TopNewBestAskShowJobs

theaiquestion

85 karma · joined March 31, 2023

submissionscomments
theaiquestion··on Bambu Lab A1 Mini
There's a ton of different methods of multicolor printing each with different pros and cons - simplified

* nozzle that mixes from multiple inputs * multiple hotends on the carriage * swappable hotends * single hotend - retracts and switchs to a different filament, few methods here, this is what the AMS and pallete are **

You can do it with multiple heads, the problem is ooze. You'll need to either be able to retract the non-used printhead so that it doesnt ooze onto your print, or be able to swap it mid print. Either way here you'll still need a method to prime it. Some materials may break down in the heat of the hotend too while they sit there unused just baking.

There's some methods that make the blocks smaller, like using the infill as the priming section.

theaiquestion··on I mirrored all the code from PyPI to GitHub and analysed it
Because most company's/people are (AFAIK) under the impression that training on public data will fall under "Fair Use" because it's substantially transformative, in the case that it isn't then you've already agreed to it on github.

It's a fallback clause, "fair use" is irrelevant if you've already given github permission to use it. By adding that clause you can no longer argue that it's not fair use to use the code you put on github after agreeing to their terms.

theaiquestion··on An image preprocessing tool to protect artworks from AI-for-Art based mimicry
Funny enough the first one I've found that works consistently with this one(however unusably destructive) is the "whirl" filter in GIMP.

Previous adversarial methods I've beat with simple noise/blur. This one is fairly resistant to basic filters.

theaiquestion··on An image preprocessing tool to protect artworks from AI-for-Art based mimicry
Just for context LAOIN's largest dataset is 5.8 Billion images

And you'd be spending that compute time for what will be a tiny, tiny,portion of that

theaiquestion··on I only lost 10 minutes of data, thanks to ZFS
Did you switch to a mac or to a windows machine?
theaiquestion··on Show HN: Lottielab – Create product animations in the browser easily
Yeah I was in school and wanted to try making games - but that was the exact year that "flash is dead - HTML5 is going to replace it"

And.... Yeah HTML gave us the canvas? No editor, no animations, no input.

The real innovations were with ES5/2015. It utterly destroyed the "mess around and make some basic games".

It's utterly laughable that "html5" was somehow the "alternative". We went from an editor that anyone can quickly create animations, games, etc to nothing - for over 10 years! And yeah to my knowledge we're _still_ not back to what we used to have.

theaiquestion··on The Death of Infosec Twitter
Has any of the OSINT/INFOSEC crowd moved to mastodon or somewhere other then twitter?
theaiquestion··on fMRI-to-image with contrastive learning and diffusion priors
I think the method of merging the pipelines via img2img should use controlnet. Possibly needing to be finetuned specifically for this, although existing controlnet models might work fine for this.

This is exactly what you'd want to use controlnet for - mapping semantic information onto the perceived structure.

theaiquestion··on Kata Containers: Virtual Machines that feel and perform like containers
Depending on your use case there's potentially negligible startup time. On the scale of single digit seconds to less then half a second depending on how much work you put into optimizing it. For some applications this will be too slow (mainly the type where you boot a container per request, although flyio seems to make it work), I think for a _lot_ of applications this wouldn't be noticed.

Kata gives you a few different options for what/how you'd like to boot including firecracker.

This isn't exclusive to firecracker but if you stay lightweight you can have vm's booting under a half second if you're using slim images.

https://jvns.ca/blog/2021/01/23/firecracker--start-a-vm-in-l...

I honestly think for a lot of people, vm's with the convenience/orchestration tools of containers make more sense for a lot of general use cases simply because of the security benefits. The convenience still needs some work though.

theaiquestion··on Congress is racing to regulate AI. Silicon Valley is eager to teach them how
Delaying public usage by decades?
theaiquestion··on Llama.cpp: Full CUDA GPU Acceleration
Compete with Llama.cpp? Like transformers llama [0], exllama [1] (really fast), or litllama [2] ?

exllama is really memory efficient and really fast

[0] https://huggingface.co/docs/transformers/main/model_doc/llam...

