C33 Technology

Minesweeper 2 - The First C33 Game

team@c33tech.com (Mike Flynn) — Sun, 27 Sep 2026 10:45:00 +0000

I like Minesweeper. Yes, that Minesweeper. The game that’s been around since at least Windows 95, where you click around flagging mines. To be sure, it’s a deeply flawed game, but it’s a fun, when not frustrating, way to pass a few minutes especially over the last 15 years or so when I’ve had it on my iPhone anywhere I was. I’ve played countless games on planes, in lines, during commerical breaks, etc.

Up until recently, all was well in the world but then it started to happen: The enshitification found even my sweet Minesweeper app…well, apps. App after app was junked up with ads and bizarre design choices. I don’t mind paying for an app, but I don’t want you to share my data and shove ads in my face just for a 3 minute game of Minesweeper!

Sometimes a weekend project is thurst upon you, and now with a little AI help you can actually finish some of them.

A few months ago I quietly made and launched Minesweeper 2 on the app store.

Here’s the key features: It’s $2. No ads. No tracking data. It looks like it belongs on iOS.

There’s other stuff too though:

A new way to play: Timed Trials where hitting a bomb won’t immedately kill the game, but instead just give you a time pentalty. This smooths over the worst part of the game when you’re stuck in a true 50/50 chance or if you’re just learning.

There’s also some fun stats, a classic mode that makes it look like the old Windows version, a couple of easter eggs and, new with v1.1, an Apple Intelligence-powered coach for when you do hit a bomb.

So far a few people of found and bought it, and I hope they love it, but that wasn’t the point. I wanted to make a pure version of this game for iOS, and I have. If other people buy it, that would be great (Seriously, just $2!). Maybe I’ll sell enough to cover that (yearly) Apple Developer subscription I had to buy, but at least I got my game back.

HackerSkill

team@c33tech.com (Mike Flynn) — Thu, 20 Aug 2026 09:00:00 -0700

The easiest way to ruin a Hack The Box night is to paste the box name into a chatbot and get the whole solution back. You get the root flag and you learn nothing. HackerSkill is our answer to that. It’s a Claude Code plugin that sits next to you while you work a target and coaches you through it instead of spoiling it.

It’ll run the whole lifecycle with you, from recon to foothold to privilege escalation to rooted, and it keeps the notes as you go.

Key Features:

Tiered hints. Every hint starts as a nudge. Ask for “more” and you get the technique. Ask it to “spoil it” and you get the answer, but only because you decided you wanted it. You can also say “no spoilers” to lock it to nudges until you say otherwise.
It takes the notes for you. Each target gets a targets/<name>/notes.md that acts as the source of truth, with scans in scans/ and loot in loot/. Walk away for a week, run /pwn lame again, and it picks up where you left off.
Hybrid execution. The safe, boring enumeration (nmap, gobuster, whatweb, smbclient) just runs. Anything with teeth like exploits, auth attempts, or shells will stop and ask you first.
A real playbook. Service by service techniques are built in, and it reaches out to HackTricks, GTFOBins, and exploit-db when a box needs something more specific.

Getting started is two commands in Claude Code:

/plugin marketplace add C33Tech/HackerSkill
/plugin install pwn-challenge@C33Tech-HackerSkill

Then point it at a box and go:

/pwn lame 10.10.10.3

That creates targets/lame/, drops in a notes template, and kicks off the first nmap scan. Run /pwn on its own to list your targets, and /pwn done to wrap up a box once you’ve rooted it.

It shells out to the usual offensive security toolkit (nmap, ffuf, feroxbuster, enum4linux-ng, nxc, evil-winrm, the impacket suite, hashcat, and Metasploit) and it degrades gracefully. Whatever you don’t have installed, it just won’t try to run.

One deliberate constraint: this is for authorized lab, CTF, and engagement targets only. Point it at an IP outside the known lab ranges and it’ll stop and ask you to confirm you have authorization before it does anything. That’s the boring answer, but it’s the right one.

HackerSkill is entirely open source and free to use. If you run it on a box and it coaches you badly, either too much hand holding or not enough, let us know! That feedback is what tunes the hint tiers.

How Not To 'Pilet' a Kickstarter

team@c33tech.com (Mike Flynn) — Sun, 12 Apr 2026 22:45:00 +0000

The Pilet, created by Soulscircuit was billed as an “opensource, modular, portable mini computer” and “Where retro meets modern, powered by Raspberry Pi 5.” They had full pictures and renderings, details, and videos of people using a prototype. For someone in the market for a portable computer like this, it was a really good sell and unlike a lot of Kickstarters, this looked really close to shipping.

The campaign ended up collecting CA$ 1,264,707 from 2,777 backers.

