maxrmk

5 ms·

31.
After ten years, will Apple give up on privacy?(14.ms)
2 points by maxrmk 2y ago|1 comments
32.
by maxrmk 2y ago
I can only speak for myself, but I would absolutely rather work burned out in tech than work in the trades again. Drywalling was absolutely destroying my body, mind, and will to live. Mobility is also a factor - I’ve had high stress tech jo
33.
by maxrmk 2y ago
I feel very fortunate to have worked other jobs before my first in tech. My last was hanging drywall and I will never forget how awful it was. I haven’t loved every part of every tech job I’ve worked, but I’ve always chosen to be there. I t
34.
by maxrmk 2y ago
If it's Mustafa vs Sam Altman, I know where I'd put my money. As much as I like Satya Nadella I think he's made some major hiring mistakes.
35.
by maxrmk 2y ago
I think this probably isn't as big of a deal as people are making it out to be. But I find a certain kind of joy in Mozilla being judged on the worst possible interpretation of their terms of service, since they do that to others _all
36.
by maxrmk 2y ago
Ok this is way too long to read in one sitting, but it looks incredible? I've been looking for resources like this for """real""" model training at scale. The authors have worked on BLOOM (the first >10
37.
by maxrmk 2y ago
I think this misses the mark. We know LLMs can learn facts. There are lots of other benchmarks full of facts, and I don't expect that saturation of this benchmark will mean we have AGI. The missing capabilities of LLMs tend more in the
38.
by maxrmk 2y ago
I'm not sure what they intended this to apply to. LLM based systems don't change their own operation (at least, not more so than anything with a database). We'll probably have to wait until they fine someone a zillion dollars
39.
by maxrmk 2y ago
I _think_ that's a factor of using rule based reward functions and not actually a feature of GRPO? The original formulation of GRPO from deepseek math uses a neural reward model that is trying to predict human rankings of responses, an
40.
by maxrmk 2y ago
There’s some disagreement over whether or not GRPO is the important part of deepseek or not. I’m personally in camp “it was the data and reward functions” and that GRPO wasn’t the key part, but others would disagree.
41.
by maxrmk 2y ago
I read this to get up to speed on RL for LLMs. If you have limited time, I’d recommend reading the entire first chapter covering the basics and some terminology and then section 5.4 on RL for LLMs. I struggled a lot with the first chapter,
42.
by maxrmk 2y ago
Seconding the other commenter, I’d love a more detailed version of how you do this. I looked into doing something similar with voice memos and couldn’t figure it out.
43.
by maxrmk 2y ago
Could be the case, I’m not familiar with their specific tokenizers. IIRC llama 3 tokenizes in chunks of three digits. That seems better than arbitrary sized chunks with BPE, but still kind of odd. The embedding layer has to learn the semant
44.
by maxrmk 2y ago
Cool! Contrary to some of the other posters I think this definitely counts as deanonymization, or at least is close enough. How anonymous would satoshi be today if we had his location to within 250 miles? Repeated applications of this attac
45.
by maxrmk 2y ago
Yeah that’s my understanding of the root cause. It can also cause weirdness with numbers because they aren’t tokenized one digit at a time. For good reason, but it still causes some unexpected issues.
46.
by maxrmk 2y ago
This is clever, and I got a good laugh out of their example video. The demo UI of "Double click here" isn't very convincing - I bet there's a version of this that gets people to double click consistently though.
47.
by maxrmk 2y ago
I'm vaguely amused by the headline of "requires no external power" right above the image of it sitting on top of (and plugged in to) a giant portable battery.
48.
by maxrmk 2y ago
As much as I hate it, I use twitter to follow a bunch of people who work at fair/openai/etc and that's been a pretty good source. There's also a "daily papers" newsletter from huggingface, but it's pretty
49.
by maxrmk 2y ago
sure, but once it's trained there isn't a running maintenance cost
50.
by maxrmk 2y ago
I was wondering about this one too... > At best, they say, Orion performs better than OpenAI’s current offerings, but hasn’t advanced enough to justify the enormous cost of keeping the new model running. wdym "keep it running"?
51.
by maxrmk 2y ago
This is really interesting. Based on their wikipedia I can see they collect a lot of RF traffic - are IMEIs identifiable with the raw data captured that way? I'm surprised they are not encrypted. I say this as someone who knows nothing
52.
Show HN: Simple Synthetic Medical Data(demo.talcapi.com)
2 points by maxrmk 2y ago|0 comments
53.
by maxrmk 2y ago
Totally– This was in the early days of open source .net, when it was still called .net core. We had just moved onto GitHub and I don’t think microsoft had quite figured out how to manage that organizationally. There is a point of time there
54.
by maxrmk 2y ago
I thought linus had lost it for a second there, until I saw it's just named after him and not something he created. I generally disagree - I think that having many developers with a shallow understanding of the whole codebase scales le
55.
by maxrmk 2y ago
That ethos runs through everything the WA team does. I learned a bit about it when I was at fb and I was beyond impressed by how few servers WA used for its core infrastructure. It's a really well engineered product.
56.
by maxrmk 2y ago
> That's 50 people serving the needs of 1,000,000. Doing a little bit of division, that means that every single member of the Flutter team is responsible for the needs of 20,000 Flutter developers! That ratio is clearly unworkable f
57.
by maxrmk 2y ago
Tried it! I really like the idea, but I think the clue generation could use some work. Every clue ended in "in games", and honestly most of them were not really game related to start with. For example the clue "Place in games
58.
by maxrmk 2y ago
There's a newer version of this model that takes a really cool RL based approach: https://arxiv.org/pdf/2408.08152
59.
by maxrmk 2y ago
other talc founder here. ask us anything!
60.
Show HN: SearchBench – A Benchmark for AI Search(github.com)
6 points by maxrmk 2y ago|0 comments

More ›