michannne (u/michannne)

michannne commented on 'Ultramassive' black hole discovered – bigger than the majority of galaxies bbc.com/news/uk-england-t... · Posted by u/achow

taylodl · 3 years ago

Au contraire! There's plenty of mass to feed such a black hole, the real question is "when you run things backward" in the lambda-CDM model why isn't the entire universe a black hole? Or why wouldn't it be the case that there wasn't a primordial mega star that quickly (in universal time) collapsed into a supermassive black hole and the remnants of the supernova is the mass of the remainder of the universe and form the CMB radiation?

michannne · 3 years ago

This is a cool idea I've never heard before. The universe we're in is the result of a cosmically massive star going supernova.

michannne commented on The RWKV language model: An RNN with the advantages of a transformer johanwind.github.io/2023/... · Posted by u/T-A

tarruda · 3 years ago

You should be able to run Llama 30B q4 on a RTX 3090. Check the table here: https://bellard.org/ts_server/

30B/q4 requires 20GB of RAM while 3090 has 24GB.

michannne · 3 years ago

Seconding this.

Llama 30B 4-bit has amazing performance, comparable to GPT-3 quality for my search and novel generating use-cases, and fits on a single 3090. In tandem with 3rd party applications such as Llama Index and the Alpaca LoRa, GPT-3 (and potentially GPT-4) has already been democratized in my eyes.

michannne commented on Ask HN: How are you using GPT to be productive? · Posted by u/yosito

michannne · 3 years ago

I'm currently working on a game despite not being very familiar with Unity. When I have some feature I need explained, or running into some bug in the game, I'd usually use Google and find some forum post that hopefully has a similar problem to mine and go from there. With GPT, I can just tell it the problem and get an answer, without scouring through multiple posts or having to constantly readjust my query. Yes, the information it produces might be incorrect, but I think the (time needed to verify the solution provided) is less than the (time needed to produce a search query that gets the articles you want * time needed to read each article and see if it helps me).

Here's an example regarding grounded characters sticking to walls:

prompt: Thanks! I added walls and ground to my game, as well as detection so that if the player touches the ground, they will be in a grounded state. However, when the touch the vertical face of the ground/wall, they stick to it. Can you tell me what I'm doing wrong?

response: It seems like the issue you're encountering is related to the physics and collision detection in your game. The player is likely detecting the wall as the ground, and it's causing the grounded state to be triggered when it shouldn't be. To fix this, you can modify your ground detection code to take the angle of the surface into account.

Here's a solution using the Physics.Raycast method:

Assuming you have a LayerMask called groundLayer for ground detection, make sure the ground and walls are assigned to the appropriate layers in the Unity Editor

[...further details on using angle detection on collision, as well as accurate sample code]

I can even ask it to give me the reference for the _Physics.Raycast_ method:

response: Certainly! Here's the official Unity documentation for Physics.Raycast:

[URL to official docs filtered out by HN]

Physics.Raycast - Unity Manual The Physics.Raycast method allows you to perform raycasts in your scene to detect collisions with 3D objects.

[...further information on how to use it, including an accurate code example pulled from the docs]

This just drastically reduces time-to-solution, especially when you consider you can lazily leave out information and still get a good response.

In short, I treat it as though each session is a fine-tuned search engine and get a lot of value from it, but it doesn't completely replace forums or Youtube videos, for one it can't argue with itself to provide conflicting but valid alternate solutions - I'd have to argue with it/ask it to produce those.

michannne commented on Facebook LLAMA is being openly distributed via torrents github.com/facebookresear... · Posted by u/micro_charm

anubiskhan · 3 years ago

Seems like at that point one might as well just pay onlyfans? Although I suspect it is not far off until gpt and deepfaking get combined to produce completely generated onlyfans (or some other competitor if this is against their ToS, I have no idea).