Their risk statement was pretty generic with risks that are valid and apply to all projects, but also projected some clear confidence in their plan. This tracked after watching a video of people using it!

While we’ve planned our timeline meticulously, delays are always a possibility in hardware projects. We’re collaborating with trusted suppliers and have set realistic lead times, but unforeseen issues may still arise. Additionally, global events and other unpredictable factors could cause delays outside our control.

But this isn’t a story about a Kickstarter that was well planned, marketed, and executed. This is a story about a Kickstarter that was well planned, marketed, and not executed well at all.

Shipping Costs

As the original delivery date got closer, emails started going out to ask about add-ons, shipping, and options such as case color. Shipping was high. Very high for some. At a minimum, the team at Soulscircuit collected another $50 from each backer, but likely much more as people added batteries, NVMe drives, or other add-on items.

At least it was close to shipping! Right?

Communication Slows

If you aren’t familiar, Kickstarter has a mechanism for creators to post updates on their progress. Kickstarter highly encourages frequent updates, no less than once a month. Here is the timeline for updates across the life of the Pilet project, with the time since the last update noted:

While still largely positive, the comments on each update started outright begging for more frequent updates.

On “Project Update #16: Showcasing the Display and Firmware”:

It has been a month since the last update. We could use a quick message about how the manufacturer selection is going.

Please post an update at least once a month! This is once again starting to feel like a campaign that will never produce any result.

You’re Just Now Finishing the Keyboard?

As the updates slowed further and the deadline was farther and farther in the rearview, the lack of substantive information and the fact that no one has seen a new picture of a working prototype began to be more and more obvious.

On “Project Update #18: Showcasing the Keyboard for Pilet 5 and Pilet 7”

Deadline pasts… months gone… year past… but no delivery of product. Hope things are going [sic] fine… and we will get the product soon

It’s been another month. Aren’t we due another update about how you are still looking for a manufacturing partner? Or another update on hardware that is being developed that was portrayed as settled work in the campaign?

5 months late and you’re just now finishing the keyboard?

Y’all in dire need of a good project manager 😂

The group of commenters that were focused on it being a scam, certainly common on Kickstarters, were growing but still a vocal minority.

The Pilet Pivot

From the campaign’s “Key Features” section:

Raspberry Pi 5: Pilet is powered by the latest Raspberry Pi 5, bringing performance boosts alongside advanced features like PCIe and NVMe support.

From “Update #20: Moving to Compute Module 5” open April 11th, 2026 (a full 2 months since the last update):

Over the past year, we’ve been developing a version of the Pilet based on the Compute Module 5, in parallel with the Raspberry Pi 5 design. After extensive testing and iteration, we’ve decided to move forward with this approach.

This update has not gone over well. The comment thread is filled with variations of this:

This is turning into a joke, lads. I’d originally hoped that we’d receive what we’d paid for in the not-too-distant future, but those hopes have been dashed. I already own a Raspberry Pi 5, which is why I backed this project. If I’d wanted something based on a CM4 or 5, I’d have bought a uconsole or something similar.

The fact that you guys were wasting time developing a separate project, when the one we backed was already months off schedule, basically kills any trust I have in either of you or this project. The Kickstarter appeared to show a device close to being ready, which was due to ship approximately five months after the end of said Kickstarter. The reality appears to have been quite different, with no deliverable in sight.

If you want to work on this new version, and eventually sell it, then that’s fine. Refund us our pledges first, though.

As it stands, the project remains in limbo. The creators certainly are able to see the feedback, but no one has been able to get a response from them. Kickstarter has yet to step in or respond to any “Trust and Safety” submissions. It’s full-on rebellion in the comments, but no one is getting their money back, so the temperature keeps going up.

It is absolutely astonishing to think the Soulscircuit team thought that this update would be met by anything besides anger.

Lessons

Kickstarter is a gamble, but there has to be some trust there, and it can’t be that much of a gamble, or the customers will never return. Things can happen. The Tump admin’s tariffs no doubt made hardware manufacturing harder and more expensive. It seems likely that they at best fibbed when it comes to their marketing materials, as the device in those photos and videos has never been seen again. Despite all that, though, they could still have the crowd on their side. There was a path. They had to communicate.

Obviously, nothing beats delivering a project on time, but failing that, frequent updates and responding to the community are damn close. Why are you working on the keyboard five months after the delivery date? What happened to those demo devices we saw last year? These are questions that could have been answered in a matter of minutes. Posting an update weekly, even if it just says “We’re still working hard! Larger update soon.” would take seconds of a day.

Do I think this project was/is a scam? No, I don’t think so, but I do think it was run by two guys who are talented technically and complete and utter morons in management and communication. I’m sure they consider themselves very smart, and many would agree, but being smart isn’t just about being able to create your own circuit boards. Smart people know their gaps and when they need help.

