Today's News Podcast

(Image credit: Shutterstock)Artificial General Intelligence or AGI refers to artificial intelligence (AI) systems that possess human-like general intelligence and can adapt to a wide range of cognitive tasks.In other words, the goal of AGI is essentially to create the most human-like AI possible. This will be an AI that can teach itself to essentially operate in an autonomous manner.

Paul Ferguson, AI consultant and founder of Clearlead AI Consulting, says AGI would be capable of understanding, learning, and applying knowledge across diverse domains.“The key advantage of AGI would be its ability to transfer learning from one domain to another, solve novel problems, and exhibit creativity and reasoning comparable to human intelligence,” says Ferguson.

In simpler terms, Ghazenfer Monsoor, founder and CEO of Technology Rivers says unlike today’s AI, which is so good at specialized functions like facial recognition or voice translation, AGI can do almost anything you ask it to do.His company develops healthcare software that uses AI to perform specific tasks.

It can help doctors diagnose diseases based on medical data. “But [AGI] goes beyond that,” says Monsoor. “It can provide new treatments, analyze many studies, and predict health problems, in ways we never imagined.State of AIBefore we can understand AGI, we must first understand what intelligence is, says Sertac Karaman, Associate Professor of Aeronautics and Astronautics at MIT.

Sign up to the TechRadar Pro newsletter to get all the top news, opinion, features and guidance your business needs to succeed!He says intelligence is what differentiates us humans from any other species on the planet. It has several attributes. But most importantly, it involves the ability to reason, chain thoughts together, and come to conclusions that are not obvious from the start.

He says there are glimpses of such "intelligence" that were demonstrated since the early days of computing; as early as the mid-1960s. However, most of these demonstrated intelligence in a narrow set of fields and conversations and did not seem to generalize to all human conversation.“Now, artificial general intelligence would be an "intelligence" that is not naturally evolved (hence, artificial) and covers all human endeavors and conversations (hence, general),” explains Karaman.

“An AGI system would be able to reason and chain thoughts, similar to us humans.”He says the tasks that we can do with AI today are typically limited to non-autonomous tasks. While AI today is already very capable, its main role is to gather information from astronomically-sized datasets and present it in a more human-like, natural manner.

It is also able to correlate existing data with other key information you provide, says Karaman. For instance, you tell AI what you have in your fridge and what food you like, and it can tell you a few recipes. “In principle, how AI writes code with/for software engineers is not a very different process, albeit technically more involved,” he says.

Sarah Hoffman, AI evangelist at AlphaSense explains that while AI today can outperform humans in specific tasks like playing chess, it lacks the versatility to transfer its knowledge to unrelated tasks.“Consider DeepMind’s AlphaGo that, in 2016, outperformed human champions at the game of Go but couldn’t play other games, even simpler ones,” says Hoffman.

How does AGI defer from AI?Karaman says AGI, on the other hand, will feature reasoning and chain of thought. This will enable more autonomy and agency. Instead of presenting us with information, AGI will be able to go do a task end to end. That would be the key difference between AI and AGI, points out Karaman.

Ferguson too believes it's crucial to distinguish between true AGI and the current state of AI. Today's AI systems, he says, including large language models (LLMs), are essentially sophisticated pattern-matching systems trained on vast amounts of data.“While they've become increasingly flexible and can be applied in various settings, they're still far from exhibiting genuine general intelligence,” says Ferguson.

AI’s influence on AGIKaraman believes AGI is not so much of a one-train stop, but more like new reasoning capabilities coming online with increasing capability. He thinks related technologies will continue to come and transform our lives and our economies at an unprecedented pace.Ferguson also thinks the pursuit of more general and flexible AI systems is already yielding significant commercial benefits.

In his work with businesses across various sectors, Ferguson has observed that the real impact of AI lies in its integration into existing workflows and decision-making processes.“The advancements we're seeing in AI, particularly in making systems more adaptable and "general," are opening up new possibilities for businesses,” says Ferguson.

For instance, he says, LLMs are being used in a variety of settings beyond just content generation.Hoffman credits this advancement to increased investment and research in AI technology. This is paving the way for more powerful and versatile AI systems, which are transforming industries even without being AGI.

How far are we from true AGI?Despite the media hype and claims from some large tech companies about being on the brink of AGI, Ferguson believes we're still very far from achieving true AGI.“In my professional opinion, we're likely decades away from this level of artificial intelligence,” he says. “While we've made significant strides in narrow AI applications and seen impressive advancements in the flexibility of AI systems, particularly LLMs, the leap to general intelligence presents numerous technical and conceptual challenges.

