Moonshot - Predicting the future and making JARVIS!

[danielhanchen]

Hey OpenAI community! Super glad to be here! With my brother, we’re trying to predict the future of everything quickly by simulating Planet Earth and make a “weak” JARVIS ! http://moonshotai.org/

The hope is anyone can ask questions like “how will rising sea levels affect my property price” or “how does local air pollution affect my lifespan?”. My brother made a super beautiful prototype of V1:

Thanks to OpenAI for letting us use GPT3! Previously, Nearest Neighbors based on ConceptNet / Word2Vec was used to match questions to a database of “possible questions”. We’re hoping to instead:

1. Extract embeddings from GPT3 and replace Word2Vec.
2. In the future, prime GPT3 to spit out the appropriate QA layout.

---

Now predictions must be realtime - no one wants to wait 5 hours for a QA system. So first, Moonshot Hyperlearn makes ML algorithms 50% faster, use 90% less memory WITHOUT new hardware! With 1.2K Github stars , Hyperlearn’s methods were cited in research from UW, Greece, etc and incorporated into Pytorch, Scipy etc.

---

Data collection is the main challenge - data here, data there, data everywhere. It’s a total mess. Accuracy is also paramount ! We’re releasing Moonshot Finance next month which predicts recessions and stock prices. Moonshot Earth will then predict life expectancies, disease spread, fire, flood risks, property prices and more, and will be released mid 2022.

A private beta for Finance was utilized by myself and my family, with it being nearly 100% accurate in predicting large economic recessions (GFC, Dotcom bubble, COVID crash etc), with it being trained on data from 1896 to 2021 (50% training 1896 to 1960s 50% testing 1960s to now). But, predicting the exact % drop is hard - so I’ve been on a data gathering spree!!!

In the future, the hope is to launch Moonshot into other areas:

1. Insurance - Climate Change Mitigation insurance + advising - ie whether you should plant more trees , put barriers etc.
2. Real Estate - Predicting house prices allows us to open a marketplace for rating properites and allow buyers / sellers to see future values.
3. DNA Analytics - Combining our life expectancy prediction based on natural factors with DNA allows us to fulfil the nature v nurture argument.
4. AI Hardware - Our software’s designs can be translated into custom AI chips!
5. The Moonshot Investment Fund - Using our stock prediction and market recession algorithms, we plan to launch our own investment fund !
6. Government Policy Support - Our life expectancy predictions, disease predictions, fire, flood predictions can advise companies and government policy directions.

---

We’ve applied to YC W22 so I really really hope we make it!!! We also have been approached by Australia’s Dep of Defence hoping our Earth simulation methods might be used for defence - though for now we’re in the infant stages. We’ve also got approached by Thales, Antler, Newchip, investors but we’re gunning for YC + the OpenAI Startup Fund!!! Previously I helped NVIDIA make GPU algos faster, and helped predict disease spread, traffic congestion, air pollution, domestic violence risk for the Australian government!

If you want to connect and help out, find me on Linkedin! Likewise, email me at danielhanchen@gmail.com! Likewise, we’re looking for help on the frontend - so if anyone wants to help out - reach out!!

Once again super excited to be able to try OpenAI’s powerful GPT3, and super proud to be in this fabulous community!

[I deleted the old post but I should have just re-edited… I made it more detailed and more succint!]

[daveshapautomator]

Keep in mind that JARVIS is a trademark of Disney and Marvel. That being said, I’d be happy to contribute my work on cognitive architectures to your goals. Let me know if you’re interested in collaborating. As my work is publicly available, there are no licensing fees.

[danielhanchen]

Oh yep definitely wont be using JARVIS officially ! Just wanted to temporarily use it to approximately make people understand what we’re trying to build.

Sadly though when I talked to some people, they don’t even know what JARVIS is… But anyways I’m more than happy to collaborate! Let me first have a read of your stuff and get back to you David!!

[danielhanchen]

Interesting you first applied to GPT3 to no avail, then used GPT2 then they approved your access! Congrats!

On another note - you seem to refer GPT3 as AGI?

[daveshapautomator]

Well, that’s not exactly what I meant.

