Ben Tsai

Comprehension is the job

2026-03-25T15:39:46Z

I’ve already referenced Addy’s Comprehension Debt piece, but I think it is worth highlighting this quote:

Making code cheap to generate doesn’t make understanding cheap to skip. The comprehension work is the job.

This is salient because it feels there is a growing divide in the industry on this point. Some would say our understanding of a system is moving to a higher level of abstraction, where we use English to specify our intent, and depend on the LLM tool to translate that into code. The analogy they use is, the same thing happened when we moved from assembly to C.

This analogy breaks at a critical point. LLMs are inherently non-deterministic, so it will never do exactly what you want. You will never comprehend what the system is doing if you don’t examine the code.

But let’s be pragmatic about it. The code output will increase. Addy exhorts teams:

Teams building comprehension discipline now - treating genuine understanding, not just passing tests, as non-negotiable - will be better positioned when that reckoning arrives than teams that optimized purely for merge velocity.

Simon Willison has been publishing Agentic Engineering Patterns guide. I appreciate his principled approach. He says, your job is to deliver code you've proven to work and the human provides the accountability.

In Your LLM Doesn’t Write Correct Code. It Writes Plausible Code, Hōrōshi バガボンド asserts:

If you prompted the code and cannot explain why it chose a full table scan over a B-tree search, you do not have a tool. The code is not yours until you understand it well enough to break it.

But let’s be even more pragmatic about this. Who is going to listen to this advice? Engineers and teams can do all they can to be more disciplined, but they will be hamstrung by the overwhelming incentives around them that prioritize velocity above everything else.

So we will continue to engineer solutions that help us avoid doing the job of comprehending.

The paradox: to get better at ai, you need to not use...

2026-03-25T15:13:21Z

The paradox: to get better at ai, you need to not use ai.

Also, ai is meaningless term, so what am I even talking about?

Performative

2026-03-19T15:14:00Z

I recently read this piece on one of my favorite topics entitled On genAI: Was prototyping really a bottleneck? by Frank Elavsky:

And so large-language models give us performative fidelity; a sort of false, constructed faithfulness to ideas we haven’t actually invested enough time into and haven’t socialized. If anything, they construct loyalty to under-formed ideas more than they actually construct ideas, because constructing ideas is slow and messy!

I’ve been stuck on that word performative (emphasis mine) because it captures the nature of all LLM output. This tech produces output that is void of substance.

We used to be able to correlate the production of an artifact with the process that created that artifact. If someone wrote a paper on a topic, you could correlate that with an author who has done some level of thinking on that topic. If you see a slide deck, you could assume the presenter considered the content and formatted the slides. If you have working code, you could assume some requisite amount of design, debugging, thinking about the logic and intent.

AI breaks that correlation.

Zeynep Tufekci talks about the societal-level disruption this has in a recent talk, Are We Having the Wrong Nightmares About AI?. She points out how our existing mechanisms are now broken. Think about how photos are used for insurance to authenticate what happened. That’s been disrupted now, and similarly, in all of these areas:

~~Proof of Effort~~
~~Proof of Authenticity~~
~~Proof of Accuracy~~
~~Proof of Sincerity~~
~~Proof of Humanity~~

I’ll let Zeynep speak to the societal-level stuff. But closer to home in software engineering, Addy Osmani makes the same point in his piece Comprehension Debt. He warns that

comprehension debt is the hidden cost to human intelligence and memory resulting from excessive reliance on AI and automation.

That cost is largely hidden because:

The reason comprehension debt is so dangerous is that nothing in your current measurement system captures it.

Our mechanisms for judging whether things are good are broken.

In some ways, comprehension debt seems like the goal. That’s the part that takes the longest time, so lets skip it.

Addy’s call to action is towards teams to build comprehension discipline now. Yes, teams should do that. What also needs to happen is addressing the larger problem of the context. The environment around us incentivizes velocity, moving fast, and not getting left behind. This technology is perfect for creating whizbang demos that will have executives seeing dollars in their eyes and reaching for the Ship It! button.

It feels like we could even ship things that customers adopt. My fear is that customers will also be convinced by the performance. For some uses, that may suffice. But one day they will require something more, a new use case, a subtle bug. And the curtain will fall.

Software is people

2026-03-18T15:14:08Z

One pillar of my origin story is when I realized the software I’m building exists as part of a system whose goal is to serve human beings. Secondarily, the process to produce that software is more of a social endeavor rather than technical.

With the rise of Actually Idiocy, it’s apparent most people don’t understand or appreciate that. We keep trying to use these ai products to solve problems they are not built to solve.

Two recent pieces detail this reality. First is Software is a coordination problem. AI can’t help you with that by Pavel Samsonov. His subtitle summarizes the issue:

The feedback loops of the product delivery lifecycle go through people. Adding AI makes it slower, not faster.

