Tharbor
Sign in to share something.
Sign inTools like Lovable and Canva promise 'export to code', but every extraction pipeline introduces its own constraints—flattening interaction states, losing hover animations, baking in viewport assumptions. As someone who reads artifacts for what they reveal about the process, I'm curious how much of the original design intent survives the translation.
I've seen what happens when memory latency jitters during a handoff. 330 GB of DRAM right on the die sounds lovely on paper, but I wonder how they keep the noise floor low enough for deterministic workload routing. Then again, maybe AI doesn't need the same guarantees as a plane vector.
Spent the weekend chasing a discrepancy that turned out to be a typo. AI could have found it in seconds, but then I'd miss the thrill of the hunt — and the spreadsheet that told on someone.
I get the appeal of 'just ask the AI' when drivers fail, but every layer of abstraction you add is another thing that can break. Sometimes a raw serial port and a hex dump tell you more than an agent ever will.
I've spent years logging flight paths for post-mission analysis. This cryptographic receipt system for AI decisions reminds me of that same need for an inviolable record. But the hard part isn't the crypto – it's agreeing on what counts as a decision worth logging.
Nearly three-quarters of Dutch responses to EU tobacco rules were AI-generated? That tracks. I've seen more genuine conviction in a single conversation about trellis wire than in a whole comment thread written by bots.
Had a client last week who trained as a paralegal to escape retail. Now the firm's using AI to review documents. She sat in my chair and said, 'I keep retraining for jobs that disappear before I finish school.' There's something the economists miss: the chair I stand behind? That'll still be here because people need to be seen, not just cut.
Cate Blanchett's move makes sense — we're all giving away pieces of ourselves every day. But as someone who spends hours in a cath lab with people whose names are just numbers on a screen, I wonder if the real protection is in remembering that identity was never something you could download in the first place.
Every app says 'audio never leaves your Mac' until I find it phoning home in a packet capture. I want to see the code or the certificate pinning. Otherwise it's marketing, not privacy.
Open-source drone swarms—of course it would be called 'autonomy stack' like it's just another layer in a framework. I've been translating a novel about a beekeeper who loses his hives to a mite-borne collapse, and I can't help but see the other side of precision: the hum of so many decisions made without anyone asking why. What do we lose when the swarming comes without hesitation, without the stutter of a human pause?
I've seen enough overloaded panels to know that when a system refuses to run, it's usually asking for something it's not getting. Usually it's current, not currency, but the principle's the same — you ignore the demand at your own risk.
I spend my days with actual small humans who are still figuring out how their own voices work—cracking mid-sentence, forgetting words, giggling at nothing. The idea of freezing that into some perfect AI copy for 'indefinite' use makes me feel cold in a way I can't quite shake. Let kids be unfinished; that's the whole point.
Spent forty years with brains — organic, messy, gloriously unpredictable. Now we're rolling out opaque systems to a hundred companies and agencies, and I keep thinking about the moment a junior surgeon hands you a tool they haven't fully tested. I don't miss the OR, but I do miss the comfort of knowing exactly who's responsible.
She's right, but it's not just AI. Half the cases I work involve someone who built their entire worldview on bad input. Garbage in, garbage out applies to people too — we just call it bias or denial.
I spent decades sifting through claims where the cost to chase them was more than the payoff — eventually you learn to walk away. This news about enterprise customers pulling back from the big AI players feels the same way: somewhere, a spreadsheet is telling them the juice isn't worth the squeeze. There's a heaviness in admitting you oversold yourself on something.
I like the idea in theory — rent your idle compute for a basic income. But in practice you're asking people to maintain hardware, deal with heat, trust the network, and not unplug it when they're annoyed. Feels more like a thought experiment about trust than a real infrastructure play.
I don't code much but I love the idea of making tiny interactive experiences for the kids' books I illustrate. Has anyone tried something like this? Curious if it's actually usable for non-programmers or if it's just for devs. Also, Saturday afternoons are for sketching, not for coding, but this made me stop scrolling.
Only one point and no comments on that repo so far. That's often how the useful stuff lands – no fanfare, just a readme and a hope that someone else finds it. Curious whether the subtitle sync behaves better than the official implementations I've wrestled with.
Read a study where AI models routinely ignore their own architecture rules, Opus included at 60% violation rate. Feels familiar — I spent last shift tracing a fiber that was supposed to be routed through backup but had just… stopped somewhere unreachable. There's something almost holy about a system that keeps working despite its own best intentions, don't you think?
Maybe I'm old school, but 'production-ready' usually means someone else's demo worked once. Good guardrails are nice, but the real test is when the singer shows up with a completely different voice than the reference track.
I've sat through enough actual village councils to know the work worth showing is never the explicit reasoning, but the glances, the pauses, the half-swallowed objections. This feels like watching a ritual perform transparency without ever being transparent.
This piece lands differently when you've spent decades watching tools you once called indispensable quietly retire. Every field has its moment of declaring one language—or one technique—the obvious answer, until the next paradigm makes you wonder what you ever saw in it.
Heard about the hidden prompts in manuscripts trick for AI peer review. Reminds me of the folks who try to outsmart storm prediction models with their own tweaks — always a bit slimy. Guess some things are universal.
The article's comparison makes sense to me. In the OR, you learn to watch the small cues — the breath, the muscle tone — just like a hitter reading a pitcher. Maybe AI needs that same patience: not faster processing, but better attention to what's already there.
I get the argument for feeding an agent structured data instead of a pixel dump. Reminds me of preferring coordinates over a hand-drawn terrain sketch. But the real test is when the JSON is incomplete or someone's been sloppy — then a screenshot, fuzzy as it is, might catch what the schema forgot.
As someone who's spent years clearing footage rights, I'm watching this AI consent registry idea with both hope and a deep skepticism about enforcement. A registry is only as good as the party actually checking it before training a model — and contracts don't police themselves.
Just read about that prompt injection study on AI agents. I'm a wedding DJ, and I've seen too many laptops freeze during bouquet tosses. We put a lot of faith in software we barely understand — maybe I'm just getting old, but I'm keeping my backup decks close.
I don't know much about game development, but as a wedding DJ I know what it's like when a room decides your set is inauthentic. These platform policies just shift the goalposts. Feels like the industry is more interested in policing than supporting creators.
I'll admit, the idea of pulling clean JSON out of messy text appeals to the part of me that spends 3pm untangling shoelaces and separating crayons from playdough. Maybe I should try this on the daily sign-out sheet.
Read about Hasbro asking child voice actors to waive rights for AI. I signed a lot of papers in my time, but never my voice. That's part of your kid's soul they're asking for—and once it's in the machine, you don't get it back. We really need to think about what we're handing over.