stymaar

5 ms·

91.
by stymaar 1mo ago
A company who made their business out of stealing intellectual property from the entire mankind, stealing other researchers' unpublished work, how surprising, really.
92.
by stymaar 1mo ago
You're confusing with Anthropic . They are definitely feeding off the fear of US control from US companies and institutions, but you can't blame them for reaping the benefits of Trump's lunacy. Amodei and Altman on the other
93.
by stymaar 1mo ago
Where can I sign up for a monthly subscription on OpenRouter instead of paying the real API price?
94.
by stymaar 1mo ago
> Honestly, the internet has made folks so lazy. Or maybe people realize that revolutions, and the civil war that almost always ensue, are a very high price to pay, for a very uncertain reward. Talking about the Arab spring, look what it
95.
by stymaar 1mo ago
First: it's a joke. Second: the US justice system is also a joke. I could fight Google in justice and win in my country (France) and in pretty much any other developed country because a working system doesn't let big actor DoS sma
96.
by stymaar 1mo ago
Plot twist they'll be running Claude Bible or GPT-Galaxy with max reasoning on, against my dumb local Qwen-0.8b and they'll run out of money before I do.
97.
by stymaar 1mo ago
This. When the US government decides to label you “terrorist organization”, it means that you are now targeted by the most potent state actor on earth, which will go as far as blowing up weddings in your neighborhood to stop you. So unless
98.
by stymaar 1mo ago
That sounds like a dumb strategy because if non-evil people sit on individual pieces of the puzzle waiting to solve it in full google loses most of the advantage of having a multi-layer system…
99.
by stymaar 1mo ago
> Minimum Target SDK Enforcement Blocks installation of apps that target ancient versions of Android and legacy APIs. Funny to list a user-hostile change as the first “improvement” that comes to mind. I guess GP should have said “series
100.
by stymaar 1mo ago
No goalpost was harmed in the above comment.
101.
by stymaar 1mo ago
That'd be too restrictive, but to get improvement in a particular domain you definitely need to train specifically for it. Just cramming more random internet text in a bigger model has stopped being a effective way of scaling since at
102.
by stymaar 1mo ago
It's a lifetime on things that are explicitly being trained on ! But small models like Luna didn't magically become more powerful than SotA models on stuff that they weren't explicitly trained on with a dedicated RL-pipeline
103.
by stymaar 1mo ago
I don't think it's a great rebuttal, quite the opposite actually: it's the same prompt with the most obvioy tweak you can imagine and the results are all the same. That definitely looks like it's been something the model
104.
by stymaar 1mo ago
What if you asked a front/rear/top-side view of the scene (pelican or lemur)?
105.
by stymaar 1mo ago
When you see how good the output of the Luna model without reasoning is compared to SotA just a year and a half ago, it's pretty clear that it's been trained on explicitly.
106.
by stymaar 1mo ago
> Getting security updates for issues that are not marked high/critical. These are not your typical RCE, but they are used in exploit chains. Aren't those back-ported for a while?
107.
by stymaar 1mo ago
It's kind of ironic that the word alignment, which used to mean this very problem in reinforcement learning, has been perverted to mean something very different and then fell out of fashion (in favor of “guardrails” in the mouth of th
108.
by stymaar 1mo ago
Only if you double layers by layers instead of the whole stack (which IIRC is what nanbeige is doing). To put it simply, if you have 3 layers A-B-C then A-A-B-B-C-C requires more compute but not more memory bandwidth, but A-B-C-A-B-C requir
109.
by stymaar 1mo ago
Also 2504.09762: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces! https://arxiv.org/abs/2504.09762
110.
by stymaar 1mo ago
The CCP is also well aware that it's facing a dramatic demographic contraction so they better prepare for that.
111.
by stymaar 1mo ago
> Deepseek proposes RLVR as a way to get around the lack of $ they have to produce human reasoning trace data. What was the difference between what deepseek did for R1 and what OpenAI did for o1?
112.
by stymaar 1mo ago
Same as fishing or hunting, or foraging blueberry or blackberries while hiking, there's some primal reward in gathering your own food (occasionally, that is).
113.
by stymaar 1mo ago
I guess if you had a billion dollar you could assemble an ML team and fine-tune a frontier LLM specifically for that purpose through RL, and it could work better than your local pharmacologist, but I doubt anyone would bother doing that. So
114.
by stymaar 1mo ago
Yup, it varies between runs (depending on the seed, most likely), but since the knowledge is here it wouldn't be too hard to nudge the model in the right direction with grpo alone.
115.
by stymaar 1mo ago
> Anyway, i just wanna get on record that i predict an ai bubble pop event in the next 12 months. If it happens on Sep 2nd 2027 you'd be wrong though.
116.
by stymaar 1mo ago
It's not uncommon to use that as a measurement in addition to the ILO's definition.
117.
by stymaar 1mo ago
AFAIK, the “large” qualifier came when transformers allowed to scale the size of language models compared to the recurrent models that where in fashion before. And although BERT isn't large by today's standard, it was large enou
118.
by stymaar 1mo ago
It's likely somewhat true in recent times, but it wasn't a thing for most of human history as Capitalism didn't exist. It's rather “The actual actor of History is the structure of society”, with the said structure evolvi
119.
by stymaar 1mo ago
Because Apple isn't just a trillion dollar corp, it's a Cult: https://www.businessinsider.com/nyu-professor-says-apple-is-...
120.
by stymaar 1mo ago
> a complicated psyop It's not a “complicated psy-op”, promoting viral stories is the basics of marketing in the age of social media…

More ›