Lex Fridman Talks Singularity With Ray Kurzweil

by

Lex Fridman talked with Ray Kurzweil about the Technological Singularity.

Ray still asserts that computer artificial intelligence will pass the Turing Test in 2029 and the Singularity will happen around 2045. In the past, Ray has clarified that his predictions date are usually with an implied plus or minus one decade. A Ray Kurzweil prediction with a date can be plus or minus 10 years and Ray would consider it a good prediction.

I have some recent videos related to Ray and his predictions and measuring the progress of technology through 2100.

I have a video where I describe how Elon Musk will bring about much of what Ray Singularity predicts with the Teslabot and self-driving cars. Teslabot, humanoid robots, will expand labor that costs less than 50 cents per hour and enable economic abundance and rapid technological advancement.

I have another video that describes a scale for measuring the economic impact of a technological singularity.

The Technological Singularity is a predicted point when technological growth becomes radically faster.

Normal economic doubling for the past 100 years and the expected levels to 2100.

The world economy has been doubling every 25 years for last 100 years
We have 15 times as much steel and oil as the world had during World War 2. There have been four doublings of key materials used globally in the past 80 years.
Therefore, we would expect the world too double its economy every 25 years til 2100.

This would be:
$220 trillion (2011 dollars) in 2045. [$24k per person]
$440 trillion (2011 dollars) in 2070
$880 trillion (2011 dollars) in 2095.

What if the Teslabot scenario happens? What about something life Kurzweil’s singularity. Global meaningful Technology Acceleration is extra doublings.

Acceleration adds more doublings
Level 1 Extra double 2020-2045 : $440 trillion PPP GDP by 2045 [$50k per person]
Level 2 Extra 2045-2070: $1760 trillion PPP GDP by 2070 [$176k per person]
Level 3 Extra by 2095 $7040 trillion by 2095 [$700k per person]

6% per year world GDP growth instead of 3% per year for 25 years

There are 16 doublings needed to achieve a Type 1 Civilization. This means an economy that is 64,000 times larger than it is today. We are at $100 trillion world economy. $6,400,000 trillion would be a Type 1 civilization.

Real World AI would be general artificial human level intelligence. Capabilities to provide broad levels of human jobs and tasks.

Teslabots able to perform loading and deliveries to massively boost the supply chain.
Teslabots able to perform manufacturing tasks in the factory.
Teslabots able to use machines built for humans.
Teslabots able to work in factories to make factories self replicating.
Teslabots able to perform mining.

These capabilities would make economic growth massively exponential.

They would be able to rapidly lower the cost of other Teslabots and increase production from millions to trillions of Teslabots.

The wealth and production capabilities would make it easier to devote more resources to improving artificial intelligence, robotics and manufacturing and all other research. Just as our current economy that is 100 times larger than the economy of the late 18th century has the resources to accelerate science and technological research and development.

Giant.ai is a far smaller company than Tesla and they are making progress with an upper torso humanoid factory bot.

Tesla sells Model Ys for about $60000, but it currently costs them about $30000-40000 to make them. A Teslabot is 1/30th of the mass of a Model Y. It would use 1/30th of the batteries. The software is an overall cost of development. If billions of bots are produced then the cost would trend toward the cost of the hardware plus Apple iPhone-like margins including the software (say 40% gross margin). At Model Y cost of $30k then the hardware cost for Teslabot will go to $1000. $2000 with margins and software. A bot can work for 8000 hours in a year. 8760 hours in a year. $2000 divided by 8000 hours is $0.25. If you add 10 cents per hour for electricity then it is $0.35 per hour. Going beyond that is bots can work in the factory and work cheaper than humans. Currently 15000 workers in Tesla China factory. Replace all of them with $0.35 per hour bots. Reduce labor cost component. If a lot of bots can increase production rates. by 2X then all costs spread over more units. Bot-produced solar and batteries can lower the cost of energy by vastly increasing the supply. Those trends could get us to $500-1000 per bot costs and lower energy costs.

In terms of the strict definition of super-intelligence, this would be a soft take off version of super-intelligence. It would be 100 years to get to a billion times total human level intelligence. Production of quadrillions of human level bots to colonize the solar system would reach a billion times ten billion in the 22nd century.

