Engineering Manager at
@Spotify
. I write about tech, coding, and investing. Born in Brasil 🇧🇷, living in Stockholm, Sweden🇸🇪. All opinions are my own.
Inspired by the Generative Agents paper I decided to create a small sandbox game to test out a few LLM models over the weekend. Running LLAMA2-7B on this one.
Augusto e eu exploramos como ele faria a arquitetura de um serviço de streaming com limite de usuários simultâneos, ficou massa!
Curiosidade: fazemos esses vídeos sem ele ter qualquer contexto prévio, então ele realmente está resolvendo esses problemas ao vivo.
🇧🇷🇵🇹 Prazer em participar dessa série do
@RealGalego
!
Simulamos uma entrevista completa de algoritmos, vale a pena assistir não só para pegar dicas de como resolver algoritmos mas também como entrevistar e ser entrevistado.
44 / 100 - Tópicos de entrevistas Senior Engineer
Eu e o
@ltmenezes
fizemos uma 'mock interview', pra mostrar mais ou menos como funciona na prática uma entrevista de algoritmos real
Entrevista com Engineering Manager
This week Spotify turned 18 years old!
It made me look back on how 7 years ago I was in Brasil making my first two open-source projects on top of the Spotify public API: a terminal client and a playlist generator. Time flies, such an awesome team to be working with!
Came across this amazing example of Product Management today.
The need for VPC connectors for serverless products in the GCP has been a long ongoing issue. Instead of shying away from the issue, the GCP Cloud Run PM acknowledges it in public, explains their reasoning and
🇧🇷🇵🇹 Prazer em conhecer todo mundo que me adicionou ontem depois do post do
@Alekson
!
Muito massa conhecer mais brasileiros trabalhando em tech, de qual estado vocês são no Brasil?
I got my first job as a programmer because I wanted to meet other people who also liked coding.
Now I get to work on one of my favorite products and talk to people building awesome stuff every day, what a privilege.
Valeu todo mundo pelo apoio ao video de ontem! Foi muito massa ver como ele foi útil para vocês e a repercussão que ele teve!
Como muitos pediram sobre recursos para estudar mais sobre System Design queria compartilhar um dos melhores pontos de partida: o system-design-primer
Launch day🚀
I've been working on a personal project with
@souza_gusthavo
.
It all started when we wanted to find "the best steak house in town that serves Guinness" and got frustrated that search engines didn't understand this type of query, so we built
This is OpenAI's GPT-4o watching a live TFT game and providing real-time suggestions. I coded this in 10min so you can expect a lot more sophisticated tools to come.
How do you personally feel about it? Do you think this should be allowed in games?
[turn the sound on]
Last week I had the privilege of speaking about how we are innovating on advisor tooling at Spotify at Liveperson's Spark Conference at the Computer History Museum in Silicon Valley, US.
It was great to share some of the amazing work our team is doing internally in this space.
What a sad day for Brazil.
We’re slowly marching towards a more regionalised internet, where governmental reach is more important than individual rights online.
Thanks to everyone from Brazil who connected with me, us expats will continue to voice our opinions on here.
Tip to new engineers: Contrary to what social media makes you believe you don't need an expensive setup to code, most engineers don't have one.
It's just that there's a high correlation between engineers who over-invest in their setup and engineers who post photos about it.
Tive a oportunidade de conversar com o
@matos_eduardo
sobre o artigo 'The problems no one wants to tackle' que escrevi no início desse ano.
Foi massa conversar com ele sobre crescimento de carreira em tecnologia, link para o artigo e o podcast nos comentários.
Short review of my experience coding in VR.
A quick disclaimer about my setup, I'm using a quest 3 and ImmersedVR, I've tried other apps for streaming my desktop and for coding this seems to be the best one due to the passthrough and support for multiple screens.
Starting with
Let’s say AI does kill coding, instead of coding we will:
- talk to domain experts to gather requirements
- write all the requirements in English
- make lengthy explanations of what the computer/model needs to do to make sure it understands it correctly
- double-check that it’s
Jensen Huang, CEO of Nvidia, argues that we should stop saying kids should learn to code.
He argues the rise of AI means we can replace programming languages with human language prompts thus enabling everyone to be a programmer.
AI will kill coding.
I tried using Github's Copilot for a few weeks and got zero value out of it.
Frequently it tried to infer the next few lines of code, however, I usually already know exactly what I want to type for the next 2-3 lines, so reviewing its code was slower than just typing.
@kentcdodds
Libraries that have direct access to the user database should have detailed documentation on which permissions they need, I can't fathom giving a library unrestricted access to a production DB.
Finally took some time to create a gist for this, this is a very short integration that should work with any game by doing minor modifications, feel free to share/copy it.
Let me know if you create a version for a different game using it, would love to see it!
This is OpenAI's GPT-4o watching a live TFT game and providing real-time suggestions. I coded this in 10min so you can expect a lot more sophisticated tools to come.
How do you personally feel about it? Do you think this should be allowed in games?
[turn the sound on]
passed 500 users earlier this week!
