Although most estimates focus on calculations per second, the relevant parameters are:

• Calculations per second
• Memory size
• Memory bandwidth per node
• Inter-node communication bandwidth

Here I assume the WBE detail level to be somewhere between level 4 (spiking neural model) and level 5 (electrophysiological model) from the Whole Brain Emulation Roadmap.  The computational capacity required would be around 100 Exa-FLOPS (1e20) and the memory capacity 10 Peta-bytes (1e13).  Plugging these into the CPU and memory timeline calculators, gives an expected arrival time of 2026 and 2020 respectively for $1M.

Based on my models, it appears that memory bandwidth will be a significant bottleneck, while inter-CPU node bandwidth will not be.

I’ve created a spreadsheet where the two bandwidths are estimated.

The architecture I envision is 1 million compute nodes arranged in a 2-D cluster, 1000 on the side.  Each node will have 100 TFLOPS CPU capacity and 10 GB of memory.  Nodes are connected to the nearest 4 neighbors using a high speed bus.  Longer distances are covered by multiple hops.

The inter-node bandwidth is dependent on the distance that voltage information has to travel, which is related to the axon length.  Although some axons are very long, axon lengths probably follow a power law distribution, and long ones are rare.  Based on a 1KHz update frequency and  ~60 nodes within average axon distance, the required communication bandwidth is 250 Gb/s, which is close to the capabilities of current technology.

The local node memory bandwidth is dependent on the amount of synapse state memory and required refresh rate.  Based on a 1KHz update rate, this works out to 14 TB/s.  This figure is about 3 orders of magnitude higher than current technology.

The high memory bandwidth required makes it is likely that some kind of CPU/memory on-die integration will be required.  The memristor seems a good candidate.

It is interesting to think of the brain as a 2.x dimension structure, due to the non-zero depth and also due to the folds.  The question has been asked if this provides a barrier to implementation.  The depth aspect is not an issue, since dividing the brain on a 2-D grid will put most vertical columns on the same node.  Folds effectively reduce the length that an axon has to travel.  It seems that in the worst case two points that “should” be a meter apart (brain diameter when laid flat) are only 10 cm apart (brain actual diameter).   This may increase the inter-node bandwidth by a factor of 100, if most axons take advantage of the added dimensionality for routing.  Since the inter-node bandwidth is relatively modest, it is doubtful that even a 100 fold increase will make it the bottleneck instead of the memory bandwidth.

In sum, here are the performance requirements:

• 1 million compute nodes in a 2-D topology
• 100 TFLOPS per node
• 10GB memory per node
• 10TB/s memory bandwidth
• 4 links to neighboring nodes at 250 Gb/s

It seems likely that this kind of machine will be available by 2030 for $1M, and possibly as early as 2025.  It would be useful to evaluate the increase in memory bandwidth over time and extrapolate to this time frame.

9:40 – Ramez Naam: The Digital Biome

Plenty of carrying capacity for the biome – 30-300 billion people with advanced biotech.  We are using only 1/1000 of the incident energy from the sun.  There’s no reason to crash due to lack of resources with advanced tech.  Population is predicted to level off at 10 billion.

Good points, but the Singularity is likely to happen on a shorter time scale.

10:40 – Lance Becker: Modifying the  Boundary between Life and Death

Cells that experience ischemia (lack of oxygen) die after re-perfusion with oxygen, rather than during the ischemia itself.  They die due to oxidative damage and apoptosis (cell suicide) related to mitochondria. How do we prevent the re-perfusion injury?

Current approaches – CPR, AED, Cooling.  Next step cardiovascular bypass, mitochondria meds, better cooling, controlled reperfusion.

Tried cocktail of meds and cooling on 6 people that failed CPR, etc., and 50% were resuscitated.

Might be applications to Cryonics.

11:15 – Ellen Heber-Katz: The MRL  mouse – how it regenerates and how we might do the same

Wound healing – scar formation (mammals) vs. epimorphic regeneration.

