I’ve been having a lot fo conversations with my team about where AI tooling fits into our development process and specifically AI code review tools. I recently shared my thoughts with my peers and it seeme good enough to share here too.

Good morning friends! There is a lot of discussion and action lately around the topic of how to integrate AI tools into the code review process, specifically adding PR review bots into Github and GitLab. I have been talking a lot with the Device Trust team about how we work and as the conclusions we have drawn are slightly different than others are taking I wanted to share our decisions and plans for others to consider as they decide how they want to approach these advances for their teams.

There is a (possibly apocryphal) quote from an old IBM Training manual:

A computer can never be held accountable
Therefore a computer must never make a management decision

At a point in the development of any code, there comes a point in time where the engineer must present the work to the wider world and say “here is what I have produced, what do you think?” To me, this is an almost sacred moment where one feels confident enough and proud enough of what you have built to solicit comments.

In our development cycle, this is the moment when a Pull Request is opened. How the engineer gets to this point is almost irrelevant; they can write the code solo, they can pair program, they can use an AI assistant, they can throw chicken bones; It doesn’t matter. There comes a moment where the engineer takes responsibility for what they have produced and shares the code.

The same theory applies to an engineer reviewing the code of another. This is a proces where an engineer has to use the wisdom they have gained over their career to improve the code presented by one of their fellow engineers. It is a place where our craft is discussed and critiqued. It is a fundamentally human endeavor. Both author and reviewer are responsible for the outcome.

I have several concerns about the addition of AI reviews into this mix. The primary one is that there is a fear that by introducing automated comments and review into the PR, it will discourage engineers from doing as thorough a job of reviewing as they would if they know they are the final arbitrar of the process. It is too easy to pass the responsibility of the review to the tool.

For that reason and others, we intend to keep the PR review process a place where humans work together to make the shared code better.

Does that mean I see no value in AI reviews of our code? Absolutely not. I see significant value in it. But for us, that process belongs in the hands of the author of the code.

We are building and sharing skills for our AI assistants to use before the engineers open their PRs. These are being built by the team and will become part of our code repository so that they are shared with everyone to use. The use of these tools becomes part of the development loop that an engineer uses before they present their code to the world through the PR process. They are there to allow every engineer to feel more confident in the code they share with the team.

I think how you integrate these new tools into your team workflows is a very important one that should be done carefully, and in a way that best supports the work we do. We are all ultimately responsible for the work we and our teams produce and we should build our processes to encourage personal responsibility for our work.

Thanks for coming to my Friday TED Talk.

As is obvious from my recent posts about my affair with AI coding agents, I’ve been seeing a lot of benefits from these tools and having fun using them. In my day job, how we integrate AI agents into our work has also been a very hot topic of conversation with my fellow managers and with my team as a whole. We are really trying to figure out how these tools can be used in a responsible manner. I believe that these tools can be a benefit to engineers. Well, that isn’t 100% true. I believe that these tools can be a benefit to senior engineers. I have doubts about early career engineers.

In order to use these tools to develop software that is sustainable requires a level of wisdom that only comes through making a ton of mistakes. It requires building a heuristic of what good software is. It requires becoming a craftsman. And that only comes through trial and error. It reminds me of my childhood and the debates of when to allow students to use calculators in school. As a student I wanted to use one immediately to solve problems. But the powers that be prevailed and I wasn’t allowed to use a calculator to do X until I knew how to do X by myself; until I had learned how X was supposed to work so I could tell if the calculator was giving a result that “felt” right. Developing that intuition was an apprenticeship. So what happens when you no longer need apprentices? How do people learn that intuition?

That intuition is like a trapdoor function. It takes time to develop, but once you have it, it is quick to apply.

This topic has been nagging at me for a while. I am terribly worried that the advent of these coding agents is going to decimate the ranks of entry level engineers because the senior engineers can use the tools to orchestrate what is in effect an entire team of entry level engineers and use their intuition to guide and evaluate what they produce. Companies are looking at this and may decide that they no longer need to hire new engineers. In our market driven, quarterly profits are all that matter world, it makes sense. But what is it going to do to our industry in 10 years? or 20? I’ll be retired. And many of the engineers who grew up before these tools existed will be too. Will their be a next generations to pick up where we left off?

The counterargument is that the tools will keep getting better and the wisdom that I value so much about coding will no longer be required. Just like I don’t need to know anything about assembly. (I would argue, though, that I have the benefit of knowing how the hardware works at that level.)

I don’t know and maybe I am just caught in “old” thinking.

