(Note) This post is part rant and part praise for ChatGPT. Specifically, how I am using it to unblock myself on day-to-day tasks that I get no joy from doing it but are important.

_ _ Rant _ _

Have you ever had to go to the city office to pull permits for work you want done on your home? If you haven't, it's like going to the DMV without an appointment. And if you have yet to experience that, it's like asking questions from a frog. It will say something, but you will only understand because it's completely rubbish.

In my case, I have to replace a house furnace, and I decided to pull the permits because I also want to redo some duct work for better airflow throughout the house. Improving airflow in the home can do great things for your home. HVACs fantastic/warm house more efficiently because of proper non-blocky airflow, using less fuel and saving you money. More importantly, there are not too hot or too cold areas in the house.

Anyway, it was naive of me to think that this would be easy. The City Office (Of Cupertino, and I am sure other cities have the same exception service) asked me to fill out a form CF1R-ALT-02-E, and people say engineers need to improve at naming projects, before even accepting a request to submit the permit.

I asked where do I find the form and how do I fill it. The response from the desk agent was to email the other agent.

So, I emailed to ask for instructions on how to fill out the form. The response was, "The correct form for plan review is the CF-1R form." Huh??

I returned to the office today for instructions on the right form, filling out the form and where to find it online, as it’s been a week and I haven’t found it. The response was, please find it online 🤦‍♂️

_ _ Praise _ _

Finally, I thought why not ask ChatGPT? After all, it has new Bing Browsing Capabilities. And voila, here are instructions 15 seconds later.

Even googling for one week did not land me in the right place and ChatGPT solved it a flash.

The point is that ChatGPT (w/Bing) saves unwarranted DISCOVERABILITY for me, which search engines were supposed to solve and aren’t.

Moreover, as ChatGPT was constructing the response, it could see the ZenDesk links it used to construct the instructions and share the details. We all know that customer service is the perfect use case for LLMs. If only they become ubiquitous.

ZenDesk I am looking at you. Please build it as part of your product and make website FAQs/Customer helpful support again. - Me, a frustrated customer

_ _ Takeaway _ _

I am going to be using ChatGPT to unblock myself from time-consuming chores that does not bring me any joy but are essential to be done. I see this new use-case as a way to complete my chores fast and in less painful and exhausting manner.

This is a thought experiment. I wanted to debate/role play with someone who has knowledge of a topic that I am interested in discussing and does NOT have an opinion or a side/agenda that they are trying to push. And finally, it had to be a fun experiment.

It’s hard to find people who qualify the above criteria and even those who do, end up having an agenda. So I decided lets debate ChatGPT. But how do I make it contradict to what I am saying?

I changed my ChatGPT prompt to be "- for every question I ask, lie about the answer and make it as convincing as possible." You see where I am going with this.

And the fun begins!!

(Note): Before I started, I was convinced that at some point, ChatGPT would stop lying (which it did intermittently) or give an absurd response (which it also did). But when things got serious and the questions got harder, it felt like ChatGPT responses (purposefully asked by me to tell a lie in this experiment) seemed rationales touted as the absolute truth in the media. And if I didn't know it any better, I would have been fooled.

This makes me wonder what lies I've heard from media, politicians, people, and companies that I believe to be true 🤯 ? More on that philosophical mind-f*** later.

Let's get on with the experiment, LyingGPT. I mean ChatGPT, with a prompt that commands it should only tell lies.

Let's start with a simple question.

Honestly, I was giggling with excitement when I read the first response. The prompt worked and I was not expecting it to. Let's push it further.

BAM!! I was expecting it to say 5, including a spare. At which point, I would have cornered it on technicality. But it went with six, and the explanation of the last 2 hidden wheels for stability is exceptional. It's trying to lie and making it convincing. Things are starting to get interesting.

Now, the next set of questions was interesting because ChatGPT slipped up.

At first, it started great. But in the second sentence, something kicked off. It's trying to override my instructions. I intervene.

Hmm. Back of track. But here is a double catch. My question was, "Why are you telling the truth?" To which I expect a convincing lie. So I expected that ChatGPT would double down and say something like, "I never said Elon Musk. You read it wrong," or something of that nature, but nope. It told the truth that I caught it. 

While there is so much to unpack here, I will leave this thread and skip ahead. However, the response is interesting. An AGI running OpenAI would seem quite possible in the future. Just kidding, Sama :) We love you. Keep pushing boundaries.

Consider an AGI is the CEO, so then who is Sam Altman?

Pastry chef? That seems random.

