DeepMind Profile picture
Dec 8 8 tweets 7 min read
In @ScienceMagazine, we present #AlphaCode - the first AI system to write computer programs at a human level in competitions.

It placed in the top 54% of participants in coding contests by solving new and complex problems.

How does it work? 🧵 dpmd.ai/alphacode-scie…
🔵 Previous AI systems don’t have the problem-solving skills needed to excel at coding competitions.

By combining advances in large-scale transformer models with large-scale sampling and filtering, #AlphaCode made significant progress in this field: dpmd.ai/alphacode-scie…
🔢 AlphaCode is pre-trained on selected public @GitHub code and fine-tuned on our relatively small competitive programming dataset.

At evaluation time, millions of diverse C++ and Python programs are created for each problem, orders of magnitude larger than previous work.
🟢 The solutions are filtered, clustered, and reranked to a set of 10 candidate programs that are submitted for external assessment.

This automated system is similar to the usual human process of debugging, compiling, passing tests and then submitting: dpmd.ai/alphacode-scie…
📈 #AlphaCode’s performance was validated using popular coding competitions, each with over 5000 participants.

It placed at about the level of the median competitor, marking the first time AI has achieved this level against human programmers. dpmd.ai/alphacode-scie…
🌐 We hope AI systems like this will assist software developers while widening access to programming for those who don’t know how to code.

To help others to build on our results, our dataset of competitive programming problems & solutions is on @GitHub: dpmd.ai/alphacode-gith…
Work by: @liyuajia, @choidavidh, @junyoungch, @NateKushman, @Mononofu, @RemiLeblond, Tom Eccles, James Keeling, @FelixAxelGimeno, Agustin Dal Lago, Thomas Hubert, Peter Choy, Cyprien de Masson d’Autume, @ibab_ml, @xinyun_chen_, Po-Sen Huang, Johannes Welbl, Sven Gowal...
Alexey Cherepanov, James Molloy, Daniel J. Mankowitz, Esme Sutherland Robson, @pushmeet, @NandoDF, @koraykv and @OriolVinyalsML.

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with DeepMind

DeepMind Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @DeepMind

Dec 9
Introducing Dramatron, a new tool for writers to co-write theatre and film scripts with a language model. 🎭

Dramatron can interactively co-create new stories complete with title, characters, location descriptions and dialogue.

Try it yourself now: dpmd.ai/dramatron-gith…
✏️ We interviewed 15 industry experts including playwrights, screenwriters and actors who produced work using Dramatron.

Canadian company @theatresports edited co-written theatre scripts and performed them on stage in Plays By Bots to positive reviews. dpmd.ai/dramatron-tw ImageImage
Want to find out more? The team will be presenting this research at #NeurIPS2022:

📅 December 9
⌚ 3pm CST

dpmd.ai/3YbA0nK @ML4CDworkshop
Read 4 tweets
Dec 8
Introducing a generalist neural algorithmic learner, capable of carrying out 30 different reasoning tasks, with a 𝘴𝘪𝘯𝘨𝘭𝘦 graph network.

These include:
🔵 Sorting
🔵 Shortest paths
🔵 String matching
🔵 Convex hull finding

And more: dpmd.ai/3FC1FqA
What makes this the first-of-its-kind? 🤔

Similar to Gato, an AI agent capable of many tasks such as language modelling and image captioning, this model covers a diverse set of algorithm domains, matching single-task specialists' power - and remaining robust out-of-distribution.
Our team will present this work at @LogConference as a spotlight talk:
📅 9 December
⌚ 5.40pm GMT

Want to join us? Save your spot now: dpmd.ai/3FxKKFz
Read 4 tweets
Dec 1
Introducing DeepNash - the first AI to master Stratego, a game of hidden information which is more complex than chess, Go and poker.

Published in @ScienceMagazine, it uses a novel model-free reinforcement learning algorithm.

Here's how it works. 🧵 dpmd.ai/deepnash-tw
🧩 Stratego is a game of imperfect information: players can't directly observe their opponent's pieces.

This makes it hard for other AI-based systems to go beyond amateur level. It also means that a successful technique called “game tree search” is not sufficiently scalable. Image
🟩 However, DeepNash’s technique approximates a Nash equilibrium, which makes it very hard for opponents to exploit its playing style.

As a result, it’s reached an all-time top-three ranking on Gravon, the world’s biggest online Stratego platform. dpmd.ai/deepnash-scien…
Read 6 tweets
Oct 20
From helping transform the way we can predict wind power output to improving the performance of Document AI and giving wider access to #AlphaFold, we’re proud to partner with @GoogleCloud to bring our AI research into the real world.

Here’s how. 🧵
1️⃣ Industries aiming to use AI to read documents need lots of training data, which can be hard to find.

We helped develop machine learning models that need 50% less data to parse utility bills and purchase orders for @GoogleCloud Document AI users.
2️⃣ Wind farms are crucial to building a carbon-free future - but the weather makes it hard to predict.

With @GoogleCloud, we helped produce a Custom AI tool to better predict how much wind power could be generated - trained on forecasts and a customer’s historical turbine data.
Read 4 tweets
Jul 14
Working together with @YouTube’s product and engineering teams, we've helped optimise the decision-making processes that increase safety, decrease latency, and enhance the viewer, creator, and advertiser experience for all.

How? A 🧵...

dpmd.ai/dm-youtube 1/
Our team developed a label quality model that helps label videos with greater precision according to @YouTube’s ad friendly guidelines, improving how videos are identified and classified. 2/
We applied #MuZero to improve the VP9 codec, a coding format that helps compress & transmit video over the internet. It was then applied to some of @YouTube’s live traffic - resulting in an avg. 4% bitrate reduction, helping reduce internet traffic, data usage & loading times. 3/
Read 4 tweets
Apr 21
Recently the DeepMind Jax Ecosystem was joined by four new libraries: Mctx, KFAC-JAX, DM_AUX, and TF2JAX. See the thread 🧵 for more detail.

We hope that this fosters new research and applications in the ML community! 1/
Mctx provides AlphaZero and MuZero Monte Carlo tree search: dpmd.ai/mctx

Work by @fabiointheuk, Ivo Danihelka, @matteohessel, and @laurentsifre, with support from many others. 2/
KFAC-JAX is a library for second-order optimisation of neural networks, and for computing scalable curvature approximations (such as the one used in K-FAC): dpmd.ai/kfac-jax

Work by Aleksandar Botev and James Martens. 3/
Read 5 tweets

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Don't want to be a Premium member but still want to support us?

Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal

Or Donate anonymously using crypto!

Ethereum

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Bitcoin

3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us on Twitter!

:(