But it’s pretty bad because the result is ridden with invisible errors. But what counts as an error if you trust the LLM and never meticulously read the output? And how much time are you really saving if you end up spending that time proofreading?

I went back to a meeting transcript after a developer struggled to turn my AI-generated summary into a coherent ticket. Looking through the source transcript, I noticed that at one point it recorded the word "embroidery." I mean, we work for a large telecom company, the chances of me using that word are slim to none. Nevertheless, the AI summarized the conversation and included "embroidery" as an action item.

In the words of Claude himself:

It's not just X, it's Y.

I chuckled, but it made me think about all the BI slop we generate daily in the name of using AI technology. It feels helpful. Creating massive documents makes us look productive. Generating a pull request with changes across hundreds of files seems impressive. Asking AI to validate an idea feels instructive. Unless you take the time to scrutinize the results.

I can only imagine what the world will look like when we feed all this slop back into the machine.

And just now, I opened Chrome, and saw the exact same message but for Chrome:

It looks innocent enough. We all start browsing the web right away when we turn on our computers, so what's the big deal here? Well, the big deal is that everyone wants to be the front page. Defaults matter. While we all want to believe we have agency, the numbers don't lie. Most of us never change the default settings. Just last year, Google paid Apple close to $20 billion just to remain the default search engine on Apple products.

While we're all distracted by the AI race, the one race that actually matters is who remains the front page of the Internet. Microsoft is trying to claim that spot by loading its browser at Windows startup. Now Google is trying to win that space back by doing the same with its own browser.

On my Mac, Anthropic has been very aggressive about trying to load Claude automatically. I suspect every company will be pursuing this aggressively in the near future. The quality of your product is secondary to the exposure it gets. Not even benchmarks matter if you can't be on the user's front page.

Google still holds the front page of the Internet, and that makes up for its AI model not being at the top of the benchmarks.

Note there are no options to dismiss. Only "Set later"

But don't let them take this space from you. They are all fighting for your attention. I urge you to simply consider pressing the little X button they tuck into the corner. It's your computer, use it the way that suits you best.

The idea changes mid-sentence.

That's probably how most of the code I write ends up. I start with an idea in my mind. I think it's brilliant, maybe it’s a way to restructure the code, an overall design, or a specific paradigm. But when I start writing it, something else pops up. Seeing it halfway down the page, I realize it's not quite what I was imagining, and I change direction.

The same principle applies when I'm writing a blog post or a story. I feel like I have a perfect concept in my head, but the moment I start putting words down, things shift and I adjust accordingly. That's why the final text rarely matches the original vision. It’s as if writing itself is the process of refining the idea.

This process is lost when you use AI. You don't write out the details of an idea as you go; instead, you just give it the gist of it. It generates an entire wall of text instantly, and now you're stuck reading through it while the tool tries to preserve that initial draft, or its own interpretation of it at least.

You only discover flaws or attempt to pivot while reading the generated output. But pivoting inside a wall of text is much harder than making a 180-degree turn mid-sentence. If you spot a small part you dislike and try to remove it, it might break something 20 lines down, forcing you to step back and make sure everything still aligns.

Using an LLM forces you to commit to your very first thought.

When you write manually, every keystroke is a chance for the idea to evolve. When you use a large language model, you have to absorb the entire output all at once, which ends up being far more time-consuming.

Both titles and content were filled with keywords that signaled to search engines what the page was about. Articles basically catered to the search engines and only incidentally served people. On my own website, I remember editing page titles several times, then waiting a few days to see if Google noticed the changes and ranked me better. In fact, before Google introduced personalized results, I built a tool at my job to track how our website ranked for different combinations of keywords.

When personalized search results arrived, a lot of tools became obsolete. It was still useful, though, to track keywords for generalized metrics. But now AI Overviews are here. And it's not just Google's, plenty of people go straight to ChatGPT and ask the large language model their questions directly. While the information returned might be sourced from real websites, there's almost no incentive to click through to those sources.

While my inbox is still flooded with people promising to improve my SEO, I think it might not be helpful at all anymore. My traffic has shifted from mostly coming from Google to just a handful of visits. Most traffic now comes from AI bots scraping my content, and RSS readers (thank you!).

I take this as a relief. We can finally drop the act. We don't have to write keyword-stuffed titles and blog posts just to appear in search results. Large language models can understand our content just fine without it, and we won't be getting that traffic anyway. We can drop SEO.

You can finally write for yourself. Write for your audience. No need to cater to the robots.