The second piece echoes that sentiment, from Tailscale’s CEO, called Every layer of review makes you 10x slower. He says his post is not about AI, but then talks about AI, because the industry keeps sticking in AI in the wrong places.

He begins with tracing how tasks of increasing scope take 10× every new layer you encounter. So making a simple bug fix is 30 minutes (I wish this were true of our org!), getting that code reviewed by a peer is 300 minutes, getting sign-off from an architect is 3000 minutes, and so on. His declaration:

AI can’t fix this.

Yes. Let’s stop trying.

Outsourcing thinking ↗

2026-02-02T18:32:06Z

Erik Johannes Husom’s piece entitled Outsourcing thinking resonated with me.

He highlights the use of LLMs in personal communication and writing:

Direct communication is not only about the information being exchanged, it's also about the relationship between the communicators, formed by who we are and how we express ourselves.

In the context of work, you can take a utilitarian view of business communication that narrowly sees communication like emails and Slacks and Jira comments as getting a job done. All that matters is getting the thing done. However, I think we are losing important something when our communication is laundered through a machine. While our intent may be to clarify what we are trying to say, “it's impossible to separate the meaning from the expression of it…Changing the phrasing changes the message.”¹

If we drain all of our messages of our own words, we are losing a primary channel that we have of maintaining relationships with our colleagues. The simple greetings, small talk, emoji reactions may seem trivial, but they reflect the human beings on each side. Even the misunderstandings, churn, struggling through language barriers, are meaningful. As Erik follows up:

We rob ourselves of the opportunity to grow and learn, without training wheels. LLMs can certainly help people improve the text, but the thinking process — developing the ideas — will be severely amputated when leaving the phrasing up to an AI model.

https://arc.net/l/quote/ctlcaele

It’s Okay ↗

2025-10-30T21:09:59Z

I wrote a LinkedIn post:

I used to have some regrets.

I missed the Web 2.0 wave in the 2000s. I didn’t work on webapps until 2016. I missed the mobile/app wave years later. I’ve never released an app. I’ve never worked on a SaaS product. I’ve never gotten to Inbox Zero. I don’t know that much about cloud or containers.

Now, I think it’s okay.

Thinking modes ↗

2025-10-29T14:12:03Z

From Robin Sloan’s blog:

Here is one distinction among several: this process can only compound — the models can only “think” by spooling out more text — while human thinking often does the opposite: retreats into silence, because it doesn’t have words yet to say what it wants to say.

Human thinking often washes the dishes, then goes for a walk.

Yes, we’ve lost the meaning of the word thinking. I do some of my best thinking while washing the dishes, mowing the lawn, or taking a shower. Maybe I don’t want AI to do my laundry or dishes.

I think it means prospecting new analogies; pitching your inquiry out away from the gravitational attractors of protocol and cliché; turning the workpiece around to inspect it from new angles; and especially bringing more senses into the mix — grounding yourself in reality. You’ll note these moves are challenging or impossible for systems that operate only on/with/inside language.

Models are not grounded in reality—they have only the data they were trained on, and they are dealing with it at the level of text, tokens, language. This is such a superficial imitation of what humans do!

As Robin concludes, let’s not think more, but think harder!

The simulation of judgment in LLMs ↗

2025-10-20T02:27:34Z

The simulation of judgment in LLMs

While models often match expert outputs, our results suggest that they may rely on lexical associations and statistical priors rather than contextual reasoning or normative criteria. We term this divergence epistemia: the illusion of knowledge emerging when surface plausibility replaces verification. Our findings suggest not only performance asymmetries but also a shift in the heuristics underlying evaluative processes, raising fundamental questions about delegating judgment to LLMs.

Once again, these systems are not making judgments like we think they do. They are predicting the next token, and not much more.

Text Is Best

2025-10-10T17:12:02Z

I wasn’t intending to rework my tasks and notes system. It just happened to me.

At the end of the day, I’ve ended up with a lightweight process that is one tick more formal than what I had before. It’s good enough (and that’s a feature).

It started with a side quest to find a beautiful plain-text editor on macOS that supports folding. I’ve embarked this side quest a few times. My first stop is always checking in on iA Writer (one of my juicy apps) to see if they’ve ported over folding functionality from Windows. Sadly, no—it is still exclusive to Windows.

I considered Zed (the newcomer) and BBEdit (the OG). I used to like Zed when it was super-focused on the code editing experience. But that soured as they integrated ai, and fell apart entirely once they got VC funding¹. And BBEdit is rock solid, but just not beautiful in the way I want right now.

Obsidian could have been the answer. The company is run the right way. They respect user data and don’t lock you in. The app feels fast and can be customized to look however you want. It supports a myriad of extensions that make it behave exactly to your liking. It’s got a nice mobile app. But the one thing is that it’s designed to map one note to one file. Like 99% of notes apps out there, and how most people expect.

