Using AI to do my Paperwork

Two days ago I became involved in a work process that did not go as planned. As can happen in all software, a unique set of circumstances arose that kept a task from completing. Imagine running out of an ingredient when making cookies. You cannot finish the task until you go to the store and purchase what is missing. That is what happened at work.

We received an alert from our monitoring service and engineers started working on it immediately. Several service-desk tickets got created and eventually we solved the problem. Now we need to document what went wrong and the solution we used. My boss gave me the task to write up an "Incident Report." I did not want to spend hours going through the various documentation we created and so I used AI to create the report.

Our service desk software includes an AI tool that probably just links into Gemini or ChatGPT. I clicked on the link to get the AI help and provided a prompt. I told it to "Create an incident report based on the service tickets." The AI thought about it for a solid 2 minutes and then spit out a well-formatted incident report. I read through the information and thought it did a pretty good job. Most importantly I didn't have to do it. I copied the output to where I needed to put the report and headed off to dinner.

This morning I met with my boss and reviewed the incident report. He purposely didn't instruct me to use AI but hoped I would. When he asked how long it took to create the documentation, I told him 2 minutes and he appreciated that I didn't spend a long time to create the report. 

LLM Demos All Look the Same

Last week I had the pleasure of attending the Snowflake Summit in San Francisco. As you might imagine, AI played a huge part in the conference. During the product feature keynote speech, Snowflake told us the answer to every question is CoCo, or their form of AI that uses Large Language Models or LLMs.

AI is much more encompassing than just LLMs but that is what everyone thinks of when we talk about AI right now. The downside to LLMs is that all of the demos look exactly the same. Snowflake kept trying to show us new features in their product, would bring up a chat prompt and type in some sort of question. It would think about it for a bit and then spit out a text answer. Sure you didn't need to code or type SQL queries to the database but there really wasn't any difference from one demo to the next.

Last year I attended only 3 days of the conference as I had to get home and couldn't stay for the final day. I should have taken a page out of last year's playbook and only stuck around for 3 days instead of the 4 I did this year. By the last day, I lost all excitement and couldn't wait to get home. In fact, I left the conference at noon and tried to catch an earlier flight back to Salt Lake.

Ultimately I came away from the conference feeling like I wasted some of my time. When all you have to do is pull up a chat prompt and enter your question, there is no need to learn the intricacies of the product. Don't know the syntax to join data from two separate data sources, just let the AI figure it out for you. The conference can be reduced from days to hours.

When I left the conference, Snowflake asked me to fill out a survey. They provided dates for the event next year and asked if I would attend in San Francisco. Naturally I politely declined as it just doesn't interest me any more. Now if they held the conference in Hawaii, I'd be there but doubt I would spend any time at the actual Summit. When it came time to put together a trip report, I could always ask AI to do it for me. 

Artificial Intelligence Being Wrong

I recently visited Alaska to do some skiing and spent time away from my home ski areas Snowbird and Alta. First off, Alaska is a whole new kind of cold when it comes to skiing. I'm glad I opted to bring my warmest gear. 

While driving to the airport after a day of skiing at Aleyeska, I got a call from my ski buddy Jim. He said that AI told him Snowbird would be closing for midweek skiing at the beginning of April. Normally Snowbird stays open daily until it gets close to the end of the season and then it closes during the week and is only open on weekends. That normally doesn't happen until May. This is important as my buddy Jim will be coming to Utah for a week in April to do some skiing and that is kind of tough to do if the ski area is closed.

Yesterday I visited Snowbird and made sure to ask around if they announced any changes in their schedule. They have not. This begs the question how AI came up with a false schedule change. The answer is very important as it will help you understand why AI often provides the wrong answer to questions.

There are multiple ways that AI came up with an answer to when Snowbird would switch to weekend only skiing. The simplest answer is someone speculated on a date and the AI found such a posting and incorporated it into its knowledge base. This often happens in real life. Someone will state a guess as fact and we all believe the person. My dad once told me that magnets are created by hitting a piece of iron with a hammer really hard. While that may be one way to create a magnet, that is not how they are made. I recently corrected his misunderstanding. Doing the same thing with AI is a bit more involved.

