Mavo is currently trying to solve a much smaller scope of problems: can it help alleviate some stress that comes from a busy family schedule.
For my family, this means making it not only accessible to me (a developer), but also my wife (who’s not), and potentially other family or helpers (who likely aren’t). That means I’m designing it with accessibility in mind: there’s a full featured GUI so you don’t have to talk to the AI to make changes if you don’t want to, events have built-in owners so you can assign someone or even round robin things and get notified if no one has an event covered, you can share a link with the grandparents to give them details that stay up to date, and, most importantly, you get these things out of the box without having to tinker (but you can go deeper if you prefer).
I’m not aiming for it to be a fully autonomous agent, but I’m definitely thinking about ways to make it more proactive. Events in the system, for example, have an internal notification that can wake Mavo up to check on things without pushing an alert to the house unless it’s needed.
If the PhDs don't cost very much more than your equivalent of factory-technicians but still get the job done, why wouldn't you do that, at least in the blunt case before cost control rears up?
Distillation originally meant matching the distribution of the student model to the teacher model using something like a KL divergence.
When you instead fine-tune the student on the samples from the teacher, which is what people mean by distillation today, you are in effect doing a monte-carlo version of the same thing. While in theory this is higher variance, given modern setups where the student and teacher are both large and are RLd heavily (leading to a sharp teacher distribution), and given that you typically use lots and lots of samples, it ends up OK.
Distillation is used to build part of a data set for fine-tuning (loosely interpreted). Advanced model traces are useless if you don't have a base model that is good enough to be improved by them.
He doesn't have to prove that he didn't send them ... that's the prosecution's burden. (In the U.S. ... I know nothing about Canadian law.) But a competent lawyer could figure out how to prove it: get ahold of the IP address from which the messages were sent. Oh look, it was in California! (The IP address they had was from the email address with which Klayme had registered his Kik account with the mistranscribed name.)
Also a competent lawyer would have done a scrupulous audit of the chain of reasoning that led to his client, which would have revealed the username mixup.
Why do you assume that the IP that sent the messages was recorded. They the police went after the email at least suggests that it wasn't in the data that Kik provided during the police investigation.
And the reasonable doubt would come from someone else could have used my account, does the service's account ID match the account ID for my user name, nothing was found on my devices, did the IPs match my known IPs, when did the access take place vs where was I, etc.
But that is all speculation/opinion/non-first hand knowledge, you can't actually say any of those things in court without an expert witness. Only expert witnesses can testify to things that they didn't witness themselves, lawyers and lay witnesses cannot.
Also don't think you can simply get the opposing expert witness to give this kind of testimony on cross examination. During cross examination you can only ask them about things they testified to during direct (and a few other things such as their qualifications). So if they never testified anything about your IP during direct then you can't ask them about whether or not your IP matched during cross.
If you don't have the money to pay for an expert witness then you better just hope that the jury already knows about IP addresses and actually realizes that IP addresses matching was never brought up during the trial (they won't btw, the general public knows nothing about these things).
Think about it this way, all this techno mumbo-jumbo about account IDs and IP addresses may as well be black magic as far as the jury and court is concerned. So only a verified expert witness can testify about them. If we allowed otherwise then every trial would be nothing but testimony about how various magical or religious principles mean that you can't possibly be guilty (or vice versa).
They could ~halve their training costs with collusion.
Instead of each paying full price to generate a SOTA model in competition, they could share the result and split the cost. This gets even simpler if they merge.
"Why can't we just give steroids to people with muscular dystrophy?" [0]
HN Discussion: https://news.ycombinator.com/item?id=28306559
https://trevorklee.com/why-cant-we-just-give-steroids-to-peo...