2024 in review

This week we take a look back at AI in 2024 and the key trends that shaped this extraordinary year.

From Anthropic’s rise to become the leading lab, to new scaling laws that changed how systems train and think, to disruption of traditional software, business and society, and the weirder corners of AI. We assess our 117 ExoBrain newsletter stories to date and reflect on the evolution of a fascinating era. We’ll be on holiday next week but will return with a preview of 2025 in January. Have a Happy Christmas and a prosperous New Year!

o3 and the new scaling laws

The AI ‘scaling’ story took a significant turn in 2024. Early rumours about OpenAI’s Q* and ‘strawberry’ projects suggested a major leap in AI reasoning capabilities. When o1-preview was unveiled (Week 37), it productionised a new approach, shifting computational resources towards ‘thinking time’ rather than training. This model was designed to leverage reinforcement learning to enhance its reasoning capabilities, allowing it to spend more time processing and solving complex problems. But behind the scenes the costs of training at the frontier had escalated dramatically, approaching $1b per run (Week 34). At the NeurIPS conference last week, former OpenAI chief scientist Ilya Sutskever suggested that we will soon reach ‘peak [training] data’, signalling the end of the scaling era. But right on cue, o1 pro mode (Week 50) and other models like DeepSeek R1 (Week 47), along with Google’s ‘Gemini 2.0 Thinking’, suggest scaling at the point of use can take over. 

The future will not only be about training ever larger models, but about teaching smaller ones to think more effectively (and getting them to work together as ‘agents’). The race is on to perfect this technique and optimise beyond the highly structured domains of maths and coding, to healthcare, finance and beyond. As we published on Friday evening, OpenAI demonstrated and published benchmarks for their next generation reasoning model, o3, planned for release in Q1 2025. The early benchmarks look stunning. It has demonstrated above human performance in the Arc AGI Prize (Week 24) and looks very strong on software development. In early 2024 GPT-4 was getting around 3% on the SWE-Bench coding test, and o3 tops 70%! The cost of this capability looks exceptionally high, but the trajectory is clear.

Claude, your personal AI

The arrival of Claude 3 in March was the most significant release in the first half of the year (Week 10). For the previous year, GPT-4 had reigned supreme, and OpenAI seemed relatively unassailable, but Claude’s remarkable self-awareness and ability to process entire books in seconds set a new bar. By autumn, Claude 3.5 Sonnet raised the bar again (Week 26). Its superhuman software engineering abilities created a step-change in what individuals were capable of doing in code and new development tools like Cursor took off. Meanwhile an array of talented safety researchers left OpenAI for Anthropic, with the Golden Gate Claude (Week 21) and now Fake Alignment papers being for many the AI research breakthroughs of the year. 

As a result. Anthropic’s main challenge has been how to keep up with demand, with Claude’s only negative being restrictive rate limiting. Competition is as intensive as ever, but Claude is holding its own, and whilst rumours of delays and issues with the next release abound, Amazon continue to pour in money and compute (and even Microsoft are rumoured to be considering investing in the next round). Anthropic end the year as the leading AI lab. Claude, beloved by many, has shown how AI can provide huge personal productivity augmentation, a knowledge working partner, and a platform for vital safety research.

An uncertain geopolitical future

The global rush to build datacentres in 2024 continued despite the unsettling fact that every advanced chip powering our AI future comes from a single source. TSMC’s fabrication plants in Taiwan remain the only facilities capable of producing cutting-edge GPUs, from Nvidia’s powerful Blackwell (Week 12) to custom silicon from every major firm. Whilst the CHIPS Act spurred construction of new US facilities, TSMC’s own Arizona plant faces delays. More concerning still, chips must still return to Taiwan for the critical packaging step, a bottleneck that won’t ease until 2027 at the very earliest. Enter Donald Trump’s election victory, bringing promises of tariffs instead of subsidies, accusations of Taiwan “stealing” US technology, and suggestions they should pay for military protection (Week 45). Meanwhile, China (Week 35), at war with the US over access to advanced silicon watches and waits, knowing that control of Taiwan would grant unprecedented leverage over the global economy. 

This technological dependency has become the most significant geopolitical risk of our time – a single military action could instantly sever the world’s supply of AI growth. As an aging Xi Jinping faces an unpredictable US president, the decisions made could tilt the balance of global AI and economic power for generations.