Algorithmus on the mountain

Rewrite my history: AI assisted coding experiment by a non coder product manager

Changelog displayed on a computer screen showing entries with added functionality for Wayback machine time, Legacy Wayback links, and Longread detection, along with a run-level count in the log.
Version control: because someone has to remember this

I was going to start this post with, “Do you read the papers?” But of course you don’t. Nobody does. But you watch TikTok for news, or see clips of a shouty man on LBC. Something like that. And, if you’re aware of the world around you, you’ll know that things aren’t going well. Unless you are a billionaire with stakes in Artificial Intelligence. Then, well, you are pretty much treated like some kind of God. If this were ancient Rome, you wouldn’t be another minor deity. You would be a figure of cold, calculated logic and immense influence. You will be called ‘Algorithmus’, or something like that.

If you were Algorithmus, you’d look down at the lesser gods on Mount Olympus and enjoy all the excitement around AI. The hype would be your strength. Meanwhile, ordinary people sacrifice their privacy for your benefit.

The Romans didn’t have AI, but they would understand hype. That’s what’s happening with AI today. I’m not ancient Rome old, but I do remember the internet buzz in the early 2000s. The internet was overhyped and in a bubble, but after the excitement faded, people found real uses for it. It’s similar to how Roman roads were built for armies to destroy villages, but ended up helping everyone.

My memories of the internet hype, more than Roman roads, shape how I view today’s AI tools. AI and machine learning are clearly useful. The chatbot trend will likely become something practical for society, hopefully in a positive way. Lots of people already use basic AI tools to be more productive. Still, I wanted to try it myself. I was curious about “vibe coding”—the idea that you can build complex software just by giving prompts to an AI coding tool. Could I really create working software just by chatting?

I’ve worked in technology my whole career, but I’m not a coder. I did take a computer studies O-level back when that existed, long after the Roman Empire, just to be clear. So I have a basic grasp of coding. As a product manager, I understand the challenges and limitations engineers face. I once worked with a senior developer who thought anything could be built with software if you had enough time. Now, I’m not so sure, since deadlines and delivery dates matter just as much as what’s possible in theory.

Software development has changed a lot in my lifetime. It wasn’t really a formal field when I was in school, but now it’s central to millions of jobs. Some companies only build software, and everyone else relies on it to run their business. Engineers have always turned ideas from people like me into real systems. Now, the question is how much of that work AI can do, and what skills you need to guide it well.

A small experiment at work a few weeks ago got me started. I needed to make sense of one of our system logs, which are tough to read. Online tools can make them look nicer, but without context, they’re not very helpful. I built a simple tool and provided enough context to organise the log output in a way that made sense for us. That experience inspired me.

A big part of being a product manager is turning business or market needs into things engineers can build. That challenge doesn’t disappear with AI. In fact, you might need to be even clearer about what you want. Human engineers usually have some background knowledge, but AI agents don’t, no matter how much training they get. They’ll improve, but for now, they really need clear instructions.

Recently, OpenAI added more features to its coding tool as part of my plan, so I decided to try something personal and see if I could “vibe code.”

The project

Back in the early 2000s, when blogging was at its peak, I wrote often and hosted my own site. It’s still online. About ten years ago, I added a feature that shows me posts I wrote on this day in past years. Usually, that’s two or three posts.

I’ve long wanted a small app that could pull up those posts each day so I can review them and decide if they should stay as they are. Part of that depends on whether the links still work. I see those posts as a time capsule and don’t really want to delete them. But blogging was fast-paced, and I didn’t always check spelling or proofread. I want to fix that, but not all at once. I need something I can run occasionally, fix a few posts, and move on.

So I described a tool that would pull the right pages, find each post, check spelling and grammar, and look for broken links. It needed to run on my computer and create a report with suggested fixes and any links that didn’t work.