These guys are in way over their head and they still don’t seem to even realize it.

Minesweeper 2

team@c33tech.com (Mike Flynn) — Sun, 22 Feb 2026 10:00:00 -0700

Gemini's Workspace Confusion

team@c33tech.com (Mike Flynn) — Mon, 09 Feb 2026 20:51:22 +0000

Google has a long, storied history of making things overly complex and confusing for their own paying Google Workspace customers. Jeff Jarvis has made many a rant over the years expressing this viewpoint on a number of features that Google has failed to deliver for Workspace customers.

Today I’d like to join Jeff and bring up a new one that I don’t think he’s mentioned: Google’s AI packaging.

Here’s what I want. At Ten2 Media, we use Google Workspace, which is largely great and comes with some added AI features, but I also want to grant a selected number of users additional AI quota, specifically in regards to using Gemini for coding. If you have looked into this, you’re already laughing. If you haven’t, you’ll soon find a tangled mess of Gemini plans, Vertex, Google Cloud Projects, Google Cloud Assist, and so many oddly phrased error messages when trying to use these tools.

Why is this so complex? Why is it a checkbox on the user’s Workspace settings? Why is it so tied to GCP and specific Google Cloud projects?

The easiest solution seems to be buying Claude Code subscriptions instead.

Running Quackjack For Everyone

team@c33tech.com (Mike Flynn) — Wed, 21 Jan 2026 07:18:31 +0000

A while back we started to open source YouTube video comment analyzer named Quackerjack, and lately we not only updated the project with a new design, we also set up an instance that anyone can use for free.

If you need to understand how your audience on a YouTube video is responding and what they are talking about, you can do that quickly with Quackerjack. You can also setup an instance of your own, but if you just wanted a quick report on a video without doing all of that you can now do that right now on quackerjack.c33tech.com

We’d also love your feedback on the project. Over the years, YouTube has added comment sentiment analysis, but Quakerjack still had a leg (wing?) up on other aspects of comment reporting, and we’d love to hear your thoughts on other things that YouTube hasn’t added but could be helpful!

Please feel free to give Quackerjack a try, let us hear that feedback, and I you’d like to setup your own instance or add custom features to it for your company please reach out!

Taking Ebooks to the Next Level with Booklore

team@c33tech.com (Mike Flynn) — Mon, 12 Jan 2026 05:31:59 +0000

In my home office, I have a shelf with a few books. Mostly technical books with a small shelf at the top of fiction, humor, and biographical books, but what I have a far larger collection of is ebooks. Until recently, they didn’t really live anywhere except for a folder on my hard drive, a Dropbox folder, or they were left un-downloaded in a service here and there. Storing them like that makes them virtually useless.

Booklore and few hours of setup and organizing things solved that problem for me.

Calibre

I actually thought I solved this problem a couple of years ago when I ran into Calibre. The intent is there with Calibre, but as you can see from one click to their website, the execution is not. Lots of features, but a terrible interface for their app and a missing web interface that is propped up by some other users creating a related project. I love their work, but Calibre just wasn’t working for me. It is just too fussy.

Booklore

Calibre was such a tease in that it was almost the solution I was looking for that I briefly thought about making my own book database application, but luckily I dug my feet just long enough to find Booklore.

Booklore is exactly what I was looking for:

Web Interface – Apps are great, but I want to be able to access my books anywhere.
Kobo Device Syncing – Calibre web has this too, and reading on the Kobo is way better than any web reading interface.
Easy Metadata Fetching – Ebooks usually don’t have much metadata, so your database is a mess without this.
Multiple Users

It’s a young open source project, so you do need a little know-how or have access to someone else’s know-how to get it going. I was able to set up a new Booklore instance via Docker on a server I have pretty quickly, and it’s been great. The only issue I had was that an update broke Kobo syncing, but I filed a ticket on their GitHub, and the issue was patched within a day.

I’d love a feature that converts a PDF to an ePub file, improving metadata discovery for comic books would be a huge help, and the ability to generate a link to a book I own for quick sharing would be wonderful additions, but just being able to make my ebook collection into something that is usable is already a huge improvement.

If you’re interested in Booklore, check out their site and GitHub page and if you need help setting up an instance, feel free to reach out!

Updating My AI Workflow

team@c33tech.com (Mike Flynn) — Sat, 27 Dec 2025 23:33:48 +0000

It’s been nearly six months since I detailed my AI workflow back in June, and since that is approximately 12 years in AI Years, I figured it would be a good time to publish what has changed for me.

Services, Tools, & Apps

Lots of changes here on this list.

