To create a systemd service, create a file called /etc/systemd/system/my_service.service. This will create a service that will be started by the root user, but it can be run under another user.

There are three necessary components of a systemd service: unit, service and install.

[Unit]
Description=Service to run my fancy API
After=network.target
StartLimitIntervalSec=100
StartLimitBurst=5

[Service]

Restart=always
RestartSec=10
User=ubuntu
Group=www-data
WorkingDirectory=/home/ubuntu/app
Environment="PATH=/home/ubuntu/.venv/bin"
Environment="API_KEY=123456"
ExecStart=/home/ubuntu/app/.venv/bin/uvicorn main:app --my_params
StandardOutput=append:/var/log/app/stdout.log
StandardError=append:/var/log/app/std_error.log

[Install]
WantedBy=multi-user.target

First, we have to describe the service. Then we tell it when to start: after the network is up and running. We tell it to stop restarting if it crashes 5 times in 100 seconds.

Then we describe the actual service. We tell it to always restart (keeping the above mentioned limits), waiting 10 seconds between restart. We can specify a user, group, working directory and environment variables in which to run the program.

With ExecStart we can specify the program to run. If it's a simple FastAPI server, I usually write out the full path here. But, if it's something more complicated, where I might have to play around with the command more often, I create a bash file and start the service using that: ExecStart=bash /home/ubuntu/start_service.sh.

Then we redirect the standard output and the error stream to files. The last bit I have no idea what it is, but it's needed.

To start the service you have to first run sudo systemctl daemon-reload. You will have to run this every time you change the service definition file. Then you can start the service with sudo systemctl start my_service (the name of the file, without .service).

To enable starting it automatically on startup, you can run sudo systemctl enable my_service.

To see the status of the service: sudo systemctl status my_service. If it's crashing, you can look at journalctl -xe or you can check the log files you configured.

I recommend always writing out full paths, especially for the Python interpreter, to avoid any ambiguities as to what will run your program or what exactly will be loaded.

And now you know how to create a simple systemd service!

tl;dr: Deep fakes are already here, they are getting easier and easier to make and it's a cat and mouse game to detect them.

Source: ProTV

Whisper is a quite good automatic speech recognition model that is open source and can run on your own computers, provided you have a GPU.

I prefer reading to listening, so I wanted to transcribe my long list of things to listen to, ideally in automatic way.

The first step was to use Whisper to transcribe anything dropped into a folder. I recorded myself while doing this and I plan to do at least 2 more sessions of this.

Lesson learned: OBS studio uses a lot of resources and sometimes the recording has issues because of Whisper also hogging up resources.

It started in March: Make IT in Oradea, a startup incubator where I'm involved as a mentor, had its first in person pitching event in Timisoara. I was invited to be a part of the jury evaluating the submissions. I had never been to Timisoara before so it was I managed to scratch that itch of traveling to a new city. Unfortunately, I was not very impressed by the city, but at least I met some interesting people.

In April, Today Software Magazine had its first in person event in two years and I presented briefly about modern NLP methods. The presentation was an exercise of my improvisation skills, because the platform used to show the slides had technical difficulties while I was speaking, so I had to wing it.

And then June was busy. I gave a talk at The Developers, the first in person technical conference in Cluj. I talked about Large Language Models. I listened to some great talks and I talked to some people who are also in the consulting business and got some great advice (and some not so encouraging things, such as the guy who signed a contract with a client 7 years after the client heard him speak at a conference).

And then the fun started: flying abroad again. It's been almost two years since I flew and even that was a short 50 minute flight. And it's been 3 years since I have been abroad (Debrecen doesn't really count).

Through Oradea Tech Hub I was invited to mentor at the CodeX hackathon in Riga. It was a really well organized hackathon. The venue was great and it was great to feel the energy from the students (mostly). I definitely wouldn't pull an all-nighter for a hackathon anymore :))

I have to admit I didn't know much about Riga before and I was very pleasantly surprised. It's not a very big city, but it has a cool vibe and it's very clean. 10/10 would recommend, I want to go there again with Roda and Gloria.

It has many buildings in the style of Art Nouveau, like Oradea (where I live), so I finally found out what Art Nouveau is! It also has a statue of St. Roland, so that might have contributed to me liking the city so much.

The last trip in June was to Berlin, where I attended the datalift summit and I was a session chair. That was a really good conference! I met so many great people, all doing machine learning. I hope I can even collaborate with some of them sometime soon.

