HNHacker News
TopNewBestAskShowJobs

Flux159

969 karma · joined October 10, 2010

Ex-Meta, Ex-Oscar software engineer.

https://twitter.com/flux159

submissionscomments
Flux159··on StreamDiffusion: A pipeline-level solution for real-time interactive generation
I think it’s just img2img with a prompt & rcfg scale and no controlnet since theres a GitHub issue about adding controlnet support open at the moment.
Flux159··on StreamDiffusion: A pipeline-level solution for real-time interactive generation
Arxiv paper here https://arxiv.org/abs/2312.12491

I think that it's possible to get faster than their default timings for a 4090 (I have been able to get 10fps without optimizations with SDXL Turbo and 1 iteration step), but their other improvements like using a Stochastic Similarity Filter to prevent unnecessary generations are good for getting fast results w/out having to pin your GPU at 100% all the time.

Flux159··on Whisper: Nvidia RTX 4090 vs. M1 Pro with MLX
How does this compare to insanely-fast-whisper though? https://github.com/Vaibhavs10/insanely-fast-whisper

I think that not using optimizations allows this to be a 1:1 comparison, but if the optimizations are not ported to MLX, then it would still be better to use a 4090.

Having looked at MLX recently, I think it's definitely going to get traction on Macs - and iOS when Swift bindings are released https://github.com/ml-explore/mlx/issues/15 (although there might be some C++20 compilation issue blocking right now).

Flux159··on Microsoft says GPT 3.5 has 20B parameters?
Link to the paper directly https://arxiv.org/pdf/2310.17680.pdf

The system that they describe, Codefusion is interesting because it's a diffusion model for generating code rather than an autoregressive model like most LLM code generators.

Flux159··on Zephyr 7B – Mistral Finetune that responds like ChatGPT
Online chat interface here: https://huggingface.co/spaces/HuggingFaceH4/zephyr-chat

You should be able to run this locally on CPU or GPU as long as you have 16GB RAM. Less if using a quantized model like https://huggingface.co/TheBloke/zephyr-7B-alpha-GGUF with llama.cpp

Flux159··on AI Workbooks – A notebook interface for LLMs, image and audio models
ChatGPT and PaLM Chat cells do take context from previous cells of the same type (we don't mix messages between them to avoid confusion).

For something like Stable Diffusion it doesn't make much sense where the prompt is going to be isolated.

This was something that we took into account while building and designing workbooks - how will normal users expect a notebook interface to work with chat style models.

Flux159··on AI Workbooks – A notebook interface for LLMs, image and audio models
We love Jupyter notebooks, I was tech lead of an entire platform around them in a previous job - but we wanted to differentiate from Jupyter a bit by not being python first.

For AI Workbooks, we thought that the interface is natural language & standard files (images, audio) that people interact with everyday. This makes it much easier for non-developers to get started.

For developers, we also support a python SDK so that you can make a workbook from python or a Jupyter notebook using some simple API endpoints. We're going to keep improving the SDK to improve the integrations too.

Flux159··on CodeTF: One-Stop Transformer Library for State-of-the-Art Code LLM
Link to the github repo: https://github.com/salesforce/CodeTF

It would be helpful to see some Colab notebook examples of how I could use this or incorporate my own codebase with these open source coding models.

The examples show some smaller interesting prediction task & translation between csharp and java, but probably easier to try it out in Colab than having to install locally.

I would also want to be able to compare Github Copilot's autocomplete with what CodeT5 would return as a prediction.

Flux159··on Numbers every LLM developer should know
I think that it would be helpful to add a fine-tuning costs for an open source model (think LLaMA to Alpaca).

From the phrasing around fine tuning right now it seems like it's using openai's fine tuning api to determine that cost, but it's not very clear.

Also this would be helpful for other foundation models if that doesn't already exist - how much VRAM to run Stable Diffusion v2.1 at different resolutions, running Whisper or Bark for audio, etc.

Flux159··on Amazon to close all of its physical bookstores and '4-star' shops
I wonder what the internal data showed on these stores - from what I remember, these stores mostly had kindle devices and some selection of books which you could more easily get online.