Another way that AI can come up with false knowledge is through inference. We are having a particularly warm winter here in the western United States. When temperatures get above a certain point, snow starts melting quickly. This week we are expecting record highs and AI could recognize that when the average temperature gets above a certain point, Snowbird switches to weekend-only skiing. Unfortunately temperature is not how Snowbird determines when people are done skiing. They look at hotel reservations at their lodges as well as a number of other factors and will try to keep the daily schedule as long as Mother Nature allows.

The important thing to gather from this post is that Artificial Intelligence is often wrong. I don't know how many times I have asked Google a question only to have it spit back the wrong answer. While I have not kept close statistical records, I feel like I get the right answer from AI only about 20% of the time. It may point me in the right direction but I always verify the answer and augment what I am told with real human intelligence.  

How Did 2025 New Year's Resolutions Go?

It is almost time to change the calendar to 2026. I thought about posting something about New Year's Resolutions but realize most of them are broken in the first 30 days of the new year. So I thought about how I did with mine this year. As I think about it, I did pretty good but there is always room for improvement.

One of the most easily broken resolutions relates to weight loss. I find that there are smart goals and there are stupid goals in this area of my life. A stupid goal is that I want to shed a certain number of pounds. It is stupid because it doesn't include how to achieve the goal. A smart goal is setting a number of days I want to ski. Skiing is exercise but it is something I love to do. The more days I ski means I exercise more. This year I set the goal to ski over 100 days for the 2024/25 ski season. I ended the season skiing 126 days. Goal achieved!

Unfortunately I can't ski all year long. So what is a smart summer exercise goal? The past couple of years I have not done as many bike rides as I would like and so I set a goal to ride over 1000 miles for the year. I have had some years where I achieved over 6000 so my goal is very doable and stretched what I accomplished the previous year. I have one more day to go but have over 1200 miles ridden so far this year. Goal achieved!

Now while we all have that stupid goal to lose a couple of pounds, did my smart goals help contribute to what I really wanted to do? I can say, "Yes." I am at my healthiest when I weigh less than 160 pounds. The skiing and biking goals helped get me down to 156 but you don't lose weight in the gym, you lose it in the kitchen. I set some smart goals related to eating well and that is what did it for me. If you are trying to lose weight without GLP-1 drugs, the way to do it is by leaving every meal hungry. So hungry you cry. If you can do that, you will drop weight easily. Well I guess it is not so easy as you cry 3 or 4 times a day.

I also had some mental and professional goals and did fairly well with them. This coming year I have set a technology goal that actually applies to computers and technology. My recent post about a single 4K monitor not being enough has me wanting to build a new computer with two 4K monitors. There is more to think about than just purchasing the hardware. I will have to redesign my office to make the space for both 43-inch monitors. I will also need space for my work computer and its second or perhaps third monitor. That is a lot of screens to look at and so I am going to have to build up . . . and maybe down. I will be sure to update readers with how the project is going.

One goal I had for this past year is doing more with artificial intelligence (AI). Fortunately I still have a few months left in my company's fiscal year so I'm not quite out of time. I did get a work environment set up for myself and all I need to do now is a bit of tweaking to tune it for my sample use case. Ultimately I want to be able to type natural language and have it generate the complex SQL to pull the data from a sample database where information is spread across multiple tables. So instead of typing code, I want to type, "What is the most popular movie over the past year and did it effect any video game sales?" The first question is a pretty simple query to a single table. The second is significantly more complicated. I'm close to being able to ask my question but have a little more Symantec modeling to do, which is really just accurately describing the data in natural language terms. I am close though.

As you start the new year, don't create New Year's resolutions without thinking about how you did last year. If you kept some of them but not others, think what you can do to set smart goals and not just boring number ones. Get creative and figure you what you can do to become excited to accomplish what you want. Right now, I'm thinking about going to various Disney parks 4 times in 2026. If you can avoid eating all the food, you really do a lot of walking and can surpass 25,000 steps in a single day. That ought to help with anyone's diet.

Time For Annual Security Training

As a Sony employee I am required to run through an annual 30-minute training course on computer security. In the past the training has seemed to be the same as previous years. Today I ran through the training again discovering it has been updated and it seemed like a whole new course, which I appreciated.

