News for hackers. Stuff that matters.
Re:FCC upholds denial of Starlink's $900M rural broadband su... by AlexandrB on Wednesday December 13, 2023 @02:09PM (#38632586)
This is important to consider when judging Tesla's engineering capability, but not when judging Elon Musk's highly optimistic promises.

Either Elon is aware of the fact that FSD is really hard and he's lying repeatedly about being able to deliver it "later this year". Or he is not aware that FSD is as hard as you say and needs to "hit the books" and understand the problem better. I don't really see an "out" here after so many missed promises.

Re:FCC upholds denial of Starlink's $900M rural broadband su... by AnthonyMouse on Wednesday December 13, 2023 @05:50PM (#38635321)
Software timelines are notoriously hard to forecast because you're essentially trying to predict how long it will take to do something nobody has done before.

So you look at what's left to do and as long as you don't run into any unforeseen issues you say, looks like about six months from now.

Sometimes you don't run into any unforeseen issues and it turns out to be about six months.

Sometimes you do, and then it takes longer. So a year later you found an unexpected problem that delayed you by a year. Someone asks you how long you think it will take again, that issue delayed you by a year but that was a year ago, so you say, looks like about six months from now.

Eventually the estimate will be true but nobody knows when because nobody knows how to make an accurate estimate. Because you have to know how hard it is to do in order to know how long it will take, which nobody knows until after somebody has actually done it.

Re:FCC upholds denial of Starlink's $900M rural broadband su... by seadan83 on Wednesday December 13, 2023 @06:39PM (#38635806)
Just reminded of the an old trope for estimates, either:

- less than 3 months

- 3 months

- 6 months

Anything less than 3 months - is an accurate estimate. Otherwise, 3 months is the amount of time it takes to do _anything_, right? If it's a really hard problem, then we say, okay, it's 6 months. At the end of the day, those 3 months and 6 months means "NFC", which goes back to the estimation technique that everything is either a "1", "TFB", or "NFC". 3 months = TFB, 6 months = NFC

On a more serious point, we (as software engineers) generally feel we can solve almost any problem with 3 months worth of resources and effort. 3 to 6 months sometimes (IMO) can be a bit of the scariest estimate to give or receive.

For reference & more on NFC/TFB/1 estimation technique: https://pm.stackexchange.com/questions/26218/how-to-implemen...

Re:FCC upholds denial of Starlink's $900M rural broadband su... by Wytwwww on Wednesday December 13, 2023 @07:41PM (#38636321)
> Software timelines are notoriously hard to forecast

Perhaps. Of course, we all know for a certainty that this had almost nothing to do with you're describing. The people how were actually designing the software knew that it won't happen in six months. I doubt Musk wasn't aware of that, so the only alternative that he was lying.

Re:FCC upholds denial of Starlink's $900M rural broadband su... by AnthonyMouse on Thursday December 14, 2023 @02:54PM (#38646398)
> Of course, we all know for a certainty that this had almost nothing to do with you're describing. The people how were actually designing the software knew that it won't happen in six months.

How do we know this?

Waymo has had this "almost" working for years. In theory you could get from the status quo to actually working in a short period of time, if you're sufficiently smart or lucky and have the right resources.

Or you could hit a wall for years.

The things currently happening with LLMs and image generation were the sort of things many people said were many years away, if not impossible, but we got there by basically throwing a lot of compute at some clever but relatively simple ML algorithms. Nobody knows until you try.

It's obvious that his estimates are optimistic, but optimism isn't inherently malicious.

Re:FCC upholds denial of Starlink's $900M rural broadband su... by Wytwwww on Friday December 15, 2023 @05:28PM (#38659606)
> How do we know this?

Some life experience + common sense?

> In theory you could get from the status quo to actually working in a short period of time,

In theory we could build a rocket that would fly a person to the moon in 6 months.

> if you're sufficiently smart

If that's the case, you wouldn't be making such bold claims (or rather actually believing in them yourself)

> LLMs and image generation were the sort of things many people said were many years away,

Does that include the people who were actually working on them in OpenAI?

> It's obvious that his estimates are optimistic, but optimism isn't inherently malicious.

In some cases it's either malice or stupidity/delusion.

Re:FCC upholds denial of Starlink's $900M rural broadband su... by AnthonyMouse on Saturday December 16, 2023 @03:30AM (#38662738)
> Some life experience + common sense?

If you asked people with "life experience + common sense" if generative AI would soon be possible six months before it actually was, they'd have said no and been wrong.

> In theory we could build a rocket that would fly a person to the moon in 6 months.

Kennedy said it would be done by the end of the decade, people called him crazy, and then it happened. The people saying it couldn't be done were still saying it right up until the day it happened.

> If that's the case, you wouldn't be making such bold claims (or rather actually believing in them yourself)

I don't think he, or anyone, believes that it will definitely happen in six months. He's giving the optimistic end of the range, but everybody knows that.

> Does that include the people who were actually working on them in OpenAI?

Before they had them working internally? Yeah, it did. Nobody knew it would work until it actually worked.

Re:FCC upholds denial of Starlink's $900M rural broadband su... by Wytwwww on Saturday December 23, 2023 @02:14PM (#38747138)
> the AI would soon be possible six months before it actually was, they'd have said no and been wrong.

GPT 3.5 / 4 were an iteration of previous models, their performance wasn't that exceptionally surprising if you were working in the field.

Also, these are fundamentally different problems. Self-driving requires massive amounts of real-world testing and then you have all the regulatory and safety/legal stuff to do. That's simply infeasible in 6 months no matter what. Generative AI has no such constraints.

> Kennedy said it would be done by the end of the decade, people called him crazy, and then it happened. The people saying it couldn't be done were still saying it right up until the day it happened.

So? I don't see any link to what we're discussing. Kennedy didn't really have that much direct involvement in the moon program.

What we're discussing is an equivalent of the head of NASA, in ~1966 or so publicly declaring that they'll send a crewed mission to the moon in 6 months (even though everyone knew that wasn't possible no matter how hard everyone worked).

> He's giving the optimistic end of the range,

He's just saying random unsubstantiated stuff because that's just what he does all the time (or because he wants to sell more cars IDK).

> Yeah, it did. Nobody knew it would work until it actually worked.

I'm pretty sure that was an iterative process, so they knew quite a while before it actually "worked" (could be released to the public).