Google AI and DeepMind News and Discussions

[Yuli Ban]

Programming has been for a long time a high-status, high-demand skill.
Companies and businesses across industries depend at a very foundational level on the ability of human developers: People who write and understand the language of computers. Recently, with the advent of large language models, AI companies have begun to explore the possibilities of systems that can learn to code. OpenAI’s Codex — embedded into GitHub Copilot — was the first notable example. Codex can read simple natural language commands and instructions and write code that matches the intention of the user.
Yet, writing small programs and solving easy tasks is “far from the full complexity of real-world programming.” AI models like Codex lack the problem-solving skills that most programmers rely on in their day-to-day jobs. That’s the gap DeepMind wanted to fill with AlphaCode, an AI system that has been trained to “understand” natural language, design algorithms to solve problems, and then implement them into code.
AlphaCode displays a unique skillset of natural language understanding and problem-solving ability, combined with the statistical power characteristic of large language models. The system was tested against human programmers on the popular competitive programming platform Codeforces. AlphaCode averaged a ranking of 54.3% across 10 contests, which makes it the first AI to reach the level of human programmers in competitive programming contests.

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

Restoring, placing, and dating ancient texts through collaboration between AI and historians.

The birth of human writing marked the dawn of History and is crucial to our understanding of past civilisations and the world we live in today. For example, more than 2,500 years ago, the Greeks began writing on stone, pottery, and metal to document everything from leases and laws to calendars and oracles, giving a detailed insight into the Mediterranean region. Unfortunately, it’s an incomplete record. Many of the surviving inscriptions have been damaged over the centuries or moved from their original location. In addition, modern dating techniques, such as radiocarbon dating, cannot be used on these materials, making inscriptions difficult and time-consuming to interpret.

In line with DeepMind’s mission of solving intelligence to advance science and humanity, we collaborated with the Department of Humanities of Ca' Foscari University of Venice, the Classics Faculty of the University of Oxford, and the Department of Informatics of the Athens University of Economics and Business to explore how machine learning can help historians better interpret these inscriptions – giving a richer understanding of ancient history and unlocking the potential for cooperation between AI and historians.

[Yuli Ban]

We investigate the optimal model size and number of tokens for training a transformer language model under a given compute budget. We find that current large language models are significantly undertrained, a consequence of the recent focus on scaling language models whilst keeping the amount of training data constant. By training over \nummodels language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, we find that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled. We test this hypothesis by training a predicted compute-optimal model, \chinchilla, that uses the same compute budget as \gopher but with 70B parameters and 4× more more data. \chinchilla uniformly and significantly outperforms \Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks. This also means that \chinchilla uses substantially less compute for fine-tuning and inference, greatly facilitating downstream usage. As a highlight, \chinchilla reaches a state-of-the-art average accuracy of 67.5\% on the MMLU benchmark, greater than a 7\% improvement over \gopher.

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

One key aspect of intelligence is the ability to quickly learn how to perform a new task when given a brief instruction. For instance, a child may recognise real animals at the zoo after seeing a few pictures of the animals in a book, despite any differences between the two. But for a typical visual model to learn a new task, it must be trained on tens of thousands of examples specifically labelled for that task. If the goal is to count and identify animals in an image, as in “three zebras”, one would have to collect thousands of images and annotate each image with their quantity and species. This process is inefficient, expensive, and resource-intensive, requiring large amounts of annotated data and the need to train a new model each time it’s confronted with a new task. As part of DeepMind’s mission to solve intelligence, we’ve explored whether an alternative model could make this process easier and more efficient, given only limited task-specific information.

Today, in the preprint of our paper, we introduce Flamingo, a single visual language model (VLM) that sets a new state of the art in few-shot learning on a wide range of open-ended multimodal tasks. This means Flamingo can tackle a number of difficult problems with just a handful of task-specific examples (in a “few shots”), without any additional training required. Flamingo’s simple interface makes this possible, taking as input a prompt consisting of interleaved images, videos, and text and then output associated language.

Similar to the behaviour of large language models (LLMs), which can address a language task by processing examples of the task in their text prompt, Flamingo’s visual and text interface can steer the model towards solving a multimodal task. Given a few example pairs of visual inputs and expected text responses composed in Flamingo’s prompt, the model can be asked a question with a new image or video, and then generate an answer.

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

[Yuli Ban]

[funkervogt]

DeepMind builds an AI that is neither narrow nor general. "Narrowly general" or "multipurpose" might be the right term.

Inspired by progress in large-scale language modeling, we apply a similar approach towards building a
single generalist agent beyond the realm of text outputs. The agent, which we refer to as Gato, works as a
multi-modal, multi-task, multi-embodiment generalist policy. The same network with the same weights
can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based
on its context whether to output text, joint torques, button presses, or other tokens. In this report we
describe the model and the data, and document the current capabilities of Gato.

https://storage.googleapis.com/deepmind ... 0Agent.pdf

[Yuli Ban]

DeepMind builds an AI that is neither narrow nor general. "Narrowly general" or "multipurpose" might be the right term.

If my term had ever taken off, we'd have "artificial expert intelligence" to describe it. That type of AI in between narrow and general AI.

And yes, this is an astounding step forward towards general AI, though it's not quite AGI itself just yet.

As Károly Zsolnai-Fehér would say, hold onto your papers! We're maybe one or two papers away from proto-AGI. Indeed, my timeline is changing rapidly after these past few months. Before, I thought 2024 was the best year for it to happen. Even after PaLM, Flamingo, and DALL-E 2, I was resolute that it would take that long.

Now with Gato, I see that all we need is scale. If they can scale it up quickly enough, there's zero reason why proto-AGI could not be realized this year.

I'm not saying it will. I thought GPT-4 would have been announced by now, and even Starspawn0 said that something like a primitive AGI could've been possible before the end of 2020 based on existing technology. Things can happen that push things back, so even if it's possible doesn't mean it absolutely will happen.

[Yuli Ban]

[Yuli Ban]

Have you ever seen anything that is on the surface innocuous but, when you think of the implications, it sends chills down your spine?

DeepMind is hiring for the Scalable Alignment and Alignment Teams

We are hiring for several roles in the Scalable Alignment and Alignment Teams at DeepMind, two of the subteams of DeepMind Technical AGI Safety trying to make artificial general intelligence go well.  In brief,

• The Alignment Team investigates how to avoid failures of intent alignment, operationalized as a situation in which an AI system knowingly acts against the wishes of its designers.  Alignment is hiring for Research Scientist and Research Engineer positions.
• The Scalable Alignment Team (SAT) works to make highly capable agents do what humans want, even when it is difficult for humans to know what that is.  This means we want to remove subtle biases, factual errors, or deceptive behaviour even if they would normally go unnoticed by humans, whether due to reasoning failures or biases in humans or due to very capable behaviour by the agents.  SAT is hiring for Research Scientist - Machine Learning, Research Scientist - Cognitive Science, Research Engineer, and Software Engineer positions.

We elaborate on the problem breakdown between Alignment and Scalable Alignment next, and discuss details of the various positions.