Ramblings to working software

I told the coding agent to assume I didn’t know how to code and to guide me through the process. Within an hour, I had a working prototype. I was surprised at how quickly I could go from a prompt to something that worked. I could test it, provide feedback, and request changes. The next day, I added more features, and the AI even found a list of over 2 million UK place names to improve the spellchecker.

Now, I have a daily web page that links to my old posts from that date, each one checked for spelling and links, with a log file that explains what the system did. It’s not commercial software, and there’s no extra interface beyond what I asked for. But it does exactly what I need.

Encouraged by that success, I tried building a second tool. This one would download any of those blog pages, reformat the text, and let me paste the result somewhere else. The AI suggested making a Safari extension, so I needed extra build tools and had to connect it with software already on my computer.

This project was more complicated and a bit frustrating. I probably could have guided the AI better if I understood the build tools more. I wanted it to work with an existing app that has a command-line interface. The AI kept suggesting solutions that didn’t work on my setup, even though I explained what I was using. That’s when my limited coding experience became obvious.

After some compromises, I got a working version. It runs on my computer and does what I wanted. During the process, the AI just waited while I went to meetings and came back hours later. It picked up right where we left off. Since language models don’t exist between chats, there’s no one waiting impatiently. I just came back when I had time.

I can see how a skilled engineer, who knows when to step in on complex parts, could get much more done with this kind of help. But my experience shows that even someone with limited coding skills, as long as they can describe the problem clearly, can quickly build useful software and solid prototypes.

There were other small surprises, too. I asked the tool to write a user-friendly description of the software’s functionality. In less than a minute, it came up with something better than I could have written in hours. With a few more tries, it could have made full documentation. The log file it created explains each step in plain language, which I’ve always liked for systems that need to be clear to non-coders.

So, where am I on the hype cycle now? I don’t think AI will replace engineers. You still need someone to turn ideas into working solutions. Clear requirements and context are still important. But I do think that for people who understand technology, AI makes it much easier to build small, useful tools. It lets you try things that might have stayed as just ideas before. For experienced engineers, it’s definitely a superpower.

The hysteria may fade. The Roman invaders will leave. Algorithmus might end up in jail. But the useful things will stick around.

Weeknotes #106: Happy Lunar New Year

AI experiments, personal reflections, life updates, and intriguing challenges explored.

Week commencing Monday, 27 January 2025

A sea view at Seaview

Quantified Self

  • This week: Stand 7/7; Exercise 6/7 and Move 5/7. (86%). Morning walks: 0/4 (days in the office don’t count). Office days 1/5. Total steps: 58,567

The weeknotes are a bit different this week as I spent some time reading and playing with a number of AI tools, specifically using my own material—like these weeknotes—to see what the models might do with them.

Let’s talk about AI

  • Once again, there’s a lot of chatter about AI. The Chinese DeepSeek models made a big splash this week and caused US technology stock prices to plummet. It’s so overhyped.
  • Unrelated, I am experimenting with Google’s Notebook LM, using last year’s weeknotes and a collection of old blog entries from 2002/3. Even though the app says my data is not used to train the AI, I figured it’s public anyway. This is where the real power of AI will come to most people: when it can be used on their own data.
  • – I asked for a cast of characters, which came out OK, but I guess it’s a surprise for people I know IRL called Frank who the AI assumes are the subject of a drag musical. For clarity, they are not the same.
  • – Related, the AI says my weeknotes highlight “a lifestyle that values social connections, shared experiences, and cultural engagement”. I think that’s mainly because I don’t write about the hours I sit at my desk typing product requirements for work.
  • – I was pleased it said that the overall tone of the weeknotes is generally positive and reflective, but it’s missing so much information about me it made me think about how I am characterised online.
  • – In another task, the AI couldn’t distinguish between me and my father when writing about countries where I have worked.
  • In response to this, I asked ChatGPT if it knew who I was. It basically regurgitated materials I have online, which (I think) is good. Other AIs, including DeepSeek, didn’t come back with anything.
  • Google Notebook LM will produce what it calls an Audio Overview which, for me, was like somebody had created a podcast about my life as viewed through those blog entries. It was uncanny. And, yes, I think it’s very much like the fortune reader whereby much of the generic stuff sounds really personal but isn’t, but it was unnerving listening to a podcast about yourself; there was enough reality alongside the generic to make it seem like it knew more about me. It was capable of referencing cultural changes between the early 2000s and today, which made the analysis seem very real. A bit like listening to my own obituary. I can’t decide if I liked it or not.
  • There are a lot of interesting capabilities. I wonder how much longer before I decide to pay for access to more powerful models and tools?