We own the glass.

What they mean is, they own your tv. They own everything on it. You may have paid for the box, but they own the glass. They can run anything they want on it, and you have agreed, in the terms and conditions, that you will inform anyone entering your household that the TV may record them without warning.

TVs come equipped with microphones, wifi radio, and tracking software. That’s enough to monitor everything that’s happening around it.

If you haven't watched the research video essay published by Gamer Nexus, I highly recommend it. I've said it many times before, TVs are cheap because you aren't simply watching them, they are watching you and listening, and reporting back to the mothership. This info is then used to sell you more stuff.

What scares me in all this is that the solution I often tell people may no longer be true. I often tell people, just disconnect your TV from the network and you are all set. Now imagine just for a second. What if they add a cellular radio to the device? One with an esim. Then that is it. You can no longer disconnect your device from the internet. They now own the glass in your house, and have a permanent connection that you cannot disable.

They own the glass. We are merely vessels to their business.

Interviewing a senior software engineer

Biggest challenge so far in any coding project?

My biggest challenge isn't code per se. It is explaining to someone else what I am trying to do with the code. It is surprisingly hard to convert the thoughts in my head (or the code) into something coherent someone else can understand.

I still believe that. I don't even remember when we had this discussion, but I'm glad Walter has kept it up. And a great reminder of my priorities.

If you weren't coding, what would you be doing?

I'd be a writer. (not a good one, but one nonetheless)

Yep.

The lead developer stood behind me while I was working on a feature and asked me to make a div clickable. That's it. Technically, it's the simplest thing you can do. But, I hesitated. He watched from behind as I tabbed through every application on my computer, doing everything but what he asked. "Just make it clickable," he said again.

I have made divs clickable a thousand times in my then short career, I knew that html standards were merely a suggestion. But this one rubbed me the wrong way. Instead I said, "I already have an anchor tag below, the div doesn't need to be clickable."

"The fuck you're talking about? Just do it."

We got into an argument, and I ended up doing exactly what he said. A little after I pushed my code, he revisited my code and added a slight delay to my click event using setTimeout to disconnect the user action to what occurred after.

If what I'm saying sounds a bit vague, that's how it was framed at a time. We had disconnected the technical ask, clickable div, from the actual feature to make the whitespace on the page clickable. Let me make it even clearer.

Have you ever used a coupon website? You buy something on a website, let's say nike.com. When you get to the checkout cart, they have a little box that says "enter coupon code for a discount". So you scour the web to find a coupon that will help you pay less for your purchase. You land on our website where we offer a dozen coupons, but they are all partially hidden. The section says "click to reveal". When you click, some random popup appears, then the code is revealed for you to copy and paste and you go on your merry way with a discount.

Everything sounds fine so far. What you don't know is that the coupon website gets credit for the sale. A portion of the price you paid goes straight into our pocket, whether you got a discount or not. We get credit from the referral, that's how affiliate marketing works.

So why is making the div clickable, or the whitespace clickable a big deal? Because it captures accidental clicks. The company doesn't care if they help you find a good discount or not. In fact, they don't even care if the code works. They just want you to click and for that popup to open so they can get the credit.

My task was framed as a technical request, instead of outright saying "trick the user into clicking." When the Honey scandal surfaced, I was confused why it was news anyway. Deception in affiliate marketing is a standard.

Very often we disconnect our work from the real world consequences.

Back in 2019 I met this girl at a friend’s birthday party. She worked at a bank in the software development sector, and told me about the project on which she’d been working. Their system collected data from various sources on people applying for financial services (e.g.: loans) and would indicate if someone was eligible, or raise a red flag. In the latter case they would have to deal with substantial additional bureaucracy, and often times would not be able to access these services anyway.

She seemed quite proud of her work, and told me her team had demoed it the previous week in front of the whole office. They’d shown the report that the system generated for each person in the team. In her case, the system flagged her as “dangerous”, and she was not eligible for a loan. Her grandmother was from Iran, and because there’s frequently cases of money laundering or other irregularities in Iran, she’s immediately flagged too. Despite her being a Dutch citizen, born and raised in the Netherlands, and working in this Dutch financial institution, her bloodline was “too risky” for her employer to lend her money.1

I was confused, and honestly, heartbroken. This person was telling me how proud they were of their work, building a system that discriminates them based on their bloodline. I questioned what she thought about it, and she explained “well, the rules are there…” —she paused a few times— “the rules are there…”. I remained silent and she eventually finished her sentence after repeating it a few times “the rules are there to protect us”. Even though this was six years ago, every time I remember this situation it brings me great sadness. Here was a person who’d worked hard to build a system which would discriminate against them, and yet stood proudly defending their work.

