Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
Klaus23
searching PlanetScale…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
14 ms
·
61.
▲
by
Klaus23
2y ago
Yes my mistake, I read your answer to mean that you think that the model could fit into the memory with the help of efficiency gains. I would be sceptical about increasing efficiency. I'm not that familiar with the subject, but as far
62.
▲
by
Klaus23
2y ago
You fundamentally misunderstand the bottleneck of large LLMs. It is not really possible to make gains that way. A 405B LLM has 405 billion parameters. If you run it at full "prescision", each parameter takes up 2 bytes, which mean
63.
▲
by
Klaus23
2y ago
So far, the law doesn't have a majority. Even if they manage to push it through the Commission and Parliament, it's very unlikely to survive contact with the European Court of Justice.
64.
▲
by
Klaus23
2y ago
> The ICC just gave Netanyahu more strength. Israelis would likely rally around their leader. Netanyahu will be emboldened to do maximum damage on his way out. What more does he have to lose at this point? His support is waning on both s
65.
▲
by
Klaus23
3y ago
They may be using very high ratio gearboxes to get the torque they need out of the limited space available, which makes them less backdrivable.
66.
▲
by
Klaus23
3y ago
The risk also depends on where the potential pinch point is. I doubt that many people put their hand on the top of the roof when the tailgate closes. The Cybertruck's danger zone is exactly where you would put your hand if you were lea
67.
▲
by
Klaus23
3y ago
I don't think the two are at all comparable. Google's External offers program is the ability to pay a percentage to allow apps installed through Google Play to link directly to external offers. Apple's Core Technology Fee, on
68.
▲
by
Klaus23
3y ago
I didn't know about the "free of charge" clause. Apple's Core Technology Fee will be absolutely destroyed.
69.
▲
by
Klaus23
3y ago
Yes, unfortunately this happens regularly, but with open source software it is at least possible to fork it. We often see forks when there are major disagreements. Not all of them survive, but if the original is bad enough, the chances are
70.
▲
by
Klaus23
3y ago
If you want to limit the initial impact of a malicious extension, a mandatory hold or slow rollout would be more appropriate. There is no need to bother normal users if they would never inspect the code anyway. If some users want to inspect
71.
▲
by
Klaus23
3y ago
If the software you are using is so bad, or the distributor so untrustworthy, that you would classify it as malware, then I think it is time to switch to an alternative. For example, it is now quite feasible to use only open source software
72.
▲
by
Klaus23
3y ago
I don't think manual updates would solve this security problem. The new owner would just have to delay the activation of the malicious parts of the software. No one is going to check the binary of an extension or try to replicate it if
73.
▲
by
Klaus23
3y ago
This is more of a technical problem. If your update either breaks something or leaves gaping security holes, then there is no good solution. I think I would rather inconvenience a customer by turning off functionality than leave a bad vulne
74.
▲
by
Klaus23
3y ago
It was not my intention to defend large corporations and their sleazy practices. I just wanted to say that the average user cannot be trusted with an easy option to ignore updates, especially when it comes to security. Users will do things
75.
▲
by
Klaus23
3y ago
There should always be an option to turn off automatic updates (unless we are talking about a corporate network), but the option should be opt-in and require some initiative on the part of the user. If the option is presented together with
76.
▲
by
Klaus23
3y ago
Anyone who has had to administer anything user-facing will tell you that some users will ignore any warning. Updates need to be automatic and mandatory. You can give them a grace period, but you have to force the issue after a while, or use
77.
▲
by
Klaus23
3y ago
What are the trade-offs with INT8? I thought even the INT4 loss of accuracy was small and the INT8 loss almost unmeasurable.
78.
▲
by
Klaus23
3y ago
There is a leaked prototype based on llama, but the rest are their own models.
79.
▲
by
Klaus23
3y ago
> My plan is to swap a hydrogen motor in i guess once mature the tech and consequently ban diesel. Good luck with that. Hydrogen in cars is dead, and even if it weren't, the tanks would take up so much space in a car not designed fo
80.
▲
by
Klaus23
3y ago
We are talking about 85 MB four times a year to keep the application up to date and running smoothly. Don't be ridiculous.
81.
▲
by
Klaus23
3y ago
It works for me, but the selected text is superimposed on top of the normal text in a different size.
82.
▲
by
Klaus23
3y ago
The demo is pretty cool, but the mobile interface could be a parody of bad interface design. The text box at the top is hard to reach if you want to open the keyboard, which automatically closes, or press the button to send the question, an
83.
▲
by
Klaus23
3y ago
It is definitely based on neural networks: https://www.youtube.com/watch?v=RpiN3LyMLB8
84.
▲
by
Klaus23
3y ago
Sounds like they were relying solely on their neural network path prediction, which failed when the truck was dragged at an odd angle. A simple lidar moving object segmentation, which doesn't even know what it's looking at but can
85.
▲
by
Klaus23
3y ago
I was talking about a point discussed in the article. There is an attack on OTP because the attacker can change the message if there is no hash-like authentication of the message's correctness. I asked if there was a pen-and-paper way
86.
▲
by
Klaus23
3y ago
Is there a good pen-and-paper OTP authentication mechanism?
87.
▲
by
Klaus23
3y ago
1000km is pretty good, but there are some caveats. There are still major technical hurdles, such as the problem of burning batteries. They will not have the mass budget to protect the battery as well as they do now. What can only be dangero
88.
▲
by
Klaus23
3y ago
OpenAI have been sitting on GPT4 for months, and on the basemodel even longer. I would not be surprised if they did some or all of the distillation of the model with GPT4. Mixtral is 56B combined, if we subtract a little for MoE inefficienc
89.
▲
by
Klaus23
3y ago
I think it was leaked that it is 20B now.
90.
▲
by
Klaus23
3y ago
There are at least half a dozen other errors in the video, and that's without me looking for them. I would not be surprised if this was just wrong.
More ›