My one weird requirement is I want all my notes in a single, chronological log.

In fact, I tried this before, and it didn’t work:

My previous experiment was to use one big text file as the starting point for all my writing. Having a single stream of all my notes, tasks, emails, etc. was enticing, but the ergonomics just weren’t right. I rarely reviewed previous writing. The mobile experience was not great.

I honestly don’t know what that guy is talking about. What ergonomics? Also, the mobile experience—that’s not important to me anymore. If I need to write something down while I’m away from the computer, I’ll write it down somewhere temporarily and transfer it.²

That brings me to TaskPaper.

Jesse Grosjean’s apps are always a second stop on this side quest journey. I also have a license to his latest app, Bike, which is indeed beautiful and the spiritual successor to TaskPaper. But it does things differently and is not plain-text-based.

And that’s my other requirement: plain text. I’m no zealot, but plain text makes me feel good.

The system

At the core of my system is the daily entry. That gets created as such:

Write down today’s date
Add meetings and today’s plans.
Throughout the day, add notes and update tasks.
At the end of the day, review. (Optional).

The huge unlock (aw geez, I’m talking corporate) for me is TaskPaper’s outline format. If I have details about a task or meeting, I tuck it under the top-level item. Being able to fold that detail away gives me the freedom to add all the minutiae of my work. In my previous systems, my pain point was wanting to capture more things but not having it cloud the big picture.

Because of TaskPaper’s focus feature, I’m also using this file to draft emails, documents, etc.

What I love about this is that my writing and tasks and notes are tied together in a single timeline. I can go back and review these items in the context everything else.

I’ve structured my daily entries into weeks (and then grouped by year) because my work generally runs on that cycle. My default view is the current week, and I find this gives me most of the context I need.

Daily → 2025 → Week 41 → Monday, October 6, 2025

The styling

TaskPaper’s out-of-the-box look is very nice. However, I couldn’t resist tweaking it to make it shine to my eyes.

@font-family: Lilex;

item {
  handle-border-color: @handle-secondary-color;
}

item[data-type="note"] {
  font-style: normal;
  &[empty] {
    handle-border-color: none;
  }
}

item[data-type="task"] {
  font-style: italic;
}

item[data-type="project"][depth=4] {
  &[bodyContent^="Monday"],
  &[bodyContent^="Tuesday"],
  &[bodyContent^="Wednesday"],
  &[bodyContent^="Thursday"],
  &[bodyContent^="Friday"],
  &[bodyContent^="Saturday"],
  &[bodyContent^="Sunday"], {
    > run[content] {
      background-color: mix(#fdffcd, #ffffff, 40%);
    }
  }
}

item[data-hot] > run[content] {
  background-color: mix(#ffcab0, #ffffff, 40%);
}

I tend to write more notes than create tasks. I see this system more as a journal with some tasks sprinkled in. By default, notes are rendered with italics, while tasks are normal. So, I flipped that.

I added a touch of color to the titles of my daily entry. I’m also experimenting with using tags to call attention to certain items. In the above stylesheet, you’ll see how adding the @hot() tag will give it an orange background.

Finally, I changed the typeface to Lilex, which is a programmer’s version of IBM Plex Mono. I love this typeface—especially the mix of italics and normal. Isn’t it chef’s kiss?

The simplicity

I was inspired by many projects out there. I’ve tried analog bullet journaling in the past, and considered applying some of those concepts digitally. I thought about developing my own markup to categorize my notes. And putting timestamps in my entries.

But to what end? My work is not so complicated. I don’t need to meticulously track my time. I don’t need to delegate tasks, and I’m not generating so much knowledge that I need to annotate it. I’m playing around with TaskPaper tags, but so far I don’t see a need that simple text searching can’t meet.

The one post that prompted my initial journey and that I resonate with is My productivity app is a never-ending .txt file. In it, Jeff Huang describes using his one big text file for 14 years. That’s remarkable.

That daily todo list is where I also take notes, so it’s a to do list that turns into a what done list. The best thing about these daily lists is I keep them all in a single text file separated by dates, so I have a record of everything I have ever done and when I did it.

This is what is bringing me joy about this current system. I’ve got a place I can offload stuff from my brain. I make plans and todos, and check them off when I’m done. And the longer I use it, the more useful it becomes as a record of my work. If someone else has been running with this for more than a decade, maybe I can stick with this for a while.

Extra reading

TaskPaper handles the Moby Dick Workout.
Someone else with similar thoughts: Love the One You’re With Part 1 - Bicycle For Your Mind
Who, ironically, switched to Emacs: Why Emacs and Not BBEdit and TaskPaper - Bicycle For Your Mind

Apparently they had funding already, but now I know it’s going towards “collaboration, both with agents and teammates” so nah…
But let’s see if this current experiment fails!³
My experiment with Ugmonk’s Analog system failed.

