So not only job layoffs..... My entire trip to NeurIPS has just been cancelled by
@MetaAI
. I've still not been to an in-person conference during my PhD. Now im going to desperately scrounge around for some money to pay for flights and hotels, which will now be super expensive.
We’re excited to announce GriddlyJS - A Web IDE for Reinforcement Learning!
Design 🎨, build ⚒️, debug 🐛RL environments, record human trajectories 👲, and run policies 🤖 directly in your browser 🕸!
Check it out here:
Paper:
We’re excited to announce that GriddlyJS will be presented at NeurIPS 2022.
Design 🎨, build ⚒️, debug 🐛RL environments, record human trajectories 👲, and run policies 🤖 directly in your browser 🕸!
Check it out here:
Paper:
Looks like
@DeepMind
saw GriddlyJS and decided to build their own version.... I had no idea they were building this but it looks incredibly similar to my work.
At
#NeurIPS2022
, we're introducing Play Lab, an accessible digital laboratory that lets you build small worlds and set up experiments to analyse how embodied agents behave in an interactive simulation.
Try it yourself:
⌚10.30am CST
📍Booth
#609
, Hall G
PhD: Where you simultaneously think you have worked the hardest in your life and produced loads of stuff,,, but at the same time feel like you've literally done nothing for 4 years and nobody gives a shit.
After only about an hour of training. Solving pretty complex logic problems... Should probably write a paper about this. Don't like writing papers. I'll just open source it in the next few days instead.
Two things that I'm unbelievably excited about.
1) I've finished and handed in my Ph.D. corrections.
2) On Monday, I'm joining
@aimistral
as an AI Scientist
@timnitGebru
As someone with an underlying health condition that means i *very occaisionally* need to take some time off work, nothing has every made me feel more shitty/guilty/shameful than HR people.
Everyone is going on about success and career opportunities and glory and all that stuff.
Do your PhD for passion and fun. You only have one life. Your motivation should be your own, not some external validation from peers or resources or career.
As PhD applications season draws closer, I have an alternative suggestion for people starting their careers in artificial intelligence/machine learning:
Don't Do A PhD in Machine Learning ❌
(or, at least, not right now)
1/4 🧵
Hello academic twitter. Let's talk about mental health. The last few months of my PhD have been really fucking hard. I'm defeated, sad and questioning why I'm even doing this to myself. A thread...
So here is my first (first author) paper out of my PhD. I'm pretty proud of it. I also have a github repository with code examples and pre-trained models! paper: code:
Emergent strategies from RL self-play with no demonstrations. The player at the top learns to "barracks rush" whereas the bottom tries to "worker rush". (About 1h of training on single gpu/cpu + rllib+impala+conditional action trees)
The motivation for a PhD should be about learning and having fun. If your motivation is broken by "something/someone might be smarter than me" then I'd think about how to change where your motivation comes from. Knowlege is infinite. Go create more.
Another AI paradox: people are excited about LLMs, some even think that AGI is just around the corner. But some students are depressed how they can still get a PhD. Is it becoming pointless?
Some personal notes on this. (1/8)
Google presents Genie
Generative Interactive Environments
introduce Genie, the first generative interactive environment trained in an unsupervised manner from unlabelled Internet videos. The model can be prompted to generate an endless variety of action-controllable virtual
Mistral 7B is out. It outperforms Llama 2 13B on every benchmark we tried. It is also superior to LLaMA 1 34B in code, math, and reasoning, and is released under the Apache 2.0 licence.
Dear Pixel-based RL community: Do you know that running your environments on different hardware might actually produce slightly different pixel outputs (not visible with the human eye, but maybe off by 1 or 2 out of 255).
I've been working on something during lockdown I think many people will find very cool and I like to share it so far as I'm now looking for potential collaborators.
