Dear readers,
AI companies are grappling with challenges such as energy consumption, data availability, and hardware limitations in developing advanced language models. OpenAI, one of the leaders in the field, is exploring new approaches to AI model training that could revolutionise the industry. Experts suggest that the key to this future may lie in techniques that mimic more human-like thinking patterns, offering an alternative to the current focus on scaling up data and computational power.
The recent release of the OpenAI’s OpenAI o1 model represents a breakthrough in AI development, leveraging human-like reasoning and multi-step problem-solving to improve performance. This new model represents a significant shift away from traditional AI approaches focused on feeding massive data into large models. As a result of the expansion of the pre-training plateau, AI pioneers like Ilya Sutskever have acknowledged that the era of simply scaling models may be over. Now, AI researchers are exploring techniques that allow algorithms to ‘thin’ more like humans, enabling faster and more efficient problem-solving.
The growing realisation that scaling models may not always lead to better performance has sparked a re-evaluation of the AI development process. To address these problems, AI researchers are looking at ‘test-time compute’, a method that enhances AI models during the inference phase when they are actively used. This technique enables models to process multiple possibilities in real time, providing more accurate results without scaling up the model.
OpenAI’s new model, o1, is nearing these techniques, improving its performance by allowing AI to think through problems in stages similar to human reasoning. This method has proven highly effective, as evidenced by its previous test performances. Namely, the new model outperforms older models, scoring 83% against GPT-4o at the International Mathematics Olympiad. The model is unique in its ability to provide step-by-step reasoning and show human-like patterns of hesitation during the process. It also reduces the occurrence of hallucinations. However, the models have limitations, such as not browsing the internet, needing more world knowledge, and processing files and images.
The next step is for the models to perform similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology. The company aims to develop a product that can make decisions and take action on behalf of humans, which is estimated to cost USD 150 billion. Removing current kinks in the system will enable the models to work on complex global problems in areas like engineering and medicine. The o1 model, which builds on existing models like GPT-4, is a critical step in AI’s evolution, focusing on multi-step reasoning and incorporating expert feedback to guide AI through complex tasks.
Venture capital investors are taking notice of these changes, as the shift towards inference clouds could drastically alter the dynamics of AI research and development.
Follow the ‘Highlights from the week’ in its section below…
EU and UK universities begin metaverse classes
Universities across the EU and UK are set to introduce metaverse-based courses, where students can attend classes in digital replicas of their campuses. Meta, the company behind Facebook and Instagram, announced the launch of Europe’s first ‘metaversities,’ immersive digital twins of real university campuses.
Polish priest brings AI to faith discussions
In Poznan, Poland, a new chapel is combining tradition with cutting-edge technology. Created by priest Radek Rakowski, the modern chapel features an AI-powered system that answers visitors’ questions about Catholicism.
Marko and the Digital Watch team
Highlights from the week of 08-15 November 2024
The rise is attributed to a 30% increase in Bitcoin’s value over the past week amidst declining silver prices, which fell by over 6%.
President Yoon Suk Yeol has voiced concerns over Trump’s tariff threats, which could affect South Korea’s chip industry by undercutting export prices.
AI is playing a key role in the future of California’s last nuclear power plant, enabling it to overcome ageing infrastructure and compliance hurdles with innovative technology.
The Guardian’s decision follows growing criticism of X’s moderation policies under Elon Musk.
Bitcoin recently gained 11% in 24 hours, underscoring its current dominance in the crypto sphere.
X faces a May 2025 hearing in France over content payments to publishers.
Europe’s initiative to boost technological sovereignty.
Interior minister warns of cyber threats as Germany prepares for snap elections.
Explore St. Peter’s Basilica like never before.
In a significant step for digital inclusion, Chad has approved Starlink’s satellite internet services to enhance connectivity across underserved areas, marking a leap forward for the central African nation’s technological…
Reading corner
Language is shifting – words like ‘dialogue’ and ‘conversation’ are being replaced by ‘debate’ and ‘discussion’. Is this hardening of tone a sign of the times? Aldo Matteucci analyses.
DiploFoundation invited experts from participating delegations in the UN cybercrime treaty negotiations to break down the agreed draft convention and discuss its potential impact on users.
Trump’s appointment of Elon Musk as ‘efficiency tzar’ aims to modernize federal administration, amidst critiques suggesting it targets the ‘deep state.’ However, the focus should shift to the broader AI transformation poised to reshape bureaucracies worldwide. As AI begins to automate core functions, particularly text-based tasks, public institutions must adapt to these changes. This includes proactive discussions, investing in training, and identifying irreplaceable roles. Ultimately, while Trump’s motivations may spark debate, the real challenge lies in preparing governments for a future where AI fundamentally alters their operations.
Upcoming
United Arab Emirates Diplomatic Academy: Training on Artificial Intelligence From November 17-20, 2024, Diplo will partner with the Anwar Gargash Diplomatic
The G20 Leaders’ Summit 2024 will take place in Rio de Janeiro, Brazil on November 18 and 19. Brazil holds the Presidency from December 2023 to November 2024. The Summit…
25th European Diplomatic Programme: The use of AI in foreign affairs Diplo will participate in the 25th European Diplomatic Programme in Budapest, Hungary, on