I just got a bachelor’s in CS
find me anywhere in linktr.ee/papetoast
papetoast
Karma: 1,122
Yeah, Hacker News is really noisy, that’s why I just linked that specific comment. I didn’t actually read the Navy paper (or any paper regarding CO2), so thanks for the additional point.
I don’t know whether cognitive effects exist at reasonable ppms, but it seems like any real effects would be too small to worth caring for optimizing productivity, so I stopped investigating.
...did you use an LLM to write this comment. (Pangram flags your comment at 100% AI)
regardless, I simply updated to that I don’t need to worry about CO2 in most normal settings and I don’t have an opinion on whether it has a negligible effect at something like 2000ppm. Obviously at some point CO2 will have an effect on cognition just because an extremely high CO2 level (~40000ppm) is lethal.
Two notable AI math results recently
2026/07/28 | mainly GPT 5.6 Pro—David Turturean’s Proof of The 2-adic Absolute Galois Group (FrontierMath Open Problems: Solid Result):
Using mostly my voice, I solved one of @EpochAIResearch’s FrontierMath Open Problems: finding an explicit presentation of the 2-adic Absolute Galois Group—open for more than forty years, now with a full proof in collaboration with David Roe, the problem’s proposer.
Still, the problem originating in 1982 shouldn’t be taken as the sign of a major enigma. The same mathematician noted the problem was “basically attention-bottlenecked”. When submitting the problem, Roe suggested the main difficulty was that “the answer is likely to be messy”.
2026/07/30 | Tencent Hy—Not sure what is the name of this thing
Tencent Hy: For a finite set of integers (A), how much faster can (|A+A|) grow than (|A-A|)?
A 1969 theorem gave an upper bound of 2 for the exponent. For more than 50 years, the best constructions barely exceeded 1.1.
With help from our research agent Hyra and the Hy3 model, we found an explicit construction showing that the optimal exponent is exactly 2.
Paper: https://arxiv.org/abs/2607.27199
Hyra blog: https://hy.tencent.ai/research/hyra
Formal proof: https://github.com/linhaowei1/sum-diff-proof
I recently saw this on Hacker News (https://news.ycombinator.com/item?id=48787625) and I find it pretty convincing. When I ask AI, they are also skeptical about both the effect of CO2 and the quality of the paper you linked. Both ChatGPT and Claude thinks it is overselling the findings.
We’ve been studying the impact of CO2 for decades, at much higher levels than you see in office buildings and have never recorded any cognitive impact (until many thousands of PPM) until the Satish study in 2012 and a handful of other studies that Satish was involved in.
If you think about it for a second these studies can’t be accurate. If they were, you’d see differences on SAT scores of hundreds of points depending on building ventilation. You’d see huge variations between taking the SAT in the springtime when the windows are more likely to be open or in the winter.
You’d expect to see massive performance differences across nearly every metric between regions that use AC vs those that depend on opening windows.
And we do not see anything like that.
The difference between Satish-involved studies and other studies is extreme, too. Their first 2012 study tested levels at 600, 1000, and 2500ppm. In many categories they observed the 2500ppm group receiving “dysfunctional” ratings in their tests. Even their 1000ppm group saw significant drops.
This sparked the panic about CO2 levels that led to people buying CO2 sensors and thinking that not unusual CO2 levels were actually destroying their ability to think. Many conclude that this has been happening all of their lives and to everyone around them who is unaware.
It triggered a lot of follow-up studies. Confusing, some of those (like the Harvard one everyone cites) included Satish, meaning they weren’t independent despite coming from different organizations.
The really interesting thing is that many of the follow-up studies that don’t include Satish have even used CO2 levels much higher than the 2012 study that caused the panic. Here is one I grabbed at random that went all the way up to 15,000ppm and failed to find any significant changes: https://pubmed.ncbi.nlm.nih.gov/29789085/
There was also a lot of CO2 research before Satish come along that failed to find significant effects at these relatively low levels. It has been researched in the contexts of air quality for submarines and space shuttles by militaries and NASA because keeping the crew of those operating optimally is important, clearly.
I haven’t moved. I have heard some good things about Keepass but I really don’t want to handle the data syncing part. Don’t use LastPass though.
LastPass notifies users of yet another data breach (via HN)
LastPass says password vaults not affected
The information accessed was limited to standard business contact information and related customer relationship management (CRM) data, including customer names, phone numbers, email addresses, and physical addresses, as well as support case data and sales-related data.
it is just a pr i made, not merged yet
My guess is that it incorrectly included the base karma from yourself when you post/comment. Claude confirms it and found 2 more karma counting issues and 2 more unrelated problems.
Pangram made it very clear that the supporting evidence is a different model than the one that generates the probabilities. I would guess they removed it because people still keep getting confused.
I don’t think that conjecture is too famous, so it shouldn’t be surprising given the past results from AI. I see it as (effort prompt → very difficult discovery) → (non effort prompt → somewhat difficult discovery)
GPT 5.6: “Recent papers explicitly call it a “famous conjecture,” but that fame is local to combinatorial optimization and approximation algorithms.”
lol I think it is LW’s mention UI messing up my paste
Edit: yeah it is. Fixed
PSA I have been (and will keep on) maintaining a list of notable mathematical/scientific results from AI:
if there is anyone waiting for July links, I cancelled the mid-July issue because I procrastinated too much. I will make a full month issue at the end of July.
I would be surprised if your post did not get blocked by automatic moderation filtering on AI written content. For your information, you have to use an “LLM content block” for all AI writings.
This is the LLM content block.
It may also be a good idea to attach your original writings in a collapsible section, because the AI-assisted translation reads like AI, with annoying jargons, phrases, and sentence structures that people on LessWrong will definitely notice. I am engaging in this comment thread just because you are new here, and also you seem to have actually slightly edited the AI text. If it is a post, I will likely read one or two paragraphs, notice that it is AI written, then downvote and close the tab.
Edit: I’m not saying your post is bad, but statistically AI writing don’t perform well on LessWrong. Your edits aren’t significant enough that the writing looks human written at a first glance.
This is a collapsible section
...
I skimmed a couple page of that feed x 3 times over different time of day. It is hard to say whether it is positive or negative evidence that people regularly leave comments because there is still no data on how often people read older posts. Though, for the record, the ratio of comments on new vs old posts is higher than I expected, at like 1 in 5-10. I thought people are going to be at least 4x less likely to comment on old posts from expecting fewer engagements but this seems wrong.
Related: The hostile telepaths problem