jimfleming

There's also Gitxiv[0] which matches up Arxiv papers with code. TensorTalk seems to have a wider focus (world generation and Twitter bots) while Gitxiv tends to be focused on active ML research.

[0] http://gitxiv.com/

From that page:

> Supported cards include but are not limited to[...]

Can you expand on that? As far as I'm aware that's not true (anymore). It runs quite fine on AWS which uses older NVIDIA cards and I know several people use it on older-gen GPU-enabled MBPs.

EDIT: clarification

It's nice to see Leaf coming along so well. Part of me would love to be able to build models in rust.

For more benchmarks (including updated TensorFlow performance with cudnn v4) see https://github.com/soumith/convnet-benchmarks

One benefit of using cloud GPUs is the ability to train multiple models simultaneously. This is difficult to do with your own hardware at a reasonable cost, especially since a large portion of your time will be spent on hyperparameter tuning.

While you're getting started with your own hardware I'd love it if you could share any feedback on something we've been building[0] as far as cloud training goes. It's currently in free public beta and works directly with TensorFlow models so there's no lock-in.

[0] https://fomoro.com (jim at fomoro.com)

Chris Olah also has a great introduction: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Hello! I'm one of the creators of Fomoro and this is something we've been building over the last couple of weeks. I have a Macbook Pro with no GPU so I frequently use AWS to train models. While AWS doesn't have the best GPUs for deep learning they're still faster than without and as an added benefit I could train as many as I want simultaneously.

Right now, Fomoro simplifies training on spot instances, it's a little rough around the edges but try it out and let me know what you think. Thanks!

DeepMind/Google has a hosted copy: https://storage.googleapis.com/deepmind-data/assets/papers/d...

Check out NeuroAnimator[0] from 1997 (also early work from Hinton). It covers some of this using local-spaced hierarchies of neural networks that predict the deltas for the next time-step.

[0] http://web.cs.ucla.edu/~dt/papers/siggraph97-sketch/

It appears to be based on Navier-Stokes, from the abstract:

> We designed a feature vector, directly modelling individual forces and constraints from the Navier-Stokes equations, giving the method strong generalization properties to reliably predict positions and velocities of particles in a large time step setting on yet unseen test videos.

Here's the most recent video that I could find: http://pchiusano.github.io/2015-03-17/unison-update5.html

A few of the other posts have videos too: http://pchiusano.github.io/unison/

Ok, so I spent about an hour casually reading the posts, watching the demo videos and glancing at the code. I agree with your premise and I like a lot of the implementation so far. It's great to see more innovation and research in this area alongside Light Table, Eve (which seems to be going in a similar direction with spreadsheets), NoFlo and others.

I'm curious about the goals of the project. If the goal is to replace an existing language and editor pairing then it's going to be an uphill battle (see previously mentioned related projects). If it's to be supplemental with existing development tools then what use cases are you targeting?

Adding this as an optional layer to an existing editor might help work out some of the ergonomics. Performance issues aside, building on top of Atom might be a good choice since it's mostly web based like the current UI in the demos. Similarly, a custom kernel or cell type in ipython might get the ideas into daily use.

> By default AI will optimize the universe to the greatest extent possible towards it's values.

Hmm, I'm not sure I understand this line of thinking. Intelligence is complex, soft, and can even conflict with itself. Why do you envision artificial intelligence as if it were a pure optimization algorithm with high-level problem solving abilities? It seems to me that a fuzzy metaheuristic with lots of competing goals would be a better comparison and this does not lend itself to the "paperclip factory" quite as readily.

I have to agree on a few points as well:

We're much closer than we were a year ago but not as close as many think we are w.r.t self-driving cars, killer robots or even reliable speech recognition.

Safe-guards are probably a good idea before unleashing an AGI onto the Internet but those working on it are not going to be surprised one morning that their AI project suddenly became sentient or dangerous. It will be a deliberate, massive project with lots of funding and lots of smart people working on it.

Here's a rather lengthy discussion of a similar redesign that began in LA:

https://news.ycombinator.com/item?id=8923196

Perhaps linters, auto-formatters, auto-complete, etc. I could understand that it could look like "self-correcting code" to a non-programmer.

I'd love for someone who knows more to chime in but its my understanding that gradient descent does not perform well on non-differentiable functions and only finds local optima.