But the city itself... didn't impress me. I won't go again to Berlin anytime soon. The food was really good, with lots of very diverse options available, but the city itself... meh, too dirty. Sure, the city center is nice, but once you've seen one European capital, you've seen them all.

Now it's time for a bit of relaxing, so now I'm in Sibiu (where I have grandparents). It's Gloria's first longer trip away from home, so we're curious how she'll handle it! So far so good, so we're hoping for at least two more trips in the next months!

When we arrived at the Kalota Lovarda, we received a brief intro to how to ride a horse while it walks and an even shorter intro on how to ride it while it trots. Turns out riding a horse is not a passive activity, but you have to move your hips to nudge the horse to keep moving, while also lightly pulling on the reins. If you let the reins go, the horse will most likely take a break to eat some grass. Horses really love eating grass and will do that whenever they have the chance.

The first five minutes of the intro had me regretting that I signed up for activity. Do this for the rest of the day? Oh boy, oh boy. Please don't fall of the horse! Please don't throw me off, horsey!

It was an interesting experience where I learned a lot, such as the fact that horses have thick skin, so they don't care about going through bushes and thorns - but you do. Also, when going through a thicker forest, horses only make sure they don't hit trees - your knees are none of their concern, or worse, not even your head when going below some low branches.

The views were beautiful, as always, but honestly, I'm not sure if I'd rather do the trips on horses or while walking. There is a certain serenity that comes with the rhythm of horses step, which makes the scenery more enjoyable in a way. But it's so exhausting after 8 hours....

Also, horses have very different personalities and behaviors. Some horses have no problem pooping while walking, others, like my Titi, would always stop whenever he felt the need to relieve himself. Sometimes he would even go off the path, as a courtesy to the others. Some horses would avoid stepping into the mess others left, other horses had no problems marching straight through.

One of the funnier things on the journey was when the owner of the horses got off this horse for some reason and then his horse ran away. I think he spent 15 minutes trying to get back on him. It was quite funny, but it also made most of us reconsider getting off our horses.

When we got to the restaurant after 3 hours, everyone was relieved to finally get off the horse and give our bottoms a well deserved rest. But then it was time to go back, on a slightly different route. This time, even the horses were tired. The older ones would stop periodically for a quick break. But, after 2.5 hours, we did get back to the stables safely.

Unfortunately, I never managed to reliably get my horse to trot and more importantly, I never managed to learn how I should move in the saddle while trotting, to avoid making an omelette. Maybe next time? But it won't be anytime soon. I'd rather go to the aquapark.

For example, about 6 years ago I wrote a 2 line JavaScript bookmarklet for a recruiter, that reduced the number of clicks he had to do, from 10 clicks to 3, for each of several thousand candidates. It was a simple script that looked for an HTML tag, extracted a value and put it in the clipboard (from where he pasted it into a spreadsheet). The guy was so happy that he gave me a peer bonus, because it saved him hours of boring work.

Another example is when I have to replace something many times in a text document and I can quickly write a regex to do it. My dad sometimes calls me to do that for him when he is typesetting books.

This week I wanted to read a transcribed sermon that came in a .doc format. For more pleasant reading, I copy pasted it into a markdown file, quickly fixed some Markdown issues (mostly lists and headers), ran it through Pandoc and got a beautiful HTML, that was easy on the eyes. I sprinkled in some copy pasted JavaScript and I had automatic popups for Bible verses when hovering over references.

Julia Evans recently wrote about some tiny personal programs she wrote that both had marginal utility in her life and also brought her joy while writing them.

Doing this kind of programming doesn't require knowing a lot about programming, much less how to develop software professionally, but it does require you to have a knowledge of what is possible with code and either a low enough friction to get started (such as a text editor with a built-in regex search and replace), or a high enough motivation (such as wanting to avoid tens of hours of boring work) to do it.

But unfortunately, I don't know if this is something that can be learned quickly. All these examples I gave took me 5-10 minutes to implement, but that's because I already knew a fair bit of JavaScript, I've had to process Markdown in the past, I've written regexes many times for work and I've had a general awareness of Pandoc, even if I didn't use it much before. If I had to search Kagi for all of these (or if I didn't even know what to search for), it would take a lot longer.

In my opinion, this approach is much more useful than the no-code approach that is popular today (at least among startups) and it would be great to see more startups trying to reduce the friction for doing just a bit of coding, not eliminating coding at all. Grist seems like a pretty good approach in this direction.

What's your favorite quick "hack" where you used a bit of coding knowledge to bring quality of life improvements for yourself?

