Many artificial intelligence researchers see the possible future development of superhuman AI as having a non-trivial chance of causing human extinction – but there is also widespread disagreement and uncertainty about such risks.

Those findings come from a survey of 2700 AI researchers who have recently published work at six of the top AI conferences – the largest such survey to date. The survey asked participants to share their thoughts on possible timelines for future AI technological milestones, as well as the good or bad societal consequences of those achievements. Almost 58 per cent of researchers said they considered that there is a 5 per cent chance of human extinction or other extremely bad AI-related outcomes.

“It’s an important signal that most AI researchers don’t find it strongly implausible that advanced AI destroys humanity,” says Katja Grace at the Machine Intelligence Research Institute in California, an author of the paper. “I think this general belief in a non-minuscule risk is much more telling than the exact percentage risk.”

But there is no need to panic just yet, says Émile Torres at Case Western Reserve University in Ohio. Such AI expert surveys “don’t have a good track record” of forecasting future AI developments, they say. A 2012 study showed that in the long run, AI expert predictions were no more accurate than non-expert public opinion. This new survey’s authors also pointed out that AI researchers are not experts in forecasting the future trajectory of AI.

Compared with answers from a 2022 version of the same survey, many AI researchers predicted that AI will hit certain milestones earlier than previously predicted. This coincides with the November 2022 debut of ChatGPT and Silicon Valley’s rush to widely deploy similar AI chatbot services based on large language models.

The surveyed researchers predicted that within the next decade, AI systems have a 50 per cent or higher chance of successfully tackling most of 39 sample tasks, including writing new songs indistinguishable from a Taylor Swift banger or coding an entire payment processing site from scratch. Other tasks such as physically installing electrical wiring in a new home or solving longstanding mathematics mysteries are expected to take longer.

The possible development of AI that can outperform humans on every task was given 50 per cent odds of happening by 2047, whereas the possibility of all human jobs becoming fully automatable was given 50 per cent odds to occur by 2116. These estimates are 13 years and 48 years earlier than those given in last year’s survey.

But the heightened expectations regarding AI development may also fall flat, says Torres. “A lot of these breakthroughs are pretty unpredictable. And it’s entirely possible that the field of AI goes through another winter,” he says, referring to the drying up of funding and corporate interest in AI during the 1970s and 80s.

There are also more immediate worries without any superhuman AI risks. Large majorities of AI researchers – 70 per cent or more – described AI-powered scenarios involving deepfakes, manipulation of public opinion, engineered weapons, authoritarian control of populations and worsening economic inequality to be of either substantial or extreme concern. Torres also highlighted the dangers of AI contributing to disinformation around existential issues such as climate change or worsening democratic governance.

“We already have the technology, here and now, that could seriously undermine [the US] democracy,” says Torres. “We’ll see what happens in the 2024 election.”

Topics:

Google has launched a new AI model, dubbed Gemini, which it claims can outperform both OpenAI’s GPT-4 model and “expert level” humans in a range of intelligence tests.

The firm’s CEO, Sundar Pichai, revealed the existence of Gemini at Google’s I/O conference in May this year, although it was still in training at the time. But today the company has announced that it will be launching the cutting-edge model to the public.

Three versions of Gemini have been created for different applications, called Nano, Pro and Ultra, which increase in size and capability. Google declined to answer questions on the size of Pro and Ultra, the number of parameters they include or the scale or source of their training data. But its smallest version, Nano, which is designed to run locally on smartphones, is actually two models: one for slower phones that has 1.8 billion parameters and one for more powerful devices that has 3.25 billion parameters. Comparing the capabilities of AI models is an inexact science, but GPT-4 is rumoured to include up to 1.7 trillion parameters and Meta’s LLAMA-2 has 70 billion.

The mid-range Pro version of Gemini beats some other models, such as OpenAI’s GPT3.5, but the more powerful Ultra exceeds the capability of all existing AI models, Google claims. It scored 90 per cent on the industry-standard MMLU benchmark, where an “expert level” human is expected to achieve 89.8 per cent.

This is the first time an AI has beaten humans at the test, and is the highest score for any existing model. The test involves a broad range of tricky questions on topics including logical fallacies, moral problems in everyday scenarios, medical issues, economics and geography.

In the same test, GPT-4 scored 87 per cent, LLAMA-2 scored 68 per cent and Anthropic’s Claude 2 scored 78.5 per cent. Gemini beat all those models in eight out of nine other common benchmark tests.

The Pro model will be integrated into Google’s Bard, an online chatbot that was launched in March this year. The company says that another version of Bard called Bard Advanced will launch early next year and feature the larger Gemini Ultra model.

The new version of Bard will be available in English in more than 170 countries as of today, but it won’t be available in other languages or even in English across the UK and Europe. Sissie Hsiao at Google says the delay is down to regulation rather than engineering: “We’re working with local policies and regulators to make sure that we’re abiding by local laws and other such things before we launch in other areas.”

Eli Collins at Google DeepMind says Gemini is the company’s largest and most capable model, but also its most general – meaning it is adaptable to a variety of tasks. Unlike many current models that focus on text, Gemini has been trained on text, images and sound and is claimed to be able to accept inputs and provide outputs in all those formats. But the Bard launch will only allow people to use text prompts as of today, with the company promising to allow audio and image interaction “in coming months”.

Collins says that Gemini is “state of the art in nearly every domain” and that it is still in testing to determine exactly how capable it is at working in different mediums, languages and applications. “We’re still working to understand all of Ultra’s novel capabilities,” he says.

No versions of Gemini were made available for testing at the launch event, but Google showed demonstrations of the AI solving homework problems and working with live video input. It is also claimed to be better at developing software than previous models: last year, DeepMind released an AI-powered code generator called AlphaCode that the firm said could beat 50 per cent of human developers, and it is now releasing an updated version powered by Gemini that it claims can beat 85 per cent of human coders.

