AI arms race 'risks amplifying existential dangers of superintelligence'
Canadian machine learning pioneer Yoshua Bengio, author of the first International AI Safety Report to be presented at an international AI summit in Paris next week, warns unchecked investment in computational power for AI without oversight is dangerous.
"The effort is going into who's going to win the race, rather than how do we make sure we are not going to build something that blows up in our face," Mr Bengio says.
He warns that military and economic races "result in cutting corners on ethics, cutting corners on responsibility and on safety. It's unavoidable".
Mr Bengio worked on neural networks and machine learning, the software architecture that underpins modern AI models.
He is in London, along with other AI pioneers to receive the Queen Elizabeth Prize, UK engineering's most prestigious award in recognition of AI and its potential.
He's enthusiastic about its benefits for society, but the pivot away from AI regulation by Donald Trump 's White House and frantic competition among big tech companies for more powerful AI models is a worrying shift.
'Superhuman systems becoming more powerful'
"We are building systems that are more and more powerful; becoming superhuman in some dimensions," he says.
"As these systems become more powerful, they also become extraordinarily more valuable, economically speaking.
"So the magnitude of, 'wow, this is going to make me a lot of money' is motivating a lot of people. And of course, when you want to sell products, you don't want to talk about the risks."
But not all the "godfathers" of AI are so concerned.
Take Yann LeCun, Meta's chief AI scientist, also in London to share in the QE prize.
"We have been deluded into thinking that large language models are intelligent, but really, they're not," he says.
"We don't have machines that are nearly as smart as a house cat, in terms of understanding the physical world."
Within three to five years, Mr LeCun predicts, AI will have some aspects of human-level intelligence. Robots, for example, that can perform tasks they've not been programmed or trained to do.
But, he argues, rather than make the world less safe, the DeepSeek drama - where a Chinese company developed an AI to rival the best of America's big tech with a tenth of the computing power - demonstrates no one will dominate for long.
"If the US decides to clam up when it comes to AI for geopolitical reasons, or, commercial reasons, then you'll have innovation someplace else in the world. DeepSeek showed that," he says.
The Royal Academy of Engineering prize is awarded each year to engineers whose discoveries have, or promise to have, the greatest impact on the world.
Previous recipients include the pioneers of photovoltaic cells in solar panels, wind turbine technology and neodymium magnets found in hard drives, and electric motors.
Science minister Lord Vallance, who chairs the QE prize foundation, says he is alert to the potential risks of AI.
Organisations such as the UK's new AI Safety Institute are designed to foresee and prevent the potential harms AI "human-like" intelligence might bring.
But he is less concerned about one nation or company having a monopoly on AI.
"I think what we've seen in the last few weeks is it's much more likely that we're going to have many companies in this space, and the idea of single-point dominance is rather unlikely," he says.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Geeky Gadgets
a minute ago
- Geeky Gadgets
GPT-5 vs GPT-4o Comparison Guide : The Key Differences You Need to Know
Is newer always better? When GPT-5 was announced, many expected it to be a new leap forward, a model that would leave its predecessor, GPT-4o, in the dust. Yet, early adopters have been met with a surprising reality: while GPT-5 features impressive advancements in reasoning, coding, and visual generation, GPT-4o continues to hold its ground with unmatched speed, simplicity, and reliability. This unexpected dynamic has sparked a heated debate among users, could the latest iteration of OpenAI's technology actually fall short in certain areas? The answer, as it turns out, is far more nuanced than a simple yes or no. The strengths and weaknesses of these models reveal a fascinating trade-off between innovation and practicality. In this comparison, Skill Leap AI unpack the key differences between GPT-5 and GPT-4o, exploring how their distinct capabilities shape their real-world applications. From GPT-5's advanced problem-solving and creative ideation to GPT-4o's efficiency in document creation and everyday tasks, each model offers unique advantages that cater to specific needs. But which one truly delivers the best value? Whether you're a developer seeking innovative tools or a professional craving speed and simplicity, understanding these trade-offs will help you make an informed choice. The answer isn't just about which model is 'better'—it's about which one is better for you. GPT-5 vs GPT-4o Distinct Strengths and Weaknesses Both GPT-5 and GPT-4o bring unique capabilities to the table, making them suitable for different types of tasks. Understanding their core strengths and limitations is essential for making an informed choice. GPT-5: Known for its advanced reasoning and ability to tackle complex problem-solving tasks, GPT-5 is particularly effective in generating detailed visual outputs. It is a strong choice for users seeking innovation in coding, creative ideation, and tasks requiring higher cognitive processing. Known for its advanced reasoning and ability to tackle complex problem-solving tasks, GPT-5 is particularly effective in