Thanks to everyone using it and sending us feedback, we're having a lot of fun building it! Try it out if you haven't.
@GergelyOrosz
@cpaik
@theinformation
Agree especially with your latter point, I feel like a lot of this discussion is based on the assumption of exponential improvement of LLM coding tools as opposed to current capabilities.
It’s like saying that writing will die because grammarly is getting better.
This was a rough week for everyone at Spotify, but the bright side was seeing how many ex-colleagues reached out to check on me.
I'm super grateful for having worked with so many great people during my career.
Fun fact: I've come across developers who believe this is true and wrote scripts to push automated random commits every day.
Believe me, companies do not count how many green squares you have on your GitHub.
You're job hunting in 2023 with this kind of Github profile ?
Most newbies don't even understand how much work it'll take to actually succeed at this tech thin.
I started noticing how much I avoid coding in small blocks of time in favor of setting it aside for bigger ones.
Getting in the flow is way too hard and precious of a state to waste it in short sessions. I'd infinitely rather code for 10h straight than do 5 2h sessions.
We went from discussions about the ethics of having multiple jobs to no one being able to find a job in less than two years.
What a wild period to be in tech.
Two years ago we had the Great Resignation and now we have the Great Layoff in tech. The common theme is lack of loyalty.
Even during the Great Resignation, companies spent more on new hires than they did on retaining existing ones which incentivized job hopping.
I was invited by
@Alekson
to talk about working and living in Sweden as an expat, I definitely recommend his podcast if you think about working abroad.
Thanks for the chat!
Mais um episodio do Tech & Accent no ar e uma otima oportunidade pra voce praticar o Ingles e conhecer a historia de brasileiros morando em outros lugares do mundo.
Dessa vez com um cara que eu acho foda demais, ele eh Engineering Manager numa big tech na Suecia, e nesse
Fun history fact, Visual Studio for Mac never really existed, it was a rebranded Xamarim Studio that Microsoft bought in 2016.
As a C# developer I think this was a good decision, they both had diverging codebases.
Not even one week since my post and we now have the code llama release where all the showcased use cases were entire functions/files generation.
Awesome to see how fast this space is evolving.
I tried using Github's Copilot for a few weeks and got zero value out of it.
Frequently it tried to infer the next few lines of code, however, I usually already know exactly what I want to type for the next 2-3 lines, so reviewing its code was slower than just typing.
I completely bombed my first coding interview.
I had no idea of what to expect or how to prepare myself, I got every single question wrong. After that, I joined the competitive coding team in my uni to study and I still code them regularly.
Failing is part of the process.
I’m mentoring a few college seniors and every single one of them has reached out to me, freaked out about them “bombing a coding interview”. So I wanted to publicly say:
1- every engineer has bombed a coding interview.
2- you probably didn’t do as horribly as you think you did
🇧🇷🇵🇹Fala galera! Frequentemente recebo pedidos de recomendação de cursos e bootcamps para entrar na área e recentemente percebi que muitos bootcamps estão anunciando taxas de empregabilidade desatualizadas ou utilizando dados de períodos específicos que favorecem a empresa e
After this tweet I went on a long coding session that resulted in several new features for a new project that I'm working on and a PR to the open-source AI Framework that I'm using.
Short coding sessions kill productivity.
Btw, shout out to
@Haystack_AI
for the quick review
I started noticing how much I avoid coding in small blocks of time in favor of setting it aside for bigger ones.
Getting in the flow is way too hard and precious of a state to waste it in short sessions. I'd infinitely rather code for 10h straight than do 5 2h sessions.
'gpt-3.5-turbo-instruct' is 4-5 times faster than the regular 'gpt-3.5-turbo' in some tests I'm currently performing.
I recommend giving it a go for everyone who can move from ChatCompletions to Completions.
@danielhe4rt
Esse assunto é muito bacana, algumas formas: redução da taxa de juros (de forma duradoura e com planejamento), aumento da segurança jurídica e, principalmente, aumento da renda domiciliar em dólar.
Working with a perfect setup doesn't always translate to a good learning environment.
The project where I learned the most about databases had a self-hosted DB that ran out of storage every few months, we had to optimize execution plans and monitor indexes sizes to optimize them.
Just finished a new article: The Rise of Flash Startups.
This one has been in the making for a while, happy about how it turned out.
Giving Medium another go this time.
In 2011, we noticed one Spotify user answering hundreds of questions in one of our first community forums. That inspired us to launch Spotify Stars in 2013, a program designed to empower super users to help other users find the answers to their questions, along with Spotify team
"A manager's job is to serve" is great advice for managers but a bad perspective to have about your own manager.
The best manager/direct report relationships I've seen work as a two-way street of learning, feedback, and collaboration.
@RealGalego
LGTM, pode rolar uma versão que salva o “outcome_hasher” na memória para não precisar instanciar ele toda vez? Um hashmap de “outcome_hasher”s.
got featured on the
@TheRundownAI
this week!🚀
Awesome to see all the new signups, thanks everyone for registering!
@souza_gusthavo
and I are shipping new features every day, try it out if you haven't already!
