The Ensembl project (http://www.ensembl.org) is a system for genome annotation, analysis, storage and dissemination designed to facilitate the access of genomic annotation from chordates and key model organisms. It provides access to data from 87 species across our main and early access Pre! websites. This year we introduced three newly annotated species and released numerous updates across our supported species with a concentration on data for the latest genome assemblies of human, mouse, zebrafish and rat. We also provided two data updates for the previous human assembly, GRCh37, through a dedicated website (http://grch37.ensembl.org). Our tools, in particular the VEP, have been improved significantly through integration of additional third party data. REST is now capable of larger-scale analysis and our regulatory data BioMart can deliver faster results. The website is now capable of displaying long-range interactions such as those found in cis-regulated datasets. Finally we have launched a website optimized for mobile devices providing views of genes, variants and phenotypes. Our data is made available without restriction and all code is available from our GitHub organization site (http://github.com/Ensembl) under an Apache 2.0 license.

Agents should avoid unsafe behaviour during both training and deployment. This typically requires a simulator and a procedural specification of unsafe behaviour. Unfortunately, a simulator is not always available, and procedurally specifying constraints can be difficult or impossible for many real-world tasks. A recently introduced technique, ReQueST, aims to solve this problem by learning a neural simulator of the environment from safe human trajectories, then using the learned simulator to efficiently learn a reward model from human feedback. However, it is yet unknown whether this approach is feasible in complex 3D environments with feedback obtained from real humans -whether sufficient pixel-based neural simulator quality can be achieved, and whether the human data requirements are viable in terms of both quantity and quality. In this paper we answer this question in the affirmative, using ReQueST to train an agent to perform a 3D first-person object collection task using data entirely from human contractors. We show that the resulting agent exhibits an order of magnitude reduction in unsafe behaviour compared to standard reinforcement learning.

DeepMind, * Work done at DeepMind. Interpretability research aims to build tools for understanding machine learning (ML) models. However, such tools are inherently hard to evaluate because we do not have ground truth information about how ML models actually work. In this work, we propose to build transformer models manually as a testbed for interpretability research. We introduce Tracr, a "compiler" for translating human-readable programs into weights of a transformer model. Tracr takes code written in RASP, a domain-specific language (Weiss et al., 2021), and translates it into weights for a standard, decoder-only, GPT-like transformer architecture. We use Tracr to create a range of ground truth transformers that implement programs including computing token frequencies, sorting, and Dyck-n parenthesis checking, among others. We study the resulting models and discuss how this approach can accelerate interpretability research. To enable the broader research community to explore and use compiled models, we provide an open-source implementation of Tracr at https://github.com/deepmind/tracr.

I had no idea, getting interested in AI two years ago, that being involved in the field would involve such a persistent sense of unease. I started out unequivocally excited, perhaps a little naive; but over the years, the concerned voices of economists, philosophers, and the mass media have gradually seeped into me, leaving me with an ill-defined feeling of hesitation about what we're heading towards.