jagraff
Joined 370 karma
- jagraff parentThis is a good point - having your code broken up into standalone units that can fit into working memory has real benefits to the coder. I think especially with the rise of coding agents (which, like it or not, are here to stay and are likely going to increase in use over time), sections of code that can fit in a context window cleanly will be much more amenable to manipulation by LLMs and require less human oversight to modify, which may be super useful for companies that want to move faster than the speed of human programming will allow.
- Interesting, it seems that the actual surface material of walls and/or furniture makes a large difference in how long VOCs stick around, due to differences in surface area at the microscopic scale.
I have a couple HEPA filters in my house that hopefully keep particulate exposure down. Does this mean that I have to run them longer? That I need more of them continuously running to keep exposure to VOCs low?
- Polymarket gives a >50% chance of the tariffs being ruled illegal, not that they would be refunded - the market only gives a ~8% chance of the tariffs being ruled illegal AND and order to refund: https://polymarket.com/event/will-the-court-force-trump-to-r...
- I think the median country GDP is something like $100 Billion
https://en.wikipedia.org/wiki/List_of_countries_by_GDP_(PPP)
Models are expensive, but they're not that expensive.
- I think there is perhaps a different conclusion that I come to - email is not the right tool for long discussions with multiple points of disagreement, because it is, generally, a linear medium, which makes it difficult to maintain different threads without careful formatting by every author in the email chain.
I am not sure if there exists a good tool for threaded discussions with multiple different focus areas - something like git but for conversations?
- Very interesting. From my read, it appears that the authors claim that this attack is successful because LLMs are trained (by RLHF) to reject malicious _inputs_:
> Existing large language models (LLMs) rely on shallow safety alignment to reject malicious inputs
which allows them to defeat alignment by first providing an input with semantically opposite tokens for specific tokens that get noticed as harmful by the LLM, and then providing the actual desired input, which seems to bypass the RLHF.
What I don't understand is why _input_ is so important for RLHF - wouldn't the actual output be what you want to train against to prevent undesirable behavior?
- Moving, ever expanding circle: https://tixy.land/?code=%28x-10*t%2521%29**2%2B%28y-10*t%252...
- I don't know about enzymes, but bacteriophages (viruses that target bacteria) are species specific, and have been used successfully as antibacterial treatments: https://en.wikipedia.org/wiki/Phage_therapy
- Abandoning Ukraine risks emboldening Russia to continue their conquest of Eastern Europe - which, ultimately, increases the risk of nuclear war anyway, only with a stronger Russia with even more leverage. Being nuclear armed should not give a country license to seize any territory they wish.
- Every dead body you've ever seen is teeming with life. Now, it's probably impossible for a novel life form to arise in a dead body, because the existing life (bacteria) that takes over has been optimized by billions of years of evolution to consume the resources around it better than any novel self-replicator possibly could. But that's not really evidence that a self-replicator couldn't get going if there weren't much better ones already around.
- Movement without energy is possible due to diffusion. Imagine you had a room with cellophane separating one side from another. Each side has a different gas, but both gasses are at the same pressure and at room temperature. Then, the cellophane is removed. Without adding any energy to the system, these gasses will mix until the whole room is a perfect mixture of the two gasses, simply because they both diffuse through the entire room. Something similar happens to allow viruses to move through your body (and, if they can be aerosolized, through the air).
- If we could create (and safely implant) artificial ligaments that are even close to human ligaments that would be amazing for anyone with ligament injuries. Once you damage a ligament it never regrows, and current artificial ligaments don't last very long and can cause other health issues. At least for UCL rupture, the current state of the art surgery (Tommy John surgery) involves taking a ligament from some other part of the body and replacing the damaged ligament, which is obviously a major surgery and still doesn't work nearly as well as the original ligament.
- I think the idea with campaigns like these is not to actually root out spies - it's to get people to self censor because they're afraid that their neighbor is going to report them to the government. If everyone is afraid of saying any forbidden ideas, nothing "subversive" can spread because everyone lives in too much fear to spread it. Knowledge of this campaign is more important than the function of the campaign itself
- You don't "need" any medication for most colds. But people want (and deserve access to) cold medication that treats the symptoms that they find unpleasant. Also, some people are prone to sinus infections - taking pseudoephedrine when you start getting congested (before developing a sinus infection) is crucial to preventing infection and avoiding damage to your nasal passages.