Topics:

A single artificial intelligence can beat human players in chess, Go, poker and other games that require a variety of strategies to win. The AI, called Student of Games, was created by Google DeepMind, which says it is a step towards an artificial general intelligence capable of carrying out any task with superhuman performance.

Martin Schmid, who worked at DeepMind on the AI but who is now at a start-up called EquiLibre Technologies, says that the Student of Games (SoG) model can trace its lineage back to two projects. One was DeepStack, the AI created by a team including Schmid at the University of Alberta in Canada and which was the first to beat human professional players at poker. The other was DeepMind’s AlphaZero, which has beaten the best human players at games like chess and Go.

The difference between those two models is that one focused on imperfect-knowledge games – those where players don’t know the state of all other players, such as their hands in poker – and one focused on perfect-knowledge games like chess, where both players can see the position of all pieces at all times. The two require fundamentally different approaches. DeepMind hired the whole DeepStack team with the aim of building a model that could generalise across both types of game, which led to the creation of SoG.

Schmid says that SoG begins as a “blueprint” for how to learn games, and then improve at them through practice. This starter model can then be set loose on different games and teach itself how to play against another version of itself, learning new strategies and gradually becoming more capable. But while DeepMind’s previous AlphaZero could adapt to perfect-knowledge games, SoG can adapt to both perfect and imperfect-knowledge games, making it far more generalisable.

The researchers tested SoG on chess, Go, Texas hold’em poker and a board game called Scotland Yard, as well as Leduc hold’em poker and a custom-made version of Scotland Yard with a different board, and found that it could beat several existing AI models and human players. Schmid says it should be able learn to play other games as well. “There’s many games that you can just throw at it and it would be really, really good at it.”

This wide-ranging ability comes at a slight cost in performance compared with DeepMind’s more specialised algorithms, but SoG can nonetheless easily beat even the best human players at most games it learns. Schmid says that SoG learns to play against itself in order to improve at games, but also to explore the range of possible scenarios from the present state of a game – even if it is playing an imperfect-knowledge one.

“When you’re in a game like poker, it’s so much harder to figure out; how the hell am I going to do search [for the best strategic next move in a game] if I don’t know what cards the opponent holds?” says Schmid. “So there was some some set of ideas coming from AlphaZero, and some set of ideas coming from DeepStack into this big big mix of ideas, which is Student of Games.”

Michael Rovatsos at the University of Edinburgh, UK, who wasn’t involved in the research, says that while impressive, there is still a very long way to go before an AI can be thought of as generally intelligent, because games are settings in which all rules and behaviours are clearly defined, unlike the real world.

“The important thing to highlight here is that it’s a controlled, self-contained, artificial environment where what everything means, and what the outcome of every action is, is crystal clear,” he says. “The problem is a toy problem because, while it may be very complicated, it’s not real.”

Topics:

People tend to cut corners and allow trusted colleagues to pick up the slack when working as a team, in a phenomenon known as social loafing. Now researchers have found that the same thing happens when humans work with robots.

Dietlind Helene Cymek at the Technical University of Berlin in Germany and her colleagues designed an experiment to test whether humans would put in less effort when they think that their personal contribution to a task won’t be noticed. …

Compared to other AI chatbots, GPT-4 performs best on a test of legal reasoning – but it still falls short of the knowledge required for human lawyers. Early attempts to use AI chatbots in courtrooms have sometimes proven disastrous, and this finding adds to evidence that AI isn’t ready to handle the complexities of real-world legal arguments.

Artificial intelligence researchers and lawyers worked together to design LegalBench, which evaluates how well AI chatbots …

Bots are better and faster than humans at online CAPTCHA tests designed to keep them out of websites. The finding calls into question whether we should continue using these kind of security measures, given how much frustration they can cause for people.

A cat-and-mouse game has played out for almost two decades between website developers who want to keep bots out of their sites and the hackers who want to bypass those protections …

An automated robot gardener has shown its nurturing side by matching humans in trials of growing vegetables – and it did it with more efficient use of water.

Simeon Adebola at the University of California, Berkeley, and his colleagues have developed an automated plot, called AlphaGarden, that contains a mix of plant types and tested whether it can perform as well as a team of six expert human horticulturalists, each with 10 years of gardening experience, on average.

The robot and human-run plots both contained pairs of eight different common edible vegetables, such as kale, chard and radicchio. Where the humans …

The artificial intelligence chatbot ChatGPT could automate some aspects of AI training that currently rely on people. The chatbot can accurately classify and label text data used in training AI at a cost of just a third of a cent per work sample – about 20 times cheaper than crowdsourced human labour.

“We expect that in some tasks ChatGPT could replace humans,” says Fabrizio Gilardi at the University of Zürich in Switzerland. In other tasks, ChatGPT could help reduce the “amount of …

ChatGPT-powered AIs given long-term memory capabilities and personal motivations could role-play characters in a simulated town more believably than human crowd workers.

“This idea of creating believable agents that actually exhibit this behaviour – that give the illusion of realism – was something that we as an academic field wanted and have been talking about for the last four decades,” says Joon Sung Park at Stanford University in California.

Park and his colleagues from Stanford and Google developed and tested AI agents …

Robotic dogs can spot invasive fire ant nests more efficiently than human searchers – and they can safely stir up a swarm of aggressive ants by poking their nests, which can help researchers identify the notorious pests.

The red imported fire ant (Solenopsis invicta) packs a venomous sting, and causes billions of dollars in ecological and agricultural damage across the world each year. To combat the threat, researchers have been testing how a robotic dog combined with artificial intelligence can help detect the ants’ …