Someone probably realized that Amazon Go & Whole Foods were better for foot traffic (since they sell groceries & other items) and investing more into selling Amazon Go walk out tech would be a better expansion plan than their own stores.

Flux159··on NYC Subway travel time map
I like the idea as its an interesting map visualization of the subway across a time domain rather than distance.

I'm wondering why there's such a large difference in certain areas between weekday rush hour and late night though (for instance, clicking on 68th St Hunter College and changing between weekday rush hour and weekday late night is making 59th street go from 2 mins away to 12 mins away).

A few notes about mobile (specifically iPhone usage): Its difficult to use on iPhone because the first tap doesn't seem to do anything to the map - if you tap again at exactly the right place (on the dot), then it will work on iPhone. I think I know what the issue with this is: the onclick event is being replaced with the hover state when you tap once, then the real onclick is called when you tap on the dot again.

Flux159··on Ask HN: Which TypeScript codebase should I study to get better?
If you want to learn more about 3d engines (game or rendering), I would recommend BabylonJS: https://github.com/BabylonJS/Babylon.js

It has fairly good docs: http://doc.babylonjs.com/ as well as an active forum for developers and users: http://www.html5gamedevs.com/forum/16-babylonjs/

Other options mentioned are also good - the Typescript compiler (https://github.com/Microsoft/TypeScript) is good to get started with especially if you read this Gitbook: https://basarat.gitbooks.io/typescript/content/docs/compiler...

You could also search popular typescript github projects that are more closely related to what you work on, but the Typescript compiler and Babylonjs 3d engine are both well written projects with fairly good documentation in my opinion.

Flux159··on Elevation Control
I see blog posts from redblobgames posted sometimes and I really like how in-depth the explanations are. I remember reading about 2d height maps and map generation for a side project I was working on a while back and the explanations he gave for why he was using a particular algorithm were fantastic (in addition to the interactive demos):

http://www.redblobgames.com/maps/terrain-from-noise/

http://www-cs-students.stanford.edu/~amitp/game-programming/...

Flux159··on A mobile, desktop and website app with the same code
I'm assuming that you mean Linux/Windows/MacOS when you say multiplatform. I agree that for Desktop apps its sometimes okay to use the same codebase (thinking of Blender, Gimp, other open source apps that work across Desktop OSes). Other times there are apps that work significantly better because they utilize an OS's strengths (Pixelmator for MacOS).

The key difference here is that user expectations are completely different between a mobile app, a desktop app (either native or node webkit), and a web app - at the moment I would say that its rare to see any app share the same codebase (not including shared libraries/dependencies) for all of those. As this post shows, its not impossible, but there are still some tradeoffs.

If I was starting a brand new application that needed to work across all platforms, I would still consider having independent code bases sharing a library as a dependency rather than having a single codebase for everything - especially if I wanted to be sure to have the best possible experience for each platform.

Flux159··on A mobile, desktop and website app with the same code
I like this as a thought experiment - can I write JS code (specifically using React) that can be shared across a variety of devices. At the same time, I would not do this for any sort of non-trivial production application. The main reason being that you're going to end up writing platform specific logic at one point or another (basic example: Handling files on the web vs on mobile - ios / android vs via node webkit). I think that the codebase is already doing something like this with its .ios.js files.

Eventually you'll end up with some shared code and other platform specific code and you'll wonder why you don't just split up the codebase into individual applications (that can easily be worked on by independent teams). You can still have a shared dependency that handles logic, but you can iterate on each application independently.

