In this post I share some of my views and experiences with AI. It is intended to spur discussion of AI to allow all of us to share viewpoints and experiences. It is not intended to be a forum for promoting or criticizing any AI product. So, contribute your thoughts and experiences, and please use other channels for product specific messages.
In a simple way, it appears there are three points of view on AI.
- One is the excitement camp that believes AI is ready to solve most of the most difficult and vexing problems we face.
- Another is the cautious one that believes AI is capable of doing some useful tasks although it has, and may continue to have for some time, significant problems.
- The third is dismissive that believes AI is over promised and not nearly as economically viable as represented.
While AI has been around in one form or another for decades as the province of academics struggling to find a real solution to a real need with isolated behind-the-scene successes, the launch of ChatGPT and other large-language models captured the public’s attention. Before the launch of ChatGPT, AI was working for the average person providing speech recognition, navigation, and internet search to name a few applications.
I’ll offer a few observations about the current state of AI. Your views and experiences might lead you to different conclusions.
- AI is becoming extremely popular. More and more people are expecting to use it to make tasks less burdensome. Business is investing in AI in hopes of running a more profitable and resilient enterprises.
- Current AI has limitations and flaws. Large language models are prone to hallucinations (providing made up data). Speech recognition doesn’t always understand correctly. Mapping programs occasionally take someone to the wrong destination.
- AI will continue to improve both through evolution and occasional revolutionary leaps.
- AI and the legal landscape is in chaos and will remain so for decades. Part of the problem is the rapid pace of AI’s evolution far outpacing regulators ability to define acceptable and required performance. Part of the problem is assigning responsibility – does it rest with the user AI or the provider of AI?
- AI is a tool for good and also a tool for evil. It depends on the user and their intent. Bad actors already use AI to create embarrassing images. It’s inevitable that some people will use AI in new, unknown ways to steal and defraud. Can AI be built with a moral compass? Who would decide the values AI should possess? How can bad actors be prevented from circumventing the moral compass if it exists?
- AI may not be current in its knowledge. For example, ChatGPT has training cutoff dates. If you ask about current events, your prompt will either be rejected or give results that are out-of-date. Even Google for web search will not always have indexed the very latest version of all websites.
- AI is more and more the property of large, well-funded enterprises. ChatGPT-4 cost $63 million dollars to train excluding the salaries of researchers. This cost will trend quickly downward as new training solutions emerge. For general applications, it will probably never reach a truly economical level available to the average person. Still, AI for very constrained applications can be developed at a tiny fraction of this cost.
Some of my recent experiences with AI
Out of curiosity, I asked ChatGPT to recommend a camera for a hypothetical application. The AI approached the problem correctly and provided a recommendation. However, the AI used a theoretical basis rather than a practical basis. It assumed that two pixels spanning the smallest feature was sufficient when machine vision practitioners know from practical experience it takes three or more spanning pixels to achieve reliable operation. Someone who relied on ChatGPT to guide the design of a vision system would be disappointed in the results.
ChatGPT made a rather startling recommendation that smaller pixels [in the camera] would result in higher resolution. Assuming what was not stated, that smaller pixels would mean more pixels spanning the field-of-view, this statement would be correct. Taken at face value, though, it could be interpreted as recommending a camera with smaller size pixels. In that case the advice is completely wrong. Again, AI would mislead a vision system developer not having reasonable experience to interpret the recommendations wrongly.
I asked Perplexity a very similar question. The answer was more reasonable recommending four pixels spanning the smallest feature. I liked that it gave references for its assumptions. However, the recommended camera, retrieved from just one commercial source, had much too high an image resolution and would likely increase the vision system’s cost.
Note that the basic versions of ChatGPT and Perplexity are free to use. Each has a paid version that supposedly is trained with more information and should give superior results.
My second recent experiment was using AI to generate videos. I downloaded a video clip of myself to generate a personal avatar and then gave it a script from a portion of an upcoming video. I’ll give the AI credit for good lip-syncing. However, the voice I didn’t recognize as mine, and it was very monotone and droning. I suspect with more experience on my part, the voice could be improved. What was most disturbing were the avatar’s gestures. They made no sense to the flow of the script. The AI program recommends keeping gestures small and generic in the video clip used to generate the avatar. Obviously, AI has not progressed to the point where it can interpret text and create appropriate gestures. My speech training through classes and Toastmasters experience stressed the importance of good, appropriate gestures in making the talk more interesting and memorable. I expect that AI has a long way to go in this area to understand the meaning and intent of text and not just the words.
Another recent experience was riding with my brother in his Tesla with full self-driving. I was impressed despite a few hiccups. The route we took was about 50 miles through a mountain highway and city streets. The only problems were the AI’s failure to engage on the return trip. Restarting the self-driving feature cured this event. The driver sets the maximum speed for the trip. Self-driving adjusts the speed to meet traffic conditions and speed limits. Occasionally, though, for no apparent reason, it would change the maximum speed setting downward.
An interesting event my brother related was approaching a construction zone in self driving mode. There was a worker controlling traffic with a sign that read “Stop” on one side and “Slow” on the other side. The worker had the “Slow” sign showing to my brother but also put his arm out with his palm facing forward – an understandable indication to stop. The Tesla self-driving program didn’t know how to interpret this unexpected mixed signal and was proceeding through the work zone. Fortunately, my brother recognized the situation and took immediate control. Tesla’s self-driving feature would be classified as a level 3 driver assist. That means the driver must be ready to instantly take control. Level 3 driver assist is, in my opinion, the most dangerous type of automation (using AI). It gives the user the ability to completely lose awareness of the driving situation yet attempts to make the driver responsible for the car’s operation. I expect lots of litigation in this area.
I believe that AI is here now as a useful tool but with significant limitations. We should all be studying and experimenting with available AI to learn its capabilities and its limitations so we can make the best use of it. We should expect to continue experimentation and study for the foreseeable future. With thousands of brilliant and passionate people working to improve AI, there will be progress in addressing limitations along with occasional breakthroughs that allow AI to advance significantly.
We must remain vigilant in our use of AI realizing that bad actors will use AI against us including offering AI capabilities that compromise our own security. A case in point is the lawsuits against Apple and Google for surreptitiously recording (speech recognition) everything that was said in the vicinity of their devices.
What are your thoughts, views, and experiences with AI? How is it useful for you? What problems does it pose for you?
I will continue my study of AI applications and technology and share them with you periodically on this blog.