Will SpaceX using technology from Open AI

[IRobot]

In my experience, there's a lot to be said against modern commercial software development methods. If I were an astronaut, I'd have died during launch twenty years ago. Unless resurrection were possible, in which case I'd have died during launch thousands of times.

The problem is not with modern software development methods.

The problem is that the average quality of SW developers has dropped dramatically due to market demand of millions of SW developers. This gave rise to a generation of SW developers who are not engineers. They know how to code, they know more algorithms and design patterns, but they are unable to "engineer".

Regarding SCRUM and other modern processes, I always say that the project idiosyncrasies define the process, not the other way around.
Waterfall can work perfectly fine if there are no expected changes of scope or milestones.

And actually some of the modern processes were created to account for bad engineers in the projects. One example is the abusive use of story points in SCRUM, which tries to solve the problem that modern SW developers can't estimate.
Story points should be used in some projects, but not always.

[Frogstar_Robot]

Waterfall can work perfectly fine if there are no expected changes of scope or milestones.

IOW, since scope and milestones always change, waterfall never works well.

[tdperk]

As I recall, The Moon is a Harsh Mistress was written back in the day when people thought that if a computer was upgraded until it was big enough, it would spontaneously become intelligent.

There are still people that basically think like that...

For that matter, do we actually need intelligence to run the rovers? A dog's level of autonomy is probably good enough.

You seem to think that animals do not possess any intelligence at all. Actually dog-level intelligence would be considered pretty damn advanced AI. I bet that any AI researcher would give his right hand for artifical system with mental capabilities similiar to dog's, since it would be light years beyond anything we have now.

Because the qualifier used is big enough, by definition they are correct.

[tdperk]

(domesticated dogs have been shown to give up and defer to owners on problems that are well within their capability to solve).

And they are wise to do so, it's easier.

Let the big honking supercomputer on two legs worry about it, while you lick err yourself.

[cro-magnon gramps]

This a very simplistic explanation of the future of computing and AI
as I've learnt about in the past Winter.... So no flame wars please....

AI Streams
Narrow AI does what it is programmed to do...
General AI does what it can learn, in a very specific area, to do...
Singularity AI can do what it thinks of to do, or is asked to do...

Classical Computers
Transistor Based, using silicon and copper etc
to be replaced in the future by newer technologies like:
Photon Transistors, using fiber optics and light
and more esoteric types like DNA/Bio based computing...

Quantum Computers
These use Quantum Superposition, and Quantum Entanglement with Q-Bits
These will NOT in the near future appear in our tablets, lap tops, Desk Tops,
cars, cell phones, refrigerators or Satellites etc They will be the new Main Frame
Super Computers, home to the Singularity...

Computers at the Nano Scale
This will eventually encompass all types of computers...
Their components will be smaller than a pin head,
BUT
They will scale up to integrate together for use in AI's various types of computers
and applications... the scary part is that I can't find how fast that will happen...
there are vague predictions, but not what I would call firm enough to quote...
These would be used first within the next 50 years in Bio Computing (Neurolink) and
Classical Computing tablets, lap tops, Desk Tops, cars, cell phones, refrigerators
or Satellites etc

My point is, that while we have grown used to the Moore's law pace of growth in computer
development and complexity, we are on the verge of an exponential jump that I don't see
anyone preparing for... We know that Elon has two companies he is associated with Tesla and
Neurolink, that at the moment heavily directed towards AI and it's integration into areas of our
lives; as well as a connection with a company that is actively developing General AI and possibly
Singularity AI... we don't know what level of integration he is using these various developments in
computers and AI in his other companies... it's like he has his own Skunkworks(1) for want of a
better term...

To bring it to the reason for this essay, Gwynne Shotwell's notions about future developments at
SpaceX involving Interstellar Propulsion, make more sense if one can posit a cross pollination of
ideas, technology and engineering based on advanced computer capabilities from EM's various
companies... he has had 16 years to get this far, so given 22 more years of development, and
their potential use of the new tech in EM's present and future companies, how farfetched is the
idea that they could be working at the early stages of Interstellar Propulsion for a launch of a
unmanned prob by 2075... based on all of EM's different companies, it is obvious that his
imagination and it's reach are not limited by anything, but the Physics of the Universe...

