Lucas Hrabovsky

I gave an updated version of my talk about the "A.I." hype wave a few weeks ago. I think of myself as a technology optimistic and enthusiast, yet many folks are asking me why I'm against "A.I." I don't think of myself in that way; I'm pointing out that the algorithms being referred to under the marketing banner of "A.I." have existed and been improving for decades, not months. History teaches us that these technology hype cycles only get in the way of figuring out the true utility of innovations to improve the human condition. Time is starting to prove my point. Take a look at the updated version of my talk from DjangoCon US. We've all learned a lot since then, and it has only been six months.

19

Have you ever wondered how video game Auto-Tiling in games works? Excalibur.js contributor Justin Young is on🔥with this latest post about authoring tile maps with aesthetically pleasing tiles! The algorithm is surprisingly grokkable and low effort to implement 🤯 https://lnkd.in/gsVHwc5V #gamedev #typescript #javascript

4

3 Comments

BDD vs. Traditional Frameworks: A Comedy 😂 Act 1: The Skeptics: "BDD? Gherkin? More work for no reason! I'd rather wrestle a kraken than write in plain English!" Act 2: The BDD Evangelist: "Hold on, my friends! BDD is like a well-organized toolbox, not a kraken's lair. Let me show you the light!" Act 3: Separation Anxiety (Solved): "With BDD, test cases and code are like a happily divorced couple – living separate lives but still getting the job done. No more Frankenstein test methods!" Act 4: Reusability Rhapsody: "Steps are like LEGOs – build once, use everywhere! Traditional frameworks? More like a jigsaw puzzle with missing pieces." Act 5: Hooks – The Heroes of the Hour: "Logging, cleanup? BDD's hooks swoop in like ninjas and handle it all behind the scenes. No more code clutter!" Act 6: Onboarding Nirvana: "Newbies, rejoice! BDD specs are your onboarding manual. Learn the features, copy the steps, become a testing superstar in no time!"

3

1 Comment

32-bit floats instead of SemVer! Obviously a horrible idea... or is it? https://lnkd.in/ghybmg4v (My searches turned up nothing like this but if it's already been invented, please let me know and I'll add a "prior art" section.)

10

8 Comments

If you've ever had to deal with inconsistent UX, multiple generations of design systems, and the perpetual back and forth between UI and backend developers - there's hope, you are not alone. And let me just say it out loud - there's no "full stack" development - it is simply impossible for an engineer to stay up to date on all the development across JavaScript, frameworks, Node.js, Kubernetes, Java, etc. You either have deep and somewhat narrow, or broad and shallow. These problems are quite common for everyone who has inherited an existing product (which is like 99.9% of us). This blog post by Kyle Shevlin about building a tightly-enforced composable design system resonated with me. #UI #designsystem #react https://lnkd.in/gStEWHtC

10

This looks like a great project from my friend Haig Douzdjian. I hope that once he finishes the overview he posts some more details about how he processes the implicit feedback (without giving away all the jewels, of course!). Human-in-the-loop feedback is a critical need for #LLM-based systems, but just like with traditional tabular #ML it's really hard to get enough feedback (aka labelled data). Building models on the side, to watch a conversation with an LLM and evaluate the success of each prompt interaction, seems like a great idea. I have to wonder how to cut through the noisiness of the #data well enough to make it valuable to the model. Hmmmmm!

5

1 Comment

Julia Wu and I recently led a “Speedy Routes Initiative” at Cedar to tackle timeout issues and improve site reliability. It involved improvements to monitoring, uWSGI and nginx configurations, Django views, and AWS infrastructure, all via a coordinated cross-team project. Get the full story in our latest Cedar Tech Decode Blog post below! #growwithcedar

49

2 Comments

Sup fam!! Relationships are everything!! They're the core of your engineering platform. Your team's design docs, software patterns, pull requests, and GitOps workflows all need to pass through a complex network of relationships before becoming meaningful to stakeholders. Before that, they're just bytes sitting on Jeff Bezos' disks. How are your relationships doing? Which relationships need some light touching up? Which ones need some serious investment? How about that one person on that one team? How about on _your_ team? Are you tracking these gaps with the same effectiveness you're tracking tech debt? It takes an intentional effort to build meaningful connections and to show peers you're a stakeholder in _their_ success. It's worth your time -- not to mention you'll feel great!! Want to optimize your engineering platform? Start by investing in your relationships!

38

It is CRAZY how much weaker OpenAI GPT-4o is at coding from the completions API than GPT-4. I'm not sure if this is the intended result, but very surprising nonetheless. The training recency is very evident, i.e. it knows about much more modern things, but the difference between its logical output is great with GPT-4 just thinking in a much more sophisticated fashion.

5

1 Comment

Last night I released my #RustLang crate steady_state="0.0.16" About it: https://lnkd.in/gq-w86k3 (give me a star) I hope to clean up docs and tests for a 1.0 release next month. Right now I am asking for feedback. #Makers #inventors #builders #programming 38

1

So far summarization has been the killer AI app for me. Others like code generation have been less impressive in my experience. But it’s really really good at distilling down text to a useful summary. A few places I’ve seen that include a slack bot that summarizes a slack thread, forums that summarize long topics, incident.io incident summarization, and asking ChatGPT to summarize or pull information from a document I upload to it.

