What I Learned At Code BEAM America

This week the BEAM community proved to me that there is remarkable talent and energy leading innovation with Elixir, Erlang, and other BEAM based tools. Not only was every presentation excellent, I was also consistently impressed with the technical depth of conversations that I had and the general level of expertise of all attendees. I loved how many presentations used live prototypes to show off their ideas, it made the energy of the conference very electric and the benefits of this technology concrete for all to see.

Main Takeaways

Working demos are powerful

A real MVP that solves the hard part of a problem well is more valuable than any document, blog post, idea, or conversation. The most exceptional talks all leveraged live demos or some form of working code, and they undeniably presented the benefits of whatever idea they were talking about.

Many talks also mentioned how working demos have changed the direction and culture of a company because it is so easy for people to make progress after something is working.

I started to realize that many MVPs fail when they solve the easy part of a problem and assume the hard part can be solved later. The most successful demos were very basic but all energy had been focused on the real challenging technical feat.

Stand on the shoulders of friendly giants

Many teams struggle to maintain systems over time that depend on too many packages. Other teams waste valuable time rebuilding complicated subsystems and end up with second class versions of popular open source libraries.

I learned that the most successful teams over the long term carefully select a few key dependencies that they are prepared to partner with over the years. This results in an optimal velocity that limits tech debt while maximizing leverage of existing technology.

Other notes

Every talk I attended was incredible and I encourage you to go check out what sounds interesting from the program when they are made available online. Here are the notes I felt compelled to write down while I was listening:

Designing LLM Native Systems

Sean Majiarity

• LLMs can be integrated best when using structured outputs, this limits the exposure to failure.
• Think socially: do not let an LLM break someones trust in your business.
• LLMs negatively shift the burden of "systems thinking" onto the reviewer, especially in code review situations. I liked this concept of identifying the important work of a developer as systems thinking.
• Engineers should share LLM chat history as context for code changes.
• Do not offload your strengths onto an LLM, offload the unimportant work.

Investigating Production Issues

Jenny Bramble and Adrian Dunston

• They mentioned the book Accelerate for CD which looked good.
• Development is expansive while debugging is restrictive.

Phoenix sync with ElectricSQL

James Arthur

• Honestly need to build some stuff with Phoenix.Sync. It solves a lot of problems really well.
• He also mentioned linear project management which looks like a cool Jira alternative.

Fault-Tolerant Machine Learning Operations

Chelsea Troy

• Model performance measurement gold standard is user feedback.
• Make sure production data is always similar in distribution to training data over time.
• Ask specific and evocative questions for better feedback.
• Telemetry interpretation can be a guessing game.
• Make it as low effort as possible to run experiments and understand your impact.
• Classic supervised models can find highly important features which you can break down into use cases.

AI for Worker Collective Action

Saiph Savage

• An important binary classification problem is "isolated vs systemic" for issues.
• Surveillance systems are depressing, combat with empowering users to observe the platform.

The Socio-Technical Elements That Make Good Platforms

Charity Majors and Fred Hebert

• An RFC should be easy to write because you have already proven the idea.
• Distribute system understanding and have team members present their understanding to reach consensus.

Other random notes

• Look at Boundaries lib.
• Break Ecto code into structure/mutation/query
• Look at Machete and Exunitspan.
• Read The Art of Creativity