Life

  • I am trying to arrange broadband for a place in Spain while not speaking Spanish nor understanding the format of many things. It’s more stress-inducing than I imagined.
  • On Wednesday, we passed into the Lunar New Year of the Wood Snake. At my Tai Chi class, we all wished each other a happy new year.
  • This week, I learned the government has a Hammersmith Bridge Taskforce. The taskforce met 16 times between 1st October 2020 and 25th November 2021. The bridge is still closed. Pretty terrible such vital infrastructure isn’t funded so it can be repaired.
  • Our first visit to the Isle of Wight for 2025. Delighted by how lovely it was.

Media

AI-sourced views on HS2 Northern leg cancellation

Some of the text on this page was written by 'AI' using gpt-3.5-turbo. I have corrected some US spellings to UK spellings in the response as the prompt for British English seems to have partially ignored

Back at the start of the year, I shared a piece of generative AI that I had created to try to understand how a mainstream tool worked in a specific, more technical context. I think it’s great but, at this stage, it was a generalist.

One of the limitations of the models used is that they have been trained on data to a fixed point in time and many are not able to look-up new data or information against the wider knowledge of the internet. Poe is a collection of AI tools and I noticed a bot had been released that is capable of conducting a web search.

So, with this week’s news that HS2 is to be scrapped I thought I’d try out the this AI to see what it came up with. I am quite impressed with both the output and the referencing but it is limited to a couple of sources. In particular, the references to the Construction Plant-hire Association (CPA) are to a single 185 word statement. Nonetheless, I thought interesting new capabilities from AI.

The Prompt

Summarise responses to the cancellation of HS2 to Manchester. Use British English.

The AI generated response

The cancellation of the HS2 leg to Manchester has generated various responses. Here is a summary of the key reactions:

Disappointment and Concern

  • Many individuals and organisations have expressed disappointment and concern over the cancellation [1].
  • The Construction Plant-hire Association (CPA) described the decision as “extremely disappointing” and a significant blow to the UK construction sector [3].
  • Former Prime Minister David Cameron criticised the cancellation, stating that it was the wrong decision and would make it harder to build consensus for future long-term projects [2].

Impact on Plans and Confidence:

  • The cancellation has raised concerns about the impact on people who had made decisions based on the HS2 plans, such as moving or selling their homes [1].
  • The CPA warned that the decision could undermine confidence and impact future investment in the construction sector [3].

Alternative Funding and Projects:

  • Prime Minister Rishi Sunak has promised to use the £36 billion of savings from scrapping the HS2 line to fund other transport schemes [2].
  • The government’s commitment to completing the HS2 line to Euston and diverting funding to other projects aimed at boosting transport infrastructure in the north has been highlighted [3].

Learn more:

  1. Tell us: will you be affected by the scrapping of the HS2 leg to Manchester? | HS2 | The Guardian
  2. LIVE: HS2 Manchester link cancelled by Rishi Sunak – updates and reaction – Manchester Evening News
  3. CPA Response to Cancellation of the Manchester Leg of HS2 | Construction Plant-hire Association

Follow-up

Interestingly, WordPress (which is used to maintain this site) now includes an AI assistant that can review what has been written on this blog. I generated a review of this page. One of the pieces of feedback was, “Include diverse perspectives by including reactions from a wider range of individuals and organisations”. It amused me that an AI was, effectively, telling another AI to do better.