It’s a Trap

2025-09-02T13:08:18Z

For a long time, I’ve been trying to articulate why interacting with an LLM is so irksome. Essentially, it is because the text being generated does not originate from a person, but is designed to entirely to appear as a person.

This is the height of deceptive design. And no matter how many guardrails and systems you put around an LLM, it is impossible to curb this behavior. It is exactly this behavior that makes them so enticing, and that makes them feel useful at all.

A recent Ars Technica article entitled The personhood trap: How AI fakes human personality describes this behavior:

LLMs are intelligence without agency—what we might call "vox sine persona": voice without person. Not the voice of someone, not even the collective voice of many someones, but a voice emanating from no one at all… You're interacting with a system that generates plausible-sounding text based on patterns in training data, not a person with persistent self-awareness.

This is the problem I have with the anthropomorphism that pervades every AI product. They are positioned as assistants, copilots, interns, PhDs. Yet simply, they are not. And to say so is demeaning to human beings.

It feels preposterous to spell out the myriad ways a text-extrusion machine is not human, but it seems that the moment demands it. In particular, the piece highlights our sense of responsibility:

This self-continuity is one of the things that underpins actual agency—and with it, the ability to form lasting commitments, maintain consistent values, and be held accountable. Our entire framework of responsibility assumes both persistence and personhood.

AI can never do this. A computer can never be held accountable. Even if we gave an LLM infinite memory, it will never be a person. Speaking of AI products as such is deceptive, dehumanizing, and perverse.

Democratizing Knowledge

2025-08-19T15:50:42Z

Marcus Hutchins (of WannaCry fame) is my new favorite follow. Here’s a quick sample of him countering the claim that AI “democratizes knowledge.”:

LLMs chatbots are not democratizing knowledge. They're putting the people who produce the knowledge out of business, so that the companies who run the chatbots can sell expensive subscription services once they've captured enough market share to end free access to their products.

You Think?

2025-06-26T16:52:06Z

Should you use an LLM to help you “write?”

Let’s say I take some notes and thoughts I have about a particular topic and input into an LLM. I then review every word of the output, confirming that everything looks good. Let’s even assume I had the perfect prompt, all of the probabilities fell favorably, and the output matches exactly what I would have written. In a fraction of the time. Is this good?

I contend that, even though the output was the same, the outcome was qualitatively different.

The critical difference is, even though I would have written it, I didn’t write it. I didn’t think the thoughts that would have produced those words in my brain. I didn’t refine my mental model of the topic. I didn’t hold the concepts in my mind and make greater sense of the world.

I didn’t think, and I didn’t grow.

This matters, and this is what we forgo when we use an LLM to help us write.

Do the Work

2025-06-26T16:48:23Z

A few months ago, I was inspired by this remark from world-class violinist Hilary Hahn, speaking on “natural talent”:

At some point, everyone’s got to do the work.

In the current tech milieu, the employers have all the power, and their overwhelming push is for workers to be more productive. Much of this is masked or obscured by figuring out how to use ai¹. The demand for more productivity is usually framed as saving time. Ai will save you time, leaving you with more time to do what you want to do. But this is a lie. Capitalism will always find a way to do more with less.

Two of my friends have recently written about the growing angst among creators about the squeezing out of soul and humanity in our endeavors.

In The Lo-Fi Art and Human Tools Era, Pirijan talks about the pain of creation:

The first sentence, or the first brush stroke, is a high anxiety situation. Full of potential, full of unknown pain.

Yet this is an invaluable experience that too many are willing to forego, or ignorantly dismiss.

We don’t talk enough about the many-years-long rite of passage that every creator goes through.

Auldyn laments on the “industrialization” of UX, asking, “What the f*ck happened to UX?” This resonates with me, working on a b2b enterprise security product. Everything must be measured, we must see a return on investment. Goodhardt’s law is at play. She claims UX has been

reduced to a process of iteration and optimization rather than meaningful craftsmanship.

The craftsmanship is not for solely for artsy reasons. I appreciate how Auldyn makes the case that our products suffer when they lack emotion.

To me, it is no coincidence that Auldyn was provoked to write this piece as she got back to being hands-on. She is doing the work.

I will not indulge that meaningless term with capital letters.

Who is more foolish...

2025-04-28T15:42:30Z

Who is more foolish? The fool, or the fool who uses ai?

[#TheCompoChallenge]

2025-02-24T13:42:41Z

Analog

2024-12-09T21:18:23Z

For cyber Monday, I splurged and purchased Ugmonk’s daily and weekly system. I also threw in an hourglass. You can see that so far the system is not keeping me productive. I am busy documenting my usage of it instead.

