Compared to playing a game which would definitely use both GPU and CPU, instead of mostly just CPU?
37,790 karma · joined September 7, 2019
> Social: 2.2 hours of usage on average
> Productivity: 1.5 hours minutes [sic] of usage on average
How is this accurate? What kind of "Productivity" tasks are people doing inside a Quest if it uses more power/battery than gaming? Also, "Social" drawing more power than "Gaming" too? Something seems weird.
The idea behind OSS is that you're able to modify it yourself and then use it again from that point. With software, we enable this by making the source code public, and include instructions for how to build/run the project. Then I can achieve this.
But with these "OSS" models, I cannot do this. I don't have the training data and I don't have the training workflow/setup they used for training the model. All they give me is the model itself.
Similar to how "You can't see the source but here is a binary" wouldn't be called OSS, it feels slightly unfair to call LLM models being distributed this way OSS.
Since Mistral is just a 7B parameter model, it's obvious that you won't be able to have it straight up write accurate code, it's simply too small for being able to accomplish something like that, unless you train the model specifically for writing code up front.
I guess if all you're looking for is a model to write code for you, that makes sense as a "hello world" test, but then you're looking at the wrong model here.
What you really want to do if you're looking for a good generalized model, is to run a bunch of different tests against it, from different authors, average/aggregate a score based on those and then rank all the models based on this score.
Luckily, huggingface already put this all in place, and can be seen here: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderb...
This Mistral 7B model seems to earn itself a 3rd place compared to the rest of the 7B models added to the leaderboard.
Edit: As mentioned by another commentator, this also seems to be a base model, not trained specifically for request<>reply/chat/instructions. They're (or someone) is meant to fine-tune this model for that, if they want to.
> Initial commit
> @brson - committed Mar 29, 2012 - 0 parents commit 4825529 - Showing 1 changed file with 1,522 additions and 0 deletions.
https://github.com/jdm/servo/commit/48255297b8958fa559d11704...
The talk: https://www.youtube.com/live/e3Y1C695CIw?si=dD8_tbyEezwei8As...
> The “level of disruption to the manufacturing environment was such that we needed to change everything in our factories” to scale up recycled PET use, he said. “After all that, the carbon footprint would have been higher. It was disappointing.”
How you calculate this? Unless you're factoring in acquiring the hardware, you can usually get away with outsourcing the training of llama2 to rented hardware, and then run it on owned hardware, so with lots of executions, using llama2 locally should most definitely be cheaper in the medium to long term, compared to paying for the training + execution of fine tuned GPT3.5
- Make a simple and complete game, with the full flow of "Loading > Starting > Playing > Finished > Published" as soon as you can, and do this for each "concept" you want to explore
- Try to focus on one concept you want to learn per game made
- Don't give up on the current game but force yourself to go through the full flow, even if the game itself is trash and you're 110% sure no one will play it.
- Accept that the first 10 games will be absolute crap and show almost no progress, but after that you'll see that you improve
- Note down flows and architectures that works for you when implementing those games, iterate on those ideas in future games
What engine you use is almost irrelevant, as long as you pick one and stay with it. Also, it depends on what type of game you're trying to develop.
- Unity 3D if you want something quick and basic that will be easy to find people to collaborate on in the future
- Unreal Engine if you want something complete and complex, and are fine with a slightly steeper learning curve, but also easy to find collaborators in the future
- Bevy if you want something experimental that focuses 100% on code (for now), with a large focus on parallelism and be able to write code with Rust
- Godot if you want something similar to Unity Engine but FOSS and not as feature complete (yet).
- Finally, write your own game engine in whatever language you want if you don't actually want to create a finished game but you'd rather fall into rabbit hole after rabbit hole trying to create a better engine. Lots of fun and educational, but not productive if the end goal is a published game.
Similar to this attack then, seems it can take up to 30 minutes on a AMD system, and up to 215 minutes on a Intel system, and it's still not 100% accurate.
Tell me when the rest of the company aligns with you and has started to show any results in providing a good experience for people to do machine learning with AMD. As it stands right now, there is so much tooling missing, and the tooling that's there is severely lacking.
But, I have a faith. They've reinvented themselves with CPUs, multiple times, so why not with GPUs, again?
Yes, super basic, but so is most features of Apple software. Simple, basic and without customization, just like how many Apple users like their software.
Amazon has implemented an algorithm for the express purpose of deterring other online stores from offering lower prices.
[redacted]
Rather than trying to compete, Amazon uses
[redacted]
Ultimately, this conduct is meant to deter rivals from attempting to compete on price altogether-competition that could bring lower prices to tens of millions of American households. As a result of this conduct, Amazon predicted, "prices will go up."> Plus of course everything is tested together
That makes sense, Apple's QA department seems to be held together with tape for the last decade or so, as every new update seems to break something, so maybe they're trying to make it as easy as possible now in order to recover.
Aha, that should be working fine?
(defn add [x y]
(let [add clojure.core/+]
(add x y)))
> I watched a video and it does seem rather complete, but [1] indicates there is no debugger?
I think debuggers tend to be used as a library across many different editors, rather than the editor/plugin providing that functionality. Personally, I don't use debuggers much as the functions I write tend to be small and evaluating small executions with the repl tends to reveal the issue quickly. Sometimes when refactoring others code I've used https://github.com/philoskim/debux to various degrees of success.
I do think cider (https://github.com/clojure-emacs/cider) has stuff regarding stepping debuggers, but I'm not sure how common it is to use it. Maybe other Clojure users can fill me in :)
> . I also don't see a profiler mentioned
Yeah, as you said, the Java ecosystem basically covers that. For OSS stuff, I use VisualVM, and for professional stuff I use YourKit, both of them work well with Clojure and points out my user-space code with ease. And I've never been paid anything for actually writing/maintaining Java code, so even with that, seems I'm able to use those tools just for Clojure :)
> As an aside, by "continuations" did you mean "restarts"?
Ah yes, of course. The condition system and restarts :) Thanks!
I just want the new changes to Safari and Notes while I don't care about the other changes, why not have Safari be a normal application downloaded from the App Store, like the rest of us pleebs have to do when publishing apps?
> I accidentally shadow an important function at least once an hour when working in Clojure
You mean that you accidentally "overwrite" (declare again) a function with the same name as the one you're now declaring, but you didn't mean to?
If so, I'm not sure how that'd be possible. In Clojure you're always in a namespace, and you'd only have one function with that one name in that namespace, and if you're declaring a function with the same name in a different namespace, you now correctly have two functions with the same name, but in different namespaces.
Unless your editor<>repl connection somehow ignores the current ns, I'm not sure how you'd end up in that situation.
> SLIME is better than any Clojure dev setup I've seen
This I'm also curious about, what exactly SLIME gives you that for example Conjure for neovim wouldn't already? Maybe something about continuations perhaps? That seems to be the only feature I've seen from Common Lisp (besides actually being able to compile to binaries) that I'd love to have in Clojure.
For the ultimate version of this, once locally trained and running LLMs becomes a bit more feasible (meaning, closer to GPT4), it would be a fun exercise to train various models with personas and let them hash it out until they all agree on something.
One model could be trained on HN, another on a specific Facebook user group (like "Moms against video games"), another on Twitter on so on, then you let them come together until they've reached consensus (or not).