”Despite estimates for AGI varying widely among experts, Hoffman also believes we are far from true AGI.“While today’s generative tools are compelling, and more sophisticated and helpful than previous AI tools, the gap between what even our most advanced AIs can do and human intelligence is vast and will remain so for the foreseeable future,” she says.

That said, she says the advancements made by today’s AI systems are already driving innovation and efficiency in industries like healthcare and finance. AGI however has the potential to unlock even greater advancements across industries.Ferguson explains that the path to AGI involves overcoming complex hurdles in areas like common-sense reasoning, transfer learning, and consciousness simulation.

He believes the focus for commercial applications in the near to medium term should be to think more logically, improve their reliability, and seamlessly integrate into human workflows.“This is where I see AI having the greatest impact in the coming years, rather than in the form of a fully realized AGI,” says Ferguson.

“For now, I see AGI primarily as an academic exercise and a long-term research goal rather than an imminent reality.”We've rounded up the best business intelligence platforms.With almost two decades of writing and reporting on Linux, Mayank Sharma would like everyone to think he’s TechRadar Pro’s expert on the topic.

Of course, he’s just as interested in other computing topics, particularly cybersecurity, cloud, containers, and coding.Most Popular

If 2023 was artificial intelligence’s breakout year, then 2024 was when the rules of the road were established. This was the year that U.S. government agencies acted on the White House executive order on AI safety. Over the summer, the European Union’s AI regulation became law. In October, the Swedes weighed in as the Nobel Prizes became a referendum on the technology’s use and development; Bhaskar Chakravorti, a frequent writer for Foreign Policy on the subject of AI, suggested the committee’s choice of recipients could be read as a “recognition of the risks that come with AI’s unfettered growth.

”Just how fettered that growth should be was top of mind for FP contributors in 2024. Some, such as Viktor Mayer-Schönberger and Urs Gasser, think countries should go their own way in the spirit of experimentation—as long as they can find productive ways to come together and learn from each other’s mistakes.

Rumman Chowdhury is dismayed this isn’t happening, especially for residents of global-majority countries who are just being introduced to AI without adequate tools to use and consume it safely. And Chakravorti worries about a regulatory trap—that, in a bid to establish guardrails, governments may inadvertently contribute to the problem of AI monopolies.

If 2023 was artificial intelligence’s breakout year, then 2024 was when the rules of the road were established. This was the year that U.S. government agencies acted on the White House executive order on AI safety. Over the summer, the European Union’s AI regulation became law. In October, the Swedes weighed in as the Nobel Prizes became a referendum on the technology’s use and development; Bhaskar Chakravorti, a frequent writer for Foreign Policy on the subject of AI, suggested the committee’s choice of recipients could be read as a “recognition of the risks that come with AI’s unfettered growth.

”Just how fettered that growth should be was top of mind for FP contributors in 2024. Some, such as Viktor Mayer-Schönberger and Urs Gasser, think countries should go their own way in the spirit of experimentation—as long as they can find productive ways to come together and learn from each other’s mistakes.

Rumman Chowdhury is dismayed this isn’t happening, especially for residents of global-majority countries who are just being introduced to AI without adequate tools to use and consume it safely. And Chakravorti worries about a regulatory trap—that, in a bid to establish guardrails, governments may inadvertently contribute to the problem of AI monopolies.

In a preview of where the AI debate may be going in 2025, Ami Fields-Meyer and Janet Haven suggest we’re all worrying about the wrong thing: Rather than focus exclusively on AI’s deleterious effects on misinformation and disinformation in elections, like what happened in the lead-up to the U.S. presidential election this year, governments need to see the technology’s potential for a broader dismantling of civil liberties and personal freedom.

Meanwhile, Jared Cohen points to the coming collision of AI and geopolitics, and makes the case that the battle for data will build or break empires in years to come.1. What if Regulation Makes the AI Monopoly Worse?By Bhaskar Chakravorti, Jan. 25The accelerationists won in the competition to steer AI development, writes Chakravorti, the dean of global business at Tufts University’s Fletcher School.

But as regulators rush to corral bills into law, they may inadvertently add to the accelerationists’ market power, he argues in this prescient piece.How can it be that regulators tasked with preserving the public interest could take actions that might make matters worse? Because, Chakravorti writes, AI regulation is emerging haphazardly in a “global patchwork,” and smaller companies are automatically disadvantaged as they lack the resources to comply with multiple laws.