The machine learning consultant part has changed a lot as well. When I started off a year ago, I had many ideas of what I would do: trainings, online courses, consulting, train lots of ML models. Drawing the line at the end of the year has revealed that I did few of the things I thought I would do and that actually, the most profitable things were things I never thought I would do. I barely did any teaching, I did some consulting, but developing a full stack proof of concept for a startup was by far the best thing financially (and really fun too). Hmmmm.... maybe I should update the list of services I offer.

I started out as a machine learning consultant. But almost all of my work has been related to natural language processing (aka: working with text). Hmmmm.... maybe I should niche down to just NLP.

Some things didn't change. I still have a pathological fear of picking up the phone to call someone. Luckily, most of my clients came to me, instead of me having to go to them.

I thought I would have to do a lot of marketing, lots of blog posts, tweets, Youtube videos. I had one good blog post that brought in two clients. All the other clients came through referral and word of mouth. Hmmmm.... who knew networking is that important? And maybe, just maybe, my blog is not as important.

For a long time I was a lurker on forums, rarely posting anything. This year I've discovered several communities where for some reason, I started being more active. Hmmm... maybe I'm not that introverted.

The biggest change was becoming a daddy. I now have a very different reason to look forward to every day: Gloria. Of course, it comes with it's own challenges, including some curveballs, but boy, is it great.

There are some negative changes as well: it was the first year in a long time when I didn't fly at all :(

Here's to more changes in 2020 v2 and to being nimble, as God leads me!

That was quite fun and I got some very useful feedback from it, so I thought I'd share this video, as a way of learning in public.

First lesson: my green screen doesn't play nice with my IKEA chair that has a mesh in the back. My apologies for the awful looking webcam overlay.

And some development lessons:

• I go with the mouse several times to the menu to select "Format code". I should learn the shortcut for that, or even better, I should set PyCharm to auto format the file when saving.
• I spent a lot of time on figuring out the parameters for the first call and on formatting them as a dictionary. I could have sped up understanding how to send the parameters by URL decoding the provided example and for the formatting I could have used a multi cursor for quicker editing.
• BurpSuite/HTTP Toolkit was recommended as a way to explore HTTP APIs.
• When I was writing the client.register_payment call I had to hover a lot over the function definition to see the order of parameters. Copying the function definition would have been faster and it would have made it easier to define keyword arguments, which are clearer for a function with so many parameters.

Thas was quite fun and useful. Thank you Michael and Catalin for the feedback!

So since 2019, I haven't set big, public yearly goals, even though I still try to carve out some directions for myself every year.

But I am glad to report that despite focusing less on such goals, I've managed to create two habits that have been on my list for a long time, through very different means.

Working out

During the last 52 weeks (so 1 year basically), I have done 113 workouts, meaning just a bit more often than 2 times a week.

While I was working at Google, I managed to work out quite often, but there it was easier because I had to go down only 2 floors from my desk to get to the office. After moving to Romania, I would have had to go to a gym. I had one attempt to do that, but my car didn't start just when I wanted to go to the gym and it was raining so I couldn't go by bike, so I gave up. Then I bought barbells, dumbells and a squat rack, but, for 2 years, I used them at most 10 times.

But last year in October, after much internal mulling, I realized I need to change how I approach my body and my health. I need to start prioritizing it, and not just in a "New Year's Eve resolution" kind of way. My health has to be a priority and that means other things have to go. For this, I had to learn how to workout, what exercises to do to achieve the outcomes that I wanted. It also meant closing my work laptop at 6 PM and getting into my gym clothes. Other times, instead of enjoying a slow morning, I would lift some weights and get sweaty. To make more efficient use of time, I would sometimes work out during the mid-week online church service (I still ask myself if that's good or bad).

The ongoing pandemic helped a bit, because I didn't travel so much, at most a couple of days here and there. But even then I would try to get in a bodyweight workout.

The conclusion remains the same: I need to prioritize my health, keeping an eye on the longterm. I still have much work to do in this area, especially on the nutrition side, but even there I'm slowly learning better how my body works and how I need to fuel it.

Memorizing Bible verses

This goal was on my list every year from 2016 to 2018 and every year I would utterly fail it. I would learn 1-2 verses and that's it.

But in 2019, someone from Brazil visited our church and told us how some young people from his church each memorized a book of the New Testament. And I decided, I'll do the same.

What had changed was that I had learned more about the science of memorization. I knew of some people who had memorized vast amounts of information, so my first step was to inform myself what would be the best way to do that. I tried putting that into practice and I've never stopped. For two years I have been using Anki to memorize Bible verses. I have memorized all of Galatians and I'm halfway through Ephesians.

