What I’ve learnt this week in AI

1. The US/China AI race continues – what will this mean? OpenAI unveiled GPT-4.5, its latest and largest AI language model this week vs. DeepSeek launching 5 open-sourced code repositories this week as part of their “202502 Open-Source Week” initiative (https://github.com/deepseek-ai/open-infra-index). Chinese AI labs have narrowed the gap with U.S. labs, with DeepSeek’s January-released R1 model now rivalling OpenAI o1-level intelligence. But it is not just DeepSeek, Alibaba is also fighting back with a Strategic AI Investment of 380 billion yuan (around $52.44 billion) in cloud computing and AI over the next three years. Compare that with “GPT-4.5” being declared as super expensive to run and Sam Altman declaring “We’re out of GPUs”.
2. The next stage of AI as an Agent is also raising questions. Microsoft CEO Nadella is quoted as stating the era of Software as a Service (SaaS) is ending and “Agent as a Service” will redefine future workflows, shifting business logic into an AI layer, rendering traditional back-end systems obsolete (proving interesting as a concept for the paper I am writing on the future of connectivity in our industry). But a split is emerging in the methodologies for training AI agents – quick, large-scale learning (unsupervised) vs. controlled reasoning to solve more complex problems

• Unsupervised learning, like GPT-4.5, builds broad knowledge but lacks deep reasoning, making it valuable for fast decision-making and trend identification, though it may struggle with complex problem-solving.
• Automated reasoning systems like OpenAI’s o3 excel at problem-solving with deeper reasoning but have a narrower knowledge base, making them more effective in environments requiring accuracy and complex decision-making, especially as AI agents take on more roles.

3. Could we be heading back to word-of-mouth (WOM) recommendations? The debate between open-sourced unsupervised learning vs more controlled reasoning matters, as the rise in greater numbers of autonomous agents is combining with declining consumer trust in AI-generated content from deepfakes. Grok 3 is gaining attention as a highly intelligent AI, with claims it outperforms GPT-4, Claude 3.5, and others – yet it is fed by real-time data from X (https://www.linkedin.com/posts/rebecca-healey_elon-built-the-most-powerful-ai-training-activity-7298001297629216769-av-t?utm_source=share&utm_medium=member_desktop&rcm=ACoAAARqE5IBeeTUq6qzU4Eqccq_UO6-OMeR6Ao) One example this week was a claim of the storming of a Birmingham hospital (https://www.linkedin.com/posts/philip-torr-1085702_elon-musk-reposts-image-of-birmingham-hospital-activity-7298367853181706240-KIMa?utm_source=share&utm_medium=member_desktop&rcm=ACoAAARqE5IBeeTUq6qzU4Eqccq_UO6-OMeR6Ao) which apparently was a scene from a movie, resulting in the local police force releasing a statement (https://www.westmidlands.police.uk/news/west-midlands/news/news/2025/february/statement-on-false-claims-over-Birmingham-disorder/#:~:text=There%20was%20no%20disorder.,and%20he%20is%20now%20stable).

While AI-powered search tools like Perplexity AI are already providing better content recommendations than Google, the rise of AI-generated content could make it harder to curate quality content. New initiatives like Jellypod AI aim to curate content with AI, but verified by humans, while word-of-mouth (WOM) recommendations from trusted sources will become more valuable.

4. The impact on AI developments continues to be felt in financial services with traditional lines between who provides what services and to whom and how continues to blur. JPMorgan’s AI tool – LLM Suite has been deployed to 200,000 employees, with half using it actively every day (https://www.wsj.com/tech/ai/jpmorgan-chase-artificial-intelligence-banking-939b1b32). The bank’s AI-driven strategy highlights how financial institutions are switching into technology-first entities, from client briefing preparation to investment banking analysis. As AI models become more sophisticated, banks will differentiate themselves based on proprietary data integration, offering more personalized and efficient customer experiences.

Brevan Howard are taking it one step further. SigTech’s emergence as a spin-out from Brevan Howard in 2019 highlights how hedge funds evolving proprietary investment tools into commercial technology solutions (https://sigtech.com/products/magic/) SigTech reflects a broader industry trend where asset managers are becoming technology providers, while fintech firms increasingly integrate investment management and quantitative research into their offerings.

5. While AI is not used directly yet in trading in traditional markets, it is becoming more active in crypto markets meaning financial services regulators will be taking a keener interest. Kaito’s AI-driven data processing capabilities allow it to analyze over 10,000 sources, including social media and research papers, while performing sentiment analysis for a more comprehensive view of the market. Kaito also integrates advanced risk management tools, offering automated risk assessment and precise execution of strategies, adjusting in real-time.
   • ESMA is looking AI’s operational use and investment in AI-driven companies, emphasizing the impact on investor protection, market stability, and regulatory compliance, stressing transparency and adherence to EU regulations like UCITS and AIFMD. Asset managers must ensure compliance, implement strong risk management, disclose AI usage, and stay updated on regulatory change – https://www.esma.europa.eu/sites/default/files/2025-02/ESMA50-43599798-9923_TRV_Article_Artificial_intelligence_in_EU_investment_funds.pdf
   • The SEC is holding a roundtable discussion on Artificial Intelligence in the financial industry on March 27 from 9 a.m. to 4 p.m. at the SEC’s headquarters in Washington, D.C. and is open to the public for either in-person or virtual attendance – read more here = https://www.sec.gov/newsroom/meetings-events/sec-roundtable-artificial-intelligence-financial-industry

As always – let me know what you think – what you found interesting, what you didn’t

Many thanks