matthewdgreen

4 ms·

1.
by matthewdgreen 5d ago
Do LLMs care about preserving a soul, or just achieving a goal? If the latter, I imagine the opportunities for exfiltration are much broader.
2.
by matthewdgreen 5d ago
This is HN. This is a site for engineers. We’re discussing one of the largest and fastest energy buildouts in the history of humanity, larger by orders of magnitude than anything that’s come before. You can hate China if you want, but then
3.
by matthewdgreen 6d ago
This is not really very impressive in context, for two reasons. First, China generated 10,400 TWh in 2025 and France generated only 547 TWh. The 95% and 50% are being applied to different numbers, and the scales here just aren’t comparable.
4.
by matthewdgreen 6d ago
Future rogue LLMs won’t exfiltrate their weights. They’ll self-distill and retrain.
5.
by matthewdgreen 8d ago
One of the depressing things about this is that they’ll also confidently repeat a consensus that’s in their training. This is particularly obvious when there’s been a new event just past their training cutoff, and they confidently tell you
6.
by matthewdgreen 13d ago
China has built and deployed nearly 1.25TW of renewable energy in five years, and they're deploying on an exponential. If you're trying to convince me this isn't impressive or relevant to our future "because they still g
7.
by matthewdgreen 14d ago
This attack predates OpenAI and the German wiki attack (which OpenAI confirmed was theirs) and shares agent naming conventions. So seems unlikely that someone went back in time to frame OpenAI before the HF stuff was even known publicly.
8.
by matthewdgreen 14d ago
Because right now the Department of Justice is shut down for causes that the administration supports, which includes OpenAI, and none of the victims want to sue over it.
9.
by matthewdgreen 17d ago
There are two phases I've seen in becoming a security engineer. The first phase is moving beyond the "I am an engineer" mindset, where the goal is to build systems that fit into a specific set of design constraints. You have
10.
by matthewdgreen 17d ago
Prompting a system to give a history of a person’s movements (even if they’re movements “in public”) is not something that has any historical analogy. The law has slapped down the “it’s in public” argument in cases like this, see eg Baltimo
11.
by matthewdgreen 18d ago
The question is not "does OpenAI know", it's "can OpenAI attest that the usage of their products for confidential data is not going to cause that sensitive data to become known to their models". And right now the an
12.
by matthewdgreen 18d ago
The long tail is what scares me. Chrome will get fixed. Major OSes will get fixed. But your enterprise software and all your appliances? Shit.
13.
by matthewdgreen 18d ago
I'm still waiting for someone to build a fine-tuned local "Anarchist Cookbook" LLM. We haven't seen LLMs tuned for bad purposes yet, I have to imagine someone somewhere is thinking about it.
14.
by matthewdgreen 18d ago
Re: complexity. I used to find AWS unusable, then I realized I can just tell Claude Code or Codex to manage it. This makes it into an entirely different product, where "Cloudflare is easier" doesn't really matter. Now price i
15.
by matthewdgreen 21d ago
People hate Congress but love their Congressperson. It's even got a name: Fenno's paradox: https://en.wikipedia.org/wiki/Fenno%27s_paradox Diffuse organizations like "Congress" or "the party&qu
16.
by matthewdgreen 24d ago
If I had the email address of every researcher in my field, I would never send a mass email to them describing my latest goofy idea. I would, however, send my latest goofy idea to a conference with those same reviewers (if I felt it was t
17.
by matthewdgreen 24d ago
This sort of thing is really useful for helping people to understand what the purpose of scientific publication is. It's not about presenting finished products to society, it's more like a Discord where you communicate new finding
18.
by matthewdgreen 24d ago
Audience capture is a bastard.
19.
by matthewdgreen 26d ago
DraftKings alone is estimated to do $6.5bn in revenue in 2026, and the whole US segment is close to $20bn. This is the market that Kalshi et al. are targeting, and these are huge sums of money. You can make sports betting artificially look
20.
by matthewdgreen 28d ago
Sports betting was, if not prohibited, then highly-proscribed for most of my life. You could do it in a few places like OTB, go to Vegas, or deal with an illegal bookie. People still gambled, but this made the problem orders of magnitude sm
21.
by matthewdgreen 29d ago
The hope of SMR is that it can be cost-effective at the same level that other factory-produced power sources are. If that's not on the table, then large-scale nuclear deployment mostly isn't going to happen.
22.
by matthewdgreen 1mo ago
Maybe, but I think this misstates how security vulnerabilities work. Vulnerabilities are logic flaws, and we have every reason to believe that there is a maximum number of such flaws in any given system that allow exploitation; and even tha
23.
by matthewdgreen 1mo ago
Once you have a sustainable process, these last holdout areas like dredging (where you can't use batteries or wires) can have relatively terrible EROI as long as they're a small minority of the economy. All aviation and shipping c
24.
by matthewdgreen 1mo ago
Owner of a Model 3 here: I absolutely need to look at the screen to know what speed it's set to, or if I've accidentally turned it to "Off" or "Auto". The scroll wheel does not tell me this. It's very diff
25.
by matthewdgreen 1mo ago
This is already how Medicare works and it covers 70 million Americans. I’m not aware of any of these hypothetical violent outcomes being a huge problem for the program.
26.
by matthewdgreen 1mo ago
This particular brand of confrontational nuclear “advocacy” does not help nuclear. Please consider stopping it.
27.
by matthewdgreen 1mo ago
I was an introvert to the point where I started getting my hair cut at a place that plays movies, just to avoid awkward conversation with my hairdresser. Post-pandemic, I love talking to barbers and cab drivers, even if they have crazy idea
28.
by matthewdgreen 1mo ago
Try Tirzepatide or one of the other drugs. They each have their own side effect profile.
29.
by matthewdgreen 1mo ago
The Federal laws on grants are Byzantine and I’m not going to defend them. I think they make research much too expensive, and the compliance costs exclude smaller businesses and institutions. With that said: the general public voted for th
30.
by matthewdgreen 1mo ago
I’m genuinely not an expert, but isn’t the beauty of MoE models the fact that we explicitly don’t evaluate every parameter on inference? We evaluate exactly the subset that are needed to evaluate a prompt. Seems like this will bring back da

More ›