The current list is:

Services

GitHub Copilot - $8.33/month
Gemini Plus - $20/month
OpenAI API Usage - $1/month

The big change here is moving from ChatGPT to Gemini. The reason is pretty simple, for the same $20 I feel like I get a better model (Gemini 3 Pro and Gemini 3 Flash are both great) more bang for my buck.

Copilot is still there and still a mainstay for autocomplete as is some OpenAI API usage. OpenAI still has the best APIs and because of that I’ve been slow to move off of them for personal and C33 projects.

Apps & Tools

Zed is still my code editor of choice and I’m still in love with it. Gemini doesn’t have a Mac app, in classic Google form, so I’m using the web interface for ad hoc needs. I’m also using the Gemini CLI interface increasingly both inside of Zed so I can use my subscription rather than pay for API usage, and as a way to quickly fire off a question on my Mac without having to open a web browser.

Uses & Benefits

Not a lot of changes here. I still lean on AI a lot of image generation and manipulation (which is another reason why I moved to Gemini), but I do use LLM, specifically Gemini, more for web search-type use cases since I trust the Google search index way more than I do the OpenAI index.

Coding use case is still very much the same. I view the AI as an assistant that reviews code I’ve written, auto-completes code as I go, and occasionally used to bounce and vet ideas with.

Active Experiments

I’m looking in to using the CLI more for new use cases like document review.
I’m also trying to use Google’s NotebookLM more going forward.
Deeper integration with my Home Assistant powered smart home. I already use OpenAI for my voice assistant, but do I move that to Google as well? What else can I do with it? Can I have it automate more?

A New Year's Redesign

team@c33tech.com (Mike Flynn) — Sat, 27 Dec 2025 18:27:42 +0000

We have a new look here at C33 Tech and I hope you like it.

Key features:

An increased focus on our consulting services and easy ways to reach out for a chat about how we can help.
Bringing projects to the front by featuring them at the top but also in a recent projects section.
A featured insight can highlight older posts that might relevant based on the latest news.
Adjusted dark mode color pallet to be a little more consistent with what people expect from a dark mode design.

We’re excited about these adjustments and hope to put more content and projects on the site in 2026!

(Oh, and whatever you do, don’t click the pi symbol at the bottom right of the page.)

Happy Holidays

team@c33tech.com (Mike Flynn) — Tue, 02 Dec 2025 01:12:03 +0000

Don't Be Like Heroku

team@c33tech.com (Mike Flynn) — Tue, 18 Nov 2025 17:26:16 +0000

Heroku, the cloud hosting platform that used to be kind of nice but after being purchased by Salesforce is a shadow of itself, sends me emails. I had an account for a previous consulting client, and even though they are moved off of Heroku my email is still on some mailing list. I try to be good about unsubscribing to things I don’t want to keep my inbox focused so I decided to finally kill this Heroku newsletter.

To start, there isn’t an unsubscribe link anywhere on the email, just a “manage preferences” link. Strike one.

Here are strikes two and three:

Don’t do this. I know there’s some marketing person in your ear telling you that you have to do these stupid tricks to keep your newsletter email list as long as possible, but what’s the point if they are only there because you tricked them?

Maybe tell the marketing person to write a better newsletter? Maybe send it less often?

Dark patterns like this aren’t helping you. They are making you look bad to your smartest customers.

Vibe Coding Regrets

team@c33tech.com (Mike Flynn) — Fri, 22 Aug 2025 05:20:07 +0000

“Makes me think vibe coding is just roleplay for guys who want to feel like hackers without doing the hard part.”

via AssafMalkiIL from Reddit

Link:https://www.reddit.com/r/vibecoding/comments/1mu6t8z/whats_the_point_of_vibe_coding_if_i_still_have_to/

AI Hype Comes Back Down to Earth

team@c33tech.com (Mike Flynn) — Fri, 22 Aug 2025 04:18:05 +0000

AI’s hype has been out of control for a long time now. Truly. It’s a shame too, because there’s interesting and useful functionality here. This isn’t the crypto boom which only birthed NFTs and accelerated ransomware, there are real benefits here, but none of this was even close to “AGI” (or even real “AI”) so there was a segment of the technology industry that sat and waited for days like the last few to come.

Gary Marcus:

Because expectations were through the roof, a huge number of people viewed GPT 5 as a major letdown. By the end of the night, OpenAI’s street cred had dramatically fallen. On the question of “which company [will have] the best AI model at the end of August”, a Polymarket poll charted OpenAI dropping from 75% to 14% in the space of an hour.

I’m still genuinely surprised they finally used the name GPT-5 on what was released. There’s good stuff in 5 but unless they were truly high on their own supply, they had to know this was coming.