Adding in molecular nanotechnology, say in the 2030s would supercharge the Teslabot technological singularity to a semi-hard takeoff.

30 thoughts on “Lex Fridman Talks Singularity With Ray Kurzweil”

  1. Moravec’s Paradox is almost as bad as the tyranny of the rocket equation. They both won’t down without a fight.

  2. Impossible to predict something like that pointless even but maybe we can agree it is a matter of when and not if at least

  3. It would be interesting for Brian to compare Kurzweil date for singularity vs Karyik Gada’s date if 2064, which he discusses in his ATOM thesis.

    • Yes good point

      Kartik Gada

      can be found easily on youtube he puts the singularity at 2064 plus or minus 8 years at present and has a series of videos explaining his reasoning they are worth a look for those interested in this subject.

  4. A technological singularity is sufficient implementation of a new technology that will begin to make such irreversibly huge changes in human society that it is difficult, if not impossible, to foresee what will come after. Each one usually appears in about half the time it took for the one to occur after the one previous to it. Bear in mind that singularities aren’t events like light switches but, instead, can span a substantial period of time.

    PAST SINGULARITIES (3 most recent)
    1815 – The Industrial Revolution (240 years after printing press)
    1935 – Electronics & Computers (120 years after previous)
    1995 – World Wide Web (60 years after previous)

    FUTURE SINGULARITIES
    2025: Full Automation (Cognition based) capital-based income grows hugely and in inverse proportion to wage-based earnings as a percentage of all earnings. (30 years after previous)

    2040: AI – More than just AI, this is SI, synthetic intelligence, in that it is not a workaround to achieve results similar to what a human could produce, this is the real deal, it’s just not made of animal flesh and it might not have any internal motivations (no lizard brain conflict) (15 years after previous)

    2047: Biological (advancements that make average lifespan estimates useless) (7.5 years after previous)

    2050: Mind-to-Mind? Man-Machine? Nano replacement of cells? (3.25 years after previous)

    2052: Singularity (with a capital S) (1.6 years after previous)

    2053: … in 1.3 years
    2053: …. 7 months
    2053: ….. 14 weeks
    2053: …… 7 weeks
    2053: ……. 24 days
    2053: …….. 12 days
    2053: ……… 6 days
    2053: ……. 3 days
    2053: …….. 36 hours
    2053: ……… 18 hours

    “Life moves pretty fast. If you don’t stop and look around once in a while, you could miss it.”
    — Ferris Bueller

  5. There is normally a very nearly perfect correlation between wealth and available energy.
    When talking about scaling bots to billions of units, it may be interesting to look at the energy budget and compare to human beings.

    A human brain is a lot more energy efficient at computation than current computers. Exactly how much is hard to assess but a human brain burns 12 – 20 watts of which one or few are for computation. How much computation? Hard to calculate but there are numbers like 1 exaflop floating around. Most energy spent by the brain seems to be for communication and metabolic processes.

    To emulate a human brain would require many datacenters full of hardware consuming gigawatts (ish).

    The energy expenditure when it comes to AI and humans is two-fold. One overwhelming part is for training and the other is for inference. This makes it harder to do the calculations. For example: A factory worker bot will need a lot of initial training but can be cloned thereafter. Further training is needed only for adaptations, calibration and minor changes.

    Humans work somewhat similarly but our brains can’t be cloned and things happen on a larger time scale. We do most of our training for the first 20 years at huge cost and then use the skills for another 50 years with muss less cost. One would guess a general intelligence bot would have to work on much larger time scales than a simple factory bot.
    One example is the Tesla DoJo trying to solve the relatively simple and scope limited self driving problem. It collects data from the world and trains for years and years at high cost in the hope of eventually being able to clone a working solution.
    Here in Sweden, the conditions are nothing like the ones in sunny California. We see that the self driving software has huge problems when confronted with things like: a low hanging sun blinding the cameras, snow erasing the contours of the road and clogging the sensors, road maintenance work etc. etc. This indicates there is a need for quite a lot of adaptability (continuous training), even for a narrow scope like self driving.