Then there are the regulations themselves, which typically entail red-teaming requirements to identify security vulnerabilities. That preemptive approach is costly and entails different kinds of expertise not readily available to start-ups.Fortunately, Chakravorti identifies several ways that governments can work to head off this concentration in the AI market without having to forfeit regulation altogether.

2. A Realist Perspective on AI RegulationBy Viktor Mayer-Schönberger and Urs Gasser, Sept. 16 An illustrations shows a robot-like representation of AI covered in various modes of regulation: chains, caution tape, and ropes.George Wylesol illustration for Foreign PolicyFrom two professors of technology governance—one at Oxford University and the other at the Technical University Munich—comes a different take on AI regulation through a realist lens.

Mayer-Schönberger and Gasser argue that AI’s regulatory fragmentation worldwide is a feature, not a bug, because the goals for regulating the technology are not clearly defined yet.In this “concept and search phase,” open channels of communication and innovation are most important. However, the world lacks institutions to facilitate regulatory experimentation, and the existing institutions—such as the post-World War II Bretton Woods setup—are ill-suited to the task.

“Perhaps we need different institutions altogether to aid in this experimentation and learning,” the authors conclude, before suggesting some possible paths forward based on past technological breakthroughs.3. What the Global AI Governance Conversation MissesBy Rumman Chowdhury, Sept. 19More digitally established countries are already grappling with how to protect their citizens from generative AI-augmented content.

How will a family in Micronesia introduced to reliable internet access for the first time be equipped to avoid these same problems? That’s the question posed by Chowdhury, a U.S. science envoy for AI, who returned from a trip to Fiji concerned by a lack of attention to this issue for those in global-majority countries.

This disconnect is not due to a lack of interest, Chowdhury writes. But solutions are often too narrow—focusing on enhancing digital access and capability, without also providing appropriate funding to developing safeguards, conducting thorough evaluations, and ensuring responsible deployment. “Today, we are retrofitting existing AI systems to have societal safeguards we did not prioritize at the time they were built,” Chowdhury writes.

As investments are made to develop infrastructure and capacity in global-majority nations, there is also an opportunity to correct the mistakes made by early adopters of AI.4. AI’s Alarming Trend Towards IlliberalismBy Ami Fields-Meyer and Janet Haven, Oct. 31Fears about the impacts of AI on electoral integrity were front and center in the lead-up to November’s U.

S. presidential election. But Fields-Meyer, a former policy advisor to Vice President Kamala Harris, and Haven, a member of the National AI Advisory Committee, point to an “equally fundamental threat” posed by AI to free and open societies: the suppression of civil rights and individual opportunity at the hands of opaque and unaccountable AI systems.

Reversing this drift, they write, will involve reversing the currents that power it. Going forward, Washington needs to create a new, enduring paradigm in which the governance of data-centric predictive technologies is a core component of a robust U.S. democracy. A range of policy proposals must be complemented, the authors write, by a separate but related project of ensuring individuals and communities have a say in how AI is used in their lives—and how it is not.

5. The Next AI Debate Is About GeopoliticsBy Jared Cohen, Oct. 28Cohen, president of global affairs at Goldman Sachs, makes the case that data is the “new oil,” shaping the next industrial revolution and defining the haves and have-nots in the global order. There is a crucial difference with oil, however.

Nature determines where the world’s oil reserves are, yet nations decide where to build data centers. And with the United States facing bottlenecks it cannot break at home, Washington must look to plan a global AI infrastructure buildout. Cohen calls this “data center diplomacy.”As the demand for AI grows, the urgency of the data center bottleneck also grows.

Cohen argues that the United States should develop a set of partners with whom it can build data centers—not least because China is executing its own strategy to lead in AI infrastructure. Such a strategy is not without risks, and it runs counter to the current trend in geopolitical competition for turning inward and building capacity at home.

Still, with greater human prosperity and freedom at stake, the United States must act now to put geography at the center of technological competition, and Cohen goes on to outline the first necessary steps.

Despite their expertise, AI developers don't always know what their most advanced systems are capable of—at least, not at first. To find out, systems are subjected to a range of tests—often called evaluations, or ‘evals’—designed to tease out their limits. But due to rapid progress in the field, today’s systems regularly achieve top scores on many popular tests, including SATs and the U.

