A deep discussion of the developments with AI and the upcoming 2025 Tesla Dojo 2 and 2026 Dojo 3 chips. But AI needs compute and data.
Won’t AI reach a limit as we use up human created data?
What if you can use mass production of space telescopes to monitor the whole galaxy down to objects the size of 10-100 kilometers in resolution?
Sending space probes would take decades to reach the nearest star and would take 100,000 years to get across the galaxy and lightspeed and sending back data would take another 100,000 years.
NASA has had many studies looking at placing one meter telescopes at the gravitational lensing areas. This allows a telescopes to leverage the gravity of the sun to make a telescope 100 billion times more effective. You would also need a solar coronograph to block the light of the other stars will enable the telescopes to directly image planets and moons and details of quasers and black hole halos and other objects.
NASA was working on a one third scale prototypes of the solar sails. Components like the sails have been built. This mission would be about 6 kilograms. A full scale mission would be less than 50 kilograms.
Galaxy Monitoring Array
What has not been considered by NASA is production of the lightweight one meter telescopes at scale. What kind of scale? The scale of all cars on Earth and beyond?
Gravitational lensing enables looking that objects directly on the other side of the sun. If we had one scope for every solar system in the galaxy then we would be able to constantly image objects in every solar system in the galaxy. We would only need to go about 20 times faster than we have up to today. Solar sails that swing around the sun could achieve these speeds. We need to improve the heat resistance of the materials with materials that we have now.
Getting out 20 times further than Pluto lets us use the Sun for this 100 billion times improvement. A halo of telescopes around the sun is the Galaxy observation array.
SpaceX has 6300 Starlink satellites before the fully reusable Starship increases capabilities by thousands of times.
Mass production of satellites could go to the millions and outnumber the 5 million global cell towers.
How about billions of Starlink sized space satellites? There are 2 billion cars in the world.
Starlink sized space telescopes could be made at scale and placed beyond the orbit of Pluto.
The telescopes and solar sails would be about 50 kilogram but slightly improved versions could get the mass down to 20 kilograms. This would be 100 times less than a car. Building and deploying the same mass of space telescopes as the world’s cars would be about 200 billion telescopes.
The same mass of the world’s cars in space telescopes would be 200 billion telescope probes. This is about the number for each solar system in the galaxy. They would be like remote probes. Going out 3 light days to monitor 4 to 100,000 light years and beyond. Learn the galaxy and the Universe in detail with a flood of information for AI analysis and learning.
1000 SpaceX Starship fleets would launch 250,000 tons every two years to Mars but 1.5 million tons to low earth orbit. The solar sail telescopes vehicles would not need to be refueled from orbit and would not need to be carried by the Starships. They would use the solar sails to fly around the sun and get speed.
The earth orbit Starship fleet could be expanded to 10,000 to 1 million SpaceX Starships by keeping production rates up.
The world has 30,000 large commercial passenger planes and those planes cost about $350 million each. The SpaceX Starship with booster can have $250k raptor engines and an overall cost of about $20 million.
A 30,000 SpaceX Starship fleet could lift 45 million tons to orbit every week. This would be enough lift capacity for 1 billion one meter light weight telescopes.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
I’ve said it before, the difference between data, and information is legion. So many 1’s&0’s, , the “data” is meaningless without context. For many decades our NSA has intercepted almost all electronic communications from everyone to anyone. Where really very good at doing “that”. Trust me. (hell, better us then “them”) That said, even back in the 1960’s, we soaked up so much “data” we STILL could not anticipate objectives, motivations, or intention in intelligence, which IS THE WHOLE God damn point. Knowing what your enemy can do is critical. Knowing what they will do, under certain conditions, can be the difference between life and death. Details matter.
Trying to understand what a person says in an electronic intercept may be straight forward. Bu as I was taught, a good translator does not translate your native language into another. You THINK and LIVE in that language. You FEEL the culture from which that language “emotes”. It’s the difference between being a tourist, and a resident. But their motivation’s, intentions and objectives? Not so easy, and this is a problem across not just cultural and social “gulfs”. I have problems with having rational conversations with some of my fellow Americans. When one gets their “news” only from a source that reinforces their ego, but provides little “news” (Defined as events and consequences) This can be a problem.
So with all this data, were soaking up how to deal with it? One way might be to “package” conceptual concepts as an isolated linguist artifact. Think of an idea, expressed as words, as a Lego block, part of the larger “toy”. The toy is the consequence, The individual pieces combine to express what the toy “does”. (OK, now everyone knows why I don’t get invited to parties) I solve problems by looking at them from a proactive not reactive way. The introduction of “turbulence” into a linguistic communication can help “bracket” it, Think of a 4D grid. (3dimensions plus time).
