Hacker News: itsmeduncan

New comment by itsmeduncan in "Telegram Desktop vulnerability allowed any user's file to be stolen"

itsmeduncan — Sat, 10 Oct 2026 15:50:16 +0000

I wonder how much of Apple's announcement around disallowing full system disk access was from this as opposed to Muse, et al.

New comment by itsmeduncan in "Talorys – A self-hosted personal AI agent on Cloudflare's free tier"

itsmeduncan — Sat, 10 Oct 2026 15:49:13 +0000

I think the reality is the goal posts are shifting on owning your own data, and apps. The idea that self-hosted is not on your _own_ hardware is crazy to me.

The Atoms Are Already Deployed

itsmeduncan — Fri, 25 Sep 2026 13:35:12 +0000

Article URL: https://itsmeduncan.com/2026/09/22/the-atoms-are-already-deployed/

Comments URL: https://news.ycombinator.com/item?id=49844450

Points: 1

# Comments: 0

New comment by itsmeduncan in "The Economics of Open-Weight Inference"

itsmeduncan — Tue, 22 Sep 2026 18:37:24 +0000

The finding that flips the usual story is the sparse-MoE reversal: on gpt-oss-120b the A100 comes in cheaper per output token than the H100, because the workload is latency-tolerant and the MXFP4 experts fit in 80GB. Chip age stops predicting cost once you decouple "newest silicon" from "cheapest inference" and let price-elastic work route to whatever's efficient.

The point about no salvage value: the paper is really pricing a rental service, not the devices, and even there the 5-year A100 term holds 80% of the one-month mark. The reason there's no clean resale market yet isn't that old GPUs stop earning. It's that the workloads keeping them earning (batch, RL rollouts, long-horizon agents) are the ones that tolerate latency, and the people running those tend to rent rather than buy secondhand.

The part the paper understates: rented A100s aren't the biggest pool of already-paid-for, latency-tolerant capacity. That would be the Apple Silicon and flagship phones sitting idle, where the capex is fully sunk and there's no hourly meter at all. Same economics, one step further. The deciding factor is who gets to make the deployment decision, and open weights are what move that decision to the operator.

New comment by itsmeduncan in "Heretic removes restrictions from language models"

itsmeduncan — Mon, 21 Sep 2026 18:35:08 +0000

The most compelling use in this thread isn't edgy content, it's the boring legitimate work hosted models refuse by default: reverse-engineering a camera you own, a PoC for a CVE on your own network, decoding a protocol to connect your own software to a printer you bought. A policy layer tuned for the median user turns into a wall for the person doing real security or repair work on their own hardware, and a local model with no refusal layer just does the job.

New comment by itsmeduncan in "macOS 27: Workaround to avoid downloading AI models and save storage"

itsmeduncan — Mon, 21 Sep 2026 18:34:57 +0000

There's a distinction worth drawing here: "on-device" and "local" are not the same thing, and this thread is really about the gap between them. A model Apple pushes to your disk that you can't inspect, swap, or remove, that runs when the OS decides, is on-device in the literal sense and gives you none of the control that makes local AI worth wanting. Local, in the sense people actually mean it, is that you chose the weights, you chose when it runs, and you can delete it. Forcing a 14-29GB model you can't uninstall is the inverse of that. The storage complaint is real, but the deeper one is that "we put AI on your device" quietly became "we put our AI on your device."

New comment by itsmeduncan in "Hister: A private search engine for the pages you visit and the files you keep"

itsmeduncan — Fri, 18 Sep 2026 13:37:13 +0000

[flagged]

New comment by itsmeduncan in "GitHub is having trouble counting things"

itsmeduncan — Wed, 16 Sep 2026 22:12:42 +0000

My scheduled for 8am PT GH actions workflows don't run until anywhere between 6-8 hours later. I gave up, and run them locally now. It is just to create a tag, and push to a specific place. A super simple problem that was solved for so long simply with GH Actions that just doesn't work anymore.

New comment by itsmeduncan in "Giving up on smart rings"

itsmeduncan — Tue, 15 Sep 2026 16:50:35 +0000

I was an early adopter of Oura, and have recently stopped wearing it. I've reverted to the Apple Watch & EightSleep for all of the tracking I could ever need. I found the experience with Oura to be overly optimistic numbers wise, as well as borderline dark patterns in the app to drive my anxiety around getting good scores.

New comment by itsmeduncan in "Why does Opus 5 feel worse to work with?"

itsmeduncan — Fri, 14 Aug 2026 22:47:18 +0000

This is interesting take. How do you think the split will show up? Harnesses/models will be designed for human consumption, and those for machine consumption? I'm building getwhelk.com and it is a cool thought-experiment for me.

New comment by itsmeduncan in "Nvidia Nemotron 3.5 Lightning and NeMo Switchyard"

itsmeduncan — Tue, 11 Aug 2026 21:48:04 +0000

Me too. I think there are a few waves we can ride here. Let's collaborate?

New comment by itsmeduncan in "Claude Code is leaking real email address as a User-Agent string in curl command"

itsmeduncan — Tue, 11 Aug 2026 14:34:19 +0000

Agreed, and run your own inference too. Open source/weight models are getting so great.

New comment by itsmeduncan in "Improving GPT‑5.6 Sol in ChatGPT, expanding GPT‑5.6 Luna access for free users"

itsmeduncan — Fri, 07 Aug 2026 14:22:39 +0000

How long can OpenAI (and Anthropic) incentivize, and keep costs artificially down for tokens to depress investment in private/offline LLMs? It's interesting to watch the cost curves, and burn to see where it stops being venture subsidizes, and the frontier open-weight models become as good.