Business Insider:

Dario Amodei, the CEO of the AI startup Anthropic, said on Monday that AI, and not software developers, could be writing all of the code in our software in a year.

“I think we will be there in three to six months, where AI is writing 90% of the code. And then, in 12 months, we may be in a world where AI is writing essentially all of the code,” Amodei said at a Council of Foreign Relations event on Monday.

It’s been 5 months. Anyone want to take this bet today?

Fortune:

Despite the rush to integrate powerful new models, about 5% of AI pilot programs achieve rapid revenue acceleration; the vast majority stall, delivering little to no measurable impact on P&L. The research—based on 150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public AI deployments—paints a clear divide between success stories and stalled projects.

I would love to see a Venn diagram of CEOs who went all in on AI without actually understanding it and CEOs that told everyone that back to office was essential for the “energy” after making big profits throughout the pandemic.

I can’t say this enough: There are real benefits to LLMs, both the frontier models (assuming they figure out the pricing model) and the local models, but a statistical model based on predicting the next token is not intelligence. It’s just not, and that was the hype a few years ago. Lately it’s risen to absolutely unhinged heights, and for what? To sell a $20 a month subscription? If a bubble pops, it won’t be the fault of the technology (Crypto) or the culture not being ready for it yet (Web 1.0), it will be because a group pseudo-technologists lost either their morals or their grip on reality en mass and took a segment of the press with them.

Cults are such a fad these days.

Hacking the STLToday Paywall v2025

team@c33tech.com (Mike Flynn) — Tue, 05 Aug 2025 02:38:59 +0000

Welcome back to our continuing series of breaking new STLToday paywall versions.

Want to hear something crazy? We’ve been doing this for over 10 years now! Or to put it another way, STLToday.com, the website for the St. Louis Post Dispatch, has been using a paywall for over a decade, and they haven’t been able to make one that isn’t cracked in a matter of minutes.

That being said, the latest version that launched in 2022, was the strongest yet. It took me a little longer to figure out and it required a Google Chrome Extension to defeat it.

Will the 2025 version, that I just noticed earlier today, build on that success and finally prevent me from reading their content for free?

Nope.

I found a way to get around it, and remove their aggressive ads in…I don’t know…let’s round up and call it 5 minutes.

No Google Chrome extension this time either. No bookmarklet, no code at all actually. I just turned off Javascript.

Here’s the bullet points

They are using some new libraries that are loaded from an external site in addition to scripts loaded locally.
I started to parse through them, but noticed that they are still loading the whole article and then deleting the article and drawing the paywall banner after the fact.
Like the 2022 version, they are doing this quickly. Maybe even faster than before.
…but since the send the whole articles, including the page styles, I thought that maybe the easiest thing to do was just prevent javascript, the code that runs on webpages in your browser, from running at all was a short cut.

It was!

Normally killing all javascript would cripple it, but to their credit, and the detriment of their paywall, that’s not the case here. The site is nearly fully functional without javascript, at least in terms of functionality for the reader. Ads don’t work, their user tracking isn’t working, and their paywall doesn’t work. Also images don’t load on their section pages. Not a huge deal. The images load just fine on the articles though.

Here’s how you do this, in Google Chrome (these instructions will work for other browsers, too just in different settings windows):

Go to the Privacy and Security settings and then the Javascript settings. Here’s a link directly to the Chrome settings page.
Leave the default behavior as it is so that all pages can use javascript.
Below that add a couple of domains to the list of pages that can’t use javascript:

https://stltoday.com
https://www.stltoday.com
https://bloximages.newyork1.vip.townnews.com

That’s it! Here’s a screenshot.

We did it again! Questions? High fives? Other various accolades? Get at me on Mastodon or via email

Go Cardinals (maybe next year)!

My Current AI Workflow

team@c33tech.com (Mike Flynn) — Mon, 09 Jun 2025 06:08:53 +0000

Sometimes, after rebutting someone’s extreme exuberance over AI or cracking a joke, I feel like people may assume that I don’t like AI or I never use it. That’s not the case at all! I do use it and I integrate it (or advocate for its use) in my CTO role at Ten2 and when consulting but I do take care to use it in situations that make sense, such as text generation, and not places where it doesn’t make sense, such as decision-making.

Because of all that, I thought it was long past time that I share my AI workflow. Specifically, what tools and services I use, how I use them, and what I think they do for me.

UPDATE: Check out the latest changes to this workflow here.