Is the answer to “restrict” the tools to be used only by people who have already learned how to write software? That could work, but it will never happen. If these tools become embedded in our lives, do they become the teachers? The new masters of the trade raising the next generation of apprentices?

I guess we’ll find out.

Just a quick followup to the last post sharing what has been going on. I’ve been keeping up the pace of development and still having a blast building more apps.

• Grafo - This has been released, had a version 1.1 released, and I have a version 1.2 in App Review as we speak. Having a well defined API both provides a nice limit on what can be built but also a great roadmap for features that could be built. The question only has been what features of the Zabbix web app make sense to have in a mobile context. As I am a home user (even if I am an ambitious home user), I don’t know how the systems are used in larger organizations, but I have built what I think I would want.

• Elekto - This is still in development but is coming soon. There is a huge community of people (myself included) who like trying to optimize miles and points as a way to travel. You can go really far down this rabit hole with spreadsheets and such to keep track of the best way to optimize things and there are apps out there to help you do it in a very rigorous way. But for the casual user (or the spouse of a less casual user), having a simple tool to determine the best credit card to use in a specific scenario can be useful. That is Elekto.

• claude-consolidate Given that I have been using claude across multiple projects and multiple computers, I wanted a system to consolidate and manage all the different settings, exceptions, MCPs, and skills across my environments. This is a bit bespoke to me (hence it not being public right now), but it has been helping me maintain some sanity. I have a similar one for the desktop version of Claude.

• I’ve also been writing some ansible scripts to bring some consistencey to my various home servers. Nothing earth shattering but they save me time.

A couple of factors have gotten me back to building things in my spare time. As my work has transitioned back to more management and less coding at work, my desire to do it on my own time has increased. But if I am being honest, the bigger reason is claude. It has done a few things; for bigger projects like the iOS ones below it has allowed me to act more as a product manager than a developer and build things that I have in my head without having to re-learn modern Swift and SwiftUI. But it has also lowered the barrier to entry for turning manual processes into applications and scripts. This is the one I feel is more transformational on a day to day basis. Where before I would have been mildly annoyed doing a manual process, but not annoyed enough to automate it myself, now using the robots I can spend far less time automating it for myself.

So what have I been up to?

iOS Apps

• Amiko - You may remember Amiko from a long time ago. It is an app I hand built to help keep track of when you last saw your friends and remind you to keep in touch. The app languished over the years without updates and eventually got pulled from the App Store. Rather than trying to update the existing ancient codebase, I started fresh and built a new app from scratch using modern practices. It is almost ready to be released.

• Fantazio - One of my COVID projects to help Dungeon Master’s run tabletop games using digital maps. It also (are you sensing a trend) languished and while it was still on the store, it was outdated. I was able to update it to run on the latest devices and add some new features.

• Grafo - This one is still in progress. I use Zabbix as my home server/network monitoring tool. I wanted an app that used the API to show you data on your mobile devices. It is coming along and I’ll probably try to release it in the coming weeks.

Random Things

• iot_monitor - I run all my IOT devices on a highly restricted segmented VLAN with really restrictive firewall rules. Gathering the data for this used to be an exercise of sitting with tcpdump and trying to capture all the data I needed. iot_monitor is a small tool I can use to watch the network and it captures just what I need for my firewall rules. Is it useful for anyone else? No idea.

• today_overview - I have a local LLM server running that I use for more private data and API things. This program is able to collect a bunch of iCal feeds, extract today’s events, then forward them on to your local LLM to give you a summary of your day. I have this emailed to me every morning so I can get a quick sense of what my day will be like.

• A bunch of ESPHome device extensions. Soil moisture sensor 1 and Soil moisture sensor 2. I’ve been playing with building my own monitoring hardware to go with Home Assistant and sometimes it isn’t supported directly.

Tons of shell scripts

• A shell script to take an ebook file, convert it to the formats I need and import it into my Calibre library without having to use the Calibre UI which is…not pretty.

• A shell script to automate backing up various docker volumes because there was no way I was going to do it by hand.

• A whole bunch more I can’t remember right now.

And if I am being honest, I wouldn’t have done any of this if I didn’t have access to the newer LLM based coding assistants. I don’t know what this means for the profession as a whole, but I can say it has made me more productive for fun little projects.

