Some readers will remember the old cartoon The Jetsons. This promised a future with flying cards, robot assistants and helpful computer tools. We have or are getting very close to the robot assistants, and the latest artificial intelligence offerings seem to be the automated helpers. Missing to date are the flying cars. That may have changed with the new Jetson ONE, a single person flying car I saw a demonstration of in a recent video. It looked good, seemed to fly with good stability and landed without any issue. You can find the demos with a simple search. The craft has vertical take-off and landing capability. However, I shudder to think of what thousands of these might look like in the skies above a city without some serious improvements in driving and collision avoidance.
- In the world of artificial intelligence, the two common terms you'll hear most often these days are Generative AI and Predictive AI. The first of these is the AI that will generate results like a story in text, pictures or music. For this to work, it needs access to huge amounts of examples, typically scraped by the internet. It also needs a workforce that labels and identifies the pictures and other data. This is often done in a third world nation. Think of any sports shoe factory in a cheap labour country. One example is the Be My Eyes app used to help the blind identify things. It started as a sea of volunteers but now also includes an AI that is available 24/7. It isn't as good as a human helper yet, but it can only improve. Other examples include platforms like ChatGPT and Midjourney.
- The second of these, Predictive AI, takes a whole pile of existing information and tries to make a guess or forecast as to what the next thing might be. Chatbots use this approach, as do stock tracking and prediction software. The main issue here is the predictions are only as good as the training it receives. What is often missing is the context and culture associated with situations. Imagine a model trained in the US that is asked questions about China, for which it has limited data and training. In the latter case, none of the cultural elements may have been included and if the question is generic, then the results will always be from a US perspective. Such systems are also limited in seeing how their predictions might change future outcomes. Finally, such systems can be easily tricked if you know the expectations and training methodology.
- The new AI assistant Manus, from the Chinese startup Butterfly Effect, went viral earlier this month. Unlike the examples above, this is being described as the first truly autonomous AI agent. I mentioned the advent of such systems in an earlier article, but as a reminder, they are designed to complete complex multi-step tasks using the web after getting prompted just once, functioning like a highly-skilled assistant. After reading about this, I applied for and was kindly granted access. The first thing you notice when entering the website are a set of demonstrations.
- I ran the seven-day visit to Japan example, and since it was coming up for me, a similar version for four days of sightseeing in London. When I copied and pasted, I'd mistakenly left in the desire to see a Nara's deer (in Japan). Manus noted this and asked me if I wanted to create a separate itinerary for Japan, look for deer in London, or ignore this. As an IT guy, I loved how it shows its work as it processes. Unlike the Japan example, this was a brand-new request so it took longer to process. The task used websites, videos, pictures and other sources. The result was a detailed itinerary with consideration for wait and travel times, train and bus information, a description of attractions and even a screenshot of a map showing their locations. I tried to imagine how long a travel agent, or worse myself, might take to put this together from scratch and decided I probably would have given up and booked a few tours instead.
- According to those behind Manus, it combines the functionality of multiple AI agents and currently outperforms OpenAI's recently released deep research agent, which is also meant to produce detailed reports by searching the web. At this stage, Butterfly Effect is not an OpenAI or DeepSeek competitor because they don't technically build AI models. For the time being, Manus has been built on top of existing models, in this case Anthropic's Claude 3.5 and modified versions of Alibaba's Qwen. More later after testing.
- To round out the week. Elon Musk in an interview said he expects 90% of cars will be auto-driving in 10 years and the majority of work will be done by machines and robots. He also put the AI Skynet scenario at around 10-20% by the same time. The last one is still too high for my liking, so let's hope AI designers are trying to minimise this outcome.
James Hein is an IT professional with over 30 years' standing. You can contact him at jclhein@gmail.com.