[1] https://github.com/turboderp/exllama

[2] https://github.com/Lightning-AI/lit-llama

EDIT: Or do you mean cuda? Because yeah, it's such a shame AMD's Rocm is so bad even geohot gave up. it's examples don't even run without crashing.

https://github.com/RadeonOpenCompute/ROCm/issues/2198#issuec...

theaiquestion··on StyleDrop: Text-to-Image Generation in Any Style
I don't mean to be rude, and to be clear I wish they did release the weights, but what did they lose here?

You don't approve - so what? Releasing the weights doesn't make them money.

I'm skeptical LLaMa is even useful for facebook commercially at least, they don't make money on it, and I doubt anyone developed brand loyalty, more then likely everyone will use whatever the next, best open model is regardless of who makes it.

Llama and SD still don't come close to midjourney/chatgpt/claude when you look at ease of use and infrastructure cost. These "99% the performance of chatgpt" are laughable if you use them (which I have extensively).

> We are done inhaling vapor.

Okay? What were you about to pay for to begin with here?

EDIT: Just to add, it's not like we got nothing from this, this can likely still be something to try with SD.

theaiquestion··on At last, the Raspberry Pi shortage is finally coming to an end
> How. Many. Times. Have we heard it already?

AFAIK what they've been saying so far has at best been "we're able to continue making them at our current pace". They've also been saying they were focusing on commercial partners over consumers (which IMO is better because commercial customers need some level of assurance of future stock to actually use pi's in their designs).

Neither of those things particularly screams "Regular customers will soon be getting pi's in their hands"

December 2022 They said they expected better stock this year, which appears to be accurate.

I'm more curious to see if they can manage to make a PI 4 successor given how many years the pi 4 has been around already.

theaiquestion··on StyleDrop: Text-to-Image Generation in Any Style
I feel like people don't understand what this is, and what google has historically done.

This is research for some math to improve the ability to do a style transfer, and they show it on their own text-to-image generator muse, which they have also published the structure of.

This is what they have typically done, and this is what they didn't do with Bard.

No, they did not release waits for it, you cannot run this on your own computer. But they typically didn't release weights for things like Lambda or Imagen either AFAIK.

This is not a product. This is not a tool for you to use. This is for researchers.

The point of this paper is not to let you run it on your computer. It's to allow other researchers to implement and build on the methods described in the paper.

theaiquestion··on HBO Max new Captcha system
Sorry I should clarify, I don't mean append-only graphs,

I mean what people call "blockchain" in the cryptocurrency sense as actual projects - there's so much stigma largely because the motivation of most of the projects appears to be "making money/investing" and not actually solving a technical problem appropriately.

If github was like this there would be a "fee" for making making commits, this fee would be paid in some proprietary coin, initially created with an ICO/airdrop. Suddenly the motivation is holding these coins because developers will need to make commits right? And the more developers that make commits the more the coin is worth, so surely you should buy and hold them right? This will be a feedback loop of endless money! Oh and it'll be a DAO so the more coins the more voting power you get too!

^ This is what I mean, where the focus is on collecting some "coin/token" - this leads to both a lack of focus on the actual problem being solved, and the problem of people associating it with a ponzi scheme.

I'm not picking a fight with distributed graphs themselves, I don't like it when they're tightly coupled with "value" that can be traded as a fiat.

theaiquestion··on Nvidia Announces DGX GH200 AI Supercomputer
Well if you want to compare them, Nvidia's has over 10x the memory and has historically sold > 1 units of their HPC products outside their company.

I'm not sure the aesthetic of what are essentially cabinets in a warehouse is that much of a consideration with HPC?

theaiquestion··on Removing support for forwarded ports
from the websites themselves actually.

with a VPS as your vpn, reddit/google/facebook/whatever all see you from a single ip, one that might change even less then your ISP's, all of your alternative accounts will all share this ip as well, and 0 other people use that ip address. basically data collection and alternative account identification becomes dead simple because your ip is basically your universal id.

you stand out as an individual, part of the "security" with things like mullvad is that you share that ip

security in depth of course, fingerprinting and stuff still exist, but if you have such a clearly unique ip address you have 0 chance

theaiquestion··on HBO Max new Captcha system
I'm not saying PGP or cryptocurrency because both of those have issues and the moment money is involved everything is foobar'd

But essentially allowing people to make "identities" via cryptography and then use a reputation system. Preferably by allowing people to follow/whitelist/favorite people across websites.

