# of issued leaf certs isn't a great indicator of propagation, user-impact, revenue etc.
1. 59 million of those 60 million could be very short validity, duplicate enrollments from same private key/csr, etc.
2. I'm still not certain that that figure isn't including for example private roots/intermediates they've issued from.
That said they're a low volume (in terms of root/int) provider so I don't think the impact of them exiting will be great.
credulous in the sense that it is an an engineering feat.
Caveat, I'm way more a car person than IT person, and dislike Tesla.
Hagerty or some similar channel had a 45m breakdown of the car with a tesla rep, and as a standalone single vehicle that can do what it does, it's a marvel. I really couldn't believe the aero effects and paneling/tightness came from the same company that last year officially "exited the car business" and sold a stainless steel half truck.
That said it's so obviousley a marketing peice/distraction. But I am thinking about the engineers who's saw this as their chance to show that yes, they can design and bring to prototype a car with the fit and finish of a ferrari prototype with the added instanity of a very impractical gas cylinder rocket system.
"LLM" literally converse/compute via statistical inference.
It is given sets of paramaters/data and infers a response based on it's acceptable confidence level. The bottleneck for this is compute power and model/prompt/rails/integrations.
The reason LLM stays consistent/can be somewhat trusted to stay on rails is that it's aware of both it's context/scope and it's confidence about the context.
Anything else would be....original thought and exposition? I guess. You could say that our brains do the same, IE; unconcious statistical analysis leading to inference based on confidence.
You converse through electrical signals and muscle contractions, but that will not make you an expert in either.
Similarly, a language model trained to emulate speech, follow instructions, and solve problems will not implicitly have an understanding of its own statistics. The research doesn't back that up and it just doesn't make any sense if you think about it.
The only thing LLM produce that could be considered its "confidence" would be its token logit values, but even that is a sketchy connection.
> It is given sets of paramaters/data and infers a response based on it's acceptable confidence level. The bottleneck for this is compute power and model/prompt/rails/integrations.
Does the output of the model have insight into those details? Just because our brains function based upon electrical potential doesn't mean we have innate insight into it.
Yes, this is what is user-level marketed as context memory.
Certain models of course force this context memory, some don't, some allow you to toggle it's "trigger level".
For example Anthropic, use haiku for low memory context queries, research. Then Opus/Fable for compiling/analysis/design, which relies on the LLM inference constantly going back into context memory.
Another example is OpenAI/GPT. The "free" tier forces context memory using the "highest available model" per your subscription (or lack thereof), and upon hitting its limit, drops you down to a low context model, and then ultimatley blocking that chat entirely.
(I appreciate this discussion and am not trying to inform you, I'm sure you already know, but typing these things out helps me understand my holes in understanding.)
Excuse my ignorance, but how does context memory provide an insight into a model's inner workings? I mean this:
> It is given sets of paramaters/data and infers a response based on it's acceptable confidence level. The bottleneck for this is compute power and model/prompt/rails/integrations.
or whatever the LLM claims its confidence is on any given output?
This is an incorrect conclusion. Your statement about statistical inference is correct but it doesn't explain why your initial statement, especially "is perfectly able to articulate it".
A good litmus test for this is, typically any artificial neural network, perceptron or even a simple linear regression model do statistical inference based on weights and selection based on confidence level. It's a generic characteristic, which doesn't explain anything specific to LLMs.
> The reason LLM stays consistent/can be somewhat trusted to stay on rails is that it's aware of both it's context/scope and it's confidence about the context.
Post hoc ergo propter hoc.
> I guess. You could say that our brains do the same
I understand that it "feels" the same from a certain perspective, but we don't know that.
I totally agree with you. there's a cost benefit analysis we make when purchasing something and I can imagine the case where I'm ok paying 50c an hour for a remote game.
