Human First

We are surrounded by AI blogs, it is everywhere. Nobody talks about what they are creating to add value for the end user. AI Pictures, AI emails they are used everywhere and they are devaluing human interactions. It’s useful to know what someone else is trying to say and when you put a layer of AI on top, you are averaging out the conversation. Somehow using a probabilistic model of all the conversations past makes things better.If AI is to be a successful technology in personal life and in business, it should be INVISIBLE, it’s use shouldn’t be a goal, it’s use should help humans and not make whatever we do more complex and less human. Instead you can see it EVERYWHERE.

Rather than an AI generated picture here’s a picture of my dog Friday

Happy Accidents

There are many documentaries about music making and a lot of them talk about the digital revolution. Studios used tape to record a track and through very complicated analog mixing techniques were able to create the ‘perfect’ record. Then there’s also artists like Daniel Johnston who prove that music doesn’t have to be perfect to be beautiful.

Diagram of an Analog Tape Recorder

When the DAW came out, music was locked in on a click track with perfect tempo, it made recording music so much easier. But a lot of the character of those old records, the mistakes and happy accidents that musicians talk about to this day got heavily reduced. Some bands that didn’t have the money to afford studio time, would practice for weeks so they could record their album live in a day. There’s a reason the early albums of a lot of bands are better than the later albums.

The DAW also had a lot of positives, musicians could start recording at home, be their own producer, it lowered the price of entry drastically. It also started a whole industry around digital music (EDM).

In recent big tech keynotes there are multitudes of demos that show how you can use AI in photo editing to improve the angle of your pictures or remove the uncle you don’t talk to anymore, but the uncle and the angle are part of the story of that picture. In the past having really nice pictures meant you hired a professional to take those because it was a special occasion. People have started using Polaroid cameras again there’s a purpose of having an artifact that you are unable to modify once it’s created. Many a movie has been ruined by reducing actors to sentence speakers and using too much editing or special effects.

For AI to be successful its main focus should not be to make human things less human.

Customer Focused

Who is the Customer? This may sound like an easy question, but it’s often a little more complicated than it sounds.

If you are YouTube for instance, you have the content makers, you have advertisers and you have viewers. All three are some type of customer/end user with often conflicting demands. Using AI to ‘copy’ an existing video that required a lot of time to put together maybe legal but it devalues a lot of work creators put into making their videos, should it be your job to reduce the use of AI for those use cases? I think so. On the other hand someone doing a 3 hour podcast on YouTube can use AI to read through the transcript and create timestamps for users to jump to specific conversations in their podcast, this is a way better use case, it doesn’t devalue someone else’s work and improves the experience of the listener.

As a consumer, I would like to have the ability to filter out any videos that are primarily AI generated, I’m ok with an infographic here or there but I do value the work people put into it. Using AI to detect AI is a great use case. The same goes for online content. Having a filter on LinkedIn to filter out AI posts would be a massive win. A lot of companies use AI to entice users but they are often achieving the opposite.

For AI to be successful it should allow us to reduce the noise in our inboxes and online not increase it. It should make software better for the end user and not worse.

Accessibility

Our technology is not friendly for people with disabilities, most of the sites out there are still horrible. There are a lot of anti-patterns where you have to prove you’re not a bot (and no I don’t ever want your cookies, stop asking). We are making the internet worse to use for a whole group of users that was already struggling.

AI can be a powerful tool here, you can take an image of a website and have it described to the end user. There are cameras now with screen readers, home devices can talk to their end users, the possibilities are endless.

Picture of Lucy Greco talking about her Camera with a screen reader. You can find more information about Lucy at https://accessaces.com/

For AI to be successful it has to improve accessibility for ALL users.

Creators

As more and more automation tools like OpenClaw are introduced and website visits don’t provide advertising dollars we keep looking for new ways to introduce revenue streams and on the consumer side for ways to avoid them altogether. OpenClaw is a workaround to the problem that most information on the web is not easily available through APIs.

The subscription model is broken, people can’t afford to pay 5–15 dollars a month for every sub stack or youtube channel (patreon) they like. There should be a subscription model that does proper revenue sharing, that elevates content created by humans and discourages AI created content.

A great example of how not to use technology is the Honey extension by Paypal that robs creators of hard earned affiliate link dollars. I can see a ton of AI search agents consuming articles or videos online doing the exact same thing.

There has to be a simple open standard that allows content creators to choose who consumes their content and if it’s an agent it should be through an agent friendly API with some form of attribution / compensation / revenue sharing, either through affiliate links or some type of subscription model.