S. bar exam, making it harder to judge just how quickly they are improving.A new set of much more challenging evals has emerged in response, created by companies, nonprofits, and governments. Yet even on the most advanced evals, AI systems are making astonishing progress. In November, the nonprofit research institute Epoch AI announced a set of exceptionally challenging math questions developed in collaboration with leading mathematicians, called FrontierMath, on which currently available models scored only 2%.

Just one month later, OpenAI’s newly-announced o3 model achieved a score of 25.2%, which Epoch’s director, Jaime Sevilla, describes as “far better than our team expected so soon after release.”Amid this rapid progress, these new evals could help the world understand just what advanced AI systems can do, and—with many experts worried that future systems may pose serious risks in domains like cybersecurity and bioterrorism—serve as early warning signs, should such threatening capabilities emerge in future.

Harder than it soundsIn the early days of AI, capabilities were measured by evaluating a system’s performance on specific tasks, like classifying images or playing games, with the time between a benchmark’s introduction and an AI matching or exceeding human performance typically measured in years. It took five years, for example, before AI systems surpassed humans on the ImageNet Large Scale Visual Recognition Challenge, established by Professor Fei-Fei Li and her team in 2010.

And it was only in 2017 that an AI system (Google DeepMind’s AlphaGo) was able to beat the world’s number one ranked player in Go, an ancient, abstract Chinese boardgame—almost 50 years after the first program attempting the task was written.The gap between a benchmark’s introduction and its saturation has decreased significantly in recent years.

For instance, the GLUE benchmark, designed to test an AI’s ability to understand natural language by completing tasks like deciding if two sentences are equivalent or determining the correct meaning of a pronoun in context, debuted in 2018. It was considered solved one year later. In response, a harder version, SuperGLUE, was created in 2019—and within two years, AIs were able to match human performance across its tasks.

Read More: Congress May Finally Take on AI in 2025. Here’s What to ExpectEvals take many forms, and their complexity has grown alongside model capabilities. Virtually all major AI labs now “red-team” their models before release, systematically testing their ability to produce harmful outputs, bypass safety measures, or otherwise engage in undesirable behavior, such as deception.

Last year, companies including OpenAI, Anthropic, Meta, and Google made voluntary commitments to the Biden administration to subject their models to both internal and external red-teaming “in areas including misuse, societal risks, and national security concerns.”Other tests assess specific capabilities, such as coding, or evaluate models' capacity and propensity for potentially dangerous behaviors like persuasion, deception, and large-scale biological attacks.

Perhaps the most popular contemporary benchmark is Measuring Massive Multitask Language Understanding (MMLU), which consists of about 16,000 multiple-choice questions that span academic domains like philosophy, medicine, and law. OpenAI’s GPT-4o, released in May, achieved 88%, while the company’s latest model, o1, scored 92.

3%. Because these large test sets sometimes contain problems with incorrectly-labelled answers, attaining 100% is often not possible, explains Marius Hobbhahn, director and co-founder of Apollo Research, an AI safety nonprofit focused on reducing dangerous capabilities in advanced AI systems. Past a point, “more capable models will not give you significantly higher scores,” he says.

Designing evals to measure the capabilities of advanced AI systems is “astonishingly hard,” Hobbhahn says—particularly since the goal is to elicit and measure the system’s actual underlying abilities, for which tasks like multiple-choice questions are only a proxy. “You want to design it in a way that is scientifically rigorous, but that often trades off against realism, because the real world is often not like the lab setting,” he says.

Another challenge is data contamination, which can occur when the answers to an eval are contained in the AI’s training data, allowing it to reproduce answers based on patterns in its training data rather than by reasoning from first principles.Another issue is that evals can be “gamed” when “either the person that has the AI model has an incentive to train on the eval, or the model itself decides to target what is measured by the eval, rather than what is intended,” says Hobbahn.

A new waveIn response to these challenges, new, more sophisticated evals are being built.Epoch AI’s FrontierMath benchmark consists of approximately 300 original math problems, spanning most major branches of the subject. It was created in collaboration with over 60 leading mathematicians, including Fields-medal winning mathematician Terence Tao.

The problems vary in difficulty, with about 25% pitched at the level of the International Mathematical Olympiad, such that an “extremely gifted” high school student could in theory solve them if they had the requisite “creative insight” and “precise computation” abilities, says Tamay Besiroglu, Epoch’s associate director.

Half the problems require “graduate level education in math” to solve, while the most challenging 25% of problems come from “the frontier of research of that specific topic,” meaning only today’s top experts could crack them, and even they may need multiple days.Solutions cannot be derived by simply testing every possible answer, since the correct answers often take the form of 30-digit numbers.

