Bram Geron

@bram.xyz

275 Followers 331 Following

Rust, product. Nushell&jujutsu fanboi. Live in đŸ» Berlin, building an API for APIs at @directus.io . Opinions my own. Used to do research on programming languages. Genocide is wrong. @bgeron on other networks.

A cursed Rust item
Lefos emailing me: Hello Bram, Happy Monday! A fresh week is yours — may it be full of energy, good work, and moments worth savoring. Berlin mornings are best with a little momentum, so here's to a wonderful start. Make it a great one. Shine bright, Lefos
German Reichstag looks wrapped in white paper
Serhii Tatarintsev's post on LinkedIn: Recently, I had conversation with a candidate that completely changed the way I think about interviews. Let me explain. At first, it seemed like they are not a particulary good fit — not a lot of experience in our field, no familiarity with our tech stack. That is, until I asked one simple question: “Do you have any references or recommendations?” “How about a reference from my good friend Ben Franklin?” — candidate winked at me and dropped a neatly arranged stack of bills onto the table. Needless to say, he was hired on the spot. So, sometimes it’s not the candidates that fail the interview — it’s you interview process that fails to ask the right question that would let them shine.
Link screenshot
A bean bag is dropped in freefall until it hits a table below. The camera is also in freefall, showing a bean bag at rest and instead the background moving, until suddenly the table slams into the bean bag from the bottom
Me right outside clouds (visualizing clusters) of Bluesky accounts centered around Danish journalists, French politics, and Japanese artists
Google Street View photo of an Asian restaurant in Berlin called “Papa No”
Screenshot of a Chaos Communication Congress talk called "Don't look up" that explains that much of satellite communication was unencrypted, showing a cleartext XML capture of full credit card payment details
Page of the book; about a modeling technique called Business Model Canvas. It concerns 4 main dimensions: 1 Value Propositions; 2 Customers; 3 Infrastructure; 4 Finances. Each is subdivided, for instance within Infrastructure, you model “what do we do (to make our value props possible”, “what do we need”, and “who do we need”.
Kagi Translate offering multiple GenZ translations for “I’m going to the shop, can I get something for you?”, including “Boutta hit up the store, u need anything?”
Aforementioned quote in context: it’s about ‘70s behaviorists within the field of artificial intelligence
Slide with quote from Stroustrup ‘94 Design and Evolution of C++: “The Simula67 INSPECT statement was deliberately not introduced into C++. The reason for that is to encourage modularity through the use of virtual functions.”
Slide with a fragment from C.A.R. Hoare’s 1996 paper called “Record handling” showing structural pattern matching (like Rust match on enum)
Screenshot of portion of linked article: 3. The FDE role at OpenAI Colin Jarvis is Head of Forward Deployed Engineering at OpenAI. He initially joined in 2022 as a Solutions Architect, and became the Head of Solutions Architecture, last year. At the beginning of this year, Colin presented a business case for setting up a Forward Deployed Engineering Team, to have a more targeted offering. The group started with two FDEs, and today there are more than 10 at the company in: US: New York and San Francisco Europe: Dublin, London, Munich, Paris Asia: Tokyo and Singapore OpenAI still has a “Solutions Architect” (SA) role. Both SAs and FDEs help with the challenge of LLMs being complicated for customers to work with. FDEs are more hands-on and typically work with more ambiguity than SAs traditionally do. SAs are more like advisory roles: they rarely write code on customers’ infrastructure, and usually build minimum viable products (MVPs), or proofs of concept (PoC) with anonymized or offline cuts of data. FDEs are much more hands-on: they write code directly on customer infrastructure and use customer tooling for this. FDEs also align with OpenAI’s research objectives, and work with customers who can most likely build solutions that help advance the research direction at OpenAI. FDEs need to work with more ambiguity than traditional cloud solution architects. FDEs help advance the product roadmap. In one case, OpenAI worked with a voice customer on call center automation, and FDEs created evals on the voice model. Initially, the model wasn’t performing well enough for the customer to commit to deploying it. So, the FDE team went back to OpenAI’s research department with data and worked on improving the model’s performance for voice use cases, leading the customer to be the first to deploy the advanced solution in production. OpenAI’s Realtime API was also improved by this work. In the end, it was a win-win:
Kagi answering “was 2000 24 years ago” correctly and clearly
Newsletter excerpt. API generator startups Many other companies offer SDK generators from OpenAPI specifications. Here are three VC-funded startups that build tools which make a breeze of building and maintaining OpenAPI-based SDKs. Coincidentally, all were founded in 2022 and have each raised between $13-28M in funding, to date. Stainless: is used by AI providers such as OpenAI, Anthropic, and Meta (for Llama models), and platforms like Cloudflare and Mux. Languages supported: TypeScript, Python, Go, Java, Kotlin. The promise of Stainless is not having to hire the “usual” one engineer per SDK to build/maintain. Stainless charges $2,500/month for 5 SDKs, supporting up to 150 endpoints per SDK. Speakeasy: generates both SDKs, API docs, Terraform providers, and end-to-end tests. It is used by Mistral AI and Vercel. Languages supported: TypeScript, Python, Go, Java, C#. Speakeasy relies on its own OpenAPI vendor extension to add support for things missing from OpenAPI like pagination, streaming, and retry automation. Speakeasy charges $600/month per language, supporting up to 200 endpoints. Languages supported: TypeScript, Python, Go, Java, C#. Fern: used by Square, LaunchDarkly, Pinecone or Cohere. Languages supported: TypeScript, Python, Go, Java, C#, PHP, Ruby. Fern supports the API specification not only as OpenAI, but also as AsyncAPI, gRPC, OpenRPC, or its own Fern Definition for the best experience. The company charges $600/month per SDK for up to 150 endpoints. Other vendors include APIMatic, liblab, and Microsoft’s Kiota.
Newsletter excerpt. RED and USE Once you've defined your SLIs, there are two frameworks I rely on: - RED: Rate, Errors, Duration. These focus on user-facing performance. - USE: Utilization, Saturation, Errors. These focus on system health. You track both because they answer different questions.
An example of Kagi Translate in use. The paid translation app offers multiple variant translations, interprets the tone of the source, and relates to the tone of the main translation. You can hover over fragments of the output and see variants of each fragment, with explanations of the subtleties of variants of that fragment
German poster by Berlin city government featuring an intersex death metal band leader and software engineer, reading “Ich bin eindeutig inter.“
Search results in standard library rustdoc, showing how to create an OsStr from bytes either with unsafety or in a platform-specific way
Three-panel meme. First panel, someone says: "Many top scientists are on the autism spectrum". Second panel someone behind them says: "So technically autism causes vaccines." Third panel, the first person looks back in wtf
A glass beaker of coffee standing in direct sunlight on a countertop. Whereas normally coffee looks brown, now the coffee looks bright: the bulk is rusty red but the edges look yellow, almost green.
A Unix pipeline with two programs is shown. The source program computes the disk size of a bunch of directories and outputs a JSON array of objects. The sink program pretty-prints the JSON table. Computing the disk size takes a while, and you can see which directory is being analyzed because the result for that directory is empty while it is computing.
Someone is slowly typing JSON into a terminal program. The JSON is an array of objects. The program gradually renders the JSON input as Rust debug output, and as a table. The fields of the Rust struct are printed even though they are missing in the JSON input. The example program is called "live".
Slide from a presentation. This slide is about principles behind the Agile Manifesto. It investigates the principle of working software being the primary measure of progress. No it is not! Business value is the primary measure of progress. Because when you implement the wrong features, then there is no value in working software.