The course had the usual warnings against clicking on links in e-mails and verifying URL's before going to the sites. This year, they provided some more details that helped understand how different character sets can be different than the regular Latin letters we are used to in English. The example they provided is that a Cyrillic V looks like the letter B. Someone could then create a mischievous website using the Cyrillic V for something like the Better Business Bureau and you wouldn't know you are going to the wrong site. Having a concrete example like that really helped underscore how subtle character substitutions can cause havoc.

A new entry in this year's training highlighted mobile device security. I prefer a full-sized keyboard and so if I can keep my phone in my pocket and use a computer instead, I do. I am not normal though as the average person uses their smartphone 6 hours a day. My usage is down around an hour per day. The training pointed out a number of helpful tips to keep from clicking on malicious links that could open your device to malware and other bad actors. I decided that by doing as much as I can on my computer, I reduce my risk for security issues. Should I get a nefarious text, I now know what to look for.

Finally the training had a section on how artificial intelligence (AI) can be used to create more realistic e-mails. Bad spelling and grammar used to be dead-giveaways of scam e-mails. Now those e-mails can be created to sound exactly like your supervisor or manager. They also warned against voicemails that can sound like the people you work with. That is a sobering thought.

The point of this post is that there are some persistent thieves and crooks trying to get access to your computer, online accounts, and smartphone. It is a good practice to review security best practices to remind you to remain vigilant against those bad actors. If you are not required to run through a 30-minute training session, you might want to find a trusted resource on the Web and do your own training.

Using AI Appropriately

I had the pleasure of helping develop the PlayStation 5 video-game console when I worked in Research and Development at Sony Interactive Entertainment. One of the advantages of the hardware is that the file system is fast enough so it is not necessary to keep multiple copies of digital assets. What does that mean? When you create a virtual world, it is made up of a number of digital assets such as trees, rocks, buildings, or any number of other things we find in the real world. Most of the time, those assets are used over and over again. For the sake of an example, a forest is made up of a lot of trees. There may be only 5 different tree models created for a game and then they are reused many times to give the illusion of a forest. With the PS4, game designers would copy those 5 trees hundreds of times. That isn't necessary on a PS5 as you can have just the 5 tree types and point back to each one on the file system any time you need it. Assuming each tree requires 4KB of memory (an arbitrary value pulled out of thin air) and you have 1000 of each tree in a game, the PS4 would require 4MB for each tree used while the PS5 only needs 4KB. Theoretically PS5 games should be significantly smaller than for other game consoles. This becomes very helpful for games like Red Dead Redemption 2 that required 2 Blu-ray disks for the game. For physical game disks, it saves a bit of money and for digital downloads, they don't take as long to put on your system.

So what does that have to do with using AI appropriately? It boils down to why game developers only make 5 tree models. Most gamers don't slow down and look at every tree in a forest to see if it is different than all the others. So why take the time to create more than 5? If you can throw the problem at a generative AI program, you can let it create hundreds of different trees. This provides a level of uniqueness currently missing in a lot of games. This totally eliminates the benefit of the PS5 over other gaming consoles as its games once again become bloated. The reality is that while the PS5 doesn't require duplication of digital assets doesn't mean game studios are using the feature. You will still find copies of digital assets spread throughout quite a few games. Why not make the games more unique?

I have been thinking about this issue for the past couple of days and came up with another area where the use of AI should be applied. I walked through my kitchen this morning to get myself some breakfast. My floor is a high-quality laminate with a simulated pine surface. While pine is great looking, it is a very soft wood and makes horrible flooring that is easily scratched and dented. Using a high-quality laminate allows it to be almost indestructible yet look beautiful. The downside to a laminate is that there are only about 5 patterns on the boards. They repeat quite often and if you look closely you notice a lot of the boards are the same. The laminate floor company could use AI to create 100 different patterns instead of just 5. This would create a much more unique floor.

Yes these two solutions for generative AI don't seem to be high priority. After all if they were, companies would spend the effort to make their products more unique. The beauty of using AI is that products can become more unique and special without humans having to spend more time making it happen. You will still need those humans to fine-tune what is generated but their time can be spent being creative and not doing repetitive and mundane tasks.