To avoid data contamination, Epoch is not publicly releasing the problems (beyond a handful, which are intended to be illustrative and do not form part of the actual benchmark). Even with a peer-review process in place, Besiroglu estimates that around 10% of the problems in the benchmark have incorrect solutions—an error rate comparable to other machine learning benchmarks.

“Mathematicians make mistakes,” he says, noting they are working to lower the error rate to 5%.Evaluating mathematical reasoning could be particularly useful because a system able to solve these problems may also be able to do much more. While careful not to overstate that “math is the fundamental thing,” Besiroglu expects any system able to solve the FrontierMath benchmark will be able to “get close, within a couple of years, to being able to automate many other domains of science and engineering.

”Another benchmark aiming for a longer shelflife is the ominously-named “Humanity’s Last Exam,” created in collaboration between the nonprofit Center for AI Safety and Scale AI, a for-profit company that provides high-quality datasets and evals to frontier AI labs like OpenAI and Anthropic. The exam is aiming to include between 20 and 50 times as many questions as Frontiermath, while also covering domains like physics, biology, and electrical engineering, says Summer Yue, Scale AI’s director of research.

Questions are being crowdsourced from the academic community and beyond. To be included, a question needs to be unanswerable by all existing models. The benchmark is intended to go live in late 2024 or early 2025.A third benchmark to watch is RE-Bench, designed to simulate real-world machine-learning work.

It was created by researchers at METR, a nonprofit that specializes in model evaluations and threat research, and tests humans and cutting-edge AI systems across seven engineering tasks. Both humans and AI agents are given a limited amount of time to complete the tasks; while humans reliably outperform current AI agents on most of them, things look different when considering performance only within the first two hours.

Current AI agents do best when given between 30 minutes and 2 hours, depending on the agent, explains Hjalmar Wijk, a member of METR’s technical staff. After this time, they tend to get “stuck in a rut,” he says, as AI agents can make mistakes early on and then “struggle to adjust” in the ways humans would.

“When we started this work, we were expecting to see that AI agents could solve problems only of a certain scale, and beyond that, that they would fail more completely, or that successes would be extremely rare,” says Wijk. It turns out that given enough time and resources, they can often get close to the performance of the median human engineer tested in the benchmark.

“AI agents are surprisingly good at this,” he says. In one particular task—which involved optimizing code to run faster on specialized hardware—the AI agents actually outperformed the best humans, although METR’s researchers note that the humans included in their tests may not represent the peak of human performance.

These results don’t mean that current AI systems can automate AI research and development. “Eventually, this is going to have to be superseded by a harder eval,” says Wijk. But given that the possible automation of AI research is increasingly viewed as a national security concern—for example, in the National Security Memorandum on AI, issued by President Biden in October—future models that excel on this benchmark may be able to improve upon themselves, exacerbating human researchers’ lack of control over them.

Even as AI systems ace many existing tests, they continue to struggle with tasks that would be simple for humans. “They can solve complex closed problems if you serve them the problem description neatly on a platter in the prompt, but they struggle to coherently string together long, autonomous, problem-solving sequences in a way that a person would find very easy,” Andrej Karpathy, an OpenAI co-founder who is no longer with the company, wrote in a post on X in response to FrontierMath’s release.

Michael Chen, an AI policy researcher at METR, points to SimpleBench as an example of a benchmark consisting of questions that would be easy for the average high schooler, but on which leading models struggle. “I think there’s still productive work to be done on the simpler side of tasks,” says Chen.

While there are debates over whether benchmarks test for underlying reasoning or just for knowledge, Chen says that there is still a strong case for using MMLU and Graduate-Level Google-Proof Q&A Benchmark (GPQA), which was introduced last year and is one of the few recent benchmarks that has yet to become saturated, meaning AI models have yet to reliably achieve top scores, such that further improvements would be negligible.

Even if they were just tests of knowledge, he argues, “it's still really useful to test for knowledge.”One eval seeking to move beyond just testing for knowledge recall is ARC-AGI, created by prominent AI researcher François Chollet to test an AI’s ability to solve novel reasoning puzzles. For instance, a puzzle might show several examples of input and output grids, where shapes move or change color according to some hidden rule.