Docs ->
Code ->
I love Gridworlds, also I'm really happy that I'm going to be spending summer working on them at Facebook AI Research with
@HeinrichKuttler
. I'm going to learn so much and its going to be awesome :D
Here's some games I trained (and am playing) using my "Neural Game Engines" paper. Games are trained from pixels and generalize to any size grid. Also they all train to this quality on a single 2080ti in about 30 minutes.
@levelsio
I'm confused, On the quoted thread it says you have to work for a company "abroad" and has income requirements for that job. Does this mean that you cannot work for a *Korean* company as a nomad, but you *can* work for a foreign company as long as you earn enough?
@BlackHC
@DeepMind
@MetaAI
Rejected from Deepmind twice, Nvidia + Microsoft didn't even reply to my applications. Remember that its the science and engineering that makes you happy, not the people you work for :)
The absolute best thing you can do when you are writing up your PhD thesis is get a gym membership. Go early in the morning to destroy yourself physically. If your legs hurt you dont want to get up from the desk.
Weekend project Success: reimplemented a statistical version of WaveFunctionCollapse that uses Jax and some tricks to learn logical rules via-gradient descent
🔥Happy to announce some big updates to Griddly!!🔥
In the just release version 1.6.0 we have:
🚀RLLib 2.1.0 Support
👀Multi-observers
😎 Partial observability highlighting
🌴2 new environments
Tonnes of bug fixes! 🐞🪲🐛
I'll explain what each of these mean in the 🧵 ->
Version 1.0.0 of Griddly is out!
* Support for RLLib!
@raydistributed
* Single-agent and multi-agent IMPALA on every environment (thats like 30 games each with multiple levels and configuration options).
* More documentation!
* Loads of testing and bug fixes!
* MORE GNOMES
New version of Griddly with 4 new RTS games and significantly better and simpler multi-agent and RTS support. I'm actually going to start writing agents now I promise
It's been a while since I've posted about Griddly.. So I'd like to finally release something I've been working on and off for a while... a multi-agent version of
@danijarh
's crafter environment:
Working on getting RLlib compatiblity with Griddly and will have some examples up soon. Here's a small teaser of what you can do on a single CPU+GPU in 10 minutes using IMPALA. (Agent has to save the butterflies from the spiders)
Optimizing is super fun. Griddly can now produce vectorized game states at a monstrous 60-70k FPS. Also can render 1000x1000 pixel game states at about 4K FPS. All on a single core, single GPU.
If you're interested in game engines and how to build them to be optimal for RL research. (Or if you want a Griddly sticker) Look for my badge!!
#NeurIPS22
#NeurIPS2022
Some prompt engineering tricks I used
1. Spaces around each value in the coordinates
2. Ask it to be careful with calculations
3. Ask for a complete list of actions
Since Grid-Worlds are all the rage at the moment, It's a good opportunity to talk about my recent work with
@vwxyzjn
on Griddly:
We provide not 1 or two baseline reinforcement learning experiments for the Griddly environment... but 156!!
I'll also be hanging out at
#NeurIPS2023
. Also happy to chat about engineering and open source! (or if you want to hang out and go to watch some music in Frenchman street 🎷🎺🎶)
After spending far too long trying to have a decent build system for C/C++ with multiple targets/platforms I can successfully say that I found the ultimate tool.
Giving up.... Just give up.
Maybe I'm just having a low in my PhD and it will all get better. Is this a normal thing to happen? Does everyone feel this shit at some point?
Either way I just needed to vent. Thanks for coming to my therapy session.
Last week I tweeted to the academic community on here about my current struggles with PhD research. I want to informally thank everyone who reached out with their experiences. You have all made me feel better and more motivated to push forward.
I've never submitted a paper to a conference like NeurIPS, but there's at least one paper accepted there that's cited my work, so I think thats a huge win.