A new podcast episode from
@devsfronteiras
just dropped, this time hosting my friend
@thiago_dmont
.
Make sure to check it out if you speak Portuguese and are interested in moving to Sweden, lots of great insights.
Nessa aventura o
@fabriciocarraro
falou com o
@thiago_dmont
, mineiro com passagens por algumas startups brasileiras que, após como remoto para uma empresa nos Estados Unidos, recebeu uma proposta para trocar de emprego se mudar para a Suécia. 🇸🇪
Today we are rolling out audiobooks for Spotify Premium users in the US!
Glad to finally be able to share this amazing product we have been working on.
@Carnage4Life
The tricky thing about finance is that most things in life reward you for learning and putting more effort into it, but not finance.
The simplicity of the answer tricks people into thinking there’s more to uncover.
Auto-completing the current line of code or the next 2-3 lines is going to break experienced programmers' flow state more frequently than it's going to help them. Instead coding copilots should focus on inferring bigger chunks of code, for example, entire functions or files.
With the current growth rate of compute and the optimization investments from companies and open-source orgs, interpreted langs runtime speed is soon going to be a non-issue when compared to compiled ones
It's incredible how much more organic traffic generates for me compared to medium.
For context, I published the same article on both platforms and I only shared the link to the medium one on my social media, still won.
Just finished a new article: The Rise of Flash Startups.
This one has been in the making for a while, happy about how it turned out.
Giving Medium another go this time.
I still think coding copilots are a great value proposition, we just need to be honest about their current state and provide feedback to these companies.
This is a great glimpse into the complexities of being a manager/lead, each generation views work differently and requires different types of support.
You must truly invest time to understand different perspectives, especially if you are working with a international team.
What is the latest generation of software engineers - GenZ - like at the workplace?
We go into this underexplored topic, with help from observations from more than 60 "older" tech professionals. Read the findings here:
🇧🇷🇵🇹 Prazer a todos que me seguiram hoje por recomendação do
@Alekson
! Escrevo a maioria dos meus tweets em inglês mas podem ficar a vontade de me marcar ou mandar DM se quiserem conversar ou perguntar qualquer coisa
@levelsio
People overestimate how important it is to have the best product. Sometimes having the 10th best product in the market is already enough for you to harvest a good chunk of profit as long as it’s a big enough market
Do you remember how MySpace was the gateway for kids to learn how to code?
WebXR has the potential to be the next one, imagine being able to code how your room looks like and sharing that with friends.
Starting the year with one of my favorite quotes on programming:
"You have to believe that software design is a craft worth all the intelligence, creativity, and passion you can muster.
You need to care. You need to play. You need to be willing to explore.".
@brunocroh
Inline code completion é um péssimo meio para LLMs.
Foi intuitivo fazer ele como a primeira ferramenta de copilot devido como esses modelos funcionam, mas ter que escrever metade de uma linha de código e ter que revisar a outra metade é uma má ideia, ela tira você do flow state.
On a side note, this is already a non-issue for most web dev today. Most enterprise applications on the web are IO-bound where code execution speed has little to no impact.
AI automation is better for individuals than it is for established companies.
Some have already started to notice this, but the main discourse still focuses on how negatively AI can impact the job market.
A thread about a more optimistic perspective on the future of work 🧵
@GergelyOrosz
@cpaik
@theinformation
You nailed it with "distraction free", they are useful but they do feel like a distraction sometimes.
Hopefully once people start using it more we can have more discussions on how they can be used instead of these extrapolations of what they could be in the future.
Manager comp in tech is a slippery slope and most companies struggle to get it right.
If they make more than ICs, then you drive people that shouldn’t be managers into the field. If ICs make more, then you drive all the technically capable people out of managing.
An interesting detail on manager compensation versus individual contributor compensation. Who makes more, almost always?
I had assumed (correctly) that managers do. At most companies.
But the closer you get to the top of the market (to the “top tiers”) the less true this is!
It's important to put this in perspective while discussing labor displacement because of AI, we are still a few orders of magnitude away from a reasonable context window size for complex tasks and that's only about the context window, not even the reasoning itself.
Just released a hack project I've worked on for the past few days.
Fences is an LLM-augmented OpenAPI spec interpreter, it automatically categorizes endpoints, generates request bodies, and learns based on past req/responses.
Try it out on 'pip install fences-api; fences'
@Carnage4Life
In my opinion neither, we are still seeing reflections from last year’s environment.
These companies probably made these decisions last year and postponed the execution for after the holiday season.
Python's GIL removal is going to be huge in the backend space, not only it's going be easier to write parallel code but it will also make the existing backend libraries more reliable.
(1/8)
@TheAnkurTyagi
Absolutely, similarly to the Feynman learning technique writing forces you to deeply understand a subject. It’s worthwhile to write even if you have zero readers.
@elliot_dylan_
@RealGalego
Valeu cara!
- Sim, ele vai ser i/o bound.
- Nao existe nenhuma solucao 100% segura para esse problema, isso que torna ele interessante. O mac id, assim como **qualquer** informacao do dispositivo do cliente, pode ser alterada/forjada por ele.