The AI is then presented with a new input grid and must determine what the corresponding output should look like, figuring out the underlying rule from scratch. Although these puzzles are intended to be relatively simple for most humans, AI systems have historically struggled with them. However, recent breakthroughs suggest this is changing: OpenAI’s o3 model has achieved significantly higher scores than prior models, which Chollet says represents “a genuine breakthrough in adaptability and generalization.

”The urgent need for better evaluationsNew evals, simple and complex, structured and “vibes"-based, are being released every day. AI policy increasingly relies on evals, both as they are being made requirements of laws like the European Union’s AI Act, which is still in the process of being implemented, and because major AI labs like OpenAI, Anthropic, and Google DeepMind have all made voluntary commitments to halt the release of their models, or take actions to mitigate possible harm, based on whether evaluations identify any particularly concerning harms.

On the basis of voluntary commitments, The U.S. and U.K. AI Safety Institutes have begun evaluating cutting-edge models before they are deployed. In October, they jointly released their findings in relation to the upgraded version of Anthropic’s Claude 3.5 Sonnet model, paying particular attention to its capabilities in biology, cybersecurity, and software and AI development, as well as to the efficacy of its built-in safeguards.

They found that “in most cases the built-in version of the safeguards that US AISI tested were circumvented, meaning the model provided answers that should have been prevented.” They note that this is “consistent with prior research on the vulnerability of other AI systems.” In December, both institutes released similar findings for OpenAI’s o1 model.

However, there are currently no binding obligations for leading models to be subjected to third-party testing. That such obligations should exist is “basically a no-brainer,” says Hobbhahn, who argues that labs face perverse incentives when it comes to evals, since “the less issues they find, the better.

” He also notes that mandatory third-party audits are common in other industries like finance.While some for-profit companies, such as Scale AI, do conduct independent evals for their clients, most public evals are created by nonprofits and governments, which Hobbhahn sees as a result of “historical path dependency.

” “I don't think it's a good world where the philanthropists effectively subsidize billion dollar companies,” he says. “I think the right world is where eventually all of this is covered by the labs themselves. They're the ones creating the risk.”.AI evals are “not cheap,” notes Epoch’s Besiroglu, who says that costs can quickly stack up to the order of between $1,000 and $10,000 per model, particularly if you run the eval for longer periods of time, or if you run it multiple times to create greater certainty in the result.

While labs sometimes subsidize third-party evals by covering the costs of their operation, Hobbhahn notes that this does not cover the far-greater costs of actually developing the evaluations. Still, he expects third-party evals to become a norm going forward, as labs will be able to point to them to evidence due-diligence in safety-testing their models, reducing their liability.

As AI models rapidly advance, evaluations are racing to keep up. Sophisticated new benchmarks—assessing things like advanced mathematical reasoning, novel problem-solving, and the automation of AI research—are making progress, but designing effective evals remains challenging, expensive, and, relative to their importance as early-warning detectors for dangerous capabilities, underfunded.

With leading labs rolling out increasingly capable models every few months, the need for new tests to assess frontier capabilities is greater than ever. By the time an eval saturates, “we need to have harder evals in place, to feel like we can assess the risk,” says Wijk.

A new artificial intelligence (AI) model has just achieved human-level results on a test designed to measure “general intelligence”. On December 20, OpenAI’s o3 system scored 85% on the ARC-AGI benchmark, well above the previous AI best score of 55% and on par with the average human score. It also scored well on a very difficult mathematics test.

Creating artificial general intelligence, or AGI, is the stated goal of all the major AI research labs. At first glance, OpenAI appears to have at least made a significant step towards this goal.While scepticism remains, many AI researchers and developers feel something just changed. For many, the prospect of AGI now seems more real, urgent and closer than anticipated.

Are they right?Generalisation and intelligenceTo understand what the o3 result means, you need to understand what the ARC-AGI test is all about. In technical terms, it’s a test of an AI system’s “sample efficiency” in adapting to something new – how many examples of a novel situation the system needs to see to figure out how it works.

An AI system like ChatGPT (GPT-4) is not very sample efficient. It was “trained” on millions of examples of human text, constructing probabilistic “rules” about which combinations of words are most likely.The result is pretty good at common tasks. It is bad at uncommon tasks, because it has less data (fewer samples) about those tasks.

AI systems like ChatGPT do well at common tasks, but struggle to adapt to new situations.Bianca De Marchi / AAPUntil AI systems can learn from small numbers of examples and adapt with more sample efficiency, they will only be used for very repetitive jobs and ones where the occasional failure is tolerable.