generating detailed visual outputs. It is a strong choice for users seeking innovation in coding, creative ideation, and tasks requiring higher cognitive processing. GPT-4o: Designed with speed and simplicity in mind, GPT-4o is ideal for straightforward tasks such as document creation and rapid content generation. Its reliability and ease of use make it a preferred option for everyday applications. Performance in Key Functional Areas The differences between GPT-5 and GPT-4o become more evident when evaluating their performance across specific tasks. Each model demonstrates distinct advantages depending on the nature of the task. Coding: GPT-5 outperforms GPT-4o in handling intricate programming challenges. Its ability to generate detailed, executable code makes it a valuable tool for developers working on complex projects. GPT-5 outperforms GPT-4o in handling intricate programming challenges. Its ability to generate detailed, executable code makes it a valuable tool for developers working on complex projects. Data Visualization: GPT-5 offers significant advantages in creating interactive and visually rich dashboards, making it a strong choice for data analysts and visualization tasks. GPT-5 offers significant advantages in creating interactive and visually rich dashboards, making it a strong choice for data analysts and visualization tasks. Document Creation: GPT-4o excels in producing cleaner, more organized documents with fewer formatting errors. This makes it the preferred model for professional writing and administrative tasks. GPT-4o excels in producing cleaner, more organized documents with fewer formatting errors. This makes it the preferred model for professional writing and administrative tasks. Processing Speed: GPT-4o consistently outpaces GPT-5 in terms of speed, particularly for tasks that do not require deep reasoning or complex problem-solving. GPT-4o consistently outpaces GPT-5 in terms of speed, particularly for tasks that do not require deep reasoning or complex problem-solving. Accuracy: GPT-5 shows slight improvements in reducing hallucinated outputs, such as inaccurate citations, but both models continue to face challenges in this area, highlighting the need for further refinement. Is GPT-5 Better Than GPT-4o? Watch this video on YouTube. Take a look at other insightful guides from our broad collection that might capture your interest in ChatGPT 5 comparison. Creative Applications and Ideation For creative tasks such as storytelling, brainstorming, or email drafting, GPT-5 demonstrates clear advantages. Its outputs are more structured, detailed, and contextually relevant, making it a powerful tool for ideation and creative writing. However, the improvements over GPT-4o, while noticeable, may not be substantial enough to warrant a switch for all users. Those who prioritize creativity and innovation may find GPT-5 more appealing, but for simpler creative tasks, GPT-4o remains a reliable option. Automation and Workflow Optimization GPT-5 introduces enhanced capabilities for automating complex workflows. Its advanced reasoning skills enable it to independently break down multi-step tasks and execute them with minimal user input. This makes it particularly valuable for professionals looking to streamline agentic tasks such as project management, workflow optimization, or multi-layered decision-making processes. By contrast, GPT-4o, while efficient, lacks the same level of sophistication in handling intricate automation tasks. User Experience and Practical Usability The user experience offered by GPT-5 and GPT-4o varies significantly, depending on the complexity of the task. GPT-5 includes a streamlined 'auto' mode designed to assess task complexity and adjust its approach accordingly. However, this feature occasionally misinterprets user intent, leading to less satisfactory outputs. On the other hand, GPT-4o provides a more predictable and user-friendly experience, particularly for everyday tasks. For users who prioritize consistency, simplicity, and intuitive interaction, GPT-4o may be the better choice. Challenges and Limitations Despite their advancements, neither GPT-5 nor GPT-4o is without flaws. GPT-5's improvements, while meaningful, do not represent a dramatic leap forward akin to the transition from GPT-3 to GPT-4. Both models continue to struggle with certain tasks, such as generating accurate YouTube thumbnails or avoiding hallucinated citations. These limitations underscore the ongoing challenges in AI development and the need for further refinement to address these persistent issues. Making the Right Choice The decision between GPT-5 and GPT-4o ultimately depends on your specific needs and priorities. Each model offers distinct advantages that cater to different use cases: Choose GPT-5: If your work involves advanced reasoning, complex coding, creative ideation, or tasks requiring detailed visual outputs, GPT-5's capabilities make it the better option. If your work involves advanced reasoning, complex coding, creative ideation, or tasks requiring detailed visual outputs, GPT-5's capabilities make it the better option. Choose GPT-4o: For tasks that prioritize speed, simplicity, and reliability, such as document creation or rapid content generation, GPT-4o remains a strong contender. While GPT-5 introduces notable advancements, it does not deliver a new upgrade over GPT-4o. Instead, the choice comes down to weighing the trade-offs and selecting the model that best aligns with your goals and workflow requirements. Media Credit: Skill Leap AI Filed Under: AI, Top News Latest Geeky Gadgets Deals Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.