 

Using Napkin.AI to Streamline Presentations

There is a lot of focus on artificial intelligence (AI) in the media right now. Some of it is good and some of it is bad. Yesterday I had a colleague ask me how I use AI in my job right now. I had to confess that I know lots of areas where it can be used, such as summarizing lengthy articles or generating complex code, but that I don't feel the need to use it. Then I thought back to when I created my last presentation. That is a classic example of where anyone can use AI to help.

I sit through a lot of presentations and hate it when someone creates overly-wordy slides. It makes for a very uninteresting presentation and is often called "Death by Powerpoint." When I find myself putting a bunch of text on a slide, I take a step back and try to figure out how to replace all the words with a simple picture. Sometimes that can be rather difficult and so I told my colleague that AI is a great way to streamline a presentation. We have an internal engine at Sony that we are encouraged to use. That keeps our confidential text from falling into the wrong hands.

My colleague asked if I had ever used napkin.ai, which is a website specifically designed to take presentation text and turn it into a picture. I immediately logged in and gave it a shot. I didn't have to create a new user or anything as I just logged in using my Google account. Then I copied some text from a personal document I had open and it generated several images for me to select. It worked amazingly well.

Should you find yourself trying to create a presentation and have a slide filled with text, I highly suggest you give napkin.ai a test. It may just help things become more easier to understand. Of course be sure not to input confidential or sensitive information as there is no guarantee it will remain so. 

Playing with Large Language Models (LLMs)

Yesterday I participated in a hands-on lab to integrate large-language-model (LLM) technology to help generate complex SQL queries to look into our data. The lab only lasted an hour and I learned a lot during that time. It also highlighted one of the use cases for LLMs for more than just a toy to help write high-school papers.

When ChatGPT released, it garnered a lot of attention. While everyone seemed impressed, I didn't see a lot of people asking how the technology can be used for good. Automatic code generations seem to be one of those good uses that can really help us in our daily lives. Rather than spending countless hours writing thousands of lines of code to do something important, you can specify what you want and have created and the LLM will generate the code for you. That is what yesterday's lab explored.

For those that don't know SQL, it is a very simple and almost English-like language used to query databases. If you have a table that contains information about customers you can use something like this to query the database:

SELECT firsname, lastname, first_purchase_date FROM customers;

This will list out all customers' names and date when they made their first purchases. It is pretty simple, right? Well it gets complicated once you try to filter that data or include info from other tables in the database. While most of my queries are 25 lines or less, it is not uncommon for me to write 500-line ones. This is where an LLM can really help streamline my work.

In order to keep the lab under an hour, all the data and most of the code was created so I just needed to put it in the right place. I loaded the data into an online database server and downloaded the code to my laptop. The only modification I needed to do to the code was update the connection string to point to my online database. Then we ran through the code see how it worked.

The code included a file to describe the 3 tables in the lab and that turned out to be the secret sauce. I closely looked through the description file and realized that it would take about an hour per table to duplicate this file for the database I work with on a daily basis. When there are only 3 tables, this isn't a large effort. Unfortunately my database has thousands of tables and it would be a huge undertaking to use this technology in my environment.

I had hoped that the LLM would be able to look at my database tables and infer what information they contained. This is possible but unfortunately is not nearly as accurate as having a data expert provide this detail first.

I'm glad I attended the lab and I learned a lot. It also demystified what is going on with this particular LLM. While I hoped I could use the output of the lab to make it easier for non-data scientists to query the database, I did come away with a strong understanding of the underlying technology.

Generative Artificial Intelligence

Today I am attending a seminar internal to my company about generative artificial intelligence or GenAI for short. It is very interesting and while I have played with the technology before, today we get to dive into a number of different topics. In my last session, we talked about the ethics of GenAI and acceptable use within my company, Sony, and how bad actors might be able to use it.

Yesterday I wrote about scam websites. I found it interesting that we would cover how someone could use GenAI to scam people. A few years ago my dad got an e-mail from a neighbor traveling in Europe who got into a bind and needed money. Wanting to verify the story, my dad called his neighbor. Yes, the neighbor was in Europe but no, he didn't need cash. Now using GenAI, the bad actor could call my dad using a fake voice that sounds exactly like the neighbor. What can you do to prevent such a scam from happening to you?