We don't see it as our responsibility to question the system. We are happy to do the work as long as we don't see the immediate consequences. I feel this whenever I see flock cameras mounted in the street. Someone had to design them, write the code, test them, ensure that they work to spec. Someone had to install them. These are people doing a good technical job, yet I'm sure they wouldn't be happy if they were tracked by the very system.

This is why I was mad when Anthropic wrote their manifesto saying they are against tracking citizens, as opposed to non citizens. You don't get one without the other. We track everyone then discriminate later.

This is why I'm against age verification systems, no matter how technically impressive the solution is. Because in the end, I'm the one who has to upload my ID to a system that I cannot trust.

With AI, it's even easier to disconnect your day to day work from its consequences. I understand it's much easier to ignore these things when you are just getting started in your career, but eventually, you have to voice your opinion.

I've taught my mother and her friends to check for that tag, but most, if not all, of the videos they share with me don't even have it. Even though they're clearly AI-generated.

If deception is something we want to prevent, and our parents can't tell real videos from fake ones, wouldn't it make sense to put a warning where it's actually effective? Just like movies have warnings beforehand, why not add one before these videos? It would take no more than two or three seconds, and it's entirely controlled by the platform. If someone is willing to watch the content after the warning, more power to them. But for those being deceived, this would actually help.

Those are the words YouTube places on videos where it detects AI, but they're buried somewhere in the description. Why not make that notice prominent when the video starts, the same way some videos open with a "Viewer discretion is advised" notice? On YouTube or whichever platform hosting the video, the warning is not embedding into the video, just displayed before. The uploader has no control over it.

While I know most people think AI videos are just fun or slop, but I argue that they can be harmful when they deceive us. They are a profitable vector for scammers, and despite YouTube's effort to curb it, they are actively promoting it in their creative studio.

If YouTube actually wants to prevent fraud and scams, they can do something about it. Let users know what they are about to watch.

“They didn’t care that they’d seen it work in practice because they already knew it couldn’t work in theory” - Clay Shirky

I complain about AI frequently on this blog, but not because I don’t think it is useful. I use it quite frequently on my day to day. But what I hate is hype and fake narratives. In fact, I believe that the opposite of their narrative will come true.

Technology will continue to improve whether Moore’s Law becomes a relic of the past or not. It’s been called dead, yet CPUs and GPUs are becoming faster than ever. I do not believe for one second that developers with pre-LLM experience will end up on the losing side.

And if technology continues to improve, then we won’t need OpenAI or Anthropic in the future. We will be able to load large open models right into our powerful personal computers with more than acceptable inference speeds. Unless you believe that computers have reached their zeniths and it is all stagnation going forward.

The gap between frontier models and open-source alternatives is already all about specialty, and it will continue to narrow. And when developers ubiquitously have access to local models, they will have access to everything.

Right now, companies are hoping that developers will use their AI and remain within their ecosystems. They're building guardrails, imposing limits, and designing their models to serve corporate interests first. They see developers as customers, not as threats to their business model. But I’ve seen how easy it is to switch from one frontier model to the next. In fact, some developers in my team accidentally switched by selecting the “auto” mode on their IDE. They didn’t realize that every subsequent request was from a different model.

It’s the developer who will end up benefiting from this far more than the corporations will.

One developer recently spent his evenings using an AI agent to reverse engineer every peripheral within arm's reach.

From those devices, I’ve come away with a full plaintext command shell inside my microphone, a webcam whose activity LED I can switch off while it records, and a key light that hands out memory writes to anyone on the WiFi.

He documented the entire process, implemented his own firmware update utilities, and completely enumerated the functionality of devices that were never designed to be user-serviceable. AI gave him the ability to fully control hardware that he has paid for in his own terms.

Another developer created OpenLogi, a local-first alternative to Logitech’s software used to remap your own mouse button. The manufacturer was forcing users to create an online account to have access to the hardware they had paid for. OpenLogi gives full control back to the user.

This is what happens when developers have the tools and the motivation to bypass corporate control. And AI is about to make this kind of reverse engineering and alternative-building dramatically more accessible.

At the speed of large language models, an activist could create a brand new rotating messaging platform every week to avoid the prying eyes of an oppressive government. Someone else can review his own model and add some self improving features. In fact, you could explore new AI paradigms.

