Hacker News: anax32

New comment by anax32 in "Ask HN: What are you working on? (August 2026)"

anax32 — Sun, 09 Aug 2026 19:30:24 +0000

I'm building a simple model inference platform: https://marigold.run

Seems everyone is building an inference platform lately, but this one is mine! And I'm enjoying it...!

New comment by anax32 in "[dead]"

anax32 — Fri, 31 Jul 2026 08:38:41 +0000

With new tools (i.e., LLMs) it feels like software engineering is moving to an assembly process.

The best engineers will have an awareness of overall problem, architecture, how pieces should fit together and what changes when the fits are made.

The top-tier will be able to adjust components, and identify gaps which can be filled by newly created, but properly shaped components to create new emergent properties from the combinations.

This has always been a requirement, but intertwined with the expression of the vision by coding. Are those two things really separate? I don't think people can be architects without coding, but I'm an old man with old ideas!

New comment by anax32 in "Protect your right to run local AI"

anax32 — Thu, 09 Jul 2026 10:43:40 +0000

yes, that's what I thought.

"You got a loicense for that PSU?" is my biggest fear!

New comment by anax32 in "Right to Local Intelligence"

anax32 — Fri, 03 Jul 2026 08:18:03 +0000

A ban on desktop PC components over a certain power threshold (300W) or PSUs would definitely affect local AI, and Europe is not a big enough market, nor has it's own internal supply chains to offer alternatives.

This has been done in the EU/UK in the past (hoovers/vacuum cleaners) so the mechanism exists.

I'm in the UK and support your direction.

New comment by anax32 in "Age verification is just a precursor to automated attribution of speech"

anax32 — Mon, 29 Jun 2026 08:28:41 +0000

And it is focussed on social networks, which require an email address, which usually implies a device.

But instead of inserting controls around email addresses (as with paid services) or devices (as with contraband), the requirement is pushed to the application layer. It really makes no sense from a technical POV.

New comment by anax32 in "Knowledge Distillation of Black-Box Large Language Models (2024)"

anax32 — Mon, 29 Jun 2026 08:21:53 +0000

The US "product machine" is so strong. They really know how to do frictionless signup and vendor lock-in on the corporate side.

New comment by anax32 in "Ultrasound imaging of the brain"

anax32 — Fri, 26 Jun 2026 20:55:42 +0000

Can anyone explain how this ultrasound can see through the skull?

I've worked on ultrasound devices and data, the shadows from bone, and distortions caused by tissue types were very difficult. If this device can deal with those distortions it would already be useful for lung imaging.

New comment by anax32 in "Ask HN: How is GPU power draw measured at scale?"

anax32 — Fri, 26 Jun 2026 20:48:32 +0000

Yes, thank you. That's exactly where I am, and trying to gather some knowledge.

The power draw from the wall is especially important, because a spike across multiple devices at the same time can cause issues which are really difficult to debug.

Ask HN: How is GPU power draw measured at scale?

anax32 — Fri, 26 Jun 2026 10:35:10 +0000

How do people measure power usage of GPUs at large (32x) self-hosted setups or small multi-rack setups? I've seen some PDUs which collect and transmit data, but I'm unsure of the processes and if/how people do this on small builds.

Currently, I collect NVML nvmlDeviceGetPowerUsage, polled at 100ms during inference, peak and mean per request, and get this type of data:

model mean-power range (W) spread stdev

qwen3-8b 114.3-121.9 7.6W 1.17

llama-3.1-8b-instruct 104.7-122.1 17.4W 4.29

qwen2.5-1.5b-instruct 53.7-73.0 19.3W 5.23

mistral-7b-instruct-v0.3 96.2-120.0 23.8W 6.01

qwen2.5-7b-instruct 88.7-124.5 35.8W 7.73

gemma-3-1b-it 49.4-56.7 7.3W 2.13

this is per-GPU, single-card data - I don't know whether anything like per-request attribution survives at rack scale, or whether monitoring there happens entirely at the PDU/BMC level instead.

Comments URL: https://news.ycombinator.com/item?id=48684938

Points: 6

# Comments: 2

New comment by anax32 in "The unbearable cheapness of open weight models"

anax32 — Thu, 25 Jun 2026 08:48:55 +0000

Open weight and local hosting is far, far cheaper. In every respect. Even support is cheaper, over time.

However, it's difficult to sell this to businesses who want contracts and KPIs, not staff and commitments.

Regulated industries will favour the closed sources, either by choice or mandate. The interesting question is whether they will have better models, or worse models. History says they will receive a worse service, but continue anyway.

New comment by anax32 in "Running local models is good now"

anax32 — Tue, 16 Jun 2026 15:19:32 +0000

I've just made a milestone on my project, moving away from AWS (budget) to self-hosted and the local models are so much faster than in the past. Beyond LLMs, having embeddings, image, video, audio gen available is crazy.

Running locally is the bar; it's hard to make these things a service which scales.

New comment by anax32 in "Why does paper fold so well?"

anax32 — Mon, 15 Jun 2026 11:08:36 +0000

I'm not getting any work done today

thank you!

New comment by anax32 in "21 years and counting of 'eight fallacies of distributed computing' (2025)"

anax32 — Mon, 15 Jun 2026 10:36:35 +0000

* You will have logs

Always gets me

New comment by anax32 in "Report on an Unidentified Space Station"

anax32 — Fri, 12 Jun 2026 10:40:03 +0000

This was a big moment for me, but I now believe it's fictional.

Thanks Ballard

New comment by anax32 in "When AI Crosses the Line: The Matplotlib Incident"

anax32 — Mon, 01 Jun 2026 15:19:23 +0000

This stuff is better than TV

New comment by anax32 in "[dead]"

anax32 — Mon, 01 Jun 2026 14:53:23 +0000

HTTP content negotiation was a good idea which decouples content from form, but only as far as format selection.

Generative models are able to transform content between media types, and it feels like the original intention can be completed -- the server generates the appropriate form at request time, rather than serving a pre-rendered one.

A concrete example: an Image2Depth model estimates the depth of a scene from a standard image, encodes that information in the response, and returns it to clients capable of rendering depth -- 3D displays, VR headsets, and so on. The content is the same; the form is specialised to the client capabilities.

New comment by anax32 in "[dead]"

anax32 — Tue, 12 May 2026 15:03:00 +0000

I've been thinking of LLM prompting as execution, which makes context building a compile step -- different build systems, different output quality. Chunking is the simplest compiler; artefact construction that infers schema and resolves contradictions is a much richer one. The article maps the current spectrum.

Show HN: JSON-logic-path – JSON logic with jsonpath multi-value resolution

anax32 — Sat, 18 Apr 2026 19:34:15 +0000

The reference Python implementation (nadirizr/json-logic-py) hadn't been updated for modern Python. We forked it to fix that, then added a vars operator backed by JSONPath for multi-value resolution -- useful when branch conditions need to evaluate model outputs or array-valued fields.

Not part of the standard; the var operator is unchanged so existing rules work without modification.

Comments URL: https://news.ycombinator.com/item?id=47818825

Points: 2

# Comments: 0

New comment by anax32 in "Digital Archaeology: where does Bitcoin's 21M come from?"

anax32 — Sat, 18 Apr 2026 19:31:37 +0000

an interesting aside: the block rewards for each epoch are the percentage issuance of that epoch.

Content vs. Curation

anax32 — Fri, 17 Apr 2026 11:21:57 +0000

Article URL: https://bayinformationsystems.substack.com/p/content-is-dead-the-curation-inversion

Comments URL: https://news.ycombinator.com/item?id=47804703

Points: 2

# Comments: 0