“Turbulence” Can take many forms. “Asymmetric” to any “known” communication is often called “noise, or static”. Symmetric turbulence is very close (as a pattern) to, the host or target communication, but is “out of phase”. This makes the later very hard to detect, but is informatically so close to the “signal” it shadows, it’s invisible, like a mirror reflection, or echo of the target signal it “rides next to”. An asymmetric “burst” sent into any signal, can “decloak” a parasitic signal.
Believe it or not, many of these mathematical-technological principles can be applied to social, economic, and other forms of human behavior. And human control. More later.
Looking for patterns.
An AI trained to look for patterns would make use of all that data.
Might see things we’d never think to look for.
I’ve mentioned in the past that the Starlink constellation could be mined for signs of gravitational waves.
There are over ten thousand laser links between Starlink sats.
If precise tracking of Starlink position changes were made and logged, then, when a gravitational wave is detected by Earth-bound sensors, the AI could examine the logged data for corresponding signals/patterns that occurred at the time of the Earth detected GW.
Bobbers on mostly calm waters, rising and falling on passing ripples.
Starlink is a giant, 3D cage of lasers, and all that is required is designing the software to utilize the data.
That seems to be a odd transition; You start out talking about the shortage of training data, and transition into a telescope survey proposal.
I mean, sure, images from telescopes are “data”, but unless you’re training the AI to draw realistic looking starscapes, it’s not useful training data. It’s not going to help the AI learn how to do anything Earthly at all.
So, setting the AI aspect aside. The gravitational focus is a long ways out, if we’re going to deploy mass telescopes in the near term, maybe a better application would be a mass occultation study to locate every significant object in the Kuiper belt? For that purpose, near-Earth space is fine, even ground based scopes are useful. You just need a lot of them, with appropriate sensors to notice when stars blink, and computation to correlate blinks between different scopes.
I found the transition jarring as well. If all we needed was data of any kind, we could more easily build 1 billion microscopes and point them at the ground and feed all that data of bacteria going about their business into an ai. I doubt the ai would learn much from that.
Instead we somehow need to make the ais smarter so that they can help us figure out how to build the solar sails, telescopes, ships, and probes we need to explore the galaxy.
I think complaining about a shortage of data for training fundamentally mistakes the problem. Human beings don’t need petabytes of training data to learn to behave intelligently. That’s because humans actually ARE intelligent beings.
The problem of training data will be solved when the AI’s are actually intelligent. A human equivalent AI will only need as much training data as a human being gets to have human equivalent performance. A superhuman AI would accomplish it with LESS training data.
If intelligence is measured by how steep your learning curve is, today’s AI’s have learning curves that are only barely distinguishable from dead flat.
Yes. Exactly. In my opinion is both an issue of architecture and methodology. We provide AIs enormous amounts of unstructured data often randomised and do it over and over until the machine reconstruct some sort of sense. We humans have progressive and structured training. First we learn to experience reality, then we learn vasic concepts, then we build on thise basis. Even with driving we build a general understanding of reality first then after few theoretical lessons and few hours of practice we know how to drive. Granted you need practice to get better but we do not need millions of hours of footage.
While it’s not a solution to a lack of AI training data it’s a pretty cool relatively near term big project that takes advantage of AI capacity to handle data and AI based manufacturing scaling. We can know the entire galaxy, every Star system, every interesting object in a lot of detail in the relatively near term future. AI can handle that very large torrent of data. We can know definitively what earth-like planets exist in our Galaxy.
I like the project of mapping every significant object in the Kuiper Belt and before that every significant object in the Solar system that might be a threat to the Earth.
Another reverse use that will likely get funded first is looking down so everything that happens everywhere on Earth is under 24/7/365 detailed observation in different EM frequencies, and variety of sensors. Perhaps a bit more controversial but lots more applications. AI can handle and make very useful that data torrent too.
Indeed. The idea of training AIs with the infinite date of the heavens is maybe scientifically enticing and poetic, but of little practical and hence commercial value. It doesn’t mean it won’t happen, we just need to keep things in their proper context.
The real “infinite” and useful dataset will be synthetic ones and those gathered from the real end user in real activities of interest.
For example realistic image/video generators can be trained with 3D renders, generated to the state of the art of photorealism to obtain AI models capable of interpolate and generate all kinds of objects and movements with high fidelity. Or with video feeds of people doing stuff.
Same for bots. People with prosthetics and cameras can train them by doing millions of hours of repetitions of the desired to emulate activities.