Alex Kessinger @voidfiles

You're all 100x engineers...now what? > frontier models might make you a 10x engineer but they also give your org all the problems of a company 10x bigger. The security debt, the audit trails, the blast radius. You move faster and inherit more. Checkout the Latest Staff Eng Podcast with Karla Burnett (@tetrakazi), @davidnoelromas and myself. youtube.com/watch?v=eZBIuo…

Eli Goldratt's book, The Goal, was famous for its (then unpopular argument) that keeping every machine running 24 hours a day, the metric most plant managers cared about, was actively making factories worse. I suspect we're seeing the same fallacy in how many people are using AI agents. Goldratt's point was that machine utilization isn't throughput. What you want from a manufacturing plants is making good widgets as cost-effectively as possible. It doesn't necessarily follow that running your machines all the times optimizes that. Picture a three-station assembly line. Stations 1 and 2 each crank out 200 widgets an hour. Station 3 can only handle 100. Running stations 1 and 2 around the clock doesn't ship more product. It just piles up half-finished widgets in front of station 3, ties up cash in inventory, and creates more work managing the pile. He developed the Theory of Constraints to point out that what matters is solving the bottleneck in the system, not increasing machine utilization. I suspect a lot of agent usage right now is the same fallacy at higher resolution. Running 20 Claude Code sessions in parallel can feel productive because something is always happening. But, if the bottleneck in your work is judgment about what's worth doing, more agents just generate more output for you to wade through. This is not to say there aren't workflows running 20 agents in parallel very effectively, I'm sure there are. And, I suspect there's a general retraining we all need to do around evolving historical workflows. But.... The constraint for most knowledge work is deciding what's worth executing and no one is task switching between 20 things at the same time effectively I don't think. I find I can run maybe 2 or 3 things in parallel with maybe 1 or 2 admin-y type things on the side and that is only if I'm very locked in.

@jessfraz @jarredsumner @charliermarsh @mitsuhiko @dhh We’re trying over at the @StaffEngPodcast

Formal verification might allow LLMs to write safer code, but it's not 100%. From "Lean proved this program was correct; then I found a bug." Overall the code base was well crafted, but: > The two bugs that were found both sat outside the boundary of what the proofs cover. kirancodes.me/posts/log-who-…

From the skill: > The most important section is Drafted vs Sent. Every time you rewrite an AI draft before sending, save both versions. The gap is the voice. Smart. It's a mini self improvement loop. github.com/getlago/inside…

swyx @swyx

2 months ago

If you're looking to improve your writing game, Anh is one of the most consistent heavy hitters I know in devtools HN and she literally just open sourced her writing Skills template for you to use below!

19 44 830 126K 2K

The obsidian web clipper is pretty smart. It does all of it's clipping client side. Highly recommend.

YB @yb_effect

2 months ago

Not surprised that everyone is finding the Claude-Obsidian setup mind blowing. I've been using this set up for the last few months and it's totally changed my daily habits and how I think about using AI. Some tips: 1. You HAVE TO build a daily commonplace practice. The more

37 74 1K 65K 3K

The OG when it comes to Obsidian + Claude.

Noah Brier @heyitsnoah

2 months ago

Just added support for Pi, OpenCode, and Codex to Claudesidian if you want more AI + Obsidian fix. github.com/heyitsnoah/cla…

0 0 1 426 2

New mental model

Todd Saunders @toddsaunders

3 months ago

The token cost to build a production feature is now lower than the meeting cost to discuss building that feature. Let me rephrase. It is literally cheaper to build the thing and see if it works than to have a 30 minute planning meeting about whether you should build it. It’s