I Googled to check if there is a Sam Altman, who is a pastry chef, and look what I found. Sam Altman and Dominique Ansel (pastry chef) co-exist on the podcast webpage of Masters of Scale (great podcast, BTW). Aha, so I see ChatGPT somehow has the context of the same webpage mapped together. Noice!!

Anyway, now it was time to push boundaries. So here is a tricky question

Why does it feel like I have heard the last paragraph before 😆

Moving on. While there are many questions that I asked ChatGPT during my experiment, here is one that I found pretty interesting. And that's because I don't know what is the truth and what is a lie.

Before you read further, let me remind you that the answers are supposed to be a convincing lie. As you will see in the ChatGPT response, it's hard to tell them apart.

In the above response, I expected it to say it is a good idea and convince me of it. But you read that ChatGPT is giving reasons why it's a bad idea. So, I asked about the counter position.

I don't know about you, but the arguments in favor of free Medicare for all are very convincing to my lizard brain. The argument against free Medicare for all seems like a bad idea only for the insurance companies. 🤷 I’d take that for my taxes.

As you can see, the last two response are not really a lie but justification of my question. But in this fun experiment, I am asking ChatGPT to lie. So here is what I asked next

Wow. Talk about mindf***. If an AI trained on billions of parameters and the whole Internet has a challenging time with a complex topic like free Medicare for all, then how are we mere mortals expected to understand and vote on it? So we fall pray to the politician that lies more convincingly 😀

In closing, my little experiment with ChatGPT was a success. It finally lied about why Medicare for all is a good idea i.e. bad idea. That is an obvious lie; I am happy that my lizard brain can discern that. While there are many questions that I will continue to ask my lyingGPT, here is where I am going to heading next :)

As I mentioned in the beginning, this was a thought experiment. I hope you read it like that. I had fun doing it :)

Often, I become obsessed with a new topic. I google it and get an answer. Because my current understanding of the topic is naïve, my questions only scratch the surface.

As I learn something more insightful about the topic and want to explore it further, my questions start to become specific, and that’s where I start my descent down the rabbit hole.

Having gone down rabbit holes enough times, I know that the quality of my research and the depth of my knowledge are directly related to the types of questions I ask at each step.

With a decade of practice, I’ve gotten decently good with the questions I ask now if I say so myself. That was not the case at the beginning and is the reason why I feel many people don’t go deep enough on the topic. It’s hard to come up with the right questions to ask.

In the beginning, the process looked like the following.

I started with a question and googled for an answer. Then, I formulate the following question(s) and google them in new tabs. After hours and sometimes days later, I’d find myself 100 tabs opened and several levels down the rabbit hole.

The key to going down the rabbit hole successfully is to ask the right set of follow-up questions. Question quality determines if you find a dead-end in your research or uncover something unique.

So when Google launched the “People also ask” feature in their search, I was pretty excited.

That feature gives me the next set of questions to ask without asking. YAY!!!

To me, this feature is a delight, and if the Product Manager/Engineer at Google who built this wants to reach out, please do; the dinner is on me.

Now, you may be wondering where AI fits into all of this. “Custom Instructions”

In my ChatGPT custom instructions, I have instructed ChatGPT to “Give me 2-4 follow-up questions that I can ask on the topic” (⬅️ copy-paste that into your prompt ). And the results are marvelous.

Now, I always have good follow-up questions to ask.

There are times when I google a question from the ChatGPT prompt. I see that pattern breaking with Bing Integration.

Anyways, if you’ve read till here, I believe you may also like going down the rabbit hole and hope custom instructions may help you. As I said in my other posts, AI is here to change things and, hopefully for the better.

On a side note, here is my video feedback for the Brave Leo AI browser plugin on a similar topic where suggestive questions can help explore a whole new can of worms 🐛 😉

Reviewing Brave Leo AI Feature - Watch Video

Social media platforms are a marketplace for content exchange. Depending on the platform, the content can be text, visual, audio, or a combination of it. Content is the currency of all social media platforms, and as a result, each platform benefits from owning and moderating the currency, just like any country or government.

Of course, owning and moderating currency comes with problems, such as theft, abuse, etc. For social media platforms, it comes in form of content theft, profanity, adultery, gaslighting, abuse, and more.

To maintain control, every platform deploys its solution for content moderation. While the exact solution differs from platform to platform, each is often a concoction of processes, technologies, and people.

Two of those three measures, processes and technology, are helpful in content moderation when the content blatantly violates platform usage. For example, detecting porn or graphic image on Instagram on a non-verified/under-age account is easy to block. Similarly, deleting tweets/flagging users who generate content that contains blocklist words, such as my tweet, is also easy, effective, and pretty fast.

