AI Newsround – Causality and China

Causality again! 

• EY Canada has been forced into an embarrassing climb down as it used AI to create a research report on cybersecurity in loyalty schemes, which included hallucinations and made-up data and footnotes in yet another sign that AI, while useful, remains pretty dumb and needs rigorous supervision.
• The report was being used to market EY’s prowess in cybersecurity, which is doubly embarrassing as this is just the kind of thing that cybersecurity is supposed to prevent.
• This is yet another demonstration that AI, as good as it is, is still nothing more than a pattern recognition system and has no causal understanding of the tasks that it is being asked to perform.
• One can put in guard rails to try and stop it from doing crazy things, but at the end of the day, what it needs are humans to double-check its work or a software sandbox to keep it under control.
• This is especially the case as agents powered by these models become able to take actions on behalf of users where mistakes can have devastating consequences (eg PocketOS).
• Those who claim that we are on the brink of superintelligent machines will claim that these are teething problems, but these problems are a result of the way these models work, and so it is very unlikely that they can be eliminated altogether.
• In my opinion, this is what makes the investment case for some of the SaaS companies that have been chopped in half by the AI scare trade.
• If one believes that these agents will need strict supervision, then the likes of ServiceNow and Salesforce are going to thrive.
• If, however, one thinks that we are the dawn of superintelligent machines and that one can just let the LLMs loose inside one’s company, and everything will magically be great, then these companies are in real trouble.
• I know which one I believe, which is why I have a position in ServiceNow.

China’s real edge

• The debate over who is leading in the AI race continues to rage, and there is no simple answer, as in some areas China is significantly ahead, but in the one that really matters, the USA has a large lead.
• When it comes to open source, video generation and access to energy, China leads the USA and the world.
• Models such as Seedance 2.0 and HappyHorse currently top the league tables for video generation performance and quality, where access to data appears to be a key metric.
• ByteDance and Kuaishou both have a thriving short video sharing platform, which gives them access to millions of hours of video content that can be used for training.
• Not surprisingly, it is Google that is in 3^rd place, most probably due to its ability to use YouTube videos to train its Veo model.
• Furthermore, when it comes to open source, many non-Chinese start-ups are basing their AI services on Chinese models as these are some of the best and most cost-effective available.
• China also already has 3x the power generation capacity of the USA and is building more very quickly, meaning that AI datacentres can be quickly powered up once built.  
• However, where China is in difficulty is in AI silicon chips, where it is stuck at an inefficient 7nm multi-patterning process with very little scope to go any further.
• This means that as the USA advances with TSMC and Samsung, China will fall further and further behind.
• This is crucial because it means that China cannot produce enough compute to roll out AI services at scale domestically and will have to rely on importing compute from datacentres overseas.
• While this currently works pretty well, if the US Department of Commerce targets this activity the next time it updates its export rules, it could cause real problems.
• This is why I still think that the USA will maintain the edge in AI until such time that China can manufacture leading-edge AI semiconductors at home.
• This could be a long wait.