While I am convinced that this habit is very beneficial for my spiritual life, I think the key here was breaking the requirement into small pieces that can be done easily. There are very few days when this takes more than 10 minutes. It's something that I can easily do in bed, right before I go to sleep. Or I can do while waiting in a queue at a shop.

The other thing that helps is to see how good my memory can become with enough practice. There are verses that Anki estimates I'll have to revisit only in 3 years. I notice that my short term memory is better. Memorizing more things is so useful, that I have started memorizing other things as well, from my wife's phone number to PowerShell commands to my business's registration number.

More work to be done

Roland is still a work in progress. God still has much to do in my life. But I want to celebrate every win, even as I look forward to what I'll learn in the coming months.

There have been plenty of moments where the predominant feeling was "what now?". How am I going to find more clients? How to negotiate with this client? The Dip, as it's called by Seth Godin, is very real and very scary. When you draw the line and see how much you've earned over six months... you start getting serious doubts. Was it worth it? Wouldn't it have been better (and much easier) to just find a nice job?

But there are other moments: when I realize I have freedom to choose my clients and the projects that I work on; after working for a whole day on something that I love, ML, without any useless meetings; when deciding with almost complete freedom the tech stack which will be used to build the ML side of things; when I take a day off almost whenever I want, just because I don't feel like working on that particular project on that particular day. Or when I realize that I am a consultant, that my clients look to me for advice and that they actually take my advice seriously. If I say that the way they did things previously won't work and they should do things differently? They'll get to it right away.

And then there are moments when I realize I barely have time to read any state of the art machine learning papers and instead I have to learn the basics of marketing, branding, business development, communication, coaching, explaining, teaching - and to put all of this into practice. Most of my clients don't care if I'm using the latest state of the art Transformer architecture (and don't even know what on earth that is). They don't even know what machine learning is. But they need someone to explain it to them - to people who have built successful companies in their own fields - and to help them understand if it's something that they need or not.

I am thankful to God for guiding me on this new path, of which I have dreamed for a long time. Faith in his faithfulness is what has kept me steady when my knees wavered.

I am grateful to my dear wife who was willing to take this risk alongside me and has been very supportive all along the way.

I am very glad I have a good accountant who can help me with all the paperwork of the company.

I am grateful to the whole team from Oradea Tech Hub, who have helped me get my name out there, and especially to my friend David Achim with whom I did many rounds of business strategy discussions.

And I am thankful to many others who have cheered me on, who have encouraged me and who have put in a good word for me to potential clients.

I kinda split my blog into two: personal posts stayed here, anything related to machine learning goes to my new domain, which is for my consulting business. I still want to post some technical content here and I do hope I'll make it to the front page of HN again :D

I haven't had as much time to write posts because I've been busy with all kinds of other content: an in person machine learning course here in Oradea, several presentations, some about machine learning, some about quick iteration, some locally, some online. It turns I only have so much creative juice in me every day.

I've resumed my goals to blog again, but at a much more humble rate. Sometimes I'm tempted to try daily blogging, but I'm a bit afraid of that commitment and of the quality of the posts that would result from that. Some people say that writing daily turns on the faucets of creativity and you'll have plenty of ideas. But for now I'll stick to a more reasonable goal of two posts per month.

But now, I have to juggle between working for different clients, talking to leads and doing marketing or administrative tasks. My to-do list just keeps growing longer and it's getting harder to pick something to work on. Should I write a new blog post? Should I work on a video? Should I do some exploratory data analysis for a client? Should I look into preparing an MLOps report for a client? Or maybe write a blog post so that my friends know I'm still alive?

Having to make a choice about this every time I want to start working is tiring, leading to choice paralysis. Often I have to work on 3-4 tasks a day. If I context switch between them too often, my efficiency drops.  So last month I started applying a variant of time blocking, about which I read from Cal Newport.

Instead of using a paper based method like he suggests, I create an event in Google Calendar when I want to block off some time. Ideally I schedule them the day before, but sometimes I either forget or something comes up and I have to change what I'll work on for the same day. I try to create blocks of one or two hours. Shorter blocks don't give you enough time to get immersed in deep work, while longer blocks are usually too tiring. I also make sure to leave some breaks between the time blocks.

I use a separate calendar so that I can easily toggle the visibility, leaving in the Calendar app only those events which have to take place at a given time (such as client meetings) and so that the time blocks don't interfere with Calendly, a meeting scheduling service I use.

