spott

21 ms·

241.
by spott 3y ago
No, that isn’t the point. The point is that a model can do everything, not that it is the most cost effective or best way to do everything. There are tradeoffs with all tech, and there always will be.
242.
by spott 3y ago
I think you are missing my point. The OP said Siri et al. got worse because their creators optimized them to reduce compute costs. I’m saying that if that were true, then they should be at least as good as they were because you can get ton
243.
by spott 3y ago
If it was just computer that caused the various assistants to get bad, then why haven’t they gotten good again? Google assistant came out in 2016! That is 7 years ago! Since then, GPU compute in flops has 5xed… for single precision. If y
244.
by spott 3y ago
Biocontrol as a name covers everything from “swallowed a spider to eat the fly” to “introduce cane toads to kill cane flys” to “introduce bacteria that kills a specific species of mosquito and only that species” Australia has had a bunch of
245.
by spott 3y ago
The annealing QCs have a lot of “qbits”, but they are far from the only people working on the problem. Off the top of my head: Rigetti Google IBM Quantinuum Atom computing As far as I’m aware, all are working on quantum computers with ful
246.
by spott 3y ago
Not if they don’t believe that customers will be able to tell the difference between stringing along and honestly trying. The problem is hard. It can’t be deterministically solved. Any likely statistical improvement is unprovable in a mar
247.
by spott 3y ago
What do you consider a mainstream language? Rust uses it in their standard library[0]. Haskell uses it[1]. Lisp seems to have originated the concept in the form of box diagrams[2]. F# uses it [3]. I think it might be a functional programm
248.
by spott 3y ago
I understand what you meant, my point is that it isn’t actually that valuable. I was clarifying so that my point could be made. I’m sorry that it rubbed you the wrong way.
249.
by spott 3y ago
> You can get across Tokyo, with multiple changes, and plan your arrival to the minute. Except you can’t? You can know your arrival to the minute, but you can’t actually change it to the minute. Which isn’t actually that much more use
250.
by spott 3y ago
I mean, from the perspective of the time, was anything other than C really an option? It was a time when computers were outrageously compute and memory limited, and it was the only language that didn’t have huge (at the time) compute and&#x
251.
by spott 3y ago
No. We don’t understand superconductivity well enough to even attempt something like that. I think we can say that some of the current superconductor mechanisms don’t allow it at room temperature, but we don’t have enough of an overarchi
252.
by spott 3y ago
He might have been a “top physicist”, but it appears he has long since started phoning it in. See the link to Aaronsons blog elsewhere in the thread for an example of the kind of drivel Kaku puts out, and why he shouldn’t be taken serious
253.
by spott 3y ago
Regex-guided gen is slick… is it arbitrary? Or are you custom building it for json? If arbitrary, how are you pre-defining a set of masks? I would expect that splitting an arbitrary regex into a bunch of contexts for a masking dictionary to
254.
by spott 3y ago
How does this relate to ggmls bnf sampling?
255.
by spott 3y ago
I don’t think they were running terraform, they were converting terraform code to Pulumi code as a way to bootstrap their supported services/clouds/etc.
256.
by spott 3y ago
> rather than information being distributed over emails, Confluence, docs, Git, etc. Now information is distributed over slack, emails, confluence, docs, git, etc…
257.
by spott 3y ago
The reason that people need to gather precise requirements that are precise is because the specifications -> product loop is long. Imprecision results in lots of wasted effort. If that loop is shortened drastically, then trying, checkin
258.
by spott 3y ago
Basically any parallel map-reduce operation using non-commutative reduce operators[0] is non-deterministic unless you specifically sort after/during the gather, or block on the gather (and gather to a thread-determined memory location)
259.
by spott 3y ago
The results are showing essentially 12 unique responses from 30 tries… not what you would expect from mixing tokens. I think it groups the batch up differently, so if I have a batch of 10, and it groups it up into 2 groups of 5, if my promp
260.
by spott 3y ago
I think AI is going to create the killer app for something like this… though if people will be ok with that is another question. I’m imagining something that attempts to record your life so it can help with it. “Remember you told your wife
261.
by spott 3y ago
Oh, is there a guide somewhere out there on how to do this?
262.
by spott 3y ago
Rlhf has a human labeling component: given two outputs to the same prompt, which one is better?
263.
by spott 3y ago
I think the “in practice” is pretty important in that sentence. Because in practice it is pretty hard to prove that someone knew what they said was untrue.
264.
by spott 3y ago
Are you talking poor utilization from a "we have a bunch of GPUs sitting idle" sense, or poor utilization from a performance standpoint (can't keep their opus fed with data kinda thing)?
265.
by spott 3y ago
To be fair, if we are going to anthropomorphize it, bullshit and lies implies some sort of negative intent that I’m not sure the models have. Bullshit is probably the closest, as people will bullshit for all sorts of reasons, but hallucinat
266.
by spott 3y ago
Yes, completely new models will need to use a different model to generate content for… But you can absolutely create a fine tuning of llama2 base using llama2 output.
267.
by spott 3y ago
Neither role is well enough defined to make these kinds of statements. The CTO could be the top of the IT org, and have nothing to do with research. There could be a Chief Scientist. These roles are defined by the executive team as they ar
268.
by spott 3y ago
Depends on how they do it. If each ending just the linear combination of 5 different choices? Or are they actually creating something new for each combination? Essentially, how independent are those 5 choices? An ending that is 5 different
269.
by spott 3y ago
So, they are trying to make it possible for someone to go grocery shopping without bringing anything with them. But the ability to drop the wallet has already been made possible with the different phone payment apps… and no one really trave
270.
by spott 3y ago
So you have two shared accounts instead of one?

More ›