plasticchris
plasticchris on Hacker News
after thinking about it more, in the attention there is some positional encoding, so I think it's just that tokens are attending to this as well.
Makes sense. If you are reading something you need it to make progress, if you are writing you are finished.
That’s just pattern matching. Most people don’t do this level of prep so they assume you didn’t either. There’s usually no ill intent, and people are maybe a little embarrassed when you point it out, but if you are tactful it’s fine.
I think it’s trained into them so they emit a stop token with enough window left to generate a compaction summary
Usually the government too.
> only get running tallies every 50 ballots means you get vote secrecy
Not if everyone voted the same on something…
Look deeper. Or more niche/esoteric
If you ask the machine you will get the median answer and do what everyone else is doing.
in your own life, ask yourself, what do I need? what could make my life, or the lives of those I love, better?
When I was a young man reading gauss, and thinking before I turned the page, I see a result here, something meaningful! and then I turn the page, and of course it is there. There is no shame in this.
A long time ago, another student of mathematics described it to me as climbing a mountain. No single step is hard, but you must make many steps to get to the point that no one else has made a step.
and yet... it feels like we are actually a one party state most of the time.
I don’t think you got the point. The liquidator in this case is highly incentivized to keep these off the market
Seems legit, I trust them. They even filed an amicus brief with scotus
Well, if they were bought on credit, and must be returned when you go insolvent to the company that makes the chips, whose market - and therefore stock - would collapse in value should the market be flooded with used equipment, and said company already dealt with a similar situation in the past few years... I think the answer is obvious.
Why do you think your life is wasted if some computer has a smaller number next to your name?
It's foundations all the way down. The difference is people have reality as our reinforcement learning environment.
These are companies famous for following the rules when it comes to handling other people’s data and IP after all, totally a non-issue and they would never violate contract law
This is why I think llms are a killer app for Linux desktop. They’ve been trained on Linux very hard, and it cleanly solves the “how do I make it do $thing” problem since everything is open and the llm can manipulate it. For example: Sound not working? Just tell the llm.
I put knoppix on a cd and since the filtering was implemented via a proxy server that was entirely voluntary to use… also bypassed the filtering.
Once I burned a bunch of Linux boot cds with unreal tournament on them and took over a computer lab to have a lan party. That was fun until the lab monitor got spooked that his monitoring system showed all the computers as turned off when clearly the lab was full…
To focus your point: we fear that with ideas having easier replication that the noise in the signal will increase.
But this is what people said about the printing press. And the internet. Less effort to spread the memes. Has it been a net negative? Maybe in some respects. Can you stop it? Probably not.
The interesting thing is that ai actually helps me learn more effectively than ever before. Each session can become a personalized little seminar that dynamically tests me as it presents ideas and concepts and I attempt to apply and extend them.
I think we need to move past recall as the test itself, and tasks as the gate we judge academic achievement by. These were only ever an approximate measure, which is why we still have things like thesis defenses.
But imagine the final being an interview with an agent that is designed to probe your mastery? From my experience so far it seems we could measure much more directly the educational outcome and also accelerate the learning at the same time. But this assumes those involved want to learn, which widespread cheating suggests is not the case. Maybe the real answer is that we shouldn’t attempt to force curiosity but instead cultivate it where we find it, allowing those who reject it to be redirected until they find something engaging.
Anywhere the nerds or artists go, if they are good, eventually becomes overrun with such types. Just look at the state of Silicon Valley and y combinator, and gentrification in general…
It will only get harder. From the trenches in tech it’s hard enough to get people who should know better to put in the work to understand the ai output.
For education I can’t imagine. You might have to resort to actual discussion in classrooms instead of homework, with the accompanying need for both more educators and higher quality therein. Which isn’t going to be easy when education is fully under regulatory capture (at least for K-12 in the USA)
Yes and much like the technological leaps of the past, everyone’s standard of living goes up.
Because they need the local infrastructure to support them, which would be difficult and expensive to reproduce out in the middle of nowhere.
Misses the point. Now that things are easier, we have greater agency, not lesser.
Probably not when you consider the training cost and upkeep expenses, not to mention the depreciation…
Qwen’s 3.6/3.8 27b actually has some algorithmic advantage when it comes to the kv cache size needed, so it actually needs less memory at equivalent context. My experience using both in vllm supports, with considerably more overhead in context size on these models than Gemma 4 31b, and better performance in most tasks I’ve tried on both models.
My earliest coding memory is messing with the gorilla code to get an advantage playing this against my brother…