This is not a topic that I talk about very much as it does not fit in my “brand” of a tech person. But we are all made of multiples, so you get a different topic today. Prior to COVID, Tiffany and I made what at the time seemed like reasonable attempts at exercise. We had a membership at a local gym and worked with a “personal” trainer once a week. That was good, but we would use that as an excuse to go to one of our local bars for a few beers and a really tasty burger and fries. Very reasonable, but we basically undid most of the benefits of the workouts. And the trainers at this gym were fine, but they should not be considered very serious; we did not get as much as I have now learned we could have.

When COVID hit and we were all at home, I found some online classes through Classpass, which is a good service. But through that, I was introduced to my first “real” trainer, Alex Najarian. She had started a 3-day-a-week online Zoom-based workout group. It started as a combo of kettlebell work plus HIIT, but very quickly transitioned to be just kettlebell. And it was serious. We had a regular group of 6-8 people doing 50 minutes of hard-style kettlebell work 3 times a week. I spent 2 years or so working with Alex before she decided to move on and handed us off to a new trainer, Nicole Bedoya, who continued my journey into kettlebell training. Through the years of lockdown, I lost a significant amount of weight and got myself to be the strongest and best shape I have ever been, even as I rapidly approach my 50th birthday. At the beginning of this year, Nicole needed to step away, and we found a new trainer, Colleen Conlon, who has taken me even further than I thought possible.

Three days a week at 8 a.m., I make it down to our basement/theater/gym and log in to Zoom and spend an hour pushing myself with a great group of strong, motivated, and encouraging friends. It has become such a critical part of my life that it is a significant loss to me when I can’t make it. This week I had a business trip to Denver and was going to miss two days. I have tried to do our classes in the hotel gyms, but they rarely have a good set of equipment and for whatever reason I feel even more self-conscious than normal doing our workouts in a hotel gym. I was thrilled when I found a small studio, Kettlebell & Co, that had classes at a time that worked with my work schedule. It was the first time I had ever done an in-person kettlebell class. But the trainer, Ben, and the other members of their team were welcoming and made it a great experience.

A long intro to the thing I wanted to write about: why do I do all this? It isn’t always fun in the moment. The intellectual reasons are straightforward: better health, longevity, mobility, and keeping all of those things as long as I can as I get older. That said, that covers staying in good health through exercise in general. I often wonder, as I progress (often when contemplating whether I should get another ludicrously large kettlebell; 80lb bell anyone?), why do I keep pushing harder? As I started saying, I don’t really talk about what I do online or with my friends except in passing. While I certainly have some of an ego and body image goals, those don’t really drive me too much either. I’m not looking to be a big muscle-y guy, it isn’t in my genetics. So why?

And here is where we get crossover with my professional life. As a software engineer, I have sort of had tangible goals and progress I could see. But especially as I have moved back into management, those tangible goals are much harder to identify. It is very hard for me to not fall prey to imposter syndrome and get stuck asking myself, what have I really accomplished? Doing our workouts is simple in that regard. It is measurable. I can see over the course of 12 weeks where I started doing 10 rounds of 10 kettlebell swings with a 53lb bell to now doing 10 rounds of 7 swings with 95lbs of bells. It is measurable and it is quantifiable. I know what I have to do to get better and the work is understandable.

The benefits of this simple, quantifiable progress does wonders for my body, but far more importantly for my mental and emotional health.

So, I’m sorry UPS delivery person, there is probably another large sphere of iron coming my way.

In my continued musings about AI and particularly, media coverage of AI, there is a trend I am starting to see that I find a bit disturbing. It bothers my logic brain. There is a quote from the recent Atlantic article What Happens When People Don’t Understand How AI Works that is a good example. (It is a good article, you should read it regardless.)

Many people, however, fail to grasp how large language models work, what their limits are, and, crucially, that LLMs do not think and feel but instead mimic and mirror.

I want to dig into this part: “LLMs do not think and feel but instead mimic and mirror.” I would agree with the second half of the statement, LLMs are statistical models that create new content based on patterns they have picked up from ingesting huge quantities of existing content. I start having a problem with the first part of the statement: “LLMs do not think and feel.” The implication here is that humans think and feel and LLMS do not. In order to determine the truth of that statement, we have to know a few things.

• What is “thinking and feeling”
• Using that definition, do human’s do it?
• Using that definition, do LLM’s do it?

I feel like we fail right out of the gate. What is “thinking” or “feeling”? I honestly don’t know. This is a subject that philosophers have been arguing about for milenia. I think that we neither have a philosophical answer nor a scientific or medical answer to what the heck thinking even is.

So if we don’t know what thinking even is in humans, how can we say one way or another whether LLMs do or do not do it? We actually have a better idea of how LLMs do what they do than we do of how human brains do what they do.

