the only way to stay sane while working in tech is to aggressively pursue hobbies and interests that are grounded in analog mediums and/or the natural world

One day, we’ll look back and realize that uploading our entire codebase into online LLMs was insanely reckless.

AI Washing: Signs, Symptoms, and Suggested Solutions for Investment Stakeholders This report addresses the ethical concerns & risks of AI washing in finance, providing crucial questions for stakeholders to evaluate managers’ AI claims.

"These firms may use buzzwords such as “AI-driven” or “machine learning–enabled” without truly integrating these tools into their investment processes."

rpc.cfainstitute.org/research/rep...

"When using GenAI tools, the effort invested in critical thinking
shifts from information gathering to information verification; from problem-solving to AI response integration; and from task execution to task stewardship."

www.microsoft.com/en-us/resear...

Com certeza. Eu ia escrever "testar é uma arte", mas na verdade é só uma disciplina jovem. Conheci poucos devs que encontraram o sweet spot entre "o que testar", "porque testar" e maintenance weight. Imagine uma LLM treinada na maioria das codebases (que normalmente tem testes ruins) e na literatura

Gostei também da discussão sobre qualidade. Acho que existe um gap gigante de ferramentas de guard rails pra desenvolvimento. Como uma suite de static analysis que valida o output da LLM. Talvez antes e depois de chegar no usuário. E talvez até retro-alimenta o prompt.

Dado esses exemplos, eu gosto de pensar que "clareza na entrada" melhora com a evolução da UX do input e da IDE.

Como deixar pistas permanentes ao longo do código (sem necessariamente escrever texto), e como dar pistas momentâneas para algo que você quer fazer agora.

Outro exemplo bom é attach imagem no prompt multi-modal, e o modelo ser capaz de usar como contexto.

Outro exemplo bom é debugging. Tem um MCP que puxa logs do console do browser. No prompt quando você mencionar um erro, ele já puxa o log e dá um chute bom.

Parecido com o que você faria num pair: "olha os logs pra ver se tem alguma pista"

Eu especulo que a dificuldade de expressar o que você quer esteja relacionada às limitações da entrada primariamente em texto.

"Acho que temos que modificar isso aqui", posiciona o cursor num método e "marca" ele, pra "dizer" pra IDE o que você acha, assim como você faria num pair programming

Xcode console logs showing: Can't find or decode reasons Failed to get or decode unavailable reasons

Yo Xcode, you're supposed to compile Swift code, not do philosophy…

After a short era in which people questioned the value of academia in ML, its value is more obvious than ever. Big labs stopped publishing the minute commercial incentives showed up and are relentlessly focused on a singular vision of scaling. Academia is a meaningful complement, bringing...
1/2

When a product is simple to use, it’s typically quite difficult behind the scenes — attention to detail, getting all the corner cases right, high quality, managing contingencies.

But hard for competitors to replicate, and awesome for customers.

Two very good reasons.

Inventing on Principle Presented at the CUSEC conference on January 20, 2012. (https://2012.cusec.net) Bio: Bret Victor invents tools that enable people to understand and create. He has…

It's becoming a yearly ritual to rewatch this. Positive vibes and inspiration

vimeo.com/906418692

Nice post. Substage got my attention, though I'm waiting for the day that we get these LLM-apps running locally – in this case a minified GPT pre-trained on command-lines.

Awesome, thank you. Are you self-hosting as well?

That looks good 👀 I particularly like that events are anonymous / not-tracking. I'll be trying it out

What do you use for macOS + iOS app metrics and telemetry? Preferably self-hosted? I've been trying @telemetrydeck.com but I'm still looking around #macos #ios #swiftui #swift #buildinpublic

no focus firewall mention 😭

would be weird if they followed people captured on the webcam

I've chosen to avoid this wave of AI altogether. I'm not interested in becoming one with the machine. If they are willing to spend time and money training artificial intelligence, why wouldn't I do the same for the real thing.

Works with Safari, Google Chrome, Firefox, Arc and Brave.

I can also add any browser you'd like (takes a couple days for the app release to be approved)

focusfirewall.com something I've been working on since last year ♥️

(urgently need to level up my marketing game for obvious reasons, any suggestions welcome)

All true. I skipped the third because I added an onboarding wizard since early days (highly recommend it, also super easy to do with SwiftUI)

I'm still puzzled at how to continue to promote. At this point trying everything and seeing what sticks lol

working on focusfirewall.com

blocks distracting websites during work to prevent impulse browsing, doomscrolling. I wanted something passive that I could pause/turn off whenevs and that worked system-wide (had to build a network extension and go deep on network protocols)

the idea that there is only one way to do something is a ruse to make you stop trying to find your own way. when in doubt, risk failing in order to hoe your own path.

Siri Is Super Dumb and Getting Dumber Siri with Apple Intelligence gives confident but embarrasingly wrong answers to sports trivia questions, both famous (Super Bowls) and obscure (North Dakota high school hoops).

I had so much fun writing this one.

Koan #58

“Why don’t I have a garden?”
“Did you plant seeds?”
“Many!”
“Did you water the seeds?”
“No, I was too busy planting more seeds.”

And the student was enlightened.