Wikipedia seems to confirm this: http://en.wikipedia.org/wiki/Gradient_descent#Limitations

Write it all down.

What do you need to do? What do you want to do? What things don't really matter to you?

Organize it on paper or whatever medium makes sense. I like OneNote and Trello. I've found its one of the easiest ways to remove those thoughts from the back of my mind is to put them someplace actionable and consistent. A stream of consciousness todo list isn't very productive.

With regards to instapaper or readability, either:

a) Treat it as a bookmarking service, not a read-later service. Then reference it when a topic comes of importance to what you wrote down above or you're just bored.

b) Clear out all 800 articles and start over, possibly being more selective or auto-clearing them monthly.

These are both things the brain does naturally, pruning through attention and focus and long-term storage for future reference :)

EDIT: Formatting

Many indie games on Steam sell for that much, though not much higher: http://store.steampowered.com/tag/en/Indie/

The point of an MVP is to get the product in the hands of users as soon as possible to test viability and garner feedback before going down a rabbit trail that may not be successful. It isn't necessarily to start making money from customers before the product is ready—although being willing to pay for something in alpha is probably a good viability indicator its not the only one.

Many games aren't treated like products but more like hobby or art projects. This isn't a bad thing but it does mean an MVP isn't really necessary or helpful. Many games do things that achieve the same results of an MVP:

1. Screenshot Saturdays. This puts the idea in front of people even if the thing is buggy, broken or just unplayable.

2. Festivals. User testing during festivals was incredibly valuable and helped us determine what attracted players to the booth, what they enjoyed about the game, what was fun, what wasn't, etc. Here the game doesn't need to be finished.

3. Awards. Similar to festivals, an award can indicate that you're on the right track. The inverse isn't necessarily true. Winning an award does not equate to market viability.

4. General social media, spreading the word and press. If someone is willing to talk or write about your game that's a good sign.

5. Steam's Greenlight: Will people vote for this?

6. Steam's Early Access: Will people buy this before its finished?

7. Kickstarter. A very time consuming and potentially expensive endeavor that might work for MVP-like launches. Will people contribute money to the possibility of this game existing?

The point of all of these is to get out there with the game as soon as possible, even if its not ready, to test out the idea. Some require higher levels of investment than others and make sense at different points in the project.

Experience: we launched our first game earlier this year with success and followed these where they made sense.

EDIT: Formatting

Out of curiosity, does anyone know why they choose I&F instead of Izhikevich neurons[1] which model more biological spike forms. Perhaps to meet the low power consumption goals?

Also, how do convolutional neural networks model time? I thought that was one of the benefits of spiking networks and STDP.

1. http://www.izhikevich.org/publications/whichmod.pdf

Sensus (http://sensusxp.com/) is working on something similar as an external case for the iPhone. Dislaimer: I have a dev kit.

Currently, its based on esformatter[0]. Most of the work so far has gone into search and replace and we don't yet expose many of the properties provided by esformatter[1].

[0] https://github.com/millermedeiros/esformatter [1] https://github.com/millermedeiros/esformatter/blob/master/li...

You're correct it doesn't have type support which would make it vastly more useful. I honestly wasn't aware `forEach` returned anything when I wrote the example. The other example finding component dependencies is a little more realistic.

Radish is for performance monitoring whereas Reddish is like a GUI for the data itself.

As for the name, we're each named similarly to redis, its bound to happen.

Thanks!

We use real-time to refer to two different things:

If a key is deleted (or any other update is made), the interface reflects that change. Right now we throttle those updates and have some other ideas to limit them so the interface isn't overwhelmed if 1m keys suddenly expire.

We also use real-time to describe the analytics. They're still in development but in their current form are line and pie graphs of user-selected data (optionally over time). There are a number of avenues we could take when it comes to visualizations and user feedback will impact how we proceed after the initial prototype of the visualizations is released.

atm we are only supporting the latest browser versions to speed up development but if there's enough interest in supporting other versions we're happy to do so

You can either connect directly to public instances (like from redistogo.com) or use our proxy which will (before we release) support encrypted connections.

Ahh, yes - adding some now, until then you can see visual representations of the interface on the splash.

Its a hosted service. Good point though, we'll update the copy so its more clear.