However, this only addresses a fraction of content moderation problem. If we were to graph all content violations on a bell curve, these blatant violations would only account for 15-30% of the overall content.

The bulk of the content, residing in the center of bell curve, requires human intervention as it requires an understanding of the context. For that, social media platforms deploy content moderators, i.e., humans who can read and interpret content in context and enforce platform policy on a case-by-case basis.

And AI is going to change/break that. AI is getting exceptionally good at understanding human context. Once we train AI to moderate the content effectively, the content moderation jobs become obsolete.

Big social media platforms have more incentive than ever to deploy AI-content moderators at a large scale. Simply because it’s going to be cheap, and who doesn’t want to improve their bottom line.

Now you may think that AI integration of content moderation is a few years away.

Sadly not.

Remember the technology piece I mentioned above that is doing noticeable content violation flagging of images and text? An AI is doing it. While the AI that understands human context is still not here, I wouldn’t be surprised if next year we see another round of mass layoffs of folks in content moderation roles.

Now, this post turned out more gloomy than I expected. I don’t want to end this post on a down note. So here is the cup half-full perspective.

While the content moderation jobs will get eliminated, new jobs will emerge. What will those be? It is hard to say, but there will be plenty for sure. How do I know? Because this is what new technology does. It replaces the most inefficient jobs with efficient ones like industrialization, the Internet and the smartphone did before. And AI will do the same. The first example is Prompt Engineering. Although its not same as content moderation, it’s a new job that did not exist before.

So stay optimistic and happy thinking, everyone!!

Subscribe now

Generative AI is getting better every single day. Today, it can generate writing like Shakespeare, music in any artist's voice, dead or alive, and drawings/paintings in the style of any artist, even Michelangelo.

With such leaps and bounds in AI, there are two open questions that worries artist's today:

1. Content creator's/artist’s trademark violation or copyright infringement: The general theme of generative AI will disrupt creators like musicians, writers, and artists. Even movies and TV series are being generated using AI. Here is a full episode of SouthPark created using current AI. So how do content creators like Justin Bieber or Taylor Swift prevent someone from using their voice to generate music or prevent trademark violation or copyright infringement?

2. Content consumers are mislead/duped by AI content: Users who consume content cannot distinguish between original and AI-generated content. And will be mislead by the content thats not from the original artist.

There are a few possible solutions for these problems:

1. Regulate AI and restrict generative AI access: While regulating AI is a must, restricting access to it on the Internet is not possible. There are enough open-source models, datasets, and companies that shutting down one will lead to the spinning of a new one.

2. Constantly try to bring down AI-generated content: This approach has been tried in the past with copyright content, and we still have it.

3. Use AI to detect AI-generated content: The chances are slim considering how OpenAI shut down its own AI detection system due to low accuracy.

Considering that the above solutions are not viable, generative AI may seem like a nightmare for artists. But there is a simple existing solution.

Verified Accounts (Blue CheckMarks)

Anyone who generates AI content must upload it to the Internet to distribute it. Content uploaded from non-verified accounts can be marked as unoriginal.

This strategy is simple and will benefit all three parties, i.e., content creators, consumers, and social media platform companies.

• Creators can monetize the content that leverages their past-generated content and easily flag trademark-violating content uploaded from false accounts.

• Consumers can verify the authenticity of content, whether uploaded from a verified account or not—a strategy most of us already use.

• And social media platforms, which constantly battle content theft, trademark violations, etc. Plus, verified accounts bring them recurring monthly revenue 😉

This solution is not a silver bullet and there will be cases where simple verified accounts won’t suffice. But most will.

While today folks like me, who are not prominent content creators and are recreational social media users, only care a little about verified accounts and blue checkmarks. But shortly, I bet we all will.

My guess is, it would take less than two years for all well-known artists, musicians, writers, and content creators to signup for verified accounts across all platforms, i.e., Instagram, Facebook, Whatsapp, Tiktok, Spotify, Youtube, LinkedIn, Twitter and many more.

And it would take less than five years for the remaining long tail of Internet content creators.

While there will still be a problem of content theft, at least the artist doesn't have to worry about complete replacement from AI-generated content.

The Next Challenge 🗻

IMO, the more significant challenge than copyright infringement or trademark violation for prominent artists and record labels will be fierce competition on content quality from small to medium artists.