Yet in many articles I read it is just stated as fact that thinking is something humans and only humans can do. I feel like we are jumping the gun on that. Unless we define it as “thinking is an activity that only humans do” which makes it a useless definition in a context where we are discussing humans and non-humans.

I find it hard to accept these arguments until we do a better job defining exactly what thinking and feeling actually are.

mast year: A mast year for oak trees is a periodic event, typically occurring every 2-5 years, when the trees produce an exceptionally large crop of acorns.

I have just completed a rather intense hiring cycle in my role as an Engineering Manager here at 1Password; hiring 2 Staff Engineers and 2 Senior Engineers in a 4 month period. As I was wrapping up the final Senior Engineer process and was waiting for final confirmation of a signed offer letter, I continued some already scheduled interviews and an event that I dreaded happened. I interviewed another great candidate and I couldn’t hire them. I do not regret any of the previous hires, but I knew that this person had the possibility of also being great for our team.

The worst part of hiring is knowing that you have to tell great candidates “no” not because they aren’t a great fit, but because you just don’t need them for your team.

Our hiring plan includes adding 2 Intermediate Engineers to our team as well and that process had not started. My initial reaction to having another great Senior candidate was to see if I could re-jigger one of those Intermediate roles to be a Senior. But I realized upon thinking that that was a mistake, that our original plan was good and I needed to stick with it.

My mindset for hiring was set from a different time at different companies. I have been hiring engineers throughout my career and generally for smaller, less well known companies than 1Password. It was always a struggle to find enough good candidates to fill the roles we had. So, if we had been looking for 1 great engineer and 2 walked in the door, we would have been stupid to not hire both of them if it was at all possible.

Hiring at 1Password and in today’s employment climate is totally different. There are an abundance of great candidates. When we advertise an open role, within a week we have in excess of 500 applicants. 50% are apam, but even so, in these 4 roles I have hired, we have found amazing hires without even making it through the whole pool of people. The combination of the tight hiring market and 1Password’s name recognition results in us having our pick of great people. This has confounded my instincts.

I am having to shift my thinking to realize that I am not hiring during a time of scarcity. It pains me knowing that I am turning away great people but I can’t hire them all. The dark side I must avoid is to not value the team that I already have less knowing that there are more acorns out there.

I have been hearing about Tailscale for a few years now. I think it was probably first on the Accidental Tech Podcast by Casey Liss and I honestly dismissed it as another one of his wild goose chases (see raspberry pi garage door sensor). But it has popped up more than once and I recently listened to an interview with Tailscale’s Co-founder, Avery Pennarun on Stratechery that I found very interesting. His road to networking and computers felt very similar to mine although he obviously carried it quite a bit further by founding a company. I had been noodling with the idea of trying it for my use case and this pushed me over the edge.

I already have a very well developed home network so I am not the traditional user. I am some odd hybrid of a “personal” use case and a level of “business” complexity. I have two use cases that I currently use OpenVPN for.

• Remote access for my and my wife’s personal devices when we are outside the home. This is mostly to route internet traffic through our home connection when we are on untrusted network, but occasionally to access services hosted at home.
• To connect two servers that exist outside our home network to the home network for more secure and easy access.

Many people seem to use Tailscale to connect all of their devices. That was not of interest to me. I already have the physical network at home set up and segmented the way I want it and recreating that with the mesh network was of no interest to me (and would involve dozens of devices, not all of which are capable of running it). So my initial goal was to have a simpler to manage version of what I was doing with OpenVPN.

I chose to set up my OPNsense router as the central node of the network and make it both a subnet router and exit node. The big change that I made from what it seems is a “standard” installation is normally an exit node will SNAT the packets going out so that they seem to be coming from the exit node. I am perfectly happy having the 100.0.0.0/8 addresses coming into my internal network. I want the tailnet to appear just as another segment to my network so that even non-Tailscale can send traffic to hosts inside the tailnet.

In order to do this, I configured the core router with:

tailscale up --snat-subnet-routes=false

Because Tailscale was no longer doing the SNAT, I also had to configure OPNSense to do the proper outbound NAT for anyone with those addresses. And I set up my DNS server to forward requests for hosts in my ts.net MagicDNS zone to the Tailscale DNS server 100.100.100.100.

