Tesla AI5 and Trillions from Distributed Inference Explained

Artificial Intellgence training is the work to make the AI models. Artificial Intelligence Inference is running the AI to get answers and provide users with value. Tesla will be able to deploy millions of AI5 inference chips in millions of cars. Those cars will be completely paid for by customers.Tesla could sell billions of humanoid Teslabots with AI inference chips.
What will this massive distributed cloud of AI inference do?
How does the economic looks versus data centers vs distributed?
Follw Brian Wang on nextbigfuture....
Follow Brian Wang on www.NextBigFut...
Follow Brian Wang on www.X.com/@Nex...
Nextbigfuture brings you the most important science and technology breakthroughs and analysis at least once per week.
Follow Warren Redlich on / wr4nygov
Follow Warren Redlich on / @warrenredlich

[@mattsenkow6986]

I think the bots will have to perform sustained inference to a much higher degree than Tesla's EVs. Inference for planning and continuous inference for carrying out a task. Any compute needed more than that available from the SoCs in the robots or EVs will be computed on the Tesla Inference Network.

[@robbiero368]

Whilst Optimus is doing your ironing it will be moonlighting as a legal secretary

[@johannesdolch]

It is my believe that this is the real reason Elon fired 14% of employees. To show everyone who works for him that the fight isn't over. Everybody who isn't hungry anymore should leave.

[@zdme4864]

not a belief … an Elon stated fact!

[@johannesdolch]

As for the deminishing returns log effect: The guy from Microsoft said that their testing showed that the link between input and output ist linear, at least for Gpt 5

[@dominicrebelo6982]

The Us has about 70 Interstate highways totalling almost 50,000 miles. Elon once said that you could power the entire US with about 100 square miles of solar which is about 10,000 miles. Humanoid robots could build solar fencing all along these interstate highways which criss cross the entire US & they would install the solar fences on both sides. They could also place a small wind turbine on every fence post to increase the electrical energy production of the highway system 24/7. That would create more than enough electrical power to power the entire US plus some without taking up any new land & along with battery storage could be the obvious simple answer. And every state could contribute something & every state could build & install as they can afford it with Federal help. Win/Win except the energy companies will hate that idea.

[@alexchow9629]

Your units are messed up. How did you go from 100 square miles to 10,000 miles? The former is a unit of area. The latter a unit of length.

[@sb5580]

@@alexchow9629 he assume at road width of .01 mile which is about 53feet. But its sill an incredibly bad idea to actually clutter roads that way. small windmills are proven to be worthless in most area. Elon was probably talking about area i doubt he was talking about making such a system.

[@darwinboor1300]

Wiring, infrastructure, and maintenance would also be major issues. Consolidated energy farms located and designed for optimal performance would clearly be a better option.

[@sb5580]

@@darwinboor1300 I once had some consulting work for a company that wanted to put solar collectors on every lightpole. It was a worst of all worlds thing that only a regulated utility with cost-pass through economics would do. the solar panel made it harder to use radio, it was an eyesore, and having to have a guy climb a pole to get at one panel was ridiculous. A roadside solar collector maybe not so good either.

[@TheEarlVix]

Well explained and articulated.Thanks.

[@dly9014]

Specific to AI5 chips in Teslas being repurposed when not in use - is there an example of distributed computing actually able to serve inference workloads competitively vs in-datacenter architectures? There needs to be a discussion here around latency. If a distributed approach worked, we would have examples with rigs similar to bitcoin mining w/ GPUs optimized for inference working in people’s homes and making money for them. Latency seems to be the key issue - I don’t want to wait for a response from my chatbot 10x longer because a car in someone’s garage is doing the processing.

[@darwinboor1300]

Thanks, One missing important point: Most new FSD v12 subversions include minor revisions of the system of AIs that make up the "end-to-end inference AI" that goes into the vehicle. Training is more than building on prior training by introducing more data. Each training cycle is a completely new training of the FSD AI system from the ground up. Training results from previous FSD subversions cannot be carried forward. In addition, post training FSD validation is critical because each training cycle tends to randomly introduce new inference errors. These errors are not easily predicted by errors seen with previous subversions. They are one of the major negative aspects of end-to-end AI without hard coded control logic to prevent inference errors from manifesting as driving errors.