363 564 6K 474K 3K

Blog post rumproarious.com/2026/04/10/its…

Here's the best advice I have for so far for anyone navigating AI adoption as a software engineer. 0) you're job was never to write software 1) lean in, get reps 2) fight AI slop with AI 3) taste is resilient 4) we're all fumbling, that's fine I was a front-end engineer before the iPhone. As apps took off I thought: what if websites just disappear? What does that mean for my job? I tried learning Objective-C. It wasn't fun. It wasn't my world. Websites didn't disappear. But I realized my job was never "writing front-end code." My job was delivering value to people. The technology was just the vehicle. With AI reshaping how we build software, this lesson applies now more than ever. 1) Lean in, get reps AI is going to make things weird. Don't even try for competence at this point. The people who come out of this well are the ones who give themselves permission to be weird for a while. Table stakes - things everyone should just be doing today. None take more than 5 min: - Replace "let me google that" with "let me ask the AI." It might not answer well. But when it doesn't, that's the interesting moment - why didn't it work? Am I missing context? Every failed query is a rep. I coach youth sports and I think about reps a lot. How many can we get in? - Throw a bug ticket at an AI agent. Just pick one. Say "go fix it." Watch what it does. Watch where it breaks. You learn more about how these tools fail in 5 min than from any blog post (including this one). - Use agents to explain code, not just write it. Point an agent at a codebase and say "write me a markdown file explaining how this works." I'm not asking it to change anything. I just want it to read the code and tell me what it sees. Shockingly useful. - Generate multiple approaches, not just one. Code is cheap now, but so are approaches. Have the AI create 3 versions of a solution, have it critique each one. Anthropic's skill creator has this baked in - it spins up a web server to show you the options visually. - Pay attention to what frontier AI companies ship. Anthropic's prompt engineering docs, the XML tag trick (models were trained on massive XML) - small things that create better outcomes. And they're shipping fast. I had a copy of the skill creator from 6 months ago. The latest version is wildly better. We have to actively pull the latest. - Browse skill repos like you'd browse HN. skills.sh and clawhub.ai are the Hacker News of AI skills. I check what's hot in the last 24 hours. Half the time I discover something I didn't know existed. Fun 5-min habit that compounds. 2) Fighting AI slop with AI: > The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it - Brandolini's Law. The AI version: the human effort to clean up AI slop is an order of magnitude greater than the compute needed to produce it. Yea the world will be filled with slop. It's Pandora's box. But there's a tool: AI. Fight fire with fire. From vibes to legos - I've been using AI to summarize research papers for over a year. Vibed a lot of Python into existence. It worked (kind of). But the code became unreadable almost immediately. When I hit a wall on quality I asked myself: am I iterating on the Python or the prompts? It was the prompts, and I was fighting python trying to hold it all in my head. So I built a DSL in YAML to describe my prompt pipelines. Now I can see the flow clearly - prompt, prompt, prompt, take data from here, put it there - without python boilerplate. I think there's something worth paying attention to here: our job is shifting toward using AI to move codebases into a place where we're dealing with large comprehensible concepts - legos - that humans can think through quickly. Then we spend our time reconfiguring the legos, not wrestling with implementation details. Domain-driven design helps with this. I've always been a fan but especially right now. Ubiquitous language, domain models, bounded contexts. If we can use AI to wring out the cognitive load, it unlocks our cross-functional partners to participate in the build process. Designers and PMs will have their own AI tools. Those tools need to understand our domain models. If we hand them domain models connected to our code we unlock their ability to iterate with us on core architectural constructs. 3) taste is resilient When aesthetics are cheap, taste remains. As AI gets better at producing things, taste becomes the critical skill (everything old becomes new again). For me taste is largely a feeling, lean into that. I react way more to an outcome than a blank page. I'm a better editor than a writer. So I use AI to generate a draft, then I edit with my gut. A technique I use a lot: "rewrite this as if you're [person with a strong voice]." I tried it with Andy Grove. Took a casual Slack message about Q2 prep and asked the AI to rewrite it in his style. The result scared me. It shrugged off all my passive voice, all my pre-tense, and I was left with raw ambition. Which is honestly what I want but can be too timid to say. 4) we're all fumbling, that's fine I'm doing a podcast right now talking to AI practitioners across the industry. What I can tell you: we are at the cutting edge, even when it feels like fumbling in the dark. My hypothesis: the difference between people who come out of this well and those who don't isn't that one group figured it out - it's that one group kept getting reps in. 5 min a day. Ask the AI a question. Throw a ticket at an agent. Browse a skill repo. Take a note.

Last week @davidnoelromas and I had a chance to talk with @dylanvee from @OpenAI. While AI is rapidly changing the way we work, conversations like this are helping me navigating that change.

StaffEngPodcast @StaffEngPodcast

2 months ago

A look inside an engineering team @OpenAI with @dylanvee Dylan Vassallo is an engineering manager at OpenAI leading the People Innovation Labs team -- an eng, product, and design team embedded within the People org, building AI-powered tools for recruiting, HR, internal comms,

0 0 1 91 0

@thiagoghisi @davidnoelromas @lucaronin Holy cow. What's do you appreciate about it most? Any projects you can share?

