Lukas Bergstrom

Just thinking about how I have seen all the works that are done under the sun; and behold, all is vanity and vexation of the spirit. 😁

@snowmaker One exception here is hard to reproduce bugs.

I realized something else AI has changed about coding: you don't get stuck anymore. Programming used to be punctuated by episodes of extreme frustration, when a tricky bug ground things to a halt. That doesn't happen anymore.

TIL higher-order rainbows exist

@awkweb @kepano ^^ knows ball

@kepano notational velocity vibes notational.net

very old tip, but still true

@benthompson @Burgos @crackitydimes @stratechery That's why we have the Defense Production Act: because private companies have freedom of contract, and we sometimes want to abrogate that freedom. AI doesn't change the calculus of rights here.

@Burgos @crackitydimes @stratechery I don’t disagree about other specific responses that can/should we use. I agree the current approach is extreme and, as I said, illegal. The point of the article is that no matter the president this is the real alignment problem: super powerful AI vs government.

Anthropic and Alignment Anthropic is in a standoff with the Department of War; while the company's concerns are legitimate, it position is intolerable and misaligned with reality. stratechery.com/2026/anthropic…

@ctatedev Tauri next please!

New agent-browser skill: Electron You can now control desktop apps built with Electron, including Discord, Figma, Notion, Spotify and VS Code Or, use it to debug your own Electron app Add it to any coding agent: npx skills add vercel-labs/agent-browser --skill electron

@escargotpro_ The party scene on the beach where he's drunk and humiliated is untouchable

Finally saw The Long Goodbye last night- having grown up on lebowskies inherent vices & other more zany takes on it i fear i like the remix more than the original. It didnt hit for me the way it should have. Also google image results for his topless neighbors are terribly bad

@emollick The launch blog post made it sound like Pro was still the highest quality option, so I'm surprised to hear this.

I had some early access to Nano banana 2. It isn't perfect but it is the first model to handle really complex images and diagrams with some consistency. "show me a where's waldo set in ancient Venice, but instead of waldo it is an otter wearing a blue striped pilots outfit."

@RamonPang A LOT of us were going to raves back in the day AND jamming Aphex Twin and the Orb in the afters. It was all part of the same scene back then. The whole narrative about IDM being something outside of dance music - much less "better" - was always bullshit.

people of my gen (younger millenial / older zoomer) are the products of what yall started - listening to electronic sound for sound-sake. No discrimination against 'high' or 'low' energy in dance music - just an appreciation of sick sequences. Isn't that the future yall wanted ?

electronic music boomers always ask - "how can you like mainstream EDM / EDC music and aphex twin at the same time?" and I feel like I covered it pretty well here

@GoodWillsmith Was gonna post about seeing Lucier but in looking something up ... uh ... holy shit

afterlife band: Éliane Radigue - oscillators Pauline Oliveros - accordion Phill Niblock - laptop Alvin Lucier - the long thin wire Yoshi Wada - earth horns Marian Zazeela - tambura Pandit Pran Nath - voice

For the impatient: tidepools.ai

I've found that journaling about my work - what my goals are, what I'm struggling with, what I'm excited about - has been enormously helpful in figuring out where to spend my time.

I'm launching tidepools - a daily journaling and task management app with a proactive AI coach. It took me a couple years to get all the pieces in place, and in the process I learned a lot about the value of writing in thinking clearly.

@tenobrus @elcarpo btw have you tried vibecoding in Retool? terrible experience. not sure how they fumbled that badly.

@elcarpo pocas de estas tienen que ver con ‘las empresas simplemente van a vibecodear sus propias versiones del saas’ btw. eso básicamente es solo retool. el resto se basan en que ~agentes personales locales van a cambiar fundamentalmente la naturaleza del trabajo

No entiendo, los AI bros piensan que todos quieren vibecodear calendly?

@brianluidog I've always wondered if there was a fad for reading Machiavelli or Clausewitz in China at some point

Did you know that in China, the prestige military strategy treatise is 36 Strategems? But it's too powerful. So when they first encountered the West, they dusted off an extremely mid brochure, called it "The Art of War", and now generations of Westerners read it at airports

@WorkflowDevKit This looks different than the guidance in useworkflow.dev/docs/foundatio… when should something be a separate step using the built-in retry mechanism, and when should it be the code example you showed?

Traditional: a retry queue, a backoff column, and a cron to sweep stuck jobs. Workflow: a for-loop with `sleep()`:

@WorkflowDevKit consider making llms.txt smaller with links to sub-pages, it's pretty huge

@ProgramWithAi The second prompt was in the same session?

When I prompt, I try to have a very clear picture of what I expect the coding agent to produce. If the output doesn't match what I expected, I rewind and update my prompt (in case the model produced something better than I expected, I update my priors). This allows me to get better at prompt engineering, and to really learn the impact of my choices. For example, here's a one-shot prompt to create something equivalent to @Steve_Yegge 's beads. This is a very "engineered" prompt, where every word has a purpose that has been empirically validated at least a couple of times. The more tired I get, the less engineered my prompts. When I start "talking" to the agent, I know it's time to wrap it up for the day. Regarding the engineering of this prompt: Green: output format Red: intent Blue: latent space navigation If I remove the words "cooperating agents", I will often end up with a CLI missing "claim" and "release", which makes it hard for slow agents to not snag each other's tasks. "tasks in a DAG" will make it used traditional planning algorithms and just regurgitate some well-know implementations. "cooperating agents" will summon 70ies-80ies GOFAI agent literature and implementation, further strengthening the prompt. "use dolt" wasn't my choice but a challenge by a friend, it didn't go all that smooth because of PATH issues. I created a skill for dolt in case I ever need it again (not enabled by default). I did about 10-20 of these over the last year (way before beads), which makes me pretty confident in this prompt. This one used 20% of the gpt-5 context window, and I then used "Use dagctl to create the tasks for a webui with kanban and graph view, use tmux and codex to start subagents and tell them to use dagctl" which created a coordinating ralph loop polling subagents every minute. TWO PROMPTS YO

@AwkwardSpud @OedipasKvass Yes! It's so much fun. The language takes some getting used to, helps to have SparkNotes or something open as you read. But you won't need the notes for long.

@OedipasKvass * went through high school without having to read a single shakespeare play should i get on the romeo and juliet grind

My students are having more fun with Shakespeare than I’ve seen any student have in like 4.5 years. Still relevant. Still great. The number of times one of them finishes a line and I just hear a football player murmur “bars” in the background is hilarious.

@martin_casado I wonder how much of the investment is driven by a belief that whoever gets to ASI first will have an insurmountable advantage. If so, I think it's irrational, but ...

The economics for AI are clearly complicated. But the crux is the following (assuming a frontier lab with LLMs, not diffusion): - if you account for the previous training run, they look great - if you account for the current training run, they look terrible - the current training run isn't in COGs (obviously) - but we all know models only have 3-6months of relevancy .... so ... either capital keeps chasing growth one training run ahead of what's currently being used. Or the market will have to rationalize the disconnect.