It's hard though not to see the writing on the wall in regards to "local compute GPU availability". Every single industry even tangentially related to compute power has "silently" moved their bus model to consumption based, despite the consumer base having 0 appetite for that. (We're not engineers/app developers who are forced to use LLM agents, we're trying to pay a service that hosts games.)
I am unsure how a medium that was a failure as a free/subscription model will succeed as a floating pay-as-you-go model.
For a period of time I would rent a windows box with gpu from aws, install my game, and play it over parsec (a Remote Desktop tool designed for low latency). It was around $0.40/hour for a high-end gaming rig for the ~40 hours I needed to binge my new game. I play new games infrequently, so the cost analysis was “a new computer for $1500, or rent one for the time I need it for $15.”
It’s just a bummer that this process wasn’t a bit easier. I recall there was some setup involved.
How often do you buy new computers? I paid something like 2500 USD for my Threadripper machine (including 128 GB of RAM and graphics card) ~seven years ago. Between now and then I paid ~580 USD (at the end of last year) to get a new graphics card and that's it. After last year's upgrade, I've yet to run into a game that it doesn't play reasonably, and I've also gotten a lot of work done with it over all the years I've had it.
The console lifecycles being like a decade long has been a fantastic blessing for us PC gamers. Most AAA games are intended to run well on those ancient almost-PCs, which means that our old actually-PCs usually play the games just fine.
It's a positive in the sense that the industry has made it marginally cheaper in time/resources to pay-as-you-go.
It's a positive in the sense that due to these industry changes (computer power going up 20x in cost.), the platforms are offering a way to let us peons play games.
It's a negative in that it's a futher commitment of these vendors to hoarde compute/GPU power from western consumers and chinese LLM vendors.
MS will routinely purchase the highest end GPU for sake of keeping this service alive, nevermind their own LLM. Good luck getting one yourself.
yep, actually in my area this will be cheaper than the gaming cafes people frequent—I think the lack of reasonably-local servers is the main thing that will block adoption.
I'm not so certain the average gamer is well-informed enough to realize what they're giving up, or that they feel strongly enough to take a stand for it if they do. I'm hopeful they will, we'll see.
But in that vein, you are in the same position as the chinese LLM providers. They cannot get hardware to host so they put the problem in the forefront by offering it open.
maybe in 2030 i'll have a chinese gaming PC with chinese local AI. would be interesting to see if they can do the same to US tech companies as they do to european car manufacturers. i'm ready for it...
AH, you clearly haven't had the "5 different chargers breaking and begging support to send a different production batch". (And paying for every charger.)
My scanwatch is beautiful, and the battery life is great. But to this day, on my 5th charger, it either charges to full in 20m, or 3 hours. Same charger, same wall outlet. Probably not great for the watches hardware.
Unfortunately I've had. I've had two different scanwatch models and ended up buying ~3 chargers each for of them. Just thinking abut USB-C and how much less different chargers types I have makes me really happy.
Worst product i've ever owned that I feel completely duped on.
It's a hard feeling to explain, when purchasing any "high cost" tech device, I like to think i'm a smart consumer and do my research.
Nope. This company is, for back of a better word, disgusting. Never in my life have I had to maintain a profesional relationship with support agents for sake of:
-charger breaking repeatedly
-Syncing breaking repeatedlty
-BIometrics are completely, completely random and arbitrary compared to other professional devices.
-time desync
-Completely, baffling missing integrations. (I can't configure a god damn alarm?)
-The big one, dynamic paywalling. Since I purchased mine, I have been locked out of the few worthwhile features that the app/device initially offered.
The scanwatch 2 is completely useless as a watch and a health device. It is marginally acceptable as a fashion device.
Yes, completely unbeleivable. I mainly think this because SpaceX has committed said engineers to delivering unfeasable amounts of payload to LEO/ISS, etc. Think on the order of "One succesful launch, delivery, and re-entry/recovery per day".
That is of course very very far away from reality (though, progressing to a feasable program faster than any space observers predicted.)