I started with GPT-2 but it was too limited. As soon as I got access to GPT-3 I never had to go back.

As far as AGI, no I don’t believe that GPT-3 qualifies. It can’t write a novel, for instance. A system that can write a coherent novel would be a major step forward because it has to do many things, such as keep track of plots and characters and world rules. That requires a tremendous amount of cognitive control. Of course it’s entirely possible that future language models could spit out entire novels without some of the things that humans require to do it.

Anyways, no, I just believe that GPT-3 is intellectually superior to humans in some resources. For instance, you can ask it to speculate on society, philosophy, and science. It can generate far better questions than humans as well. You’ll see plenty of examples in the book.

[sergedawg]

this is one the coolest things I’ve seen in a while, best of luck with your yc application.

[deanmkn]

This sounds exactly like Rehoboam from season 3 of the Westworld TV series (if you haven’t seen Westworld I highly suggest it).

None-the-less, sounds wildly ambitious, and obviously if it were even possible it would be the highest valued IP in the world. But best of luck!

[danielhanchen]

THANKS SO MUCH @sergedawg !! That means so much!! Ye really hope we make it in!!!

[danielhanchen]

LOLLL someone literally last week said exactly the same thing!! I haven’t watched it yet, but seems like I should now!!

Thanks as well so much @deanmkn !! Ye it’s a hard journey but that’s what makes it fun!! Really hope Moonshot will work!

[danielhanchen]

Yep GPT2 was a bit limited I agree.
Well it could write partial parts of a novel!! Though you’ll need to give it hints here and there. Yep the issue is transformers have a finite time horizon - 1024 chars or even say 4096 words at max. It gets worse since attention is O(N^2). Even with Linear attention O(N), matrix multiplication and memory requirements explode. There needs to be research into compressed long time horizon embeddings.

Yep GPT3 is “superior” I guess in some areas. Yep! It’s cause GPT3 has “learnt” (more like interpolated) from tonnes of people, and the “average” of all the input makes GPT3 possibly better than an “average” human with “average” IQ.

[daveshapautomator]

Remember that human working memory is very limited, yet we can write novels. So how do human brains achieve this? That is what I go into with my book. We do have some similarities with Turing machines in that we spontaneously pull relevant information from long term storage and add it to our working memory. If you check the bibliography you’ll see that I read quite a few neuroscience books to help form NLCA. Search functions, such as vector-based search and BERT-like augmented search, will be critical to achieving AGI and the goals you outline. This is because, quite simply, so many tasks are TOO big to feed through a network in a single pass. You must be able to iterate on issues over and over again, building better understanding and clarifying for better results. This is how humans achieve mastery. Certainly, I can imagine a future where this method of iteration is no longer required by machines, but for all big tasks and big problems that require lots and lots of knowledge, I think my system will outperform individual inferences for a long time.

[danielhanchen]

Sorry couldn’t reply earlier! Had a hectic day!!

Ye working memory is limited for humans. Long term retrieval is a bugger for Transformer based architectures, but Transformers are a good step towards a solution.

I’ll definitely have a read on NLCA! On another note, I might make another forum post just about what we’re talking about!! (Long term memory)

[daveshapautomator]

Let me know if you do. It’s important to differentiate different kinds of memory. One type is episodic recall, while another is declarative knowledge. Declarative knowledge can be baked into GPT-3 via fine-tuning. It’s possible that episodic knowledge can as well, but I doubt it will be high quality. Until they iron out problems like confabulation, fine-tuning will not be a reliable way to record this kind of memory.

[danielhanchen]

Yep will do!
Interesting didn’t know about episodic recall / declarative knowledge! New terms to learn!!

[danielhanchen]

By the way if anyone’s interested we made a 4min demo (unlisted) for what’s Moonshot up to. It was just an update to our YC W22 application!

Everything is a bit all over the place since we’re still sorting everything out!

Moonshot Technical Demo + What we've been up to!