    If the artificial bots are supposed to have individual built in training in their processing, this will increase energy expenditure beyond achievable anytime soon. Evolution has optimized for energy efficient solutions and it will take a while to improve on that.
    Just waiting for Moores law may not suffice and we may need room temperature quantum computers before it can happen. All that points to large centralized processing systems controlling dumb but autonomous bots as opposed to distributed, basically self replicating and evolving smart bots. More like an ant colony than the homo sapiens solution.

    • We don’t actually need the individual bots to be capable of much learning, given that in bots learning can be cloned. You can eliminate almost all of the labor involved in industrial production with the bots not having any more learning capacity than insects, provided that there’s some system in place for detecting failures and learning solutions to distribute to the non-learning bots.

      That’s the way Tesla does the self-driving learning: They look for cases where humans have to take over, and feed the training system with those situations. Then send the learning back out to the cars.

      Humans need huge learning capacity because every human has to relearn almost everything on their own. Bots? Not so much.

      At least for industry. If you start rolling them out for environments like homes, or routinely dealing with people, the learning requirements go up exponentially, because the bots will continuously be encountering new situations.

      I see a possibility of an almost feudal situation, if the robots end up taking all the labor that doesn’t require individualized learning, leading to a great deal of wealth in the hands of those who own them, and putting out of work the humans who had done that work before. What jobs would remain for them?

      Service jobs, basically.

      • Revolutionaries, unfortunately, if the government isn’t permitted to use enough bots to satisfactorily support all the non-bot owners, and providing, of course, that they aren’t simply overmatched by bot armies.

        • Why assume that the government is going to care, either? Plenty of governments have turned genocidal, and if a majority of the population were no longer contributing anything but mouths to feed and votes every two years, I think the best case option is genocide by attrition, if government has anything to say about it.

          In fact, if you look at how few governments are bothering to pursue pro-natal policies in the face of radically sub-replacement birth rates, it’s arguable that Western governments have already settled on a policy of genocide by attrition, and just haven’t formalized it yet.

          • I keep worrying some of the worst hit countries will start trying to implement the hatchery-style crèches in Huxley’s Brave New World (written in 1931). The technology is mostly already available.

            Granted, it would be fouled up and eventually fail, but what an enduring mess it could make in the meantime.

              • Except that it appears that any income below about $200k gets people into the “not willing to reproduce” range.
                It isn’t low income that produces no children.
                It’s much more likely to be the urban environments and current culture.

                Of course, urban environment and culture can also be used for a genocide by attrition campaign.

                • I believe that the stats point to higher income folks having fewer children than lower income. Thus we are looking at genocide by attrition amongst higher income\iq along with reverse evolution as the less able spawn more.

  6. what’s the economical (reasoned) value of a $60000 model Y, if labor is reduced to a value of $0.35/hour?
    what’s the share of labour cost on a model Y?

    What’s the impact on organizing stability for societies and precautions to be done with appearing ‘technological singularities’?

    • Remember that price does not vary with cost, only with supply and demand. That being said, Musk may push for slashing the retail prices of cars if labour becomes much cheaper than it is now; I’m just not sure whether he’ll have the strength to push through the council of shareholders, who might not necessarily want to pass up an opportunity for profit.

    • Labor for a Model Y is about $3000-5000 per car. A Shanghai produced model Y having lower labor costs. Teslabot lowering labor costs to 35-50 cents per hour could reduce Tesla labor costs per car by $2000-4000 per car, but more profits would come from increasing production levels from each factory. Make 2 million cars instead of 1 million cars from Shanghai and that would increase from 35% margin and $21000 of gross margin per car for 1 million cars or $21 billion in gross profit. Instead 50% margin and $30,000 per car for 2 milion cars or $60 billion in gross profit.

      • “In 2021, Tesla’s average national wage for manufacturing jobs in
        the U.S. was $21.60/hour plus benefit […]”
        Statistics for EU show an average wage of ~13.x€(2018) and average labour cost of ~29.1€(2021), if similar for US Tesla Gigafactories (Shanghai, average wage ~$24000, >15% increase since ~2010, 2022 ~650000″model3&Y”/a == 13000″3&Y”/week (30d24h 50000/3month, ~45-50s/Y), 445 robots for model Y on one production line; Nevada ~3MWh/worker per year2021; Tesla all2021: 936000/39000’line employees’ ~24cars/worker), labour cost of ~$45/hour (@$3000-5000) relates to ~65-110 working hours each car?