Flux159··on Wikimedia Maps Beta
Cool! It looks like they're using leaflet and hosting their own tile server from Open Street Map data. I really like that they opened up how to connect to their tile server (thanks to chippy for the link: https://www.mediawiki.org/wiki/Maps ). When I was looking at options for creating a side project with open street map data earlier I ended up using Mapquest's open data api because Google maps had API restrictions after a number of calls and it was easy to integrate with leaflet, but if I was making it again I would consider using wikimedia's option. (Fyi, the sideproject is a static site hosted at http://newsatlas.io - it takes a news/rss feed & attempts to parse the location client side by parsing the news article's summary text and displaying the location on a map - it was written in JS with angular and leaflet as dependencies - hosted on S3).
Flux159··on Show HN: 2D PID controller simulation
PID is a way to minimize error in a system. Its usually used in industrial settings to control mechanical devices. An example would be you want an electric motor to rotate at a constant N degrees/second. If you just apply a voltage to the motor, you probably won't get an accurate movement because of the variable characteristics of the motor, the load on the motor, and other factors. To fix this, you can add a sensor that can determine how fast the motor is currently moving or how much the motor has moved (a rotary encoder). You then compute the error and use PID control to fix the error.

There are limitations of PID and what it can and cant be used for. There are also other ways of controlling systems (like state feedback).

I might have made some errors in the above statements because I haven't done control theory in a long time, but any control theory course at a university would cover PID control.

In relation to what the post is about, it seems like they've tuned a PID control loop to perform PID error calculations based on where you click and where the circle currently is, then move the circle to that position with a force (/acceleration) determined by the controller.

Flux159··on The Architecture of Open Source Applications
This seems like a great resource to read about the architecture behind a lot of great applications and tools (nginx, scalable web architecture, nosql databases, git just to name a few). Definitely going to bookmark for later.

Related to that, I wonder if there are similar resources for learning about Blender's architecture (open source 3d modeler that has its own tiled UI system and integrates with an internal python interpreter) or the Linux kernel. I feel like those could be immensely useful for studying/understanding those large codebases for newcomers.

Flux159··on Why startups fail, according to their founders
A link to the image that matters in the article: https://fortunedotcom.files.wordpress.com/2014/09/unknown2-e...

A few things to note are that it seems like a few of those items could be combined into the same general problem in a failed startup: Not the right team, poor marketing, poor product, and disharmony on team/investors could potentially be related.

Another point is that these are obtained from self post-mortems by founders, whereas if the data were obtained from employees or investors, you might get a different story.

Flux159··on Walmart Labs: Why we run an open source program
I'm not going to comment on who is writing the article because that's not the point.

I think this is a good article pushing for more open source development at traditionally large enterprise software companies. It helps the open source community and it helps the company attract talent. It also gives good advice to large enterprises saying that they shouldn't just throw their code into open source without good documentation and community engagement. At the very least its a good start for companies that would have otherwise kept all their code internally.

Just because the article mentions Walmart Labs working on hapi (apparently an express-like nodejs framework) doesn't mean that HN should suddenly disregard the content of the article (as some of the other comments seem to be doing).

Flux159··on Ask HN: HN APIs (or lack thereof)
As far as I know, HN offers an official API here: https://github.com/HackerNews/API

But it specifically says "We hope to improve it over time, and may later enable access to private per-user data using OAuth." So I'm assuming that login, votes, comments, submission is for a future release of the API. That might be why the apps you use resort to scraping HN.

For reference, here's the blog post talking about the release of the API: http://blog.ycombinator.com/hacker-news-api

Flux159··on A Rant on VC in the Flyover States from an Arkansan that Raised $100 million
Reading the article, I was reminded of pg's essay on trying to create another Silicon Valley: http://www.paulgraham.com/siliconvalley.html

I think its great that the author is putting his own money and efforts into making northwest Arkansas a better hub for entrepreneurs. However, I think that one of the reasons that it was/is difficult to get money from investors in the area is due to the fact that many of the current investors may not have earned the money through recent startups, which seems to be a key factor in whether they would want to invest in other startups in the area.

Flux159··on The dangerous allure of the “Billion Dollar Startup Club”
The author seems to be trying to dissuade founders who are only in it for the money from starting startups solely for that reason.

On that note, I would suspect that a company whose primary interest is making the founders money wouldn't necessarily have their customer's best interests in mind and would probably never reach a $1B valuation.

Flux159··on P5.js: Processing for today’s web
Wow. At first I thought it was just a video, but its more like an "interactive intro". I remember being introduced to Processing years ago in CMU's Intro to Programming class, glad to see that it's still being used and developed.
← PreviousPage 5 of 5