27

3 Comments

Distributed tracing is a powerful observability tool, and yet, many organizations see little return for their investment. It's a familiar story: there's an initial period of enthusiasm, developers implement the tool, use it for a few weeks, and then stop using it. Meanwhile, costs keep creeping up. This excellent talk explains why that happens, and what you can do about it. Sampling can be as low as 0.01%! 🤯 https://lnkd.in/gJRdZPjY by Paige Cruz via USENIX Association SREcon24

13

4 Comments

Today, we’re publishing our biggest piece yet.   We partnered with Puzzle to write about the next generation of accounting tools.   Our CEOs, Alex Lee and Sasha Orloff, explain how AI can turn accounting from a repetitive chore into a major source of business intelligence.   They cover everything from decreasing costs to improving accuracy and enhancing customer support.   We also hear from Fouad Farhat of Soma Capital—an investor in both Truewind and Puzzle—about the current state of accounting, and where it’s headed.   📌Check out the full article here: https://lnkd.in/eUzvZjTm

18

An important part of our strategy coming to fruition.

8

AI is innovating a variety of industries across the US and Agriculture is no exception. 🐄 My new friend, Catie McVey, has created a new "Agtech" startup that uses patented 3D Modeling and Machine Learning algorithms to evaluate the health of cows by looking at specific characteristics such as hoof morphology and facial characteristics. Working with her to architect how to get source data from "The Stockyards" to "The Cloud" has been a lot of fun. I am constantly discovering new ways AI can make an impact in industries from legal, to healthcare and now agriculture! Read more about Catie's journey on RIoT - Internet of Things' blog #AI #software #agtech

9

My Recharge side project right now? Overhauling the automated E2E tests in our CI/CD pipelines. It's taught me a ton about our environment. Like a lot of fast-growing software companies, we built a ton of E2E stuff when we didn't have a whole lot of solid unit and integration testing. Over time, that's all gotten a little crufty, and it takes us longer than it should to execute individual deploys. Our objective is to get this down to a level where the whole suite can comfortably be executed by any engineer, and it can serve as a training ground for new engineers to learn the business. In doing this, I've discovered tons of inefficient code through long-running tests, features I didn't even know existed, and a whole lot of solid work that's been done over the years and flying under my radar. Totally recommend digging in here for anyone trying to learn more about the code they support.

14

Been thinking a bit more about AI and programming. In a previous post, I mentioned that I was worried about the "uncanny valley" effect. For those who don't know, the phrase "uncanny valley" was coined to refer to people's reactions to human-looking robots. We aren't bothered at all by robots that don't look like human beings (think R2D2), or robots that look a bit like human beings (think C3PO). We aren't bothered by robots that look exactly like human beings (think Blade Runner). We *are* bothered by something that looks a lot, but not quite, like a human being - it doesn't blink enough, it's skin is the wrong texture, and so on. It's that dip in the middle - where it's almost, but not quite, human - that's referred to as the uncanny valley. I'm sure someone has pointed this out before, but I think the same curve applies to the usefulness of AI. As an assistive technology, when you know it is being helpful, but can't take over for you, it's useful. Think autocomplete, or grammar checking. As something that can completely replace a human being, it's useful. You can argue that we're getting close to that with, say, image recognition (leaving aside problematic biases). There is a similar valley between these two, where human beings think that AI can take over for them, but it can't. (This effect probably already has another name, but I don't know what it is, so I'll live with "AI Uncanny Valley".) This AI Uncanny Valley can lead to funny results, as it does for the law clerks who have ChatGPT prepare legal briefs, only to have judges point out that they are citing case law that doesn't exist. It can also lead to tragic results, as it does when people trust their lives to Tesla's "Full Self Driving" mode and get killed. But the bad outcomes aren't always so obvious. One of the strategies that gets proposed for AI-driven software development is to have users describe the outcomes they want - feed the system a list of tests that have to be passed - and let the AI develop code that will make the outcomes happen. To the user, passing all of the tests would make it seem like the AI did what it was supposed to do. The problem is that human beings are really, really bad at coming up with tests. I watched a talk (see below) by researchers at NC state recently, who did a study of novice programmers. They found (to no one's surprise) that novices undertested their code in general, overtested the parts of the code that worked, and that their tests often mismatched requirements. They recommended a checklist of best practices for test developers, of the same kind that surgeons use before an operation to prevent infection. Imagine a world where this is what drives software development. Novices state requirements, badly. Code is generated that meets those requirements, but no one wrote it, and no one understands it. Code lives in that uncanny valley and no one ever notices. And we think we have software quality problems today... #AI #testing

5

2 Comments

OK so the reality of some conferences is pretty grim - very surface-level, vendor-driven content. I'm NOT a fan of 20-minute sessions... I can barely understand the problem being discussed in that amount of time. It just does not work for me. That's where UberConf makes a huge difference. 90 min in-depth sessions, on relevant topics by industry experts and practitioners. Conference size is limited to maintain a very comfortable environment without the need to walk 15 min to your next session, or it gets filled up. ooh... food is also great, fully catered for breakfast, lunch, and dinner (guaranteed great food)! https://uberconf.com/ I'm excited to join the amazing speaker line-up for a third year in a row. The opportunity to engage, share, and learn with the best (both speakers and attendees) is second to none. I've got a couple of new sessions planned, but I am most excited for: Crafting Consistent APIs at Scale: Embracing Linting and Reusable Models. Turns out that creating resources and building APIs in a distributed model is hard (who knew right?)! As teams expand and projects multiply, maintaining a unified API design can pose significant challenges. This presentation delves into the powerful combination of linting and reusable models as tools to navigate these challenges and ensure consistency across large-scale API ecosystems. Come check it out! Discount Code: uber24-travis4

22

5 Comments