The obvious answer is to do exactly what my dad did before: call the neighbor directly. How you get off the phone is up to you. My favorite is faking a disconnection but the important point is to call the neighbor directly using a number you already know and not one provided by the scammer. This is especially important when the person asking for help is a close family member. While you may feel you can tell the difference between a son or daughter and some GenAI tool, the tools are becoming so advanced, it is getting tougher to hear differences.

As with any technology, GenAI can be used for good but it can also be used for harm. You don't have to become an expert in it to keep yourself safe from potential scams. You just need to understand what is possible and use your own common sense to prevent potential fraud.

A New AI Toy: ChatGPT

Several years ago I worked on a chat engine and learned a lot in the process. At the time we had a long way to go to create a program that did a great job mimicking human conversation. OpenAI is a company that released ChatGPT in November of last year and it does a pretty good job of realizing that goal. Today I decided to sit down and play with it for a bit and see how good it is.

One of my colleagues played with it as well and fed it a multiplication problem with 2 4-digit numbers. It came close to providing the right answer but a calculator revealed it didn't get the problem correct. Unless you pulled out a calculator though, you wouldn't know because it stated the answer with such certainty.

Armed with the information that it was good at answering questions but it wasn't that great at Math, I asked a different kind of question. I thought back to my college English courses and asked why experts consider "Moby Dick" to be such an example of classic American literature. I suffered through the book in college and hated it. Then I decided to pick it up after learning to sail in the hopes of understanding it better. While I understood the vocabulary, I actually prefer Herman Melville's "Typee" over "Moby Dick." I think Mr. Melville did as well.

ChatGPT came back with an amazing answer that would have greatly improved my grade in College. The answer had 5 parts discussing the detail the author included in the story indicating that Herman Melville actually spent time on a whaling ship. It talked about the struggle between good and evil as well as the futility of seeking revenge. Having read the book twice and really understood it once, I had to agree with the answer.

I don't think ChatGPT is going to take over the world. It would need to have a better understanding of mathematics and engineering to do that. It will give college professors something to worry about though. If I was back at the University, I would immediately use ChatGPT to help me start any papers I need to write. So is that plagiarism? Probably but if done correctly, it would be impossible for anyone to tell. I'm glad I am not a professor. About they only thing they can do is move from written papers and exams to oral ones.

A Morality Setting for Artificial Intelligence

Many of us interact with artificial intelligence (AI) on a daily basis without realizing it. If you look at the recommendations provided by Netflix or Amazon, then you are being served results from some sort of AI. Most of the time, those recommendations are based on previous purchases or actions. If you watched one of the Marvel superhero movies, you might get recommended another one that you may not have seen.

Now imagine that you want to change your behavior. Perhaps you have decided that smoking is bad for your health and want to quit or that you want to stop watching violent movies. At first you may have a firm resolve to not do those things any more, then you get an e-mail telling you about a sale on violent movies or smoking products. While you may not immediately give in, you may find it difficult to resist the temptation to give in and go back to your old ways.

So how do you tell online retailers and service providers that you would like to change your behavior and to please stop sending you recommendations for products you are no longer interested in purchasing? Right now, you can't without doing something drastic like changing your phone number, e-mail address, and creating all new accounts.

This also leads to the question of if service providers should build in some sort of morality. When you go to watch the latest superhero movie, how would you feel if you received the message, "Violent crimes are up, perhaps you should watch a romantic comedy." Or you go to purchase cigarettes online (I don't even know if that is possible) and see, "Smoking is bad for you, how about some nicotine gum instead?" My instinct says nobody would be happy even though we would all be better off with such suggestions.

A Scary Thought for Halloween

This morning I answered a one-question survey about what is the most scary Halloween character. A number of the usual suspects could be chosen. As a child, a monster would have definitely have made the list. Now that I am an adult, the most scary things in my life are much different. I am scared of things like getting extra random screening at the security in the airport or having the IRS decide I didn't pay my fair share of taxes. I am also afraid of illnesses like cancer or Alzheimer's disease.