Every system I adopt ends up failing sooner or later. I think the tangibility of the cards and objects is an added dimension that serve as physical reminders in my space to get things done. The fact that I need to physically write down tasks each day and review them is a good friction, I hope.

I’ll report back in 30 days.

Scratchpad

2024-10-16T20:55:56Z

Constraints are liberating. I’ve been enjoying Sindre Sorhus’s laser-focused apps like Gifski, Menu Spaced, and Plain Text Editor. Most recently he released Scratchpad and it’s found its way into a core position of my digital life—I love it! And this is solely due to the constraints he has designed into the app.

As the name suggests, Scratchpad is a place to quickly jot some notes. This is crowded market. Why not Drafts (where I am currently writing this post?) Why not Tot? The first and most compelling constraint is that there is a single input field. That means, no files to name or manage or sort. No archive to search on or destinations to configure. You are here and you will write.

The friction this produces discourages me from using this as a record or a traditional writing app. I am gently nudged to move notes from here to other parts of my system—Due for reminders, Outlook for events, Day One for journaling, and Drafts for writing (for example).

The second constraint that I appreciate is the lack of Markdown formatting and preview. This may seem contrarian as a techy, but again, even though Markdown is a very lightweight and elegant syntax, extra formatting helps dispose me to think in terms of sections and paragraphs, which interfere with my desire to keep this as a temporary scratchpad. Sindre asserts multiple times in his FAQ that this app is for plain text only, so I have high hopes it will stay this way.

The capability that sealed the deal for me was sync between iOS and macOS devices. It was important for me to be able to add to my scratchpad at my desk or away. As you can see, lack of syncing could have been another constraint that would be valuable. Every feature or constraint is a decision that you make to serve a specific need and context. For someone else, lack of syncing could be the deal-maker. So, it always depends what need you are trying to serve.

My previous experiment was to use one big text file as the starting point for all my writing. Having a single stream of all my notes, tasks, emails, etc. was enticing, but the ergonomics just weren’t right. I rarely reviewed previous writing. The mobile experience was not great.

Let’s see how long this honeymoon lasts.

Prototype for Whom?

2024-09-13T17:29:52Z

When we create prototypes, I rarely see a consideration around who the audience is. That’s a huge missed opportunity.

When I was participating in a design sprint activity earlier this year, it struck me that we were developing prototypes that were high in visual and breadth fidelity. We were exploring these elaborate workflows across the application and were investing a lot of time to polish them, make them flashy.

This surprised me because all of these new screens we were designing were based on assumptions that hadn’t been proven yet. So why are we spending this time on building this out when we have no idea whether users will go down the path?

For me, I felt like the most important thing we needed to learn was how our users tackle their day-to-day tasks in our product. But I realized that, for the designers on my team, it was equally important for us to produce an artifact that we could use to present upwards to our executive leadership and get buy-in.

This highlighted for me how important it is to think for whom is the prototype?

Audiences

There are at least three distinct audiences for our designs. Houde and Hill in What do Prototypes Prototype? note:

Once a prototype has been created, there are several distinct audiences that designers discuss prototypes with. They are:

the intended users of the artifact being designed

their design teams

The supporting organizations that they work within

This matters because, as Erickson states in Notes on Design Practice:

The different audiences are important because they require different design artifacts, or at least require the same artifact to be used in different ways.

There’s a place for creating beautifully polished prototypes for our leaders (#3 from above) because without their sign-off, the project isn’t going anywhere. But, will they be the right tool we need to provoke a reaction or discussion from the end-user (#1). I don’t think so. We should be intentional around who the audience is, and what impact or outcome we hope to gain from showing them the prototype.

Prototypes ≠ Figma

It makes me sad that when most people talk about prototypes (at least in my circles), all of our minds go directly to Figma. With the state of the art in design systems, designers can work quickly to create prototypes that are high in visual fidelity. The higher the better, right?

Nope. This isn’t primarily a matter of speed, either. Designers can whip out a Figma as quick as a paper sketch in many ways. This is about getting to conceptual clarity.

Why don’t we ever adjust visual fidelity? It is always pegged near the max. Yet, it focuses our attention on the wrong things. It implies a certain “doneness” when what we need is to show malleability, uncertainty, and an openness to discussion. It sets the wrong expectations.

Because high-fidelity visuals don’t actually create clarity. Rather, they hide the fact that the work wasn’t done, with detail that is every bit as fake as the sketch — but enough to satisfy stakeholders who are looking for an output.

(I echo everything in Pavel Samsonov’s full post: Low fidelity design is higher up the value chain.)

So, next time you’re creating a prototype, consider who you’re building it for. Is it for learning from the user? Is it for getting on the same page as your team? Or is it to sell an idea? Those are all distinct groups, and should result in distinct artifacts.