I like hacker new's method of making new people green. And I wish I could make it highlight the big names I recognize.

The problem with this is that nobody has figured out the distribution system for how we communicate the keys - IMO blockchains are the closest but it's so difficult to mention them because 98% of them are money-grabs. PGP/GPG has struggled so hard pypi literally removed support for it.

The second problem is that what will likely happen is sites like twitter will only allow very trusted accounts and never allow new ones - effectively locking you into one account.

theaiquestion··on Show HN: Smallville – Create generative agents for simulations and games
You might consider supporting ooba's api which would give you a lot of support for different things really quickly.

https://github.com/oobabooga/text-generation-webui/

theaiquestion··on Show HN: Smallville – Create generative agents for simulations and games
> it's still not very practical to require a dedicated 24GB GPU

totally agreed, you could get away with 12GB too which is in the midrange.

That said yeah it's still not something you could make a game with yet, I'm just pointing out 300GB+ of VRAM isn't the bar for entry here, it is reachable for medium-high end consumers but that's not really including the games resources either, and most gamers aren't medium-high end so...

theaiquestion··on Show HN: Smallville – Create generative agents for simulations and games
> most models I've seen are 300GB+ and require significant computational resources to operate (think several $15k NVIDIA A100 compute nodes).

What? Where have you been the last 3 months?

> the quality of the responses from the model are correlated with how large (and therefore how much compute) the model has

There's a lot more to this including the model structure, training methods, number of training tokens, quality of training data, etc.

I'm not at all saying that Vicuna/Alpaca/SuperCOT/Other llama based models are as good as GPT3.5 - but they should be capable of this, they still create coherent answers.

You need preferably 24GB of vram, but you can get away with less, or you can use system memory (although that'll be slow).

There is a openai api proxy that might let this work without too much work actually

EDIT: It actually says in the readme they plan to support StableLM which is interesting because at least at the moment that's not a well performing model

EDIT 2: You should try the replit2.8B model - This is surprisingly good at programming - https://huggingface.co/spaces/replit/replit-code-v1-3b-demo

theaiquestion··on StarCoder and StarCoderBase: 15.5B parameter models with 8K context length
> All the code generation tools, StarCoder included, still have hallucinations.

This also includes humans. We "hallucinate" in very similar ways. For example mistaking localhost:8080 for localhost:8008 in a large config file. Attempting to use methods that were deprecated and no longer exist, etc.

IMO there's two ways to prevent this is - one is to make better performing models (architecture/training data/training amount/etc)

The other is the exact same as humans. Compile time tools that let it know immediately if it hallucinated, types, linting, tests, etc.

You just do it as a loop the exact same as a human. You write code, the compiler tells you that method doesn't exist, you adjust your code/consult the documents (also doable with agents).

theaiquestion··on Run Llama 13B with a 6GB graphics card
> It’s slow, but if I ask it to write a Haiku it’s slow on the order of “go brew some coffee and come back in 10 minutes” and does it very well. Running it overnight on something like “summarize an analysis of topic X it does a reasonable job.

I'm sorry but that's unusably slow, even GPT-4 can take a retry or a prompt to fix certain type of issues. My experience is the open options require a lot more attempts/manual prompt tuning.

I can't think of a single workload where that is usable. That said once consumer GPUs are involved it does become usable

theaiquestion··on I’m a student – you have no idea how much we are using ChatGPT
> Well, seeing the run-on sentence, I'm not surprised.

I wish you'd visit the link instead of responding to the HN title. If you click the link you'll see the actual title is

> I’m a Student. You Have No Idea How Much We’re Using ChatGPT.

subtext:

> No professor or software could ever pick up on it.

theaiquestion··on I won’t spend more than $250 on a new smartphone
I totally understand the financial benefits of "mid-tier" stuff.

We're all familiar with how "low-tier" stuff can cause more trouble then it's worth or even cost more money then if you bought something better.

But I've found the quality of life and other benefits of buying the "high-tier" stuff is sometimes incredible. I bought a 1080ti which was effectively the highest consumer GPU and that lasted 5+ years while have more memory then even most modern GPUs (11GB). I've found the noise cancelling quality at the 350$ tier headphones to be miiiiiiiiiiiiles above the 150$ pair and the 75$ tier (I own three tiers of the sony ones from a few years ago).