If I stop to think about monsters, the scariest one on the earth today is man. Think about the recent mass shooting in Las Vegas and it has totally stolen the spotlight from natural disasters like hurricanes Irma and Maria. This has me rethinking several of the projects I am researching at work. I am helping to build some really fun technology, but in the wrong hands, it could be used to do much more scary things than any mass shooting to date.

Artificial Intelligence is one of the subjects I am researching and still think it is pretty stupid. However there are some areas where great progress is being made. Pull out any of the latest smartphones and try to take a picture of someone. The first thing you notice is that someone's face gets highlighted so you can provide a name or tag to be used by Facebook or some other social media site. This means that a tiny device that fits in your pocket knows the difference between a person and a random object. Now couple that with acoustic analysis where a couple of speakers can tell the direction that person's voice is coming from. A number of toy-robot researchers are using this technology to create robots that look at who is talking in a group of people. All of that seems pretty harmless and fun, right?

Now let's take that same technology that can recognize faces and knows how to aim at a person. I don't want to elaborate any more as I don't want to give anyone any ideas. However my realization this morning left me with a sick feeling in my stomach and the thought that while it isn't quite SkyNet from the Terminator movies, it is moving in that direction. Perhaps I shouldn't be so quick to dismiss the scary uses of Artificial Intelligence. Monsters are not the things scaring me this Halloween season.

Demystifying Artificial Intelligence

One of the nice things about my job is that I get to play with some of the latest technology. Right now I am doing a lot with artificial intelligence (AI) and machine learning (ML). I remember back in the 1980's and how AI had the potential to be something great. Then it stagnated for 2 decades. Due to the low cost of computer processing power and memory, AI is making a comeback. Hopefully it has a bit more staying power this time around.

As I research new technologies it is important not to buy into the hype. I read one article today claiming that those that don't make the effort understand AI and ML will begin to look at computers as magical machines that they understand less and less. That can be true of any technology and so if you find yourself thinking of a new product as magic, that is what the marketing people want you to believe.

The more I work with the current state of artificial intelligence, the more I realize it is just really good at guessing the right answer. Most AI engines get fed a bunch of numbers (characters and words can be represented by numbers) and what those numbers mean. The engines then try to figure out what a different set of numbers mean. It then becomes very important to train the AI engine with a lot of data as well as the correct data.

What happens if you train your engine with the favorite breakfast cereals of children between the ages of 5 and 7? When you ask it what an adult would like for breakfast, you will most likely get the wrong answer. While I do like Fruit Loops, I don't eat them for breakfast anymore. What happens when you recognize the age mistake but only train the engine with data from Americans? If you ask what someone from Japan eats for breakfast, once again, you will get the wrong answer.

Currently artificial intelligence is at an interesting point. We are creating useful applications for it that can greatly help us in our daily lives. It is also important to remember that it isn't magic and is only as smart as the data with which it has been trained.

Machine Learning and Lots of Data

At the beginning of the month I wrote about artificial intelligence and how it is not going to take over the world any time soon. I have continued to play with it and have been working on a branch called supervised learning. My basic example or use case is to feed my simple program a bunch of training sentences that are categorized as questions or statements. Then I have a test group of sentences to see how well my program has learned.

I started with a very small set of training sentences thinking that a person would be able to distinguish between a question and a statement fairly easily using just these examples. My training set began with only 30 sentences and my test set had 20. After training, my program correctly identified 16 of the test set. That sounds pretty good at 80% but I really need to get closer to 100%. So I added more training data. I found a list of 800 random questions and added them as well as several pages of text from two popular books I found online: Uncle Tom's Cabin and The Old Man and the Sea. That brought me closer with 18 sentences correctly identified as statements or questions.

The statement, "I like to ski," was wrongly classified as a question while "Who is your favorite actor?" got classified as a statement. So I added more training data until I got 100% correct classification. My original training data started with 30 sentences and is now close to 1000. That seems like a lot of extra work.

Now it is time to tune my algorithm. There are some things I can do to get better results with less data. However if you plan to embark on your own supervised learning project, be prepared to collect a lot of training data. You will need it.