In all this craziness and hype around AI and language models... dont forget about all the people and experiences in your life who make it worthwhile being hyped. Go call your Mum and Dad. Reach out to an old flame. Make your favourite meal. Paint a picture. Do weird shit 💥
Playing around with JAX for the first time and the speedups you can achieve with things like vmap instead of dumb loops for things like data augmentation are just crazy
Its been a long time since I've tweeted anything Griddly related... but I've decided that I'm starting the new year by adding a tonne of tutorials/documentation. Going to try and put out 1 tutorial every week for the next 12 weeks or so. Watch this space.
Griddly Tutorial Time Week 2: Projectiles
In this week's tutorial you learn how to build a Reinforcement Learning environment where you control an alien that can throw fireballs!
Theres lots of basic mechanics explained in here so good for beginners.
MAESTRO uses multi-agent unsupervised environment design (UED) to train policies that are robust to both the choice of environment 🌍 and co-players 🤖, leading to complex emergent multi-agent behaviour.
We're excited to welcome
@Bam4d
to our team as Sr. Technical Staff! He is also the creator of , which is specifically tailored to training Reinforcement Learning agents. We can't wait to see what you will build for us! 😃
Also I'll be at
#NeurIPS2022
from this evening so hit me up if you want to talk about AI in games, AI-first game engines, or just want a drink and see some music! 🎶
🔥Happy to announce some big updates to Griddly!!🔥
In the just release version 1.6.0 we have:
🚀RLLib 2.1.0 Support
👀Multi-observers
😎 Partial observability highlighting
🌴2 new environments
Tonnes of bug fixes! 🐞🪲🐛
I'll explain what each of these mean in the 🧵 ->
Happy to share our
#ICML2021
"Revisiting Rainbow" paper, w/
@JS_Obando
!
We argue for small- to mid-scale envs in deep RL for increasing scientific insight & inclusivity.
📜Paper:
✍️🏾Blog:
🐍Code:
Thread 1/🧵
Absolutely one of the hardest things you have to do as an adult is say goodbye to a Pet. Yesterday I had to say goodbye to my friend Poppy. She was 18 years old and me and my partner spoiled her rotten for the 6 years we looked after her. RIP Pops
Using GriddlyJS we created a modified version of
@danjar
‘s Crafter environment… We then used GriddlyJS to create 100 challenging escape rooms in just few hours. We tested a strong Domain Randomization (DR) agent, but there’s a big generalisation gap.
Recently I commited a small fix to
@raydistributed
's RLLib to stop pytorch from weirdly switching to CPU instead of GPU under heavy workloads. The fix was to add a small 'sleep' into a spinlock. Explanation here:
I *think* the solution here is to always add a small amount of noise to your pixels to make sure any embeddings are robust to different hardware rendering.
Definitely something that might require further investigation!
After a long day of thesis writing I like to relax by learning terraform and building infra for the next version of that includes logins, sharing envs and persistence of models.
I am being promoted to Professor of Artificial Intelligence at
@UCL
(
@UCLCS
@AI_UCL
), effective as of October 1st. This achievement is largely attributable to the extraordinary
@UCL_DARK
students and collaborators with whom I've had the privilege to work with over the years.
1) implement algorithm 2) spend ages fixing everywhere that the tensors are the wrong dimensions 3) finally run it a few times with no tensor errors. 4) NaNs and -Inf everywhere 5) Loss function == spaghetti 6) the fear 8) breakpoints interruping sleep K) Maybe I am the loss?
Designing and building grid-world environments for RL research has never been easier! 👌
You can define the game objects 🐻, mechanics 🚡 and rewards 🏆 using Griddly’s description language.
You can then draw 📝 the levels right in the browser!
Cool thing about using Griddly + IMPALA + rllib... you can train 100M timesteps in about 4-5 hours on a Single CPU (8 cores) and single GPU. The environment will thank you.
It's strangely exciting that I'm going to hang with
@vwxyzjn
IN PERSON later in New Orleans. We've co-authored two papers together and talked loads over discord. But IN PERSON is different
#NeurIPS2022
I actually really enjoy doing hackerrank questions... but when you get those annoying tricky edge cases that are hidden and you can't work out why they fail... bit ragey