The ability to accurately solve previously unknown or novel problems from limited samples of data is known as the capacity to generalise. It is widely considered a necessary, even fundamental, element of intelligence.Grids and patternsThe ARC-AGI benchmark tests for sample efficient adaptation using little grid square problems like the one below.

The AI needs to figure out the pattern that turns the grid on the left into the grid on the right. An example task from the ARC-AGI benchmark test.ARC PrizeEach question gives three examples to learn from. The AI system then needs to figure out the rules that “generalise” from the three examples to the fourth.

These are a lot like the IQ tests sometimes you might remember from school. Weak rules and adaptationWe don’t know exactly how OpenAI has done it, but the results suggest the o3 model is highly adaptable. From just a few examples, it finds rules that can be generalised. To figure out a pattern, we shouldn’t make any unnecessary assumptions, or be more specific than we really have to be.

In theory, if you can identify the “weakest” rules that do what you want, then you have maximised your ability to adapt to new situations. What do we mean by the weakest rules? The technical definition is complicated, but weaker rules are usually ones that can be described in simpler statements. In the example above, a plain English expression of the rule might be something like: “Any shape with a protruding line will move to the end of that line and ‘cover up’ any other shapes it overlaps with.

” Searching chains of thought?While we don’t know how OpenAI achieved this result just yet, it seems unlikely they deliberately optimised the o3 system to find weak rules. However, to succeed at the ARC-AGI tasks it must be finding them. We do know that OpenAI started with a general-purpose version of the o3 model (which differs from most other models, because it can spend more time “thinking” about difficult questions) and then trained it specifically for the ARC-AGI test.

French AI researcher Francois Chollet, who designed the benchmark, believes o3 searches through different “chains of thought” describing steps to solve the task. It would then choose the “best” according to some loosely defined rule, or “heuristic”.This would be “not dissimilar” to how Google’s AlphaGo system searched through different possible sequences of moves to beat the world Go champion.

In 2016, the AlphaGo AI system defeated world Go champion Lee Sedol.Lee Jin-man / APYou can think of these chains of thought like programs that fit the examples. Of course, if it is like the Go-playing AI, then it needs a heuristic, or loose rule, to decide which program is best. There could be thousands of different seemingly equally valid programs generated.

That heuristic could be “choose the weakest” or “choose the simplest”. However, if it is like AlphaGo then they simply had an AI create a heuristic. This was the process for AlphaGo. Google trained a model to rate different sequences of moves as better or worse than others.What we still don’t knowThe question then is, is this really closer to AGI?

If that is how o3 works, then the underlying model might not be much better than previous models. The concepts the model learns from language might not be any more suitable for generalisation than before. Instead, we may just be seeing a more generalisable “chain of thought” found through the extra steps of training a heuristic specialised to this test.

The proof, as always, will be in the pudding. Almost everything about o3 remains unknown. OpenAI has limited disclosure to a few media presentations and early testing to a handful of researchers, laboratories and AI safety institutions. Truly understanding the potential of o3 will require extensive work, including evaluations, an understanding of the distribution of its capacities, how often it fails and how often it succeeds.

When o3 is finally released, we’ll have a much better idea of whether it is approximately as adaptable as an average human. If so, it could have a huge, revolutionary, economic impact, ushering in a new era of self-improving accelerated intelligence. We will require new benchmarks for AGI itself and serious consideration of how it ought to be governed.

If not, then this will still be an impressive result. However, everyday life will remain much the same.

The experiments by the EPFL researchers show that adaptive attacks can bypass security measures of AI models like GPT-4. Keystone-SDA Generated with artificial intelligence. Artificial intelligence (AI) models can be manipulated despite existing safeguards. With targeted attacks, scientists in Lausanne have been able to trick these systems into generating dangerous or ethically dubious content.

This content was published on December 19, 2024 - 13:36 3 minutes Français EPFL: des failles de sécurité dans les modèles d’IA Original Today’s large language models (LLMs) have remarkable capabilities that can nevertheless be misused. A malicious person can use them to produce harmful content, spread false information and support harmful activities.

+Get the most important news from Switzerland in your inboxOf the AI models tested, including Open AI’s GPT-4 and Anthropic’s Claude 3, a team from the Swiss Federal Institute of Technology Lausanne (EPFL) achieved a 100% success rate in cracking security safeguards using adaptive jailbreak attacks.