Services, Tools, & Apps

GitHub Copilot - $8.33/month
ChatGPT Plus - $20/month
API Usage - ~$1/month
Google Gemini - $0/month * Used through Ten2’s corporate Google Workspace or Google’s free tier personally.
Zed
ChatGPT App
OpenAI Codex
MindMac\
Ollama
Web Interfaces (ChatGPT, Gemini, etc.)
Custom Apps

Yes, I am very OpenAI-heavy. That’s not really a choice about the models themselves, as I think by and large they are all extremely close together in functionality and abilities, but more of a choice of how to consume it. I do use other models, Claude in particular, through GitHub Copilot, but the integrations OpenAI has created fit nicely with the way I want to use the models.

A couple of other things to call out:

Zed is my code editor of choice, and it’s great outside of any AI features, but it does now integrate some nice AI features. Primarily this is in the form of auto-complete/suggestions, which I have wired to Copilot, and “agentic” coding via an assistant window. I’m using API keys to power that, but I haven’t gone full “vibe coding” for any project yet. Zed also has an AI service that I could see myself moving to if Copilot ever pushed me away.
MindMac is a macOS application that has a solid idea behind it: Plug in your API keys and house all of your conversations in one place. It has been great and actually earned a place in my macOS dock…but…updates have slowed dramatically lately, which is not a good sign. I find myself using the ChatGPT app more since I started paying for their Plus service anyway.
Ollama is used to run models locally. I don’t do a ton of that now that things are a bit more mature, but it’s helpful when I want to experiment with something.
My API usage, outside of the occasional test or project, is for my home-brew JARVIS voice assistant in my home office powered by Home Assistant.

Uses & Benefits

I suck at Photoshop. I’ve learned to be barely passable with Pixelmater on my Mac. AI’s greatest gift to me personally is the ability for me to describe something and it can make generate something close enough that I can bring it to someone (or occasionally I can do it myself) and take it the last mile into something usable. The new image model from OpenAI is great and the $20/month is almost worth it for me to not have to wait as long for my image alone.

I don’t use AI for general search. I like it when a prompt can utilize live web results rather than just data from when it was trained a year ago, but I don’t want an LLM filtering results or making any kind of “decision” regarding web search results, and neither should you.

I do use it for search-ish low-risk tasks. For example: “How do I make a Paloma?”

For work, I will use it to flesh out a paragraph into something like a product spec for a task, and it’s amazing for generating test data, or for text processing. I don’t have a lot of these tasks, but I also want to spend as little time as possible on them when I do, so it’s a win there.

OK, let’s talk about coding.

I use multiple services and apps for different parts of the coding process, but it’s best to group all that usage under the idea that AI is my Jr. Developer assistant. This Jr. Dev Assistant:

Is good at some things, OK at others, and downright awful at a list of things. As the “manager," you need to know what those things are!
I don’t have to mentor or teach it anything. This saves a lot of time, but obviously would have major downsides if I was in the position where I needed to deal with staffing churn.
Like all technical managers, I am constantly working on letting myself delegate tasks to my Jr. Dev Assistant.

Currently, AI handles tasks for me like the following:

Figure out that I’m setting up this very long switch statement and type it out ahead of me. – Smarter auto-complete seems so old now, but it’s still as useful as anything AI does.
Go look around the repo and submit fixes to any bugs you see. – Mostly they are superficial bugs, but they are still bugs!
This gives me this error: . Can you find it in this file?
Make this “bottle” function that I can drop into the project myself. – Greenfield, new coding work for simple goals is pretty consistently good, so I will set up situations for “new” code when I can drop it into a more complex project.

The overarching summary is: It’s helpful, and I could definitely use it more, but I’m not worried about the livelihoods of developers.

I haven’t done much of anything in the way of full agentic “vibe coding,” mostly because I can code, but also because I haven’t needed any of those small, great for demos-type projects. If/when I do, I will give it a try, but for now I think that’s a lot of smoke, in that there’s definitely a fire, but it’s not as big as it seems from the other side of the hill.

Conclusion

AI is a powerful tool that can be used in a multitude of ways, from text generation and coding assistance to image creation. While it offers significant benefits, it’s crucial to use it judiciously and understand its limitations. It’s clear to me that the hype is outpacing what it can actually do, and that we’re already seeing the limitations of LLMs, so will I continue to call out and poke at that breathless hype monster? Absolutely. But that doesn’t mean there aren’t valid uses, and I will continue to use them, experiment to find others, and enjoy the benefits of this approach for a long time.

If you have any questions, comments, or AI-generated love notes, please send them my way!

LinkedIn is Going Great

team@c33tech.com (Mike Flynn) — Fri, 14 Mar 2025 20:28:50 +0000

🚨🔥AI will obviously solve everything because hype is never ever wrong. Do you want to know a thing that will help you?!

I have an AI thing that does stuff like, I don’t know, let’s say sales or marketing…whatever. It’s amazing though and it won’t be replaced in a matter of days by a new model. You need this! Look at this AI-generated image!!

