There is only so much talent available for designing.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable
- Thread starter AudubonB
- Start date
I bet nobody ever invites YOU over for breakfast!
Last edited:
Agree. Each floor house clean (1800 sq ft) is $120. Doesn't include dusting, laundry, or dishes. Less than five year pay back.
I'm impressed with it's honest portrayal of where our bot's development's at.
Last edited:
You must not attempt to divorce any kind of funding from the appropriate milieu’s cost of capital (CoC); in that the investment ‘world’ still is not globally homogeneous we safely can say that CoC is as determined by the US Fed, so - no, the appropriate CoC is not “free”, as it effectively had been when interest rates were approaching zero. If that remains unclear, ask “Whence comes that $9.7 billion?”; the answer is, of course, from ones who believe GS can provide a return greater than that benchmark risk-free rate determined by the Fed.
Interest rates always matter.
lafrisbee
Active Member
A CLEAN house.... YES YES YES
Cult Member
Born on the 4th of July
Are you sure? Optimus as a labour will have his cameras / eyes everywhere and this could result in recordings of human colleagues behaviour or quality of work. Entrepreneurs will profit from this information and unions will ban the use of it. That could be challenging. Will we see factories run by Optimus only?
I pay $3900 a year for that now and it’s every other week. so $25k would be a no brainer as I might get much more utility out of it Such as dish and clothes washing. Plus the windows!
Captkerosene
Member
I think that's what it costs to produce ... not what it will sell for.
CLK350
Member
HR has called you to the 13th floor. Bring your pass key with you.
CLK350
Member
Would be great (TMC devs please take note) if I could mark some members according to diff criteria. For example in this case would love if I could mark it as something like " asfasfd", as short clean version of "typical person who thinks b/c he made some $ on TSLA or is knowledgeable in some areas is COMPLETELY off in his understanding of certain areas of Tesla development/ or a troll"
Knightshade
Well-Known Member
I think the fact virtually nothing beyond L2-for-all-NA by end of year was really discussed, no RT, and no HW4, factors into why that's not getting much attention from the general public.
Not sure how you got "Robotaxi valuation is getting closer" from anything they said last night, other than the fact Dojo (coming next year) should let them rev NNs and such faster... so in THAT sense, yes, that's "closer" than if it didn't exist... but unlike many past events and unusual for Elon the presentation was almost 100% devoid of "We WILL have RTs by X date, totes for sure!" claims.
It was actually pretty fantastic to see Elon doing a presentation that was focused on "Here's the problems we've been solving, how we've been solving them, and what we're looking to get done in the future" WITHOUT crazy-short unrealistic target dates being thrown all over the place.
Nobody is going to know when generalized RTs are ready until they're ready. Until then there remains myriad possible dead ends, local maximums, compute limits, and other potential roadblocks to work around on the way there. Speaking to that was some of the Dojo discussion where it was pretty clear they had a design intent of making something to provide them max offline compute possible that would still be relevant even if they had to make serious and fundamental changes to many other aspects of the RT effort--- they wouldn't do that if they KNEW FOR SURE they're on the right path today with no major rewrites needed. Instead they designed specifically EXPECTING there'd be more of them.
The Accountant
Active Member
Yes - I know it's not free. But wouldn't a promising innovative company with start-up losses have a better chance of getting funding through private equity than interest bearing debt in this environment? What's GS going to do with the $9.7B? Sit on it?
I am happy to see this closed loop recycling of batteries between Redwood and Tesla at GF 1(Sparks).
JB Straubel is attacking a big problem. The factory is even powered by Solar panels.
I wonder what can be recycled from the ICE vehicles that are left behind.
www.teslarati.com
JB Straubel is attacking a big problem. The factory is even powered by Solar panels.
I wonder what can be recycled from the ICE vehicles that are left behind.
Tesla to use recycled materials from Redwood's closed-loop NV campus
Redwood's Nevada Campus would be capable of recycling, refining, and remanufacturing battery materials. It won't have waste streams as well.
www.teslarati.com
woodisgood
Optimustic Pessimist
To me there seems to be a large swath of gray area between having the bot contribute specifically in Tesla’s factories vs. the general purpose bot that can do a lot of great things in our homes, businesses, etc.
I have no doubt we’re straight line to the former. How we get to the latter is still a bit nebulous. Analogizing to the cars, the reason FSD can drive anywhere is that the training data comes from everywhere. The ”robots” were already out in the world collecting training data en masse because they were a useful thing people were buying. In contrast, how does Tesla collect all the speech and household data that is more pertinent for the bot? In this regard, someone like Ring/Amazon has a headstart on the ideal training dataset.
This is why I have always thought the way to the truly useful bot is via a development platform/App Store like iOS. In such a scenario, third party developers take on the costs and responsibilities of giving the bot the training environments it needs for the use cases at hand. Let’s say I’m a restaurant chain. I invest millions to buy bots, put them into a diverse array of kitchen “sets” I construct, and give them exposure to the environments and tasks they need to learn. They break a lot of dishes, waste a lot of food, but in the end depending on how good my training setup is and how much $$ I’ve invested, maybe they’ve learned how to cook in a commercial kitchen. This discrete set of training is then sold as a “skill” (like an app). Maybe it costs a ton if the training is very good. And maybe it only costs $3 if I have only taught them how fry an egg one way. I get a large share of the profits. Tesla handles training via Dojo, deployment via a skill store, and has final say in terms of safety, much like app stores. This question was posed during the Q&A and understandably Elon expressed safety concerns about opening it up in this way. The magic sauce here lies in whether Tesla can put together a development platform that unlocks an entire ecosystem of “trainers” while keeping malicious skillsets from being developed.
