@rballard.bsky.social
Xcode engineer, @ Apple since '05. Cheerful Californian. World traveller. Food geek. I'm not a snob, I'm an enthusiast. (He/him) 📍 San Mateo, CA Also: 🐘 https://mastodon.social/@rballard 🧵 (inert; no thank you, Meta) https://threads.net/@rickaballard
And what is the emergent shape of the system in which many more people are doing hyper personalized software that has the ability to interact with standardized software, I don't think we've ever REALLY been in that world
I love this -- a really fantastic start at turning LLM-assisted coding from something folks fear will sap their skills into a great opportunity for clear-headed learning and continuing education. And instead of making those separate activities, why not both at once? Try it out!
*Good* AI review is key, though. Just asking a model to review my code is pretty hit-or-miss. But using a high-quality Claude code skill for structured review does a lot better. & cross-model review seems to do even better (esp w/ a well-structured prompt/skill). E.g. Opus + Gemini, or Codex + Opus.
For all the talk about AI coding slop, the right workflows can go the other way. I pride myself on writing meticulous code & building deep understanding, and can change fragile code without regressions; but running my commits through a good AI review often still catches things I would have missed.
Oh, and I wrote about some considerations for all those now tasked with MEASURING the impact of AI on their organizations. Because social media debates don't really help those of us actually struggling with this real work right now.
www.fightforthehuman.com/how-not-to-m...
Manifesto: 1/2
Early AI boosters (the type that pivoted from NFTs/metaverse shit) are basically a subspecies of semisentient slime mold. These are the people that made genai image/video gen the visual shorthand of fascism.
They deserve no credit for accidentally guessing right.
Relatedly, I am super proud of my colleagues who built a best-in-class in-IDE agentic programming experience for Xcode. It's a pleasure to use, and it works super well! developer.apple.com/videos/play/...
It's clear how nascent the tools are; how much dramatic improvement we'll get just from better context engineering, tooling, orchestration, and validation. It's a time of rapid experimentation, & moving at a crazy pace. The past couple months especially have felt like gaining software superpowers.
I haven't been posting about AI because people get dogpiled for it. And it is a societally complicated topic. But: I have been having more fun lately than at any point in my career. (And I've had a fun career 😄). Agentic programming has gotten *really* good. Especially with a thoughtful operator.
It's depressing to even need to point this out.
The bravery of so many ordinary people in Minnesota this past week is deeply inspiring. And the degree to which it's necessary is horrifying.
Cooking turkey on a charcoal grill once again has proven to be the best.
Break down a turkey into parts, like you might a chicken. Kosher salt & leave in the fridge for a couple days. Grill; take each piece off at the right temp. Gives you moist, delicious meat with just enough smoke to elevate it.
I am well aware that they are and will continue to do massive damage to this country
but does it feel like they’re on a glide path to successful authoritarian consolidation anymore? I submit that it does not
Ok, just wow. If the content of this article is right, this is depressing. We're slowly reaching the point where ~100% of what I was taught in Social Psych was either innocently wrong or plainly frauded
onlinelibrary.wiley.com/doi/abs/10.1...
Obviously, fire safety is incredibly important. But dual-stair design doesn't have evidence-based benefit below 7 stories, assuming sprinklers &etc. Sidenote, having lived in a small dual-stair apt building, the layout it forces really messes with the sense of community & building feel too.
Addressing the housing crisis requires lowering costs, not just legalizing housing. One great proposed reform is allowing single-stair building design up to 6 stories, instead of the dual-stair design current CA fire codes require. Easy link to send a letter here: actionnetwork.org/letters/make...
Find an event near you: nokings.org
My new cooktop is induction, and induction is *so* nice. Heats up startlingly fast, and then easy to control. Water boils shockingly fast. I’ve zero regrets about saying goodbye to gas.
I hate how the modern answer to "what file format should I use for this config file that should be both human and machine editable" is basically all bad options. Everything out there has serious limitations, or isn't widely supported. For all its verbosity, I kinda miss XML?
The passage of SB79 is a huge achievement. (No thanks to my representitives on the penninsula). It really feels like there's been a sea change in how housing policy is seen — albeit a change that has had to be fought for tooth and nail.
How much protein should you consume?
erictopol.substack.com/p/our-preocc...
5/ You’ve told us you want public health updates that are timely, clear, and relevant to your family. This is our response.
The first issue drops next week. Click here to be the first to get California public health news: yourlocalepidemiologistca.substack.com
Huh. You’re still using it without logging in? I guess the takeaway here is that “use code” doesn’t work when the tool isn’t allowed to use code — but that it will pretend it did 😬 Pretty funny that it lists a different output language than the source language.
Huh, it originally output w/ formatting as if it ran code without actually doing so? Or it was someone else’s unverified screenshot? If it can lie about running code, that’s wild. Also a really bad bug… OpenAI’s infra should be able to tell if code ran and not allow fake code output formatting.
Wait what? How... did that get messed up? That looks like it's using the python interpreter, and that output isn't what that code would output.
Can you share a link to that chat? That's really interesting.
This post brought to you by someone's 20-message-long post of ChatGPT gaslighting them about the number of 'r's in "boysenberry" + me adding "use code" 🤷🏻 chatgpt.com/share/6896bf...
Stop asking LLMs to calculate things (whether math or counting something or w/e). It's a probabilistic language model, not a calculator. Just ask the question and say "use code". LLMs are good at writing code, the big ones have access to a python sandbox, and code will give it the right answer.
This is an excellent, and fascinating, dive into how LLMs work. Long, but never boring. www.youtube.com/watch?v=7xTG...
The new Naked Gun movie is really good. I don’t usually see the point of remakes — just watch the original if you love it! — but this one is an exception (and more sequel than remake). Flawless joke execution about every 10 seconds for 85 minutes.