Do you want to see the thing I made? Just follow me and comment “Why am I doing this? I’m just falling for LinkedIn reaction chum. Do I really think this thing will do something for me? Who is this guy? I have no idea, but here I am commenting a very specific comment just so he can send me a DM with a Google Docs link but then add me to 500+ newsletters and LinkedIn pages I don’t want.” below.

(If you’re interested, you can find this post on LinkedIn.)

Software Development is Dead, Long Live Developers!

team@c33tech.com (Mike Flynn) — Thu, 27 Feb 2025 20:52:48 +0000

“If you don’t learn to code, in ten years it will be like being illiterate!”

That was what someone exclaimed on a panel discussion I was on in 2013. It was a talk about bringing technology and entertainment together held in Beverly Hills and hosted by a coding bootcamp company. Two of the people on the panel were from a different bootcamp company, and then there was me, an actual technologist working in entertainment. Surely I would agree that everyone needs to learn to code, right?

Actually, I’ll disagree a bit with the other panelists on this. Do I think everyone should try coding at some point in their education? Yes. Similar to art, music, and writing, everyone, but especially children, should be exposed to coding to see if they like enjoy it or have a talent for it. But do I think that everyone should be coding all the time? No. Would I equate it to illiteracy? No. Why? Because coding is hard! Sure, you have to learn the programming language syntax, and learn how to run the code and all of the other mechanics of development, but what’s most difficult is the fact that you have to think through problems in a very structured logical way. Not everyone is wired to think like that, in the same way that not everyone can find the beat in a song, or can draw a cat in a way that other people would actually recognize as a cat.

The good news is, that you don’t need to learn to code to use technology in the same way that you don’t need to know how to take apart a carburetor to drive your car. Technology will come to you as it matures and society accepts it. If you have the interest or a goal to learn how to code or anything technical I strongly encourage you to do the work and hopefully become successful, but you don’t need to do any of that so you can use Facebook or make a spreadsheet in Excel, and that’s not going to change.

They didn’t invite me back for any more of their panels.

Here we are now in 2025, over ten years since I sat on that panel and not only do I not get the sense that anyone feels particularly illiterate if they don’t know how to scrape some Python together, the narrative has done a full 180! Don’t bother learning to code! AI will code all the things!

In a recent interview with Joe Rogan, Meta CEO Mark Zuckerberg said that AI will replace mid-level engineers by 2025. He believes AI can take over coding tasks, allowing human engineers to focus on higher-level problem-solving and creativity. Other tech giants like Google have also started integrating AI in coding processes. While initially costly, companies hope the transition to AI-generated code will become more efficient over time. This shift may reduce the demand for mid-level coding roles, pushing software engineers towards more strategic responsibilities.

Yet again I’m about to be the another wet blanket. Only the hype has changed.

Just as I did in 2013, I will start with some common ground. Yes, AI/LLMs are helpful for developers and absolutely can make a developer more efficient. At the moment, I am coding a lot as I am in a Start-Up CTO role, and I use GitHub’s Copilot and have experimented with some of the other models. The experience has been largely quite good. I absolutely see a return on the investment of the $10/month for Copilot plus a few extra dollars in API requests every month. I will also agree that people who can’t code have been able to use LLMs to generate things they wouldn’t have been able to otherwise. Seeing people dream up and create scripts for document automations, Excel macros, or Google Doc scripts is genuinely exciting and potentially a huge gain in productivity for those people. The problem is that none of the things I just mentioned are fully descriptive of what software developers do.

Developers understand and solve problems. The tech changes all the time, but the problem-solving doesn’t. People used to write assembly code, then higher-level languages came that were more approachable and allowed for more efficient development, but the result wasn’t fewer developers, it was more developers. Way more! Software ate the world, and the process of translating a problem into something that can be solved by new code was an increasingly critical skill. Writing a Python script to reformat a CSV, make a blog, or scrape content off the internet is a great demo, but most developers aren’t making little scripts; they are working on huge advanced systems that integrate into other huge advanced systems. Even if LLMs can someday create perfect code, the requirement to understand the problem enough to describe the solution you need remains a steep climb. Then there’s the matter of distributing the resulting code, which leads to potentially big questions about scale! It goes on and on and on…

Sexy AI demos from a guy sitting behind a MacBook and talking a little too slowly than you’d prefer make for a fun LinkedIn post filled with emoji, but it’s a far cry from Zuckerberg’s robot developer army he’s pitching to Rogan, and even OpenAI agrees.

In a new paper, the company’s researchers found that even frontier models, or the most advanced and boundary-pushing AI systems, “are still unable to solve the majority” of coding tasks.