The challenge is getting the bot in large enough numbers into the right environments to learn the tasks you want it to. It won’t learn how to weed your yard effectively until it has seen millions of yards and tried to deal with millions of weeds, much like FSD hasn’t gotten good at right turns until it has seen and attempted millions of them. It won’t get good at navigating your kitchen until it has tried to open different fridges, dealt with varying standards of kitchen clutter and organization, used a variety of frying pans and cooktops, etc. Since the bot is unlikely to find its way into our homes in large numbers until it is actually useful, there needs to be a way to get all of this training data, and I think that will eventually require a robust ”training” community.
I have no doubt we’re straight line to the former. How we get to the latter is still a bit nebulous. Analogizing to the cars, the reason FSD can drive anywhere is that the training data comes from everywhere. The ”robots” were already out in the world collecting training data en masse because they were a useful thing people were buying. In contrast, how does Tesla collect all the speech and household data that is more pertinent for the bot? In this regard, someone like Ring/Amazon has a headstart on the ideal training dataset.
This is why I have always thought the way to the truly useful bot is via a development platform/App Store like iOS. In such a scenario, third party developers take on the costs and responsibilities of giving the bot the training environments it needs for the use cases at hand. Let’s say I’m a restaurant chain. I invest millions to buy bots, put them into a diverse array of kitchen “sets” I construct, and give them exposure to the environments and tasks they need to learn. They break a lot of dishes, waste a lot of food, but in the end depending on how good my training setup is and how much $$ I’ve invested, maybe they’ve learned how to cook in a commercial kitchen. This discrete set of training is then sold as a “skill” (like an app). Maybe it costs a ton if the training is very good. And maybe it only costs $3 if I have only taught them how fry an egg one way. I get a large share of the profits. Tesla handles training via Dojo, deployment via a skill store, and has final say in terms of safety, much like app stores. This question was posed during the Q&A and understandably Elon expressed safety concerns about opening it up in this way. The magic sauce here lies in whether Tesla can put together a development platform that unlocks an entire ecosystem of “trainers” while keeping malicious skillsets from being developed.
The challenge is getting the bot in large enough numbers into the right environments to learn the tasks you want it to. It won’t learn how to weed your yard effectively until it has seen millions of yards and tried to deal with millions of weeds, much like FSD hasn’t gotten good at right turns until it has seen and attempted millions of them. It won’t get good at navigating your kitchen until it has tried to open different fridges, dealt with varying standards of kitchen clutter and organization, used a variety of frying pans and cooktops, etc. Since the bot is unlikely to find its way into our homes in large numbers until it is actually useful, there needs to be a way to get all of this training data, and I think that will eventually require a robust ”training” community.
insaneoctane
Well-Known Member
I'm sure that a non-zero amount of money will be invested in TSLA from people who want to "get in at the ground floor" of Tesla's robotic vision that was shared Friday. I honestly can't quantify that amount nor compare it to how much leaves TSLA because they see Optimus and AI as a "distraction", but I certainly see this as a net positive!
Last edited:
Gigapress
Trying to be less wrong
Why don't you start with being nicer to people with whom you disagree? I seem to remember being taught this in preschool.
Most of this presentation is drastically less technical compared to what we've seen in the first two editions of Tesla AI Day, and Cruise is still saying some stuff that makes no sense.
In it, for example, Cruise claims that emergency vehicles are "a great example of why we need so many sensors". They say you "need" (yes, they used that word) lidar to recognize an open door and other protruding objects and you are at risk of not stopping for an emergency vehicle in time unless you have audio input to detect the siren noise. Come on. Last I checked, deaf people are still allowed to drive. Humans can operate a car safely with nothing more than passive visual sensors and kinetic sensors from the motion feedback they get from the driving dynamics. In other words, Cruise is effectively saying "Our signal processing is insufficient to understand driving using input data streams that demonstrably are enough for biological control hardware, so we think that it's fundamentally required to add in other data streams."
Not in the Cruise presentation was any of the following:
- Socially awkward nerds who seemed uncomfortable presenting
- Large-scale real-world data pipeline and autolabeling
- Remind me, how many Cruise AVs have been built and are driving right now?
- Custom training hardware and true full-stack design with hardcore customization at every level, including chip, tile, power, thermals, file formats, transfer protocol, and more
- Math equations that would scare the audience
- Deep description of how the computation actually works from incoming photon measurements to driving outputs; Cruise showed nothing about their actual neural net model
I think @Knightshade summarized this situation well in July:
Last edited:
dhanson865
Well-Known Member
If the H100 gains 6x on a single core or even a single cabinet it's immaterial because it doesn't scale to thousands of cores like Dojo. It's that simple you are comparing apples to oranges. Dojo isn't looking at small or medium scale it's looking to be the largest scale compute cluster available.
Similar threads
- Locked
- Locked