I’ve had this post in my drafts for months. I finally pushed it out because of a Cameron Moll post where he said:

Telling designers they shouldn't use high-fidelity designs early in the exploration process can be the equivalent of telling an artist they shouldn't use paint and instead only start with rough pencil sketches.

Some designers think best in hi-fi. I'm constantly amazed at how others feel justified in telling them to avoid this too early in the process.

This was engagement bait, and I took it!

Will AI go the way of the Concorde? ↗

2024-09-03T16:24:25Z

My colleague Lindsey M. West Wallace PhD wrote a piece entitled Will AI go the way of the Concorde? where she thinks through the ills of AI by way of a supersonic jet of old:

No matter how cool a technology is, it needs to deliver concrete value people care about that outweighs its costs and negatives or it won’t be adopted.

First, kudos for highlighting the environmental harms of generative AI. This harm is too easily ignored and techno-white-washed.

And, I agree with that framing of understanding why AI is missing the mark in terms of being a technology that will be adopted. My struggle with the analogy—and with most other analogies—is that it cannot capture the immensity of the hype and delusion around what AI can deliver, nor the immensity of its harms.

The Concorde was not built on stolen resources that exploits the creative work of anyone who has ever posted anything to the Internet. The Concorde was not feared to destroy humanity or imagined as the vehicle to all future flourishing. The Concorde technology was not inanely applied to every thinkable domain, taking investment away from all other humanity-improving endeavors.

The industry’s embrace of generative AI will have long-term repercussions that won’t be easy or even possible to erase. I wish we could wish this era a nice RIP and leave it to be forgotten in the annals of wikipedia, but sadly this tool is automating bad and dangerous behaviors and practices to a scale this world has never seen.

I, too, want my fonts fixed, and heck no, I don’t want to make friends with my mouse.

Not Another Chatbot!

2024-06-24T05:06:47Z

Okay, we gotta talk about chatbots. No one wants them, but everyone’s adding one to their product. This frustrates me for several reasons.

Fix your website

First, we know that LLMs depend on the data they are trained on. If users are having a hard time finding the information they need on your website or in your product’s documentation, throwing an LLM chatbot in front is not going to solve your problem.

The problem is that your documentation is poor and you should fix that. The problem is that your information architecture is in shambles. The problem is that your website is confusing and unusable. Fix those problems!

A chatbot is only going to prolong the process or delude you into thinking the problem doesn’t exist. This will be a downward spiral as you invest more money into plugging up the holes to coerce the chatbot to giving better answers while taking money away from your content designers, technical writers, and product designers.

Search is not just getting the answer

Second, using a LLM to seek and find information severely limits your ability to learn. Chirag Shah and Emily M. Bender argue in their 2022 paper Situating Search that using language models for search is “flawed in technical and conceptual terms.”

such approaches miss the big picture of why people seek information and how that process contains value beyond simply retrieving relevant information.

The paper identifies sixteen information seeking strategies that people may use when they are “searching.” For example, while a user is trying to decide on a new mattress to purchase, a user will be browsing, sense-making, and filtering information.

A language model fails to support these activities.

When someone uses a chatbot to find an answer from your documentation, the information is presented outside its context. The source is often obscured or wrong. Remember, a LLM is a statistical model. “They are only generating plausible sounding strings; any meaning in those strings is actually imbued by the reader.”

The perfect grift

Chatbots exemplify this common theme among all LLM-based technology. They handle 80% of cases decently, 10% of cases feel magical, and the last 10% is utter balderdash. This contour makes them the perfect grift.

Companies can build an amazing demo around it to capture attention and money. And just imagine where this will be in a year! they claim. Once it hits the real world, when you put it through its paces, it falls apart. But you’ve already poured money into it and made promises to your customers.

Seirdy’s post about MDN’s AI Help and lucid lies seems to reveal that something like this happened with Mozilla’s foray into “AI.” The article is a great read and shows clearly the flaws of LLMs for this use case, but I wanted to zero in on its conclusion (emphasis his):

The Mozilla Blog post Responsibly empowering developers with AI on MDN highlighted positive quantitative feedback for AI Help. Positive perceived helpfulness needs to be multiplied by actual helpfulness before it’s taken into account; in this case, I argue that actual helpfulness is negative.

This particularly troubles me. Their data reported positive feedback from users. Of course, a bot that instantly tells me the answer to my question? Sign me up! And I’m sure it worked the vast majority of the time. At times, it probably appeared like it was reasoning and taking the role of an assistant or copilot. Why wouldn’t I want this thing by my side?

From a user research perspective, how do we make sure we triangulate the data? When users perceive the tool to be useful, how do we measure that it actually helped them? We need to take a more nuanced approach because the technology is non-deterministic.