The researchers used a newly-developed benchmark called SWE-Lancer, built on more than 1,400 software engineering tasks from the freelancer site Upwork. Using the benchmark, OpenAI put three large language models (LLMs) – its own o1 reasoning model and flagship GPT-4o, as well as Anthropic’s Claude 3.5 Sonnet – to the test.

Specifically, the new benchmark evaluated how well the LLMs performed with two types of tasks from Upwork: individual tasks, which involved resolving bugs and implementing fixes to them, or management tasks that saw the models trying to zoom out and make higher-level decisions. (The models weren’t allowed to access the internet, meaning they couldn’t just crib similar answers that’d been posted online.)

Though all three LLMs were often able to operate “far faster than a human would,” the paper notes, they also failed to grasp how widespread bugs were or to understand their context, “leading to solutions that are incorrect or insufficiently comprehensive.”

Yes, the models will continue to advance. Claude 3.7 in particular is making some noise as of this writing, but if you can focus your eyes and see through the hype, this will start to look familiar.

What a developer has done throughout a given day has been under constant change since the beginning. We’re not punching little holes in carefully ordered cards anymore. We’re not writing machine code by hand anymore, and soon maybe we won’t need to type quite as much as we do today, but none of that was ever really the job. The job of a developer is to understand how things work, diagnose problems, and use the best tools available to solve those problems. That’s not going away.

Software development as you knew it is dead. It is changing into something else, just like it always has. Long live software development!

The Availability Shortcut

team@c33tech.com (Mike Flynn) — Fri, 14 Feb 2025 18:19:25 +0600

This Shortcut allows you to quickly generate a list of your meeting availability on a given day that is either passed in to the Shortcut or selected from a calendar. The availability text is then added to your clipboard for it to be pasted in to an email, Slack, or whatever other document you are working with.

Example:

Aug 29, 2023:
9:00 AM - 10:00 AM
12:00 PM - 1:00 PM
1:00 PM - 2:00 PM
4:00 PM - 5:00 PM

Version: 1.1 (updated to fix issue with macOS 15)

Supported on: iOS, iPadOS, and macOS.

✚ Add "Generate Availability List" to Shortcuts

More info about Apple Shortcuts can be found on support.apple.com/guide/shortcuts/welcome/ios.

Apple Intelligence Summaries Are a Mess

team@c33tech.com (Mike Flynn) — Tue, 07 Jan 2025 22:25:00 +0000

Jason Snell over at Six Colors takes Apple to task over the current state of their Apple Intelligence notification summaries. He’s 100% right. They are bad, especially when summarizing news, and that’s unacceptable even with the “beta” tag.

Take a look at his included example:

A non-apology and the promise of a warning label isn’t enough. The company should either give all apps the option of opting out of AI summaries, or offer an opt-out to the developers of specific classes of apps (like news apps). Next, it should probably build separate pathways for notifications of related content (a bunch of emails or chat messages in a thread) versus unrelated content (BBC headlines, podcast episode descriptions) and change how the unrelated content is summarized. Perhaps a little further down the road, news notifications should be summarized based on the full text of the news article, rather than generating a secondhand machine summary of a story already summarized by a human headline writer.

We need to stop treating terrible AI summaries as your old angry uncle who is always raging about something that isn’t even happening (frankly we should stop placating him too). This is unacceptable and if you’re shipping, it’s on you as the developer.

Full stop.

Thoughts on Apple Bloggers and Vision Pro Coverage

team@c33tech.com (Mike Flynn) — Wed, 14 Feb 2024 17:24:04 +0000

I posted this on Mastodon, but I figured I would cross-post it here as well.

I wonder if we’re headed for a weird place with the #Apple Blogosphere (AKA the Apple Blog Mafia). All of the main players are talking about Vision Pro all the time and saying things like “I lose track of time in here!” and I wonder if they might lose track of their audience next. Not very many people have AVPs, and how much content do you want to see on a device you don’t have or didn’t want? It’s also entirely possible that I’m wrong and most of their audience are A1 Apple faithful, the type that get worked up over stock prices that they don’t own, and will keep dying for more AVP content regardless. It will be interesting to see how it plays out.

The more I think about this, the more I think my initial assumption is wrong for the bulk of the Apple blogger/podcaster audience and they like hearing about things they don’t own (see Mac Pro, etc), but I do think we’re a few weeks out for some people to tire of the constant coverage. This is especially true because the coverage is largely pretty weak with posts like “Here’s all the timer apps I’ve found that work on AVP” or YouTube video like “Is AVP good for sleep tracking?!”

Until Apple talks about a significant update, probably at WWDC, this very interesting technology is going to get really stale in terms of content. To be clear, that’s OK, but I hope that Apple Podcast/Blog mafia realize this or it’s going to be a long, boring spring for their audience.