Of course just because it's expensive doesn't mean it's better value, or even better in general, you have to be really strategic and honestly sometimes lucky, and obviously most people can't afford the "high-tier" of everything they want. But somethings I'm just really happy I went with the "high-tier".

Obviously if you're just want to watch videos and casually listen to music 20$ earbuds might be all you need. I will say few 20$ earbuds will be rated as "quality audio" even by non-audiophiles, but if that's all you need then that's great!

If you need decent quality audio and high quality noise cancelling for your work in an office and plan to use it for years, then you might consider the "high-end" because it really can make a massive difference depending on your needs.

theaiquestion··on Effect of LSD on reinforcement learning in humans (2022)
> "lsd is great, you should try it" is fine. "take this much and if you're on this is fine. i survived" is not, imho.

I guess I should have worded it differently.

The dose I took was actually ~1/2 a tab, as compared to starting with a full tab (the exact ug is basically impossible to know). Some people would view this as "wasting it". I think starting with a full tab would likely be overwhelming for some people, but I don't mean to say any ug is "safe" or recommended. It's just that a lot of people are going to do it anyway, and if they do I'd rather they have the best chance at being safe. This stuff isn't really well medically documented so it's basically all off "experiences".

I haven't used it in years, and I wasn't doing super well at the time.

theaiquestion··on Effect of LSD on reinforcement learning in humans (2022)
The most accurate method is to have it in liquid form and dose that way.
theaiquestion··on Effect of LSD on reinforcement learning in humans (2022)
> possibly I'm high but I find this rather amusing. Enjoy but don't play doctor until you get your degree.

I'm not saying it's safe. In fact I think there's significant potential for harm from careless use of psychs, and I can't recommend it from a safety perspective.

That being said, if you were going to use it - and this goes for any drug - you MUST test it. You cannot OD on 4 tabs of LSD. If it's actually 2C-B (it's half the price) then you could die.

--

If you're on the fence - don't take it.

If you were going to anyway - take it in a way that minimizes harm.

theaiquestion··on Effect of LSD on reinforcement learning in humans (2022)
I actually disagree somewhat.

So I likely overdid one particular month....

But I have (likely) permanent HPPD - particularly when I take my adhd meds I can get some motion inpainting, geometry visuals, and turning off the lights at night looks like a "shower of colors", and also white looks simultaneously red/blue/green/purple (the color you see after looking at a really bright light and looking away).

It's not currently causing issues for me but it is a real thing. MRI shows no abnormalities, eye exam also shows nothing odd. The visuals are almost identical to light lsd visuals. I'm 100% aware they aren't "real", none of my doctor's believe it's psychosis.

I'm also 100% convinced that chronic psychedelic and possibly chronic marijuana usage contributes to schizophrenia.

And I'm certain you can get real PTSD from a bad trip.

I have no idea what extent the medication combo did, but with or without antidepressents, I don't want to say it's "safe". I'm not saying it's dangerous, but I don't think it's harmless.

Also smoking marijuana with lsd makes it like 10x more intense, and I have a feeling that also has a chance at shifting the harm potential.

theaiquestion··on Effect of LSD on reinforcement learning in humans (2022)
~15/20ug is the dose an experienced person would percievably be able to notice, so microdosing is below this.

75ug is the lower end of the recreational scale. I first ever started with ~50ug (which honestly I'd recommend for beginners) and I had antidepressents in my system so it's effect is dulled a bit - and that was a giggly, eyes-full-of-wonder, oversaturated, and slight motion inpainting, wonderful experience. There was no way anyone mistakes that for being sober.

That said the come-up at any dose and any experience level is hilarious because everyone is always like "is it hitting yet?" as it progressively becomes more intense.

tl;dr - placebo could fool even some experienced users if they know they're being given a really low dose. The treatment group would absolutely know no matter how much experience you have.

https://psychonautwiki.org/wiki/LSD

EDIT: Sidenote, safety first friends. Always use two different reagents to test, and always have a plan to control anxiety in case of bad trips.

Page 1 of 2Next →