I'm not very strict about the time blocks. If I find that I'm in the flow when a block ends, then I'll continue working on it. If something else is more urgent or I'm simply in a very strong mood for another task, I'll work on that and I'll simply move the calendar event to another time.

How do you organize your time and decide what to work on?

De ceva timp mă bate gândul să trec la următorul nivel de creare de conținut. Scriu pe blog de 10 ani și îmi place asta. Unele posturi pe care le-am scris despre programare și machine learning au avut succes. Așa că m-am gândit să fac un curs de machine learning.

Pe net sunt o mulțime de resurse de machine learning, cursuri care mai de care. Și eu am învățat din ele, deci sunt și cursuri bune și foarte bune printre ele. Dar pentru început, aș vrea să încep prin a face un curs în limba română, unde nu cred că sunt suficiente resurse de calitate. Bine, practic va fi o romgleză, că abia pot să pronunț „învățare automată”, „machine learning” alunecă mult mai bine. Ce să mai zic de deep learning...

O altă lacună pe care am identificat-o e că majoritatea cursurilor sunt pentru programatori care scriu cod în fiecare zi și vor să știe folosi și unealta numită machine learning. Dar este o lipsă mare de înțelegere a modului cum funcționează machine learning și inteligența artificială în rândul managerilor și, de ce nu, a oamenilor non tehnici. Dacă te iei doar după ce citești la știri, imediat urmează scenariul Terminator, când în realitate toate sistemele de ML au slăbiciuni mari și ușor de găsit.

Asta duce la unele așteptări nerealiste din partea conducerii unor firme, care vor să devină mai „hipsteri” și să folosească ML, dar vin cu idei complet greșite, care nu pot fi făcute să meargă suficient de bine. Sper să pot să ajut și astfel de persoane.

Mulți oameni cred că trebuie cunoștiințe tehnice foarte avansate ca să folosești chestii de inteligență artificială. Dar bariera scade tot mai mult și apar aplicații și în domenii creative, cum ar fi generare de imagini sau de text și care pot fi folosite relativ simplu, odată ce înțelegi conceptele de bază.

Dacă vă surâde ce ați citit mai sus, intrați pe pagina cursului.

Sometimes, if you don’t know about a pattern and you read code that uses it, it might seem strange. Why is this extra layer of abstraction here? Why is this API broken down into these pieces? After learning about the pattern, you might learn that the extra layer of abstraction is needed because the layer that’s below changes often. Or that the API is broken into those specific pieces because this makes it easy to cover more use cases in an efficient way.

As I’ve started diving head first into the world of running my own consulting business, I’m starting to learn about a whole other world of “design patterns”, unrelated to programming. And suddenly many things that I’ve seen before started to make sense.

My friend David has been bugging me to start a community for people passionate about machine learning in Oradea, where I live, for almost two years. For a long time I was thinking, why does he push so much for this? Well, after taking Seth Godin’s Freelancer Workshop, now I know that being the person who organizes a community is one of the best ways to make yourself known.

Another example is that I saw website offering a sort of business networking thing for a very high membership cost (or at least it seemed expensive at the time). Why would anyone do that? Then I learned about a thing called alchemy network 1 and how if it’s done well it can bring great value to it’s members.

All my friends who are freelancers charge by the hour. That’s what I thought was normal. But then I heard about Value based pricing by Jonathan Stark. A different pricing “design pattern”, which aligns the incentives of the client and of the service provider in a much better way. Let’s see if I can pull it off though.

Just like in programming, design patterns help us find the correct solution faster and communicate more efficiently. The more patterns you know, the faster you can recognize a situation and react better to it.

What are your favorite design patterns?

Visits were up: 60.000 sessions compared to 10.000 in 2019. Most of my sessions were from unique visitors, because those were around 53.000, compared to 8700. Pageviews are at 73500, versus 16300.

Most of this is due to some posts that got very popular. The Moving away from Gmail post is now my most popular blog post ever, dethroning the neural network post that is 7 years old and is still getting 2000 views per year. It was on the front page of HackerNews and it got 36000 pageviews in 3 days. The Obsidian post was also quite popular, having been suggested in the Google app, getting to 8000 views. My Rust posts all got over 800 views, with the web crawler one getting over 2400. Surprisingly, how to bridge networks with a Synology NAS is a very interesting topic, because that also got 1000 views.

The Ghost platform has worked ok during the last year, but it has some small friction points, so I'm thinking about changing again. But regardless of how I'll post, I definitely plan to keep post more content.