With the endless drive towards AGI, the objective is for AI to become as close to or better than humans as possible. In the future it will become harder and harder to distinguish between AI generated content and human content. AI generated content is a massive problem for the AI industry as well, since using AI generated data can cause the model to degrade significantly. There’s articles and tools now, devoted to recognizing AI generated content. If we use AI to generate content, we shouldn’t be afraid to tag it as such.

For AI to be successful it has to provide revenue and incentives to REAL content creators.

Distance between the Developer and the Customer

If you read the Agile Manifesto you understand that the biggest problem they were trying to solve is the relationship between the developers of the software and the end users of the software. In a way a whole industry was founded on achieving the exact opposite. We are measuring developers by LOCs generated and have turned something that was all about putting the human first into something that is highly process driven instead.

AI is great for prototyping and creating proof of concepts, it can help drive the conversation and requirements. For instance designers can vibe code and application and hand off to the engineering team for proper implementation. You can have focus groups on software early and make changes early in the development process (shift-left).

For AI to be successful, it has to help in reducing the distance between the end user and the developer. Anything that has been generated by AI should be flagged as such.

Distance between the Developer and the Software

With AI we’re often increasing the gap between the developer of the software and the software itself rather than amplifying the software engineer as the driver of the tools (keeping responsibility of the software they create), and have AI do the tedious task of code review, look for vulnerabilities and standard violations.

I understand that we can automate simple bug fixes based on a null pointer exceptions in a log file somewhere but most software development is based on judgement. You don’t choose the best solution, you often choose the least worst solution. As an example adding a library to your codebase is incredibly convenient since it gives you code someone else maintains for ‘free’, however you are now responsible for all the upgrades and dependencies it pulls in. There’s a heavy hidden maintenance cost to relying on 3rd party libraries. A lot of software development is about compromising and AI should not be in charge of making those decisions for you.

Microsoft told its employees to stop using Claude Code, the claim is that this is a financial decision. What everyone seems to miss is that they have a very capable tool with Github Copilot that has gotten a lot better over the last months and what better way to improve it even further than to have your own employees dog food it. It is also model agnostic, meaning it is not bound to a single model provider, something we should all aspire to doing.

Claude and Codex are amazing tools but there are way better harnesses out there like OpenCode that allow you to choose and mix and match, using one model provider against another model provider for review. There’s also the problem where we made these tools so convenient to use that the end user doesn’t understand the underlying technology anymore and there for makes the assumption that it will do the right thing.

For years Developers have complained about not having enough time to write complete unit tests or refactor, with augmented development they can finally do so with a lot less effort.

Most Coding Harnesses have the ability to create extensions and skills and allow developers the ability to do things they wouldn’t have time to do otherwise. Test driven development is bound to make a come back.

Well maintained existing code bases have a leg up since they provide great examples and patterns for coding agents to draw from. Proper domain boundaries make it easier for the agent to figure out where to make changes and do so in a way that doesn’t impact every single part of the system.

Most developers sculpt their code into existence, they follow an iterative approach making judgmental decisions for things to work one way or another. Requirements or user stories often miss critical criteria that is not known or thought of at that time.

Vibe coding is a powerful tool for prototyping but also a dangerous trend for production ready code. Kent Beck calls AI a Genie, since it does exactly what you told it to do but it never does what you want. Some developers even get addicted to trying to get it to do exactly what they want over and over again, as if they’re operating a slot machine.

There are a lot of stories out there about agents of agents, etc… doing all the coding and the token cost of those projects is probably massive and the quality is probably reflective as well. Claude code for instance has been struggling with a flicker issue for months now.

For AI to be successful it should empower the developer to stay in control of the software they develop by offering insight, provide critical advice, and be a partner not the opposite.

Bespoke Software

If you go to Walmart you can buy shirts in all sizes but none really fit. There’s the opposite experience on Savile Row in London, they will measure you top to bottom and take a month to make you the perfect suit with the perfect fit.

Software has become too hard to use for a lot of end users since it’s trying to please everyone and therefore achieves the opposite. There’s a tendency to keep adding functionality, AI can help make software custom to the needs of the end user, with a simple prompt the user could make or request changes to the interface that would make it enjoy the software more, we all appreciate dark mode, but we can take things further than that. A lot of tools like emacs and vim are successful to this day because they allow a level of customization for the end user other tools don’t. A lot of software these days makes decisions for the end user, telling them what they should like rather then allow them to pick the experience most suitable to them.

A great example of customizing in the age of AI is the minimalistic coding agent Pi, it allows you to create extensions and customize itself through prompting.