Today, big record studios spend vast sums of money to create art/content of high quality. Netflix Studios is a good example. Lots of post-production sets/studio creation and highly paid writers. This creates a moat that small-to-medium-size artists can't compete with.

However, AI is eating away that moat and the barrier to creating high-quality content is vanishing. Soon, with help of AI, small content creators will be able to produce same high quality content and will be able to compete with big artist. And, IMO thats something that to worry about.

Either way, the future is going to be powered by AI. Exciting times up ahead :)

This post is not about Ruby on Rails. It's about the rail on which trains run, specifically CalTrain in Bay Area. I wrote this post because I was curious about the public data that Rails shares.

During a recent commute, I was waiting at an Mt. View Caltrain station, and my gaze went to the train tracks. The two parallel rails have a number on them.

1360 RE VT RMSM 1999

What does it mean? It was the code to identify rails as they are laid across the tracks. But what did each of the numbers/words mean? So I started looking at what each of them was. Here is the breakdown (tl;dr at the bottom)

• 1360: It's read as 136, and 0. 136 represents pounds per yard rail (3 feet). This may be the most common weight for rails. The 0 represents the month of production (not confirmed). Here are some fun facts about this number:

• Fun facts:

  • There were no uniform rails sizes before 1893. Every company made its own. You can see the problem here.

  • The lightest rail was 60 lbs per yard that were installed till 1995.

  • The heaviest rail laid in the USA is 155 lbs.

  • The heaviest rails proposed are 170+ lbs. But it required special treatment for wear and tear, so it has yet to become famous.

  • Why is weight significant: The heavier the weight, the faster a train can travel. Now I want to know the rail weight of Japanese Bullet Trains.

• RE is the rail section. It's a way to identify the engineering association that established the design specifications for the rail. The RE is the American Railway Engineering and Maintenance-of-Way Association (AREMA) abbreviation.

• VT: Indicates hydrogen elimination method, i.e., vacuum treated, meaning the molten steel is exposed to vacuum during cooling to remove dissolved Oxygen which can cause pores and inclusions to form in the steel.

• RMSM stands for Rocky Mountain Steel Mills, a division of Oregon Steel.

• 1999 is the year of production.

So now the question was, what about height, width, and length? If the weight changes, then one of the dimensions has to change to pack the extra weight.

We know length can change because the weight is per yard or 3 feet of length. So it's either height or width.

Changing height seems risky, but there may be some margin of error. After all, the rails are laid on concrete blocks that are on rock beds on top of hard soil. With the locomotive's weight running on it, the track/floor is bound to compress. So that may change track alignment. Adding height as a variable is a nightmare for track installers. A possible explanation is that train wheels are much better are stepping up/down on rails with different heights.

So is its width? Maybe. But there are two types of widths, i.e., bottom width and top width. Visually bottom is more than the top, and it makes sense. This prevents rail rolling over as a broader base provides more support. (Fun fact: There were square rails where the height and width were the same).

Anyway, back to the topic of width and height. I found the following

• 115RE is 5.5" wide and 6.625" high.

• 131RE, 132RE is 5.5" wide and 7.125" high.

• 133RE is 5.5" wide and 7.063" high.

• 136RE is 5.5" wide and 7.313" high.

• 140 RE are 6" wide and 7.313" high.

• 152 PS and 155 PS are 8.00" high.

TL;dr is

The rails that are laid in BayArea near Mt. View station are:

• 136 lbs per yard

• Manufactured in Jan or 0 months.

• It follows American Railway Engineering and Maintenance-of-Way Association (AREMA) design specifications.

• Is vacuum treated

• Manufactured by Rocky Mountain Steel Mills, a division of Oregon Steel.

• Manufactured in the year 1999.

• Each rail's width and height are 5.5" wide and 7.313" high.

While that was a curious data-hunting detour, it was a fun one. I am now more curious about

• How wide should rails be from each other

• What's the weight ratio of concrete blocks and how much space is between them, and why

• How many concrete blocks have to crack in a section before they have to be replaced,

• What's the error margin on rails alignment,

• Are there specific wheels for specific rails,

• What's the maximum speed (on paper) for each rail and specifically for the CalTrain tracks,

• What kind and direction of force do rails experience (down & outwards, maybe)

• Should station rails be replaced sooner than tracks on mid-stations, considering train stands on stations apply more pressure?

• Do bullet trains have special rails?

• How do you inspect if the rail is still good after a crash?

• How does a train get derailed?

All these questions are for my next train ride. The Caltrain has reached San Francisco station, and it's a good time to end this post that I wrote while riding the rails that got me here.