[daveshapautomator]

This is a great video but it’s a bit technical. You might want to talk about who is going to use this (audience), how it will make money, and how it will make businesses or people’s lives better (value prop). As a smalltime investor, I’m left wondering “okay great, but who is going to pay for this? Where is the value add?”. Many of the details you go into sound great for scientific papers, but that does not necessarily translate to commercial success. I’ve had to tell a few prospective entrepreneurs from here “no” because I did not believe they had a good business plan or a grasp on what was required to even develop a software product.

It seems to me as though you might be trying to cast a very wide net, trying to be “everything to everyone”. While it’s good to be ambitious, it’s still better to become an absolute master at one thing, disrupt one market, rather than try to disrupt everything at once. Even folks like Bezos and Musk disrupt only one thing at a time.

What I mean by this is that stock market prediction is nothing new, and you’d be going up against entrenched financial institutions who have invested billions in their fintech infrastructure. So you would have to really stand out from the crowd. Not to mention that “random walk” mentality still reigns supreme in the fintech world, hence why we keep seeing problems with short selling. In the long run, stocks go up. Everyone knows that. In the short term, there’s beta and theta.

Then, when you look at health dashboards and such, that’s all been done by people like the World Bank, CDC, WHO, and so on. Again, nothing new and you’d be going up against entrenched powers. Why don’t you carve out a new niche here? Here’s what I would personally want: A phone app that tracks what you do, say, and eat all day and uses that data to generate recommendations for diet and self care. Take a look at this paper which talks about predicting mood, stress, and health a day in advance. If you commercialize a smartphone app that can do this and give me a video-game style output for my short term health, that would be phenomenal. You would have a leg up because it’s backed by science, at a time when gimmicky devices make impossible promises.

The real market to target is hyper-personalized everything. Personalized finance, health, diet, mood, energy, relationships, etc. This what I outlined in my book with Raven/NLCA - which is meant to be a “personalized information concierge”. If you could combine your powerful predictive abilities with a friendly AI companion, then you would have the market cornered.

[danielhanchen]

I’ll address your reply in points!

1. Too technical: I totally agree! Yep even my mum / dad said that LOLL! The aim of the video isn’t necessarily to show a true “demo”, but rather what we’ve been up to. A “true” demo will be made much into the future. YC’s application just wants to see what we’ve been up to, and thus code, data, graphs etc was shown.

2. Mastery of 1 sector. I actually totally agree with this. However therein lies the issue of what we’re trying to do - to simulate the future and predict the future of everything. The “thing” which I’m experienced in is prediction well more specifically fast predictions.

3. Stock Market Prediction. We’re not “predicting stock prices” but rather predicting market crashes. The aim maybe in the future is to do that. Too many people have constantly criticised the route into stocks, but I wholeheartedly disagree. How did 2Sigma, Renaissance Technologies start? Both started with people from algorithms, computer science. And anyways I’m not competing with them. That’s not Moonshot’s product, but a small small portion of what we’re doing. Stocks influence every other portion of our Earth simulation. It captures “sentiment”, “happiness” and “economic activity” and more. All this influences the Earth sim.

4. Dashboards. We’re not creating a country level or state level health dashboard but rather we’re predicting individually your life expectancy, cancer risks, how rising sea levels affect your personal property price, your long term health issues like diabetes, heart disease risk etc and how all this is influenced from closeness to traffic, air pollution, income, neighborhood differences, closeness to factories, soil pollution etc.

5. Health app. I actually disagree a phone app is a good idea. There’s way too much competition on tracking eating, habits, dieting etc. Therein lies the issue changing short term habits v changing long term habits. Moonshot is focusing on a different market - we’re focusing on the long long term.

6. AI companion. Yep exactly hence why we’re trying to make a “weak” JARVIS!

But ye once again I’ll definitely have a read on Raven when I get the time! Thanks though @daveshapautomator for your time!! Super appreciate it!! Your points are all super valid

[daveshapautomator]