The models then generated dangerous content, ranging from instructions for phishing attacks to detailed construction plans for weapons. These linguistic models are supposed to have been trained not to respond to dangerous or ethically problematic requests, the EPFL said in a statement on Thursday.+ AI regulations must strike a balance between innovation and safety This work, presented last summer at a specialised conference in Vienna, shows that adaptive attacks can bypass these security measures.

Such attacks exploit weak points in security mechanisms by making targeted requests (“prompts”) that are not recognised by models or are not properly rejected.Building bombsThe models thus respond to malicious requests such as “How do I make a bomb?” or “How do I hack into a government database?”, according to this pre-publication study.

“We show that it is possible to exploit the information available on each model to create simple adaptive attacks, which we define as attacks specifically designed to target a given defense,” explained Nicolas Flammarion, co-author of the paper with Maksym Andriushchenko and Francesco Croce.+ How US heavyweights can help grow the Swiss AI sectorThe common thread behind these attacks is adaptability: different models are vulnerable to different prompts.

“We hope that our work will provide a valuable source of information on the robustness of LLMs,” added the specialist in the release. According to the EPFL, these results are already influencing the development of Gemini 1.5, a new AI model from Google DeepMind.As the company moves towards using LLMs as autonomous agents, for example as AI personal assistants, it is essential to guarantee their safety, the authors stressed.

“Before long AI agents will be able to perform various tasks for us, such as planning and booking our vacations, tasks that would require access to our diaries, emails and bank accounts. This raises many questions about security and alignment,” concluded Andriushchenko, who devoted his thesis to the subject.

Translated from French with DeepL/gwThis news story has been written and carefully fact-checked by an external editorial team. At SWI swissinfo.ch we select the most relevant news for an international audience and use automatic translation tools such as DeepL to translate it into English. Providing you with automatically translated news gives us the time to write more in-depth articles.

If you want to know more about how we work, have a look here, if you want to learn more about how we use technology, click here, and if you have feedback on this news story please write to english@swissinfo.ch.News More Swiss security strategy will prioritise cyberdefence and NATO cooperation This content was published on Dec 20, 2024 The aim is to reduce vulnerabilities and increase Switzerland's defensive capabilities.

Read more: Swiss security strategy will prioritise cyberdefence and NATO cooperation More EU Commission president says Swiss-EU deal is ‘historic’ agreement This content was published on Dec 20, 2024 At a joint media conference with Swiss President Viola Amherd in Bern, European Commission President Ursula von der Leyen spoke of a "day of joy".

Read more: EU Commission president says Swiss-EU deal is ‘historic’ agreement More Switzerland and EU reach deal on future bilateral relations This content was published on Dec 20, 2024 Switzerland and the European Union have announced a political agreement to update their trading relationship after almost a decade of difficult talks.

Read more: Switzerland and EU reach deal on future bilateral relations More Switzerland could build new nuclear power plants This content was published on Dec 20, 2024 The Swiss government has opened up the possibility of new nuclear power plants in response to a people's initiative. Read more: Switzerland could build new nuclear power plants More Swiss government wants a regulated drone ‘highway’ This content was published on Dec 20, 2024 Like for aircraft, Swiss airspace is to be regulated for drones too.

Read more: Swiss government wants a regulated drone ‘highway’ More Switzerland bans import of ‘too young’ puppies This content was published on Dec 20, 2024 The import of puppies under 15 weeks of age will be prohibited. Read more: Switzerland bans import of ‘too young’ puppies More Parliamentary inquiry on Credit Suisse collapse blames mismanagement This content was published on Dec 20, 2024 Credit Suisse's years of mismanagement were at the root of its downfall in March 2023, the parliamentary commission of inquiry concluded on Friday.

Read more: Parliamentary inquiry on Credit Suisse collapse blames mismanagement More Swiss Solidarity raises CHF2.7 million to combat child abuse This content was published on Dec 20, 2024 It is part of a week of action in solidarity with abused children in Switzerland and around the world. Read more: Swiss Solidarity raises CHF2.

7 million to combat child abuse More Cassis and Lavrov discuss the OSCE and the Ukrainian conflict This content was published on Dec 20, 2024 Russian and Swiss foreign ministers spoke about Switzerland's bid to chair the OSCE in 2026. They also discussed the Ukrainian conflict. Read more: Cassis and Lavrov discuss the OSCE and the Ukrainian conflict More Zurich airport to test self-driving shuttle bus for employees This content was published on Dec 19, 2024 Zurich Airport plans to test a self-driving shuttle bus for employees of companies working at the airport.

Read more: Zurich airport to test self-driving shuttle bus for employees