Most developers I know have a side project they don't have time to work on. With AI, they will have the ability to execute on those ideas. These days, I'm able to run Deepseek on my own $400 local machine at a much slower speed, but I'm not in a hurry. Give it a couple years, I could run something even more powerful.

I don't need a project management tool where I have to pay monthly. My actual needs are much simpler. For example, I actually like Jira, despite how much I complain about it on this blog. But I don't need to have Jira for myself on my personal projects. I can build a tool that works solely for my needs.

I can easily build applications in environments I am not too familiar with but that are more appropriate for the task. I can build prototypes in a couple hours now and throw them away if they don't match my initial expectation. I can do so much more.

Dario Amodei and others are trying to scare us with the capability of AI. They sell the fear of superintelligent systems that will render human developers obsolete or dependent. But this is not what's going to happen.

What's going to happen is we will not need Anthropic anymore. Yes, they will have the high-end hardware. But we don't need high-end hardware the same way most people did not need a professional camera. All they needed was a crappy camera with a good filter to post on Instagram.

Flickr was superior to Instagram, but the superior technology lost to the one that was "good enough" and in everyone's hands.

The funny thing in all this is that by making everything AI-dependent, by building their moats and their guardrails and their API toll booths, companies like OpenAI and Anthropic are selling foot guns. They are building the dark fibers of our era, the infrastructure that developers will eventually subvert, repurpose or simply bypass. Because eventually, we won’t ask for permission. We can just do whatever we want with tools freely available to anyone.

We own our devices. We own our data. And soon, we'll own the intelligence on our own terms, without a subscription fee and without a corporate overlord. At the very least, we will get free GPUs.

My job was data entry at a non-profit. It was my first job in the US where I didn't have to lift heavy objects. My main task was entering collected donations into a database. The data source was email attachments, which came in various formats. Excel files, Word files, and some directly embedded in the email body. The database itself was FileMaker Pro.

Every part of the process was wasteful. To receive the emails in the first place, I had to call the person who collected the donation, argue with them, and guide them through the process of emailing me the information. The donations were collected at events, and the data was often just handwritten on a note. Once I got the email, I would extract the content from the attachment and organize it into an Excel file. Then I would manually enter it into the FileMaker database, one row at a time.

In a full workday, I could process no more than a dozen entries. So I decided to improve the process.

I started from the end. Entering data into the database was error-prone. I had to enter the name, address, donation amount, and a plethora of other information into a single row of a database viewer, and I often put the wrong data in the wrong column. So instead, I created a form with validation.

I built a standardized Excel file where the people collecting donations could enter data directly into the correct columns and send it to me. This turned out to be too much to ask. Several people didn't know how to use Excel, and it became a nightmare of technical support. Eventually, I switched to a formatted Word document. Annoying, but it worked.

Lastly, I created a process. Every Monday, I would email everyone to remind them to send me their donations by Wednesday, so I would have plenty of time to enter them on schedule. The deadline was arbitrary, but it slowly hardened into policy.

My system was a success. I was entering hundreds of donations every day. In terms of productivity, I had more than 10x'd my output, which I thought was a good thing, until I started getting noticed.

What have you done for me lately?

My productivity machine's knob had been dialed to 11. But I had also set the new standard for how much needed to get done in a day. So my manager hit the reset button. The dial went back to zero, but the expectation stayed exactly where it was.

It's funny how that always happens. Increased output never turns into more free time. Instead, it becomes the new normal. Anything less looks like a drop in productivity.

These days, with AI common in the workplace, every semi-technical manager has started building apps themselves instead of taking requests to the dev team. They get a rush from how much code they can produce, and they enjoy that high for a while. But they fail to see that the productivity gain only exists at the very beginning, before the app is actually in use. Once people start using it, you don't get to regenerate the app from scratch every time you want to add a feature. Instead, you have to add things slowly and carefully, without breaking what already works. All fields start green, but they brown eventually.

When you get an extremely productive teammate, you start looking at the rest of the team as time-wasters. In fact, once you get used to what that teammate produces, you start treating it as the baseline and eventually you ask, "What have you done for me lately?"

I was let go from that job in a little celebration, right after I'd trained my replacement. One morning, my manager simply announced that it had only ever been a summer job, and summer was ending. So my last day was set. I was too green to point out that this had never come up during hiring, and that I'd actually started well before summer began. They brought cake on my last day.

What they failed to see was that my process was still very much manual. My replacement was let go when she couldn't manage more than a dozen entries a day. So much for productivity gains.