Generative AI

Some of the text on this page was written by ‘AI’ using text-davinci-003

Yesterday, while on a break between meetings, I asked one of the currently popular AI tools to write something about how advertising might be impacted by AI technologies with a focus on the space I am currently working in: addressable advertising for broadcasters and pay TV operators. I also asked it to highlight some associated risks.

I asked a very open-ended question so I was not expecting anything specific about any individual technology. I tried a few times to see what alternatives were suggested. In general, what I got back was wonderful marketing speak: better targeting for advertisers, more relevant ads for viewers and more money (because of both of those things) for broadcasters. If I was starting a new ad tech company then I’d probably use this text. But I didn’t learn anything. Responses from the current models are very generic. I’d love to try the same question on an industry-specific data set trained with fewer buzz words and generalisations.

The prompt:

Write three paragraphs explaining why AI will improve advertising technologies with a focus on technologies for video and connected television. In the first paragraph outline the benefits for advertisers. In the second paragraph explain the impact on broadcasters and viewers and in the last paragraph outline some risks associated with AI and advertising. Use British English.

The response:

The use of Artificial Intelligence (AI) in advertising technologies has the potential to revolutionise the way brands, broadcasters, and viewers interact. For advertisers, AI technologies offer the potential to increase the effectiveness of their campaigns by making them more personalised, targeted, and cost-effective. AI-enabled technologies allow advertisers to identify the right viewers, track their behaviour and preferences, and deliver customized ads that are tailored to their interests. AI can also help advertisers better understand the impact of their ad campaigns, identify potential opportunities for improvement, and optimise the ads for maximum performance.

The introduction of AI in video and connected television advertising technologies also has the potential to enhance the viewer experience. AI-based technologies can be used to deliver personalised, targeted, and interactive ads that are tailored to the viewers’ interests. This can help broadcasters create more engaging and entertaining advertising experiences for their viewers, as well as offer brands the opportunity to create more immersive and compelling ad campaigns.

However, there are some risks associated with the use of AI in advertising technologies. For example, the use of AI could lead to the emergence of new forms of advertising that are intrusive and manipulative. AI-based technologies could also enable companies to track and analyse consumers’ behaviour in order to deliver ads that are tailored to their preferences. This could lead to the creation of a ‘surveillance economy’, in which companies have access to an unprecedented amount of personal data. In addition, AI could be used to create ‘fake news’ and other misleading content, which could undermine the trust and credibility of brands.

Source: https://platform.openai.com/playground/p/tcPXeuzqQmIXTsTrsT1y0Q20?model=text-davinci-003.

And She Said, “Life in plastic has been fantastic”

But really, the job I’d really want, is writing the answers to the daft questions people ask.

Earlier, on Facebook:

Voice recognition systems – Amazon’s Alexa or any other – must be one of the most fascinating areas of computing to work on. There are a lot of very smart brains working on artificial intelligence systems that enable Alexa to answer the millions of questions asked of her every day.

But really, the job I’d really want, is writing the answers to the daft questions people ask. “”Alexa, what’s the first rule of Fight Club?” or “Alexa, do you really want to hurt me?” Anybody who job means they get to write “.. and happy birthday to Barbie whose life in plastic has been fantastic for 58 years” must be having a blast.

Mashable has a story today about a woman who asked Alexa about any connections to the CIA – at which point the system appears to crash. Obviously proof that all conspiracy theorists are correct.

The person with that fun job was swift off the mark. Now, somebody’s waved their neuralyzer & Alexa’s been updated and claims to work for Amazon. If it was me, I’d update Alexa’s sense of time and stop responding to my happy ‘good morning’ at 10pm with an equally up-beat ‘good morning’ and make her grumpy and tired.

Or maybe that would be a bit too human.