HN user

topynate

1,086 karma
Posts5
Comments283
View on HN

The first two people you mentioned co-wrote a paper titled "Towards decolonising computational sciences" and certainly anyone at DAIR would also be in that ideological cluster. I don't think an absolute majority of the signatories have those sorts of associations, but a good many do. What's puzzling me is why so many are Dutch?

It's not perfect but it does make speech easier to understand, which is the ostensible purpose. There's a slight uncanny valley effect. Latency requirements are very tight, and Indian English is actually timed differently to American or British English, so it's a hard problem.

The company should consider targeting RP instead of General American. In my experience attempting the latter with a slight Asian tinge makes you sound like you're trying to be a WASP, while speaking the former with the same small errors makes you sound like you studied PPE at Oxford.

That's possible in general but not for this application; a chat interface to an LLM isn't very useful unless you can tell it whatever you want—including GDPR personal data—and then pick up the thread of conversation later.

If the topic is tissue damage from sharp ice crystals, it's pretty handy to draw the distinction between cooling methods that cause that and ones that don't.

I'm a little surprised that Shenzhen doesn't seem to be churning out ITAR-busting anti-jamming systems. The tech is pretty old by now and the market is there.

Space-based geoengineering is technically superior to stratospheric aerosol injection, but the latter is practical almost immediately, and is likely to be much cheaper no matter what. In my view that outweighs concerns about the ozone layer.

It has also been pointed out that launching enough space-shades with chemical rockets would itself deposit a considerable amount of particulate matter in the upper atmosphere, so to be really atmospherically 'clean', one would need to complement geoengineering with a radical new launch technology.

Perelandra's devil – the Un-Man – struck me on first reading as an excellent early depiction of a hostile, alien form of intelligence, superior but purely instrumental. Lewis was very early in working out the implications of that – nowadays the Rationalists and a lot of others would agree that there can be entities with superhuman intelligence that don't intrinsically value their intelligence, and that such beings would have almost irresistible persuasive ability if given the opportunity. (Lewis differs in also giving the Un-Man genuinely supernatural abilities with which it attempts to overawe the protagonist.)

"it regarded intelligence simply and solely as a weapon… Thought was for it a device necessary to certain ends, but thought in itself did not interest it. It assumed reason… externally and inorganically…"

No frost on the sample either. The only way I can think of to fake this in camera is to make the "sample" out of a strong magnet, and make the "magnet" a hollow shell concealing a chilled piece of YBCO!

What's the disclosure that breaks the NDA this tweeter is supposed to have had? I searched every quote in the thread and they're all already public. Tweet #11 shows as "unavailable", but the archived version at https://threadreaderapp.com/thread/1681673980682305541.html refers to a "Code Yellow" that did leak, but in March: https://www.businessinsider.com/shopify-issues-code-yellow-f...

I tried to find people complaining about being left on hold by shopify support and got nothing on twitter. All the complaints are about the account payments being on hold, i.e. shitty but par for the course risk/compliance issues.

The remaining thread content is the sort of inference and speculation that anyone with the right political bent could have made, given a few hours to research the company.

[dead] 3 years ago

That's impressive, although the second iteration isn't a uniform improvement over the first. Did it work?

From their discord:

It would be interesting to know why you chose those FLOPS targets, unfortunately it looks like the models are quite under pre-trained (260B tokens for 13B model)

We chose to train these models to 20 tokens per param to fit a scaling law to the Pile data set. These models are optimal for a fixed compute budget, not necessarily "best for use". If you had a fixed parameter budget (e.g., because you wanted to fit models on certain hardware) you would train on more tokens. We do that for our customers that seek that performance and want to get LLaMA-like quality with a commercial license

Which is the point made elsewhere in these comments, e.g. https://news.ycombinator.com/item?id=35344192, and also usefully shows how open Cerebras are. They're pretty open, but not as much as they would be if they were optimising for filling in other companies' moats.

The proximate cause is that Substack was founded by people less likely to censor, who have so far kept control of the company. It may be more meaningful to say that Substack was founded in 2017, and Medium was founded in 2012, and the sort of censorship we're talking about was much more salient in 2017 than in 2012, so Substack was in some sense selected to be more against that sort of censorship.

Trying to work out what's going wrong in the cases this article describes, such that the principle "you can't patent something that exists" apparently doesn't apply. Is it that an independent breeder develops a certain strain, say red salad, and then a big company makes a modification to that strain, and patents the modification, such that the independent breeder is prevented somehow from continuing to develop red salad? That still doesn't make perfect sense to me but it makes more legal sense than "company got patent on what you already made, you can't make it anymore", which is just not how patents work.