David Metcalf
dlmetcalf.bsky.social
David Metcalf
@dlmetcalf.bsky.social
infosec, risk management, software/enterprise arch/design, climate, tech, AI/ML, health security, pandemics, clean air, novid.
You're welcome to believe whatever you like, or you can go read the research yourself (there's hundreds of studies on it) and/or reproduce them. But either way, it's still too small to have any meaningful impact.
October 10, 2025 at 7:21 AM
I ended up building a purifier that sits in my lap (or on a table top, as air curtain for meetings) and quietly does around 300m3/h CADR (with approximately laminar output). 1.4kg with 4h replaceable battery. But, sorry, it's not available. Maybe later.
October 6, 2025 at 5:17 AM
Very well fitted respirators are extremely effective. But most people, even among the most cautious, still remove them at times. E.g. meals or rehydrating on long-haul lights (unless drinking shakes through a Sipmask etc). Or in bad weather, with limited options, eating indoors at restaurants.
October 6, 2025 at 5:08 AM
E.g. If you're visiting a poorly ventilated GP clinic, sitting in the waiting room with 6 other people, then that portable is 150 times too small. The outlet area is also too small and not laminar flow, so entrains unfiltered air along with it too.
October 6, 2025 at 5:01 AM
That portable purifier does approximately 1.8 litres per second on max.

ASHRAE 241-2023 specifies 30 l/s/person (108m3/h) equivalent CADR (total of fresh & filtered) for transportation waiting areas, eg airport boarding lounge. 15 l/s/p for an office. 45 l/s/p for a healthcare waiting room.
October 6, 2025 at 4:54 AM
Should be good for about 150m3/h. (>20x more clean air).
October 5, 2025 at 8:55 AM
Cut a Starkvind panel in ½. Arctic P14 Pro on back. 140mm finger grille (4x M3 nuts & screws). $2 PWM fan module (speed dial). USB-C to 12VDC PD module. Fold cardboard box (join w/ hotglue or borax doped wheat paste). Seal & waterproof cardboard w/ shellac, beeswax & metal stearate dissolved in IPA.
October 5, 2025 at 8:51 AM
Sorry to tell you, but its 4 cfm CADR, equivalent to 6.8m3/h, is pointless. Too small to have any (meaningful) real world impact.

Any random high static pressure computer fan with small replacement filter panel strapped to it, USB fan header adapter, and phone power bank, gives 10x better output.
October 5, 2025 at 5:36 AM
"But we didn't think the face eating leopards we sold out to would eat *OUR* faces!!"
October 5, 2025 at 5:22 AM
Like the antivaxxers who say, "But we're not against vaccines!". Yet they are. Public health was dismantled from the inside. By cranks who sold out and sided with pro-disease, anti-health cookers. Now crying about their jobs after they've lobbied against their own field. 🙄
October 5, 2025 at 5:21 AM
No great loss, since pro-COVID 'let it rip' charlatan Andrew Hayen was Deputy Head of School (at least until June 2023 anyway, and continued having a job there).

People like that are why the healthcare system collapsed.
October 5, 2025 at 4:37 AM
Perhaps they meant seasoning, sprinkled everywhere?
July 17, 2025 at 8:51 AM
(Or maybe we're talking at cross purposes and those using it uncritically (as opposed to critically) should assess; if you're the one encouraging the critically question part 🙂).
June 4, 2025 at 10:38 PM
Why do you think it means that? You should critically assess your understanding of the word "critically".

LLMs are extremely useful tools. I use them daily. Just learn what they actually are first, so you know how to do it appropriately, and have reasonable expectations.

bsky.app/profile/dlme...
Not interesting. Sad that ppl anthropomorphise, have irrational expectations & get emotional abt outputs of linear algebra over matrices trained w/ lossy fitting stochastic decent on undisclosed datasets (often synthetic) & evals; probabilistically inference sampled to out of distribution prompting.
June 4, 2025 at 8:49 PM
This video is a year old now, and much has happened since, but it's still an excellent overview of what these AI systems are (especially large language models):

youtu.be/zjkBMFhNj_g
[1hr Talk] Intro to Large Language Models
YouTube video by Andrej Karpathy
youtu.be
June 4, 2025 at 11:39 AM
But if you don't learn what the underlying abstractions of the technology are, how it works, and its limitations; then you're likely to be surprised and sometimes very disappointed too.
June 4, 2025 at 11:27 AM
It is possible to train models to imitate some reasoning (in a sense) and to recognise mathematical logic. E.g. huggingface.co/deepseek-ai/... can now prove maths theorems at expert competion level.

You can also train models use external tools (e.g. executing code its generated).
e.gm
June 4, 2025 at 11:27 AM
To paraphrase Andrej Karpathy (I hope, but all errors mine): all LLM output is hallucination and sometimes those hallucinations happen to be useful.

I find them often enormously useful tools, but never lose sight of that.
June 4, 2025 at 11:09 AM