the8472

42 ms·

751.
by the8472 3y ago
> And yet the moment they do that some lawyer submits a bunch of hallucinations to a court and they get in the news. That's the lawyer's problem, that shouldn't make it OpenAI's problem or that of its other users. If
752.
by the8472 3y ago
The difference to doomsday weapons is that we can build the weapons first and then worry about using them. With an AGI building one alone might be sufficient. It could become smart enough to unbox itself during a training run.
753.
by the8472 3y ago
Browsers don't have TCP/UDP/etc. sockets, so a WASI impl running in one can't provide that either.
754.
by the8472 3y ago
> It's a chance to create a superior operating system runtime interface. But they're not doing io_uring style syscall queues.
755.
by the8472 3y ago
Try to extrapolate. If nobody has an open port to which a connection can be established, how will the network work? Trackers don't enable hole-punching, existing peer connections do[0]. And hole-punching is hardly a reliable measure to
756.
by the8472 3y ago
It's a pretty big change, so yes, some impact is unavoidable. But the GPL could still operate on the margin of free vs. mandatory fees. If the company doesn't want to go through the headache of paying each open source project it u
757.
by the8472 3y ago
Meh. Patents used in standards have similar mandatory licensing and many standards can be used for things the inventor dislikes. And you can donate the money if you find its source unpalatable.
758.
by the8472 3y ago
We could go a step further and make licensing mandatory. Anyone can use the work, but they have to pay a part of their profit to the originator if they do. As a side-effect this completely legalizes non-profit fan works.
759.
by the8472 3y ago
That's just saying how people tend to use it, not what the technology does. Namespaces/Cgroups can be used both ways, to isolate a single process (or even partially-isolate individual threads) or entire userland stacks. The questi
760.
by the8472 3y ago
I read the quote you're replying to as "you can also do this under linux", i.e. it didn't dispute that freebsd jails can do it. But it's not a distinguishing feature of jails.
761.
by the8472 3y ago
Calling "VM-like" is not helpful because containers have also been called that, it also doesn't explain anything. Is it running a separate kernal image? Does it do hypercalls to the host kernel? If not then I wouldn't ca
762.
by the8472 3y ago
That cuts two ways. If everything they said was on the record there would be no point in putting on a show since it will immediately turn out the show is all fake. Imagine if politicians acted like normal people, negotiating things in nuanc
763.
by the8472 3y ago
Loading assets JIT for the next frame from NVMe hasn't been a thing for 20 years though. Different kinds of latency floors. What I was asking is whether LLM inference can be structured in such a way that only a fraction of the weight i
764.
by the8472 3y ago
Even an A6000 tops out at 48GB while you can attach terabytes of RAM to server-class CPUs.
765.
by the8472 3y ago
> I don't think there was a market for it before LLMs. At $work CGI assets sometimes grow pretty big and throwing more VRAM at the problem would be easier than optimizing the scenes in the middle of the workflow. They can be optim
766.
by the8472 3y ago
You can also add journal_async_commit,noauto_da_alloc > maybe disable flushes with https://github.com/stewartsmith/libeatmydata overlayfs has a volatile mount option that has that effect. So stacking a volatile ove
767.
by the8472 3y ago
It's all tradeoffs.
768.
by the8472 3y ago
Afaik true rebalancing isn't in the works. Some limited add-device and remove-vdev features are in progress but AIUI they come with additional overhead and aren't as flexible. btrfs and bcachefs rebalance leave your pool as if you
769.
by the8472 3y ago
> zfs is technically great It's only great due to the lack of competitors in the checksummed-CoW-raid category. It lacks a bunch of things: Defrag, Reflinks, On-Demand Dedup, Rebalance (online raid geometry change, device removal, d
770.
by the8472 3y ago
Well speak of the devil. It happened. https://www.theregister.com/2023/05/09/intel_oem_private_key...
771.
by the8472 3y ago
memory, internal iteration and scratchpads can and have been added. Foundation models don't include this yet, but it's not unexplored territory. Online learning is more difficult but I'd argue that various fine-tuning and con
772.
by the8472 3y ago
I think Figure 8 in the GPT4 report already implies this for RLHF, i.e. the base model is fairly well-calibrated about its confidence but the RLHFed one becomes less calibrated which ought to make confident, incorrect statements worse. Idk
773.
by the8472 4y ago
> Yes, you can get compromised through firmware updates. It's not just about updates through regular channels but about evil maid attacks with signed malicious firmware. This vector would be avoidable if you could sever the trust re
774.
by the8472 4y ago
You cut off the "including the firmware signing keys" part, that's critical. Otherwise the vendor could be coerced or subverted to sign malicious firmware which then subverts your system at runtime. Only when you can bring yo
775.
by the8472 4y ago
If you could replace all the vendor keys (including the firmware signing keys) with your own then TPM could make sense. But today's TPMs don't support that.
776.
by the8472 4y ago
That doesn't change the fact that reducing carbon emissions would help with the problem .
777.
by the8472 4y ago
Surely those abductions would be exciting in some sense.
778.
by the8472 4y ago
> but no system is perfect, right? If your data isn't going to be exhaustive or unbiased samples you could at least show the omission rate. E.g. "there were 100 tenants in period X, 1 left a review".
779.
by the8472 4y ago
> but that would not really help them But it would? It wouldn't solve the problem, but it would be part of not making it worse.
780.
by the8472 4y ago
Martian invasion fleet disguised as a moon. https://www.project-apollo.net/mos/mos215.html

More ›