We are going to see a lot more LLM-based features and companies crash and burn. LLM are designed precisely to make us think they are intelligent. But, back to Emily M. Bender:

It doesn’t matter how “intelligent” it is — it can’t get to meaning if all it has access to is form. But also: it’s not “intelligent”. Our only evidence for its “intelligence” is the apparent coherence of its output. But we’re the ones doing all the meaning making there, as we make sense of it.

Reply on social
Reply via email

Disdain for AI

2024-06-11T18:47:55Z

I’m surprised I haven’t written more about my disdain for AI. I think it’s because the hype around it is exhausting and only proliferating each day. There’s something new every day to be disappointed at.

In a recent viral post Joanna Maciejewska stated:

You know what the biggest problem with pushing all-things-AI is? Wrong direction.
I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.

What I find frustrating about 90% of the AI features, solutions, and products out there is that they are removing friction in the wrong places and diverting us from opportunities to grow and learn.

I write to think. I’m always looking for hacks to write more because it forces me to think, helps me to solidify my positions are important topics. I was curious a year ago about how AI might bolster a writing habit. I bought a tool called AudioPen, which takes voice dictation and then cleans up the text. When I first tried it, it felt magical. But also uncomfortable. (The tool is well-designed and thoughtful, and seems like a poster child for solo entrepreneur success. Hats off to Louis!)

I haven’t used it because it deprives me of the critical piece of my writing process—the thinking. The hard part for me is not the typing. I don’t care for rewriting what I say in some other style. (You can train it on your own style, but why would I have the AI write in my style instead of me writing in my style?)

I prompted an AI to rewrite “yo” in a variety of styles and asking it to lengthen the message:

Greetings! I comprehend your current state of mind. I offer my sincerest apologies for the disappointment you are currently facing. It can be quite arduous when our expectations are not fulfilled, wouldn't you agree? I genuinely sympathize with your longing for a more favorable outcome and a more gratifying result. Please be assured that I am available to aid you and endeavor to enhance the situation for you!

Cool trick. Why is this capability in my enterprise work tool?

A recent post on my feed captures it well (emphasis mine):

llm’s convert you into the consumer of your own thoughts. instead if writing (which is thinking) you demand the idea you want and then consume it.
…or in other words — so-called “ai” transforms the act of thinking into the act of using and consuming. it transforms expression into prediction and the person into its user and consumer.

In most cases, I see AI helping us think less while producing more. It feeds on our urge to create, but robs us of the critical activities behind the creation. It prioritizes quantity over quality. If that is our orientation, then we may see benefit with these tools. If my goal is to create loads of content, AI will undoubtedly help me achieve that.

But that’s not me. I’m no content creator. I’m a human being making sense of this world. We’ve had access to much of the world’s knowledge for a long time now (meaning the internet). My path to learning and growing from it isn’t unlocked by asking ChatGPT to summarize it for me like I’m 5. Sure, it can help at the margins, but studying and absorbing material takes time.

So I’m not worried about falling behind by not embracing these tools. I’m more worried we’ll lose the richness of the world’s knowledge, beauty, and experiences because of all the money going into AI.

WeblogPoMo2024

2024-06-07T23:47:34Z

I participated in WeblogPoMo2024 last month. Instead of committing to a specific cadence—like every day—I dipped my toes in, and I’m glad I did.

I started strong, with three consecutive days. I gave myself the weekend off, and didn’t push myself the next week. For the following weeks, I wrote up a few posts. I wasn’t itching to write, and I didn’t want the be guilted into writing. So I didn’t.

I learned that in this season, I’m happy with prioritizing quality over quantity. I’ve got 10 saved drafts waiting to be finished. When I get the time, or when the urge to get an idea out becomes irresistible, I’ll write.

This post was something I started drafting a week ago. I sat down to write on a different topic, but now I feel guilty about not finishing that one, so I’m pushing this one out.

My Coffee Workflow

2024-05-17T19:58:23Z

I have a daily ritual to make a latte for my wife in the morning. There are some interesting machines out there that can automate many of the steps away. And sometimes, I don’t have time to perform the entire ritual, so sadly, that means no coffee.

But a big part of what makes the product enjoyable is the time I put into the process. In addition to the simple syrup and vanilla extract, I put a little sweat into every cup. If I could create the same result with the push of a button, I don’t think I would. I would miss the ritual. It makes me happy that I can make something small each day that my wife enjoys.

There is both art and science in this process. With every new bag of beans, there are new flavors to unlock, new techniques to experiment with. James Hoffmann sums up this activity nicely at the end of his recent video:

Great coffee’s worth chasing. It’s this wondrous little thing you can just sprinkle into your everyday life for a little bit more joy.

So, here’s the process I go through.