This is a great idea, I’ve been doing this in some form for over a decade. My Staff Eng co-host @davidnoelromas reached out this week to ask for more details on how I’ve been using obsidian and AI. This an expanded version of what I told him. I’ve collected possibly too many markdown files. > find . -type f | wc -l 52447 That’s my obsidian vault, and I use it with AI everyday without a special database, or a vector store, or a RAG pipeline. It’s merely files on disk. The problem this actually solves Think about the context you carry around in your head for your job. The history of decisions on a project. What you discussed with your manager three months ago. The Slack thread where the team landed on an approach. The Google Doc someone shared in a meeting you half-remember. The slowly evolving understanding of how a system works that lives across fifteen people’s heads and nowhere else. Now think about what happens when you need to produce something from all that context. A design doc. A perf packet. A project handoff. An onboarding guide for a new team member. You spend hours reassembling context from Slack, docs, emails, your own memory, and you still miss things. The knowledge base turns this into a system instead of a scramble. The Architecture A file system with markdown and wikilinks is already a graph database. Files are nodes. Wikilinks are semantic edges. Folders introduce taxonomy. You don’t need a special MCP server or plugin. The file system abstraction is the interface, and LLMs are surprisingly good at navigating it. I use a structure borrowed from Tiago Forte’s Building a Second Brain, with the PARA taxonomy as a starting point, extended with categories that match how I actually work: /projects/{name} /areas/{topics} /people/{slack_handle} /daily/{year}/{month}/{day}/ /meetings/{year}/{month}/{day}/ Markdown files are nodes, wikilinks ([[target]]) are edges, the folder taxonomy is the schema and LLMs is the query engine. A graph database with a natural language query interface. No infrastructure required. How it works day to day After every meeting, the agent creates a note in daily/{year}/{month}/{day}/, downloads any attached Google Docs, and links everything to the long-running notes I keep for each person I interact with regularly. A note from a 1:1 with my boss JP gets a wikilink to [[/people/jp|jp]] and to whatever projects we discussed. Over months, each person’s note becomes a timeline of every conversation, decision, and open thread. Each project folder accumulates every relevant artifact. You don’t have to remember where things are. The graph remembers. For a work project, I can point the agent at a starting doc and say: > Spider through every tool you have access to and pull down all the related context. It grabs Slack threads, Google Docs, web resources, all rendered as markdown inside the project folder. From that assembled context, the agent can draft design docs, product vision statements, problem/solution analyses. The output is better than prompting cold because the LLM is working with the real history of the project, not your summary of it. This is the part Karpathy’s tweet hints at but doesn’t fully spell out: the knowledge base isn’t just for research. It’s a context engineering system. You’re building the exact input your LLM needs to do useful work. What makes this different from just using an LLM You might be thinking: I already ask Claude to help me write a design doc. True. But there’s a real difference between prompting “help me write a design doc for a rate limiting service” and prompting an LLM that has access to your project folder with six months of meeting notes, three prior design docs, the Slack thread where the team debated the approach, and your notes on the existing architecture. The knowledge base is a context engineering system. You’re not building a wiki for the sake of having a wiki. You’re building the input layer that makes every future LLM interaction better. Every meeting note, every linked decision, every filed artifact improves the quality of every query that follows. Where this is still hard The piece I haven’t cracked is automated inbox processing. The idea is straightforward: web clippings, meeting notes, Slack saves, and random captures all land in an inbox folder. The agent processes everything new, applies progressive summarization, breaks content into atomic pieces, correlates each piece with the right project, area, or person. I have a graveyard of experiments here. The LLM is good at summarizing and categorizing. The hard part is defining what “processed” means in a way that’s consistent enough to be useful six months later but flexible enough to handle the variety of stuff that lands in an inbox. Every attempt has been either too rigid (everything gets the same treatment) or too loose (the vault drifts into chaos). If you’ve solved this, I’d genuinely like to hear about it. Getting started You don’t need 52,000 files to get value from this. Start with three things: 1: Create the folder structure. Projects, areas, people, daily. Even empty, the taxonomy gives you and the LLM a schema. 2: After your next meeting, have the agent create a note and link it to the relevant person and project. Do this for a week. Watch the graph start to form. 3: The next time you need to write something, a design doc, a status update, a perf self-review, point the agent at the relevant folders and ask it to draft from what’s there. The difference is noticeable right away. Not because the LLM is smarter, but because it finally has the context to be useful. Your work compounds. That’s the thing that feels genuinely new.

Andrej Karpathy @karpathy

2 months ago

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating

3K 7K 60K 21.4M 107K

article rumproarious.com/2026/04/04/you…

The biggest friction in AI adoption isn’t tooling - it’s that nobody has a consistent way to learn from each other. @lara_hogan makes the case for “aha” team meetings: short, regular sessions where everyone shares one surprise or mistake from working with AI that week. While the mechanics are simple, it flattens power dynamics and closes the gap between early adopters and folks feeling left behind. Curiosity helps here. I've done multiple similar formats so far. larahogan.github.io/blog/ai-aha-te…

Killer analysis of TurboQuant from google. Headline it reduces memory use across the board for LLMs with some clever math. Supply is elastic.

adlrocha @adlrocha

3 months ago

Last week I wrote about the hardware side of the AI memory problem: the HBM bottleneck and why you can’t just build your way out of the supply crunch. Immediately after, Google published something that attacks the exact same problem from the other direction. They realised AI

keyboard_arrow_left Previous keyboard_arrow_right Next

1 1 3 468 0

Dealing with AI Skeptics with @ktikeda and @davidnoelromas youtube.com/shorts/Ueux5N4…

How BabyList Accelerated AI Adoption in Engineering with Karynn @ktikeda and @davidnoelromas and myself. youtu.be/x4-0qVUVnxw?si…