catawampless

Karma: 17

catawampless 28 May 2026 0:34 UTC
1 point
0
in reply to: kman’s comment on: leogao’s Shortform
Even ignoring AI, current expectations are that world population will peak in ~60 years at a population of ~10 billion. This seems totally workable with relatively incremental improvements in technology for resource management, energy extraction, and habitat preservation. This is only bolstered by much of our future energy supply growth coming from non-extractive solar rather than fossil fuels, which we continue to find new sources of regardless.
When you add on the fact that almost all population growth in the next 50 years is due to occur in sub-Saharan Africa, where there is plentiful land and the only limiting factors are energy supply to tame it, we’re actually pretty well situated re: overpopulation concerns.

catawampless 26 May 2026 16:35 UTC
10 points
2
on: Cognitive Security as an AI Safety Cause Area
If “realistically simulating human impacts in silico” means something like high-fidelity models of how humans respond psychologically to AI interaction, that seems like one of the higher-risk cases of safety research also being capabilities research. A good sim for “how AI affects humans psychologically” is a good sim for “how to affect humans psychologically, using AI”.
The hardest part of CogSec evals is figuring out an eval meaningful enough to detect dangerous effects, while not being meaningful enough to hill climb on. Maybe eval is too benchmark coded, when actually what we want is comprehensive user surveying, similar to Anthropic’s What 81,000 people want from AI, but conducted independently, tied to usage data in a privacy-respecting way. One of the genuinely good uses of AI here may be qualitative research at scale: preserving the nuance of how people feel about and relate to AI, rather than immediately collapsing it into 1–5 survey responses.

catawampless 19 May 2026 20:35 UTC
3 points
0
in reply to: davekasten’s comment on: davekasten’s Shortform
From my research, the current meta is the Luggable https://www.cleanairkits.com/products/luggables. Basic insight is that by going from a 99.97% removal HEPA filter to a 93% MERV-13, you can run your filters at a higher speed, quieter, with much better power efficiency, and cycle air through the filter much more quickly due to the higher fan speed. Most HEPA grade filtration misses the mark because they’re far too loud running at full blast, so a quieter, faster, less efficient filter actually does a better job. The LuggableXL is about as loud
The one missing piece is if you live in a high VOC area, you’ll want some sort of activated carbon filter. The main options I’ve seen here are:
- An AustinAir, expensive, but has 15 lbs of activated carbon, recommended by CleanAirKits
- An AirFanta 3Pro
- You can tape an activated carbon sheet to the inside of your Luggable
Personally, I have an Aerolamp + LuggableXL running at all times in my house, and the combination has served us extremely well according to our air quality sensor and anecdotally not getting sick after parties anymore.

catawampless 17 May 2026 20:25 UTC
1 point
0
on: Is ProgramBench Impossible?
Downstream unit testing
Instead of autogenerating the unit tests to fill any gaps, we can see if software that depends on it still works.
Simon Willison discusses an early version of this employed by the StrongDM team. For context, they build Digital Twins of all the software their system depends on, and have agents run QA testing continually against those digital twins. To verify correctness of the digital twins, one technique they use is verifying client libraries that make use of their dependencies still work. Direct quote from the article:
I did have an initial key insight which led to a repeatable strategy to ensure a high level of fidelity between DTU vs. the official canonical SaaS services:
Use the top popular publicly available reference SDK client libraries as compatibility targets, with the goal always being 100% compatibility.
I could see a version of this where a benchmark like ProgramBench is built, but the bar is that libraries that make use of those dependencies still pass their integration test suite with a swapped out implementation

catawampless 16 May 2026 20:34 UTC
5 points
0
on: Models finding software vulnerabilities is not the primary source of cybersecurity risk
The social/logistical aspects of cybersecurity vulnerabilities will accelerate greatly due to AI. I’d expect the response from tech-savvy organizations will be to increase the pace of software delivery—a long standing trend for other reasons. Continuous deployment, forced autoupdates, focused research on fraud and suspicious activity detection.
The main risks are around organizations that structurally cannot increase their pace. Think banks, aviation, medical systems, drug manufacturing, areas where because the risks of vulnerabilities/defects has historically been extremely high, we intentionally require verification and slow down their development pace. If a vulnerability is discovered in these areas, they’re precluded from responding with a patch the way SaaS vendors can.
I hope one of the major aims of the major labs diffusion focused deployment orgs is to help these institutions in particular, it’s probably one of the higher ROI places to be involved, considering the surface area for vulnerabilities is generally smaller and a concerted vulnerability search could prevent these issues from occurring in places where we can’t respond.

catawampless 17 Mar 2025 15:17 UTC
3 points
0
on: How I’ve run major projects
I’d love to see anonymized version of one of these project documents. For example, with your plan to victory, do you assign actual dates to each concrete step? Or just assign different sections of the plan into milestones that have actual dates? What form factor works? Is a checklist okay, or should it be a spreadsheet?
Regardless, fantastic article, this type of tacit knowledge is usually only made explicit for those within the walls of major companies, so it’s nice to have as a reference.