        With average working hours ~48weeks*40h*11/13=~1600 fulltime hours availability (12h4d/2d.off? Shanghai temporary super-shifts ==15000″3&Y”/week), resulting to ~14-24 cars per year from each working person, collective ~85000 workers per Gigafactory2M (compared to 2021 all Tesla average productivity ~24cars/worker ~65-85h?/worker&car).
        (Model 3 $35000, materials cost ~$16500, Teslabot ~100h=$35-50, ~120 ModelY/a?, Gigafactory2M ~17000Teslabots, global car sales 2021 ~66M = 3Giga2M&10years, US automobiles (excl. trucks) ~100M = 5Giga2M&10years, average age of light vehicles in operation within the USA is ~12years, for (battery) electric vehicles ~3.8years now@2022, ~1/4 of personal miles in cars are work related, ~2022 ~210M drivers each ~14200mi/a = ~3*10^12mi = (BEV ~18-38kWh/100mi summer/winter) >540TWh)

        Productivity of ~85000 employees (on supply chain) on Teslabots, Teslabots on Model Y’s?

        Where’s net value input to GVA from without cheap fossil energy sources (nuclear? renewables? suitable geothermal?) and increasing amount of mining/material resources being re-/upcycled (space mining?) and what’s a necessary timeline for societies to keep stable avoiding social splitting because of unbalanced education&wealth distribution?
        What’s the value of educational work (like here?), then ($), if Teslabots release humans from manual labor?

        • that’s the original meaning of ‘money’ as a means of comparison of all aspects of social interaction and for balancing resources and demand
          (unless speculative transactions of capital (‘interest and cost’?) gets ‘self-referential’/’a financial singularity’ being a measure for value for goods&services) (?)

        • Concur. The actual cost (not the price and not the profit) of anything is determined by the amount of human labor necessary achieve it, or to obtain the funds necessary to obtain it (the latter being for things that cannot be created in modern times at any price (mostly antiquities and real estate).

          Self-replicating technologies for fully automated resource collection (to include farming), manufacturing, and transportation would effectively reduce the cost of almost anything to near zero, providing sufficient resources are available.

          The exception being that owners of land where resources are collected or grown would still seek, not remuneration, but some form of payment for allowing it. Likewise, owners of land used for factories would expect payment, as would owners of any private land used for transportation infrastructure (roads, airports, seaports, spaceports, etc.). Also, providers of liability insurance for these activities would require transportation. And, of course, owners of intellectual property (be it movies, books, music, or patents) would still expect payment.

          • Yes, ‘say’, we give value to education like real estate has (accepted stable) value, on a y2100 base?

            We might no even call it payment (for resources with “cost of *almost anything* to near zero”), but maybe valuation (of qualities), tribute or state&status, what includes being (a natural) role model/example for imitating/reproduction. We would provide a base ‘income’ for societies stability, knowing that this includes (low) percentages of inefficiencies also (me guessing, even lower share of ‘meaningless’ labor as of today’s parts of social, financial activities; lower pressure for supporting ‘financial interest’ on its own purpose) with a strong difficulty in ‘how *globally* equalize a measure for labour or invested hours into activity’, easier for goods and energy?
            Technological (or a knowledge) singularity is like a win on the lottery within cultural evolution of societies (or global community) _ where, then, to further invest this?

            Why we are far off improved (higher) education foundation on wider base (ww(w))?
            What new legalisations of law (international law(s) to be reviewed? with more advanced social knowledge, for balancing (www) cultural progress instead of cultivating laws/rules steadiness (instead of stability for societies) on its own purpose?)

            Technological singularity is a great chance, depending on personalities forming its purpose for development?

          • Gigafactories “owners of any private land”, should grow food on its roofs, covered with solar panels (greenhouse) for reduced water evaporation (and on its way for (closed loop) habitats, independent from atmosphere) (?)

          • would see China and/or Tesla supporting to build Gigafactories (now) for Pakistan, best on flooded areas on buildings on stilts, additionally adding (free) water purification and stable platforms surrounding ‘factories’ (?), not may others doing something like that (forming consciousness and generic support/supply includes United Nations also) (?)

Comments are closed.