(1) A skunkworks project is a project developed by a small and loosely structured group of people
who research and develop a project primarily for the sake of radical innovation. The terms originated
with Lockheed's World War II Skunk Works project.

[Nilof]

Possibly yes for some tasks like mining rovers or pollenization of plants in greenhouses, but not for anything critical like landings etc. Machine learning is a last resort if you're unable to make things work any other way, and are also much slower than traditional programs performing the same task. You can't directly fix flaws in the program other than by finding some way to train it to do that thing correctly, which can often end up failing.

[launchwatcher]

Possibly yes for some tasks like mining rovers or pollenization of plants in greenhouses, but not for anything critical like landings etc. Machine learning is a last resort if you're unable to make things work any other way, and are also much slower than traditional programs performing the same task.

Probably better to say"resource intensive" rather than "slower", because ML models are generally embarassingly parallel with relatively shallow data dependencies so you can almost always make them go faster by throwing more hardware at the problem.   
Plus, the underlying computations are sufficiently regular and simple that people are building custom hardware to run the models.
One such "TPU" chip has 64K 8-bit ALUs organized in a systolic array, while a conventional CPU core might only have a handful of much bigger 64-bit ALU's).

[leovinus]

A bit late to the thread but an interesting read. Some thoughts below.

Just to put my words in before this gets locked by mods (a similar thread in Advanced Concept already disappeared into the ether): OpenAI is mostly pure research, they don't focus on practical applications (except maybe AI safety). Their results are built on others in the field, and they publish their research in papers so that others can built on it. So there wouldn't exactly be "technology from Open AI", a better question is whether SpaceX will use machine learning technology (Deep Learning and Reinforcement Learning in particular) in their work.

The game player AI is based on Reinforcement Learning, its practical application for SpaceX may be robotics that can learn by following human demonstration, instead of being programmed, although this application is still in its infancy. More mature machine learning technology would be Deep Learning based computer vision, which should be able to replace anything that requires human vision.

Agreed. You could split the title question in two parts:

(1) Will SpaceX use machine learning based on statistical models in their systems on the way to Mars?

Of course. Every Silicon Valley company is doing that, as well as most companies in Europe, Japan and China. If you are in the software business then you better have engineers who can apply machine learning and statistical models to reach your goals. The 'trick' is to choose the correct tool (architecture, design, implementation, hardware) for the correct jobs though.

That leads to
(2) Will SpaceX using technology from Open AI?

Maybe. OpenAI is tiny w.r.t to the number of scientists & engineers at Amazon, Google, Apple and others. That limits what OpenAI can do but at least they are not bound by a product. There is a tendency to produce more academic than engineering results. They are experts at applying general machine learning to solve for a goal but sometimes in search of an application.

Not sure SpaceX needs it now, but aren't some of the chinese companies playing around with first stage landers that use machine learning? Or is that just a fancy way saying you've paid a dynamicist to optimize your positive feedback closed loop algorithms?

Presumably the second. There are a few people who think neural nets should be renamed to something like node-based function approximation to reduce sensationalist news articles.

Agreed. It is appropriate for us to not over-use the term "AI". We have seen these situations in the '90s, e.g. [1], where a company or even country will forge ahead to solve it all. Ambition is great but a grain of reality helps to ground us all and define better goals. In a nutshell, the Japanese 5th Generation "AI" failed spectacularly because the programming tools and technology where not powerful enough. Too deterministic, too slow, not enough algorithm innovations.
 
What we call "AI" these days also had a bad start in the 80's. The basic mathematics of backpropagation were the same but the machines were slow to process large amounts of data. Hence, a part of the current AI and what we call the (extreme) Deep Learning and applications is simply applying old mathematics on hardware that is exponentially faster.

The quantum AI will go through a similar phase once the hardware comes online and we can finally run quantum algorithms, like Grover's sorting algorithm, or prime factoring. Therefore, algorithmic research is one key to future success.