[@psdaengr911]

I don't think that this is much different than what happens with all babies learning to walk and then to drive. Think of FSD as a series of driver-ed students. Each student learns through a curriculum of curated experiences, then drives while being observed. The difference with AI is that each new generation of students is objectively faster at processing those experiences. The "school" analyzes the performance, accidents and mistakes and uses that to select a better set of experiences for the next class to learn from. Finally, the skill of the best AI student at any time can be captured and downloaded into other students.

[@davidfellowes1628]

This was really excellent. The way to compliment you adequately is to say, "we too need to up our game and to be able to use and ask the right questions. Insightful

[@psdaengr911]

42.

[@davidfellowes1628]

@@psdaengr911 thanks

[@psdaengr911]

AI can learn from any form of input. It could improve the design of chips and hardware on which it runs using iterative simulation. I see it eventually being the primary tool used for engineering products for manufacturing.

[@InspireLifeGoals]

Appreciate the content.

[@eranrichardson5874]

40:47 Ar4 is an inference chip, and cannot be compared to nvidia GPUs - that’s a completely different beast with weaker aquracy and connectivity (unlike the dojo)

[@ArizVern]

YOU GUYS ARE NUMBER ONE.

[@colinkeizer7353]

What can I put in orbit that will make money for me? Maybe beyond orbit? Will there be a combination of Optimus, FSD and SpaceX tools that a school, a determined individual or a small business can do in space that will be enabled by those incredibly low costs of delivery to LEO?

[@psdaengr911]

The cost of moving kgs of freight optimized for launch to orbit can't be directly compared to shipping kgs of random material door to door cross country. You are ignoring contents, packaging and logistics of getting freight to and from the terminals, and comparing one hub and spoke system with limited hubs to multiple private mesh systems. Energy-wise it will still make more sense to ship one cargo ship from China to San Francisco and distribute brake rotors or cars, or to send containers by jets than to send them by a "cheap" Starship.

[@colinkeizer7353]

Yeah, how many times has Elon told us solar power is the answer? Totally agree with you we will see platoons of robots installing solar arrays near his data centers. It will be fun seeing if he puts up fence row solar in the vertical orientation.

[@davidmarkmann6098]

The training computer adjusts the weights in the neutral net model based on driving data. The model is then running in the inference chip in the car which is reacting to camera, navigation and brake and steering inputs.

[@alanseymour1252]

If North Korea were to capture a Tesla robotaxi, would they be able to mass produce pirate copies (without needing the data for training compute)?

[@MrPJD17]

no

[@sagecoach]

My early 2023 Model Y with FSD just had a black bear run across a two-lane highway left-to-right just missing us by a hair. I didn't see it until it was at the car fender either, and there was good daylight with no traffic. The car had no reaction and I did too late except to report the incident. I have witnessed no response. to many common things, like speed limit signs, potholes, deer, and most road debris. I expect it to get better than me. Tesla must pick the 10,000 miles to count, I have rarely seen 100 miles clean and I report a lot. Distributed Inference is going to need more learning onboard and car-to-car territory sharing.

[@colinkeizer7353]

Do not tell the djinn to make you a milkshake. You will not like the results.

[@psdaengr911]

Someone else might.

[@williamwoo866]

thxs

[@colinkeizer7353]

I wonder if I will be able to put a few of those AI5 chips into a computer, attach the machine to a Powerwall and a Solar Roof, then rent its services out full time, maybe over Starlink. This seems like another possibility for generating my own long-term baseline personal income.

[@sagecoach]

We are going to see some rapid unscheduled disassembly of AI in most of its implementations before or if we see stability, if we survive. We will screw it up, it is what humans have been doing as long as we have existed. Really! We have already created a massive record life extinction event and it may include us. Not knowing what we are doing never stops us. I love you all, we are in this together.

[@FoundingFathersUSA]

Inference is like running a program, learning is like programming it (automatically).

[@psdaengr911]

Inferring is drawing a conclusion from data. Learning is the process of remembering data and inferences, organizing them within an internal framework of principles/rules/values inferred from previous learning. Knowledge is the result of learning that can be applied to new data.

[@psdaengr911]

Re music and art with AI being super plagiarism, the plays of William Shakespeare covered every basic relationship, situation and outcome that humans can be in. He didn't create those situations, relationships and outcomes. These were previously described by other authors. Is his combining them into stories not creative? Is Tennessee Williams a plagiarist of Shakespeare and the authors that Shakespeare had read? ?

[@ArizVern]

YOU GUYS ARE NUMBER ONE.