On a similar note I very much look forward to the day when entertainment providers leverage these so I can say to Netflix, et al, "I would like to watch a documentary about XYZ, narrated by someone that sounds like Joe Schmoe, and with the styling of SomeOtherShow".

michannne · 3 years ago

I'm not seeing how onlyfans equate to virtual AI girlfriends

michannne commented on What is ChatGPT doing and why does it work? writings.stephenwolfram.c... · Posted by u/washedup

thegabriele · 3 years ago

I keep on stumbling on this story, yet i can't seem to be able to grasp it.

We have senses so our brain has some sort of a priori knowledge of the world? We're all almost blind when we born.

michannne · 3 years ago

The point of Minsky's action is to demonstrate to Sussman that Sussman's intent is essentially "if you can't see it, it's not there", but of course the room doesn't become empty when Minsky closes his eyes - the neural net won't lose preconceptions just because you randomly wired it, you just lose the ability to see and control what those preconceptions are.

michannne commented on ‘Breakthrough’ obesity drugs that have stunned researchers nature.com/articles/d4158... · Posted by u/pseudolus

myshpa · 3 years ago

Calorie is not a calorie.

We now know, for example, that fibre is very good for our microbiome. Exclude it from your diet and you'll suffer.

> it is certainly possible to create a "healthy" everyday-McDonalds diet (but you would be changing the meal to the point it won't be recognizable)

So it is possible or not? Why the quotes?

> You can still become obese eating 3000 kcals of salad every day without exercising

I'd like to see you try (just salad as you wrote, no oils and animal based products, please).

michannne · 3 years ago

Not sure if you're being purposefully ignorant or not so I'm not going to engage this discussion, but for anyone reading, this should be basic dietary knowledge

michannne commented on ‘Breakthrough’ obesity drugs that have stunned researchers nature.com/articles/d4158... · Posted by u/pseudolus

str1k3 · 3 years ago

Wow that's some impressive logical gymnastics. If these people adhere to the "healthiest diets" they will eventually no longer be obese.

American food is shit. Most people have no clue what healthy food is, even if they intend to eat a healthy diet.

michannne · 3 years ago

This is pseudoscience. As long as you consume a certain amount of calories and get your daily nutritions, you can eat whatever you want without getting obese. You can still become obese eating 3000 kcals of salad every day without exercising. Of course, healthy caloric intake will only ensure you don't become obese, you'll still be consuming way more fat/carbs than you should if you eat McDonalds, but it is certainly possible to create a "healthy" everyday-McDonalds diet (but you would be changing the meal to the point it won't be recognizable)

michannne commented on A Review of the Odin Programming Language graphitemaster.github.io/... · Posted by u/gingerBill

michannne · 4 years ago

> This means valgrind cannot really detect memory issues as well since Odin has it’s own internal allocators. helgrind cannot work because Odin doesn’t use pthread primitives. ltrace cannot work because that just provides wrappers over every libc function, etc

Unfortunately, I've gone through too many bugs that had to be fixed through valgrind debugging that I'm unwilling to part with it

michannne commented on A Web UI for Stable Diffusion github.com/AUTOMATIC1111/... · Posted by u/feross

qudat · 4 years ago

The results in midjourney are significantly better than SD. I find it much easier to get to a good result in MJ and I've been trying to understand why. Anymore insight you could share?

michannne · 4 years ago

Good engineering. Midjourney likely has a lot going on under the hood before your prompt actually gets to Stable Diffusion. As an example you can check out this research paper [0] which seeks to add prompt chaining to GPT-3 so you can "correct" it's outputs before it reaches back to the user. There's also no rule that states you can only make one call to SD, MJ likely bounces around a picture through a pipeline they've tuned to ensure your generated image looks more reasonable.

[0]: https://arxiv.org/abs/2110.01691

michannne commented on Tell HN: A new way to use GPT-3 to generate code (and everything else) · Posted by u/goodside

b20000 · 4 years ago

the reason it does well is because you gave it a question that might come up in coding interviews.

michannne · 4 years ago

I think that's amazing on its own. Rather than having to do a leetcode hard on the fly, you can just ask for the solution. You'll still need to know the problem space in order to properly optimize it, but it can save a lot of time.