Scottish Sun
2 hours ago
- Scottish Sun
Bank-emptying Gmail and Outlook attachments overtaken by even WORSE costly email con that's much harder for you to spot
Over 3 billion attacks have been sent out so far CASH KILLER Bank-emptying Gmail and Outlook attachments overtaken by even WORSE costly email con that's much harder for you to spot Click to share on X/Twitter (Opens in new window) Click to share on Facebook (Opens in new window) BRITS are being warned to watch out – because the dodgy email attachments that used to drain your bank account have just been outdone by an even sneakier scam that's much harder to catch. Cyber experts have revealed that online crooks now prefer planting malicious links over using infected attachments - and the results are far worse. Sign up for Scottish Sun newsletter Sign up 2 The tools are so easy that even low-level scammers are jumping on the trend Credit: Alamy According to a new bombshell report by Proofpoint, the hidden traps are tucked inside emails, buttons, and even PDFs or Word docs, and one wrong click could see your logins stolen or malware silently installed. Over 3 billion attacks with dodgy URLs have been sent out and the main goal is to steal passwords. This hacking scheme isn't just being used by criminal masterminds either. The tool are so easy to get hold of that even low-level scammers can launch convincing fakes that bypass security checks like multi-factor authentication and take full control of your account. Proofpoint also uncovered a jaw-dropping 400 percent spike in a sneaky scam called 'ClickFix' – where users are tricked into clicking fake error messages or CAPTCHA boxes. These convincing cons trick you into running harmful code, opening the door to remote access trojans, info-stealers, and more. Meanwhile, QR code phishing attacks are exploding, with over 4.2 million attempts spotted in just the first half of 2025. These nasty little codes target your personal mobile – dodging work defences completely. And let's not forget smishing – dodgy texts that try to fool you. 2 MM1WHE Internet fraud concept with faceless hooded male person using tablet computer, low key red and blue lit image and digital glitch effect Credit: Alamy More than half of all SMS phishing attempts now come packed with malicious URLs, making it harder than ever to stay safe. Selena Larson, top threat analyst at Proofpoint, gave a stark warning: 'The most damaging cyber threats today don't target machines or systems. They target people.' She added that these new-style scams are designed to exploit human psychology, using trusted brands and familiar tech to lure you in – whether it's a dodgy CAPTCHA, a QR code, or a believable text message. This comes after a devastating con carried out by Chinese organised crime groups was exposed. So-called 'pig butchering' is where scammers established fake romantic and trusting relationships with victims before luring them into fraudulent investments or other financial traps. In 2023, Shan Hanes, a banker from Kansas, US, embezzled £34.6million from his bank to cover his losses, having fallen victim to a pig butchering scam. Hanes was later sentenced to more than 24 years behind bars. Usually, a pig butchering scam works in three stages – hunting, raising and killing. This involves a scammer finding a victim online, chatting to them in order to build up trust and then getting them to invest large amounts of money into fraudulent schemes. The scam works in a similar way to a traditional romance scam, where scammers approach their victims by posing as a possible romantic partner on a dating app, or as a friend via social media. The big difference though is how the scam is executed. With a romance scam, trust is based on the victim's urge to maintain a romantic relationship with the scammer. In this scenario, the scam can often last for years. Pig butchering scams though, in comparison, generally take place over a much shorter time period. The scammer, rather than focusing on trying to extract money through emotional manipulation, leans more on the victim's desire to make money together with the scammer. This can involve just a few months rather than years to take advantage of the victim. Usually, the scammer will present themselves as being financially successful and confident with a broad network and have appealing investment opportunities. Once the victim has made an initial small investment, the scammer will then try to escalate the process and push them into making a much larger financial commitment, reports.
Telegraph
4 hours ago
- Telegraph
NHS patients to be discharged using AI notes for the first time
NHS patients will be discharged from hospital using notes made by AI for the first time. Information on a person's condition, care and medication will be made by AI and checked by a doctor to help speed up the process. Currently health workers have to deal with the laborious and time-consuming task of filling out notes, leading to prolonged time on wards. A trial at Chelsea and Westminster Hospital starting this autumn will kick off the three-month pilot project using live data for real patients. Officials hope this will help get patients home, free up doctors' time and create more beds for other patients while standardising the notes. The AI is a large language model (LLM) similar to ChatGPT but has been created in-house by the NHS. The Telegraph understands patient data will not be shared with third parties or used to train the computer model. The tool, called 'AI Discharge Summaries', is one of the Government's so-called 'AI exemplars' alongside technology which promises to halve the time probation officers spend organising notes and software to support teachers in marking and lesson planning. Wes Streeting, the Health Secretary, said: 'We're using cutting-edge technology to build an NHS fit for the future and tackle the hospital backlogs that have left too many people waiting too long. 'Doctors will spend less time on paperwork and more time with patients, getting people home to their families faster and freeing up beds for those who need them most.' Peter Kyle, the Technology Secretary, said: 'This is exactly the kind of change we need: AI being used to give doctors, probation officers and other key workers more time to focus on delivering better outcomes and speeding up vital services.' The AI tool for discharge summaries promises to improve efficiency within hospitals and provide medical staff with the 'gift of time'. The software enables doctors to draft discharge documents faster by using the LLM to find key information from medical records – such as diagnoses and test results. It aims to improve what has been described as an 'outdated system' in which busy hospital staff were forced to sit down and fill out forms. The Government hopes to use AI to 'modernise' services across health, justice, tax and planning. Elsewhere, technology will be rolled out to speed up planning approvals by converting hand-written documents into data within minutes. The Government claims this will slash the hours spent by planning officers who have to manually check such documents. The technology is one of a number of projects backed by Sir Keir Starmer. Earlier this year, the Prime Minister said AI will be used to 'turn around' the economy and public services. Referring to the entirety of AI exemplars – not just the NHS – Mr Kyle said when the scheme is rolled out in full, it will unlock some £45bn in 'productivity gains'.
