khafra

12 ms·

91.
by khafra 1y ago
That's not a comic, and it's not artificial superintelligence: https://arxiv.org/abs/1606.05915 Whatever AI comes up with by 2030 is going to be much more clever and unexpected.
92.
by khafra 1y ago
Hario ceramic coffee grinders also work for pepper.
93.
by khafra 1y ago
I hope there's a mad scientist somewhere, making a cross-genetic venus flytrap that also produces capsaicin and nicotine.
94.
by khafra 1y ago
LLM model interpretability also uses Sparse Autoencoders to find concept representations ( https://openai.com/index/extracting-concepts-from-gpt-4/ ), and, more recently, linear probes.
95.
by khafra 1y ago
"LLM whisperer" folks will confidently claim that base models are substantially smarter than fine-tuned chat models; with qualitative differences in capabilities. But you have to be an LLM whisperer to get useful work out of a bas
96.
by khafra 1y ago
I mean, it's not the biggest thing you can do; you could start selling to the government, become a cleared contractor, and then you could require a USG security clearance for job applicants. I would call the on-site interview and&#x2
97.
by khafra 1y ago
Seems like a simple cybernetics problem, especially if it's point-to-point. If there's many different inverters on the same circuit, you might need something like a directed graph, but there should be simpler, more local solutions
98.
by khafra 1y ago
Need to speak whale to make this AI-powered scheme work? Sounds like a job for AI: https://news.ycombinator.com/item?id=43680899
99.
by khafra 1y ago
How should the speech coach stimulate you when it detects you using a particularly euphonously or impactfully?
100.
by khafra 1y ago
> > ... you make a request ... > Sounds like the work of mustering up instructions... like programming... Again, try Deep Research. You make a vague request, and it works with you to make it specific enough that it can deliver a pr
101.
by khafra 1y ago
Yeah, seems like chemical rockets, or a railgun or trebuchet or something to get it up to speed first would be a lot more efficient than ground to orbit with just a laser.
102.
by khafra 1y ago
Nice! This is a much-needed space for security tooling, and I appreciate that you've put some thought into the new attack vectors. I also like the combination of signature-based analysis, and having an LLM do its own deep dive. I expec
103.
by khafra 1y ago
> the need for translating vague desires from laypeople into a precise specification will remain What makes you think LLMs will never be able to do that? Have you tried any of the various DeepResearch products? The workflow is that you m
104.
by khafra 1y ago
In practice? You are issued a special, secure phone expressly for connecting to the network which is not physically connected to the internet. Many details of e.g. CsfC are unclassified; you can read the architecture pdfs.
105.
by khafra 2y ago
You can joke about Microsoft Teams not being a real messaging platform, but running it on a network that's physically separated from the Internet is quite effective at keeping random journalists out of your chat groups.
106.
by khafra 2y ago
Sounds like you're saying the distinction shouldn't be between instructions and data, but between different types of principals. The principal-agent problem is not solved for LLMs, but o1's attempt at multi-level instruction
107.
by khafra 2y ago
Need a predictive version of this that tags politicians in tweets expressing my stance on whatever is being debated, right before they use their phones on livestreams.
108.
by khafra 2y ago
I've seen grapes dropped in a bowl of vodka in the freezer, then strained out and put in a bowl after they freeze. Good way to make your co-workers chill out a little.
109.
by khafra 2y ago
Yudkowsky just mentioned that even if LLM progress stopped right here, right now, there are enough fundamental economic changes to provide us a really weird decade. Even with no moat, if the labs are in any way placed to capture a little of
110.
by khafra 2y ago
A single 3min 670nm morning exposure may be more valuable in latitudes where the sun does not rise until after 8AM.
111.
by khafra 2y ago
Broad advice I've seen for already-employed people looking for a change: Switch employers and roles at different times. e.g., Learn enough to be useful, then talk to the security guys at your company. Prove you're useful and trust
112.
by khafra 2y ago
That was a helpful list, in your second comment downthread. What are your top 3 metrics that perform the best on the greatest number of those features that make cosine distance perform poorly?
113.
by khafra 2y ago
Robin Hanson recently related the story of a firm which actually made data-driven decisions, back in the early 80's: https://www.overcomingbias.com/p/hail-jeffrey-wernick It went really well, and then nobody ever
114.
by khafra 2y ago
> Just 20+ years ago this sort of sympathy seeking broadcasting action was associated with mental health illness, like Munchausen Biproxy. Do you have a reference for the claim that the diagnostic criteria for Munchausen By Proxy (or Fac
115.
by khafra 2y ago
> I have done no such thing. I have never said that welfare receipts "uniformly" do anything. I am just saying that SOME will not use their money wisely and our choice is either to provide them with additional benefits or let t
116.
by khafra 2y ago
> An agent, or something that has agency, is just something that takes some action, which could be anything from a thermostat regulating the temperature all the way up to an autonomous entity such as an animal going about it's busin
117.
by khafra 2y ago
Right; and houses release those toxic fumes when they're on fire. If poisoning other structures with your house fire was an actionable tort, insurance would incentivize owners (and builders) to make houses that were less likely to catc
118.
by khafra 2y ago
Ok, you're guessing that current welfare recipients are uniformly imprudent and lack all virtue, rather than having some sizeable minority who are simply temporarily down on their luck. Let's grant your assumption, for the sake o
119.
by khafra 2y ago
> So he think humans are intervening once every 1-2 miles to train the Waymo Just to make sure we're applying our rubric fairly and universally: Has anyone else been in an Uber where you wished you were able to intervene in the driv
120.
by khafra 2y ago
Seems like properly-priced insurance should fix that, if you add a little bit of tort for the owners of the hazardous housing?

More ›