For AI to be successful it should allow software to meet the user where they are.

Better not More Software

As the software industry continues we get worse and worse at providing software. Somehow Office Tools are over 2 GB each and it takes about 15 GB to have all the office tools on MacOS. That is 2.7 GB for a word processor and 2.5 GB for a spreadsheet. I understand that the size of a piece of software isn’t necessarily an indicator of the quality or the efficiency of the software but we used to be able to fit an operating system AND a word processor on a floppy disk.

Screenshot displaying the absurd size of Microsoft Office in MacOS

A lot of companies use feature matrices as a way to sell their product. What’s often lost is that the end user of that software often only uses a small subsection of the functionality provided. When enterprises are considering replacing software it shouldn’t be important which features the software provides, but which features they’re actually using. Often new employees talk about software in their previous companies they liked a lot more, usually because they offered less but better.

New editors like Zed are gaining end users because they put performance, ease of use and customizability first.

For AI to be successful it should help us improve the quality and efficiency of the software we create and not add more bloat.

Reliability

We are all familiar with the Strawberry problem and LLMs. Most models will be able to count the ‘r’-s in strawberry correctly now.

However I recently asked ChatGPT to provide the answer to this simple question: How many r-s in the word “supercalifragilisticexpialidocious”.

Asking ChatGPT to count the number of occurences of the letter ‘r’ in Supercalifragilasticexpialidocious from the musical Marry Poppins, it incorrectly answered 3.

Somehow we can still fool it very easily, the reason is very simple and it has to do with the way text gets converted to tokens.

Gary Marcus often talks about the Neuro Symbolic layer in AI. It’s the tool calling part, or the part where you add a deterministic layer on top of an indeterministic LLM.

A lot of LLM tools are heavily relied on to provide answers without being surrounded by the proper context, tooling and guardrails. You can easily generate a javascript function to count the number of characters, somehow the neuro symbolic layer in a lot of the AI tools is heavily under developed since it requires a lot of time and effort to create, but it would make everything operate a lot more efficiently and reliably.

We are ASKING USERS TO VERIFY summarized search results because it is AI generated. How did we become this hostile to the end user?

For AI to be successful it should provide answers in the most efficient and reliable way possible and decline to answer if it’s not available to do so. At a minimum security controls and guardrails should be in place to prevent unauthorized access and prompt injection.

Environment

Data Centers have gotten a horrible reputation and for good reason, they generate noise. In some cases they have massive polluting power turbines right next to them and they often use tremendous amounts of water. I don’t think we have an energy problem, there’s plenty of ways to generate energy. But we do have a massive fresh water problem and a global warming problem, and we definitely have a reputation problem. It’s gotten so bad that even the governor of Texas is proposing legislation.

Most companies with climate pledges are most likely not including the amount of energy used by AI for their organization. We’re aiming for Mars without appreciating what we have on earth.

As of late, putting a price on token usage overall is a good thing, it forces end users to look for the best model for the job at the correct ‘reasoning’ level. It’s positive to see that token maxing is out the door and more time is spent on token usage optimization.

For AI to be successful, the AI industry needs to be a good citizen, respect our environment not destroy it.

Privacy

As humans we easily fall prey to fear mongering. We have never been safer, however due to social media and marketing we feel the need to have cameras on every corner of every block.

With the power of AI this offers a tremendous amount of power that can be used to control people’s opinions. Green cards can get declined to people that post negatively about an administration online. Abusers can follow or track their victims online. Peaceful protesters phones get registered and identified.

The more we use these AI tools the more they will know about us. Memories are created in a database on the agent side to be able to serve us better. However historically big tech have not been good stewards of this information. The ability to run LLM tools on device or on a local shared server in your house can help alleviate some of these problems.

For AI to be successful end users should be in control of the information they share.

Conclusion

There’s a massive opportunity for using AI to change our lives for the better and improve the quality of software we produce. How we choose to use this technology depends on all of us.

To introduce AI in any organization, the first area of focus should be quality over quantity and not using usage as a metric or a goal but rather focus on the outcomes. Less bugs, less bloat, more focus, happier customers.

As a whole, it’s clear that we need a lot more transparency in AI usage to be able to treat it as such. There are a lot of awful voices out there claiming all software should be rewritten with AI. I hope this article gave some food for thought.

Recommended Reading

Process Over Magic: Beyond Vibe Coding Faster, Smarter, and Safer Coding with AI Assistants https://pragprog.com/titles/ubaidev/process-over-magic-beyond-vibe-coding/