THat, combined with Elon signaling to investors/market that tesla no longer produces cars.....it's hard to imagine the company splitting their very, very , very over-commited engineers to a flying car. At least, they aren't commiting the point that this will be a consumer-market car, or even a halo car.
> Think on the order of "One succesful launch, delivery, and re-entry/recovery per day".
Given that they have largely relied on reusability rather than strongly ramping up production of boosters for Falcon 9, and given that they have performed 3 launches in a single day, one launch per day is clearly achievable without any further advances, and without even using Starship.
To be clear, the stated goal for Starship is one launch per hour throughout the year, which would still be a tiny number compared to, say, airline flights, but quite impressive for orbital flight.
If space-x would ever send out one Starship per hour, that would be horrendes for our atmosphere (potentially).
It would also mean nearly 900.000 tons of resources leaving earth for ever and potentially poluting our atmosphere if they have to ever reenter.
Preparing this while our planet earth has A LOT of empty space and desserts and no breakthrough in quantumslipstream or warp core tech is ridicolous dangerous for all of us :|
The launches themselves can be made roughly neutral by producing the methane and LOX from the atmosphere using solar power. For an hourly cadence, that would require as much solar as a third of all US installed commercial solar, but the US is already adding that much every few years, and accelerating, so it doesn't seem impossible. That's best-case, of course, but given that, all the combustion products would balance out propellant generation.
Reentry of dead satellites and other debris does seem like it's eventually going to be more of an issue, so that will be something to work on.
I wouldn't count on any warp drive breakthroughs, potentially ever, but the whole solar system is still out there for living and industrial space.
Even if resources don't evaporiate in our atmosphere, you need to send up a lot of resources like metals.
We are so so so far away of mining resources on other planets/moons than earth, setting this up now is just crazy though.
And lets think crazy wild: in 20 years we send up enough stuff like machines and resources to start mining and setting things up, to do what? I love scifi but it has to reach some kind of goal.
True, they are clearly top of the entire planet when it comes to succesfull LEO launches.
I can go even further and say I trust they will be the company to deliver, given the decades long historical relationship between US Space Agencies and Vendors.
That said, 2.6 test falcon launches per week, is so, so many orders of magnitude away from "~150 full payload delivery and recovery missions per week".
I struggle to think how anyone with experience in the industry (and eye balls) thinks this is aligning with reality.
> SpaceX has committed said engineers to delivering unfeasable amounts of payload to LEO/ISS, etc. Think on the order of "One succesful launch, delivery, and re-entry/recovery per day". That is of course very very far away from reality
This is a mathematically silly assertion. Over the past decade, SpaceX has scaled their launch capacity from ~20 metric tons per year to ~2,300 metric tons per year. That's more than two orders of magnitude. It's not unreasonable to imagine that they could scale a further 0.3 orders of magnitude if there were customers willing to pay for this amount of lift.
CUrrently they're averaging, what, max 3 falcons per week, usually in test configuration?
The partnership scope calls for.....~150 launches per week, delivering payload and returning the module.
Trust me, I love space and pushing the frontier at the cost of US dollars, but that is....so significantly far away.
COnsider the boeing 737-8,9,10 MAX. That program is far, far less grandiose in it's delivery scope, has an ancient legacy OEM that is very familiar with regulated contracts w/ the gov, and they can't deliver a plane within 5 years of the expected contract completion to their premier airline partner.
I can be convinced IE; some transparent internal reporting on project lead's current understandings as the project progresses, but there just isn't any statements from SpaceX/Nasa/USGOV on what we can all see with out eyes.
Planes carry billions of passengers per year, so their design and operation needs orders of magnitude more oversight than an unmanned cargo mover. Every little detail of an airliner needs to be scrutinised for the various ways it could affect safety. Rockets don't require that level of safety assurance.
No contradiction between "more than everyone else combined" and "it is unfeasable to launch 36,500 metric tons per year compared to the about-2,500 tons in the last 12 months".