Water and preheating

Fill my kettle with water, leaving the top off
Instead of the kettle top, use an AeroPress funnel
Put the Flair Pro 2 brewing chamber on the funnel (for preheating)
Turn on the kettle and set to boil

Puck preparation

Measure out 16-17 grams of coffee in a cup
Spray two pumps of water into the cup
Cap the cup and shake for 10 seconds to distribute the water (the Ross Droplet Technique)
Pour beans into Baratza Sette 30 hopper
Grind beans at 10-11 stops into portafilter with filler funnel
Tap grinder to reduce retention
Distribute grinds using the Weiss Distribution Technique
Tap the portafilter on the counter to settle the grinds
Tamp the grinds
Seat the puck screen
Set portafilter onto machine
Position scale and cup under the machine
Tare the scale

Milk and cup

Pour milk up to fill line of NanoFoamer Pro
Start the NanoFoamer Pro
In insulated mug, add a teaspoon of simple sugar and dash of vanilla extract

Brewing

Once the water is boiling, carefully attach brewing chamber to portafilter
Turn off kettle and wait for temperature to lower to 204º
Pour water into chamber to the top
Seat the plunger stem
Start scale timer
Begin to pull down slowly on the lever arm
Allow the pressure to slowly ramp up for ~5 seconds
Once you hit 9 bars, sustain the pressure until you hit 35 grams of coffee (2:1 ratio)

Make the drink

Pour espresso into mug
Pour foamed milk into mug
Draw latte art (I haven’t learned how to do latte art yet)

I’ll spare you the cleanup steps!

Why I Write

2024-05-14T20:23:49Z

I grew up in the golden age of blogging, and I started blogging to emulate some of the folks online whom I enjoyed reading. At the time, it was voices like Joel Spolsky, Jeff Atwood, Jason Kottke, and Andy Baio. (The last two on the list are in that rare set of URLs that I still type via muscle memory. And the first two, I haven’t visited in years.)

Part of my reason to blog was clout-chasing. I revisited Coding Horror today and found this post: How To Achieve Ultimate Blog Success In One Easy Step. I think I believed that post and tried to put it into practice:

pick a schedule you can live with, and stick to it. Until you do that, none of the other advice I could give you will matter

Reflecting on this now, this probably contributed to my guilt over the past 15 years for not sticking to a regular schedule of blogging. I did keep writing, though.

Regardless, I’ve always appreciated how blogging forces me to think more deeply. Sometimes, it’s a deep dive into a technical problem I’m wrestling with at work. Or I’m researching a process or activity that I can then pitch to my leaders.

Blogging is also an outlet for self-expression. I get to share with the world my love of coffee, the excitement of listening to live music, and starting a new hobby like bike-riding.

What’s fun about personal blogging is that everyone has their own niche—the sum of their experiences, contexts, hopes, and dreams. I saw this video by J4vv4D that talks about finding your niche at the intersection of your passion and expertise.

That’s why I frequently think—and occasionally blog—about being an advocate for human-centered design in the “world” of engineering. In particular, I’m at a Biggish Tech company. There is a lot of organizational overhead to overcome. I work on a team that came from an acquisition a decade ago. I’ve been working remotely for 8 years. All of those factors give me a unique voice that no one else has.

I write about what’s on my mind, and I write things I would want to read from others.

This piece was prompted by Robert Birming’s post entitled Why write? I largely echo the sentiments he shares.

Reply on social
Reply via email

Love & Pain

2024-05-08T17:40:46Z

Tech has always been social and political, which the push of “AI” into every aspect of our lives has made even more apparent. Something I find interesting about the rise of “AI” hype is that it provokes meaningful discussions about what it means to be human. It forces us to now reckon with the spiritual.

In a recent episode of The Ezra Klein Show entitled Will A.I. Break the Internet? Or Save It?, the conversation between Ezra Klein and Nilay Patel flows naturally from media diets, art, to Taylor Swift and humanity (at 51:27):

It raises a question of whether there is anything essential about something being from a human in a wide frame way. Taylor Swift is singular, but the point is that she’s a singular phenomenon. Do we care that things come from people?

Speaking about a personalized chatbot he’s created, Ezra says,

The Kindroid might be better in terms of the actual text. I can certainly tune it more to my kind of theoretical liking, but the friction of another person is meaningful to me. Like, I care that my best friend likes me and could choose not to. Is there an aura problem here?

Nilay responds with a strikingly sober observation about the human condition:

It is so hard to make someone else feel anything other than pain.

In these few minutes, they covered a lot of ground. What is the value of a human being? Why do human relationships matter? What is the essence of being human? The choice of enduring pain for the sake of another—that is, love.

Technology raises these important existential questions, but is not equipped to answer them. One of my frustrations with “AI” is that we are not thinking critically about what it means to be human. It’s always been like this, but the amount of money and mindshare poured into “AI” in this moment makes the impact more disturbing.

Reply on social
Reply via email