I set up the core router to also advertise several routes for my internal network (to subnets that I trust the clients to get to). When installing tailscaled on my two remote unix hosts I needed to explicitly tell them to use those routes via:

tailscale set --accept-routes=true

It took me less than a day of putzing with things occasionally to have it all set up. The “fixed” servers could talk to hosts in my local network. All the clients inside my network could talk to the hosts that were connected via Tailscale using their 100.0.0.0/8 addresses or MagicDNS names even though they dont have tailscaled running. And for remote access on my mobile devices, I set up the Tailscale app and they are able to connect and use my home network as an exit node.

Overall I am pleased with how easy it was to set this up. I was a little nervous of the security implications but was pleased to see the Tailnet lock function that prevents unauthorized devices from being added to your network even if Tailscale is acting maliciously or compromised.

So far so good. I still need to see how well Tailscale can punch through restrictive networks to build tunnels. But happy so far.

Two posts in a few days, huh. I saw a tweet (xeet?) that I won’t link to because I hate what that platform has become, but it talked about the wonders of the new Claude Code tool. I figured I would give it a shot and it is pretty impressive. I had a few goals I wanted to accomplish, upgrading some components of my blog engine and CSS stuff that I hate doing. So I asked it to do it for me. Below is a summary (produced by Claude) of the changes it made for me (we made together?). All this ended up costing me $1.83. A bit expensive and I could see it being very expensive if I wanted to use this all day for my job. But an interesting experiment in what is to come.

Everything below the line here was written by Claude.

Today I spent some time modernizing the technology stack that powers this blog. Over time, web technologies evolve, and it’s important to keep up with these changes to maintain security, performance, and compatibility with modern browsers.

Bootstrap Upgrade: From 4.6.2 to 5.3.3

The first major change was upgrading from Bootstrap 4.6.2 to the latest Bootstrap 5.3.3. Bootstrap 5 brings several important improvements:

1. Dropped jQuery dependency - While my site still includes jQuery for certain functionality, Bootstrap 5 no longer requires it as a dependency, making for lighter and faster pages.

2. Improved CSS variables - Better support for CSS custom properties, making theme customization easier.

3. RTL support - Bootstrap 5 has better support for right-to-left languages through logical properties.

4. Upgraded Popper.js - Moved from the older popper.js library to the newer @popperjs/core package.

The upgrade process wasn’t entirely straightforward. Several changes were needed:

• Updated the package.json file with the newer dependencies
• Modified the Makefile to reference the new package locations
• Updated CSS classes like float-right to their Bootstrap 5 equivalents (float-end)
• Adjusted SCSS imports to accommodate the new Bootstrap file structure

Jekyll Upgrade to 4.4.1

In addition to Bootstrap, I also upgraded the Jekyll static site generator from 4.3.2 to 4.4.1. Jekyll 4.4.1 includes various bug fixes and performance improvements that make the site generation process faster and more reliable.

Sass Warning Challenges

One interesting challenge during this upgrade was dealing with Sass deprecation warnings. The current version of Dart Sass (which Jekyll uses to process SCSS files) warns about several deprecated features that Bootstrap 5 still uses:

1. @import rules - These are being replaced by the newer @use and @forward rules
2. Global color functions (red(), green(), blue()) - Being replaced by namespaced versions (color.channel())
3. Other global functions like mix() - Being replaced by color.mix()
4. Mixed declarations in CSS that will behave differently in future Sass versions

While these warnings don’t affect functionality today, they indicate future changes that might impact the site. A full resolution would require creating custom Bootstrap wrappers using the modern Sass module system, but that’s a project for another day.

Ruby Version Requirements

This upgrade also highlighted how important it is to maintain compatibility with the right Ruby version for Jekyll. I discovered that Jekyll 4.4.1 works best with Ruby 3.2.3, not the newer 3.3.x versions which caused some compatibility issues with the Jekyll logger.

In Summary

These updates represent the kind of regular maintenance that keeps a website healthy and current. While users might not immediately notice these changes, they lay the groundwork for future improvements and ensure the site continues to work well with modern browsers.

The upgraded packages are now being tested on a separate branch before being deployed to production. This approach helps catch any compatibility issues or styling problems before they affect the live site.

There’s always more to do in the world of web development, but keeping dependencies updated is a vital first step in maintaining a healthy website.

There is a joke somewhere about opinions and assholes. It seems to me that writing about AI and its impact on everything has been proving that to be true. Yet, here I am, adding more to the pile. In my defense, I am writing this mostly to settle my own thinking and put down my current opinions for reference so that when I am wrong, I have something to look back on later.