Especially as the working launch vehicle is getting retired and the replacement has only had one circularized orbit, and is currently behind the development pace set by the Saturn V, N1, and even the Space Shuttle.
The N1 was cancelled after four explosions, and the fourth Starship launch was non-explody.
But the fourth N1 launch was 10 years after the N1 work was started. Starship began work under a different name in 2012, and by 2022 it had yet to launch at all, let alone reach the non-explody stage.
If you assume all the pre-2016-rebrand work was scrapped, then you can have it beating the N1; otherwise I think my point stands.
For the first few years they were just doing design work, and it's hard to say how much effort they were putting in since their focus back then was Falcon; the first Falcon 9 landing was 2015, and the first Falcon Heavy flight was 2018. For Starship they didn't even decide on stainless steel until 2019.
Comparing apples to apples, this seems to be the appropriate starting point for N1:
> Studied and designed by OKB-1 since 1959, it was the counterpart to the US Saturn V
First launch was 1969. Starship's first launch was in 2023, so it took one extra year from initial concept to launch.
Starting from first launch, N1 did four launches in three years and never succeeded. Starship did 15 launches in three years, had multiple mission successes in which they achieved orbital velocity but had planned a suborbital trajectory, and then orbited on the 15th launch. Their first mission reaching orbital velocity was in 2024.
Also worth mentioning that a lot of the Starship work focuses on full reusability, a quite difficult problem which N1 did not attempt.
That is just a one OOM scaling. It's not that difficult, especially considering that Starship just delivered the 3rd largest payload ever launched and it wasn't even at capacity.
If he was the fastest runner, won a the races and had all the world records, I don’t think failing to meet his aspirational goal of 75 would be important.
> I personally gain a lot more from astrophysicsi than from Starlink.
Good way to show your privilege.
Starlink is changing peoples lives around the world, previously the fastest they had was dialup, now they’re video calling doctors and family and doing school online.
I’ve seen it with my own eyes in rural Australia, arctic Canada and Tunisia.
You saw what? Rich people in other countries (a small amount of) having now fast internet?
Wow.
Now take my proper argument, stop nitpicking and compare a handful of peple on a planet with 9 Billion people.
My main argument was and is, that a lot more people are getting hurt by this than gain something. Loss of research and clear night sky and risk of atmospheric poisening vs. a handful of peple now benefitting from it...
No. I saw rural people (not rich) getting to access the world of medicine and education people in bigger population centres take for granted. Not to mention connecting with their families.
They were very happy.
“Handful of people” benefiting from it is vastly missing the mark.
It's a similar level of complaint to windmills spoiling the view. And if you poll people which would they rather have in the wilderness: no satellites visible at dusk and dawn, or full internet access, which would they prefer?
I can't imagine an individual who both has their health records (pertaining to some issue) and also doesn't have the ability to understand the context.
If I feed in 15 crown/root canal invoices into AI asking it it's opinion, it will tell me I have a very bad mouth. This does nothing to inform me, and will lead to the expected AI response of: "Here are some practices in your area specializing in oral surgery".
There's a reason it's advised by doctors to not share health information. practices/doctors proscribe a regimen/treatment plan, and have to assume in good faith the patient will follow it. Imagine if every doctor heard "Well, AI told me..."
I can, because I've been that person. I got a blood panel done recently. Didn't hear anything back from my doctor because everything was in the OK range. But I still wanted to actually understand the results. And yes, I could have spent time researching what each individual market meant. Or I could ask an LLM, which gives me the same information in seconds. Now obviously if I planned to take an action like starting a supplement or something based on that information I would verify it, but the LLM is still providing value there.
1. 59 million of those 60 million could be very short validity, duplicate enrollments from same private key/csr, etc. 2. I'm still not certain that that figure isn't including for example private roots/intermediates they've issued from.
That said they're a low volume (in terms of root/int) provider so I don't think the impact of them exiting will be great.
reply