@GergelyOrosz "load-bearing"
It seems natural, actually. I have noticed again and again that I start using the same words and phrasings as the people I spend the most time with and so do they. Now many share those interactions with chatbots.
@OpenAIDevs in the codex macos app, it seems like auto generated commit messages describe the _thread_ instead of the code change being committed - imo this is backwards and the message should describe the atomic code change, not by back/forth with the coding agent
@andrewchen 10 seems like too many but this could be directionally accurate. I think you’re still going to be gated by ability to run useful experiments and most orgs probably can’t make sense of the concurrent experiments 10 PM-types would want to run.
They literally did.
They got all the phones on the market. And saw it was all junk by Apple standards.
They knew this for a while but had to get a few technical things in order first.
They had to get really good at miniaturization -> iPod helps them learn that
Then they had a cross learning moment, where the company they acquired for their touchscreen technology — which was working on computer displays — they took that and brought it over to “Project Purple” -> thus multitouch was born
So now that have multitouch and know how to make things small. ANDDDD they are still looking at junk ass phones in the market.
Once they see what’s wrong with them they ask themselves — “what’s the phone we’d make for ourselves?”
They “vibed” AFTER they knew the market, they had technical pieces in place, and this is where all the software love that made the iPhone the iPhone comes together
Pulling from springboard and all the macOS work -> that’s what made the iPhone feel so good with direct manipulation of multitouch
iPhone wasn’t some isolated “do what you want” event
It was built off of decades of HCI learnings. A key acquisition. And the success and learnings of the iPod.
Learn the real stories and don’t just take a cliche like “users don’t know what they want” as an excuse to build nonsense
Instead of debating the ROI of investments in training versus inference, isn't the answer more of both? Until we find that we can't scale training of base models, fine tunes of those base models will benefit from more underlying intelligence. R1 benefitted immensely from that
@sh_reya I think this is the key point. LLMs crush when writing functions with clearly defined inputs and outputs. Even better when you can easily generate a test suite to lock in the contract. But that's already easy for experienced eng and only so much work looks like that.
@DylSell classic innovator's dilemma? We'll see who has the risk appetite to not just fork their product into true AI-first (i.e. not copilots), but also the ability to migrate legacy usage to vAI. Duolingo's much closer than Asana/ZI imo, but apples to oranges.
@pkedrosky To your point, I know very few people (can count on one hand) who exercise for purely longevity/health benefits and actively dislike exercising. It's too arduous a pill to take if there's not inherent enjoyment
@pkedrosky If you're lucky, exercise is fun and becomes a hobby. Two birds with one stone. Though exercise as a hobby can easily turn into something that is no longer healthy? Thinking compulsive exercising, going harder, longer than that which yields a positive health impact
@levelsio There's something about lifting weights in full ranges of motion that seems to heal joint issues. Surely can create them, too, if done poorly, but has benefits beyond just getting physically stronger
@bhalligan At big companies that's probably directionally true. At small companies that get a decent amount done, not possible. You'd immediately sense if only 1 person out of 10 did any work. 10 out of 100, less easy to sense but possible. 100 out of 1000, probably hard to tell / measure
@Austen Show me a large, complex codebase written from scratch by an LLM, agent loop or otherwise
Or rather, what's your definition of the modifiers large and complex?
Good examples of what I'm talking about in this post: https://t.co/6QoTcd1twx
Instead of google searching for "pretty print JSON" which would lead you to some janky, ad-riddled site from 2010, you can create a custom version tailored to you with a tweet-length prompt
ChatGPT canvas is slick, but Claude's ability to render html, css, and JS in-browser makes it more powerful for creating micro, personalized software. Not sure why canvas didn't launch with this capability as it makes that use case overly cumbersome