There have been rumblings about the power of LLMs and “AI” for quite a while but for me it became real after Github released Copilot in 2021. This is the point where as a software engineer, it had the opportunity to have a direct influence on my livelihood and lots of people started opining on how LLMs would change the world forever, be a flash in the pan, or maybe end the world.

As a geek, I was pretty curious. I made some early attempts at running some of the AI models locally on what at the time was a beefy GPU. I was not impressed. It was very difficult to get running, the results were mediocre at best, and using just the sound of the fans as an indicator it was hugely wasteful. It seemed like the biggest usecase waas making bad pictures with too many fingers. I dismissed LLMs and “AI” as a crutch that was decades if not longer away from being anything useful. I think this was partly my pride as staff engineer, partly my disdain for the search for silver bullets that would make everyone a successful software engineer, and my limited experience with LLMs at that point and their propensity for making shit up (aka hallucinating).

Then in 2023 I had an idea that started to change my mind. I still firmly believed that LLMs hallucinatory behavior was bad but there was a crossover to my role as a dungeon master for a few Dungeons and Dragons campaigns where hallucinating could be good thing. So I tried something

Write a damage liability waiver form for a dungeon where fantasy adventurers must fight undead creatures for practice. Include standard liability legalese. The name of the dungeon is The Dungeon of Darkness and the proprietor is Berala Evenfall.

Thus began my tentative use of LLMs. There is a lot of reasonable debate about using LLMs for generating creative content. I am uncomfortable that there are companies that have shamelessly used content without compensating the original authors to build billion dollar companies. But to expand the enjoyment of my players in a casual setting I found them to be very useful. From writing deeper backstories for minor NPCs to fleshing out my world building it helped me have more fun playing D&D and I think that was a good thing. I confined LLMs to making shit up in an arena where making shit up quickly was their strong point.

Tangent: I do ponder at times that there is a strange parallel between AI training and human learning. If I think of the human brain, what do we do? We are pattern matching engines. We take new situations and compare them to all of the experiences we have had in life to that point and try to figure out the most likely response to achieve the outcome we want. That doesn’t feel all that different to me from what is being done with LLMs. Even with human’s, there is a fine line between being “inspired” by the content I have freely been exposed to in life and plagiarizing it. Back to the original story.

Sometime later in 2023 I got access to use Github Copilot. And as much as I hated to say it at the time, I was grudgingly impressed. I did not see it as any kind of existential threat to human developers but it was definitely a useful tool. It often produced a good first draft of what I was working on; enough that I could more quickly fix what it had done than typed it from scratch by myself. Or to help me document what I had just written in a way that would be helpful to future developers by commenting things more completely. However, I found that it had an interesting bimodal distribution of results. Sometimes it would pre-generate a brilliant implementation of what I had barely started to do myself while at other times it would stubbornly produce absolute garbage over and over. It made me realize that my mental model for how these tools can fail is nothing like how I think of code written by other humans.

That realization is I think what concerns me most as LLM and AI tools advance. For the whole of human history, when interacting with people around us, we always have to make judgements on the veracity of what we are seeing, what they are doing, and if we should trust them. And this is even before we take into account potential malice or intentionally lying. The user experience of AI technology is to create tools that we interact with similarly to other humans. However, the failure modes of LLMs do not mirror that of people. For example, we expect a reasonable human to be able to say “I don’t know” if they are not confident in their answer (and we distruct people who lack that ability). AI tools? Not so much. See using glue to keep pizza cheese on. We have been conditioned (incorrectly I might add) by pre-LLM software to believe that computers do not make mistakes. This brings about a huge risk that our brains are just not set up to determine how much we can trust the information we get from them.

My current use of AI tools is sporadic. Copilot is part of my development workflows and sometimes it is useful and othertimes I will go days without ever using it. I have also started dabling with running some of the models locally again. The tooling has gotten far better (I use ollama and Enchanted) and being able to run the models on my Macbook Pro has significantly improved the experience. There is something magical about being able to do a query that is basically using the whole of the internet locally and offline from your laptop to find answers. It is a nice backup. And being able to have multiple models that I can run side by side to compare allows me to find the right tool for the job. Analagous to finding different human experts. I also feel that I am learning where these tools fall apart and where I need to be wary.

So where does that leave me in early 2025? My biggest error early on was forgetting how quickly technology can move. I feel like I have been watching the technology improve faster and faster right before my eyes. There was a quote I heard on the Stratechery podcast (not original I am sure) from Ben Thompson about progress that we can go “decades with years of progress and then years with decades of progress.” I feel like we are in the second situation. How long will it last? No idea. But I don’t think we are done seeing rapid growth.