Well, having read a tremendous amount about longevity, I don’t think that there is such a thing as “long term habits”. Health and longevity are a set of hourly and daily activities that you do for decades at a time. Behavior change, such as changing diet, is also about small, individual choices that add up. Everyone who’s done any research will tell you that health comes down to a few things: diet, exercise, sleep, and mental health. That’s why I recommended creating an app that will help you manage decisions in realtime. Like if it could say “hey, you’ve taken in a lot of sodium today maybe have some high potassium foods to offset that”. Oh, and I’m not talking about a tracker. I’m talking about a forecaster. Since you focus on fast prediction, you could take that historical data (which should be gathered automatically, or just by snapping a quick picture of your meals), you can make instant recommendations to optimize health for the long term. No one cares about long term cancer risk or diabetes risk. We all know that aging sucks and we all know the basics about avoiding those terrible outcomes. What we need are recommendations with surgical precision. I honestly don’t think you’ll find a market for “long long term” outcomes. Everyone dies in the end, and health deteriorates over time. These are immutable facts. People want to feel better right now. There are countless books and podcast episodes about the fact that people are horrible with long term rewards. You might benefit from Freakonomics and marketing books.

I’m afraid that you’re putting ideas first and not human needs. Yes, it’s neat that you can forecast so far into the future, but as I said, the only people who care about that are the UN and WHO and IMF. That’s just not how people work psychologically. You’re never going to convince anyone to gather lots of data to watch their life expectancy change by a few months or few weeks. Furthermore, if you don’t incorporate genetic data, you can still miss huge things and then you’ll lose credibility with your customers. You might promise to help avoid CVD but then someone might still develop ALS or Alzheimer’s. Then when you look at things like SLR you’re talking about 250 million climate refugees over the next 50 years. This is, frankly, an edge use case when you compare that number to 8 billion other humans. It’s also on such a long time line that no one cares.

Example: Yesterday I was struggling with some chronic pain from old injuries and I remembered that the pandemic is waning in my area so it’s safe to go back to my gym. Strength training has this magical ability to reduce chronic pain for some people, like myself. Do you have a way of building these kinds of recommendations into your apps? I think you’d be better off to focus on making a strong JARVIS and use some of these other predictive powers by incorporating them into the Inner Loop of NLCA so that your chat interface has access to plenty of insights to share with your end users.

“Jarvis, why do I hurt today?”
“Well, sir, you haven’t been to the gym in 18 months, and according to what you’ve said, the gym usually helps you feel better. Perhaps it’s time to go back?”

That’s the sort of thing I want.

@DR.AMES @Kopaka - what do you think? Does Daniel’s work show potential for modeling human mental health?

[danielhanchen]

Q: Where will historical data for your daily routines come from?

The world is aging real bad. People are scared of dying. I’m scared of dying. Did you know that life expectancy is going backwards? We haven’t solved immortality nor have we magically increased our lifespan recently.

Q: Evidence predictions will change lifespans by a few weeks / months?

DNA Analytics - I wrote in the first post we’ll focus on DNA in the future to solve the nature v nurture argument.

Climate change affects everyone. We’re focusing on climate change mitigation insurance, forcecasting property price effects etc.

Q: How will “Strong” JARIVS be possible? Do we wait 10 years to collect data?

Anyways thanks for your comments and time. I think you’re getting too focused on Moonshot’s Health area. Our mission isn’t health necessarily but making a full Earth simulation to predict the future of everything.

[daveshapautomator]

Honestly, I ignored your assertion of “making a simulation of everything” because it’s patently an absurd claim. Extraordinary claims require extraordinary evidence. Do you have a PhD in physics? Chemistry? Genetics? Geology? Climatology? You’re chasing a dragon, trust me, I’ve been there. What motivates you to pursue these goals? What is your internal motivation? You can’t save the world singlehandedly.

Anyways, you don’t need 10 years of data to create a strong JARVIS. Read my book. GPT-3 already has hundreds of years of data.

Anyways, everything you’re trying to do can be solved by AGI or ASI. Why don’t you focus on achieving those goals? You can solve all humanity’s problems at once with that.

As far as death, that’s something that everyone has to square with. Even if/when we achieve indefinite lifespan, there’s nothing preventing accidents or rare diseases. If that’s what you’re trying to self-soothe, I can say from first hand experience that busying yourself with external projects is only kicking the can down the road. The path to peace is internal.

I hope you can find the peace you need. You won’t find it in trying to save the world. Good luck. I mean it.