Possibly yes for some tasks like mining rovers or pollenization of plants in greenhouses, but not for anything critical like landings etc. Machine learning is a last resort if you're unable to make things work any other way, and are also much slower than traditional programs performing the same task. You can't directly fix flaws in the program other than by finding some way to train it to do that thing correctly, which can often end up failing.

Based on my experience, it is more a question of choosing the right tool for the right goal. If you can solve a goal (quality, failure modes, speed, memory footprint, embedded, network ) based on analytical mathematics and a traditional deterministic program that is maintainable and extensible, then go for it. An example is the inverted pendulum and I assume that something like that is pumped up to the max in landing a Falcon booster. However, as we push the technology boundaries, you will come across questions that you cannot design an if-then-else program for. Blends of deterministic and statistical machine learning work well in that case.

Waterfall can work perfectly fine if there are no expected changes of scope or milestones.

IOW, since scope and milestones always change, waterfall never works well.

Indeed, In Silicon Valley, I'd say a "pragmatic prototype based" process is the current flavor. Not CMM (too rigid, not flexible enough to deal quickly with change as mentioned above), not SCRUM, not Agile, not Object Oriented. The way to deliver on-time, to budget and with requested quality is to make a prototype, set a delivery date, keep an eye on the critical features and use the best software methodology you like to engineer it and deliver on time, driven by the prototype(s). Having great engineers helps of course As an example, something like the iPhone software is not build using CMM. It just wouldn't work. The caveat is of course that you will not get 100% of the 'specification', more like 90-99% in reliable way.

[snip] Machine learning is a last resort if you're unable to make things work any other way [snip]

Maybe we should rephrase that. Machine learning is not as a 'last resort' but a solution you choose to achieve a goal. Certain goals will have no other choice. Example: a speech recognizer to transcribe sentences. Back in the '60s and 70s, early deterministic approaches based on parsing languages where not accurate, not robust, not applicable to all languages. Moving to statistical models improved the scalability to other languages, extensions in functionality, more accuracy and more languages by 'just' applying more data. Suddenly, instead of a heap of if-the-else software that no engineer can read anymore, your goal becomes feasible, and the software is maintainable and extensible based on Markov, Neural and other models. It does require a different engineering skillset though.

[snip] and are also much slower than traditional programs performing the same task. [snip]

That is a bit too general IMHO. I'd say "it depends" Most of you run machine learning, based on statistical models, on a daily basis, on a device that we loosely call a "phone". Based on the correct architecture, design and implementation of the software, you'd be surprised how fast, e.g., a Neural Network can run. Based on personal experience, I have often seen machine learning solutions outpace older deterministic solutions. However, as mentioned earlier in the thread, you need to engineer it well where the quality and experience of your engineers is critical.

[1]https://www.nytimes.com/1992/06/05/business/fifth-generation-became-japan-s-lost-generation.html

[Nilof]

TBH, I think machine learning performs best at tasks where the requirements (not just the implementation) are difficult to express. If you can only define the task by pointing to a bunch of correctly solved tasks by a human and where the final solution is likely to include a large number of magic constants either way, machine learning tends to do great simply because there is no way to easily break it up into smaller tasks. If it's something that a human could easily give some formal specification for that can be expanded to any level, then clearly that's going to be more straightforwardly expressed as a traditional program.

With that said, the distinction can get blurred somewhat by things like decision trees. XGBoost-based solutions win Kaggle contests all the time, and unlike NN's they can be opened up and read somewhat like a (messy) regular program. They just don't necessarily work on all problems.

[IRobot]

(1) Will SpaceX use machine learning based on statistical models in their systems on the way to Mars?

Of course. Every Silicon Valley company is doing that, as well as most companies in Europe, Japan and China. If you are in the software business then you better have engineers who can apply machine learning and statistical models to reach your goals. The 'trick' is to choose the correct tool (architecture, design, implementation, hardware) for the correct jobs though.

Deterministic algorithms have no need for machine learning. In fact, they cannot use it. Most SW development does not need machine learning. You are exaggerating its benefits.