In regeneration – blastema – cell de-differentiation -> new limb. In mammals – regeneration in deer antlers.  Rabbit ears.

MRL mouse – they regenerate various tissues.  Inflammation response is necessary for regeneration.  The mice don’t use mitochondria (?!).  (during regeneration?)  DCA blocks glycolysis – blocks regeneration.

MRL mouse cells are lacking p21 molecule.  p21 protects against DNA damage, but blocks regeneration.  Selective blocking of p21 is currently investigated.  Removing p53 increases regeneration.

What about immortality?  MRL mouse variant lives to 3 years, die of autoimmune response.  p21 doesn’t get cancer until 16 months of age.  But they get an inflammatory autoimmune disease – die of tumors.

Separate tumors, inflammation and regeneration.

11:50 – Anita Goel: Information Processing & Physical  Intelligence in Nanomachines that Read/Write DNA

She is talking about controlling ribosomes and such from the macroscale.

Applications – includes “gene-radar” – portable diagnostics for DNA/RNA – point of care.  Also, energy transduction at the nanoscale, read/write DNA, store information, computing.

DNA Polymerase can go forward, backward (error correction), or pause.  This can be controlled from macro-scale.  Quality of output can thus be controlled.

These machines are very energy efficient.  How do they do that?

More stress on infrastructure-free diagnosis, developing world, etc. .

14:00 – Shane Legg: Universal  measures of intelligence

Yet another exponential computation graph.

2025 – 10^20 ops per second.  What does this translate to in meaningful intelligence?  He wants intelligence on the Y axis.

Human vs. Ideal.  Internal properties (“understanding”, “soul”, etc.) vs. External Properties. He is going to focus on External Ideal quadrant.

Definition of intelligence: “Intelligent system are expected to work, and work well, in many different environments.” – Gudwin

He shows an equation that takes into account success, agent, environment and Occam’s razor.

This measure is called Algorithmic Intelligence Quotient (AIQ).  It seems to order some relevant learning agents in a correct way.

I think it would be great if we had a general intelligence measure, so we can gauge our success in creating artificial systems.

14:40 – John Tooby: Can discovering the design principles governing natural intelligence unleash breakthroughs in artificial intelligence?

For scientific debates, history shows us that the truth typically lies incredibly far beyond the limits of what even the most radical scientists of an era propose (e.g. Galileo and the size of the universe).

All the pieces are in place for realizing the Enlightenment project of a rigorous natural science of human nature: reverse engineering the code.

Emphasize evolution – natural selection is the only known physical process that can push against entropy.  It only builds structures that solve evolutionarily recurrent adaptive problems.

Evolutionary psychology.  Rant against marginalization of evolution in psychology academia.

Example – avoid mating with genetically similar organisms – deleterious recessives.  But, care for genetic relatives.  An identification problem.  Computation of kinship index and adjustment of behavior based on it.

Implications for strong AI

- psychology map of humans so that it does what we mean

clues on how to achieve strong AI:

- dedicated intelligence
- pinpoint strategies
- “brilliant hacks”
- a mosaic of modules

Universe is too diverse to hold in our mind.  Minds need to reduce.

15:20 – Tooby, Goertzel, Yudkowsky & Legg panel: Narrow and General Intelligence

…
Goertzel – AIXI is “fictitious”

Legg – AIXI is a useful model, and MC AIXI is useful in practice for limited problem sizes.
…

Goertzel – if evolution can harness large amounts of computation with a simple engine, then maybe we can achieve general intelligence with a small number of more general modules.

Legg – need domain specific algorithms to be efficient.  But a general system can solve with unlimited power.

More discussion about how much domain specific algorithms you need.  Possibly start with a limited set and develop more over time.

Legg – classes of algorithms that are not exactly domain specific, but applicable to multiple domain areas.

9:50 – Gregory Stock is talking.  He is skeptical about progress in the bio realm.  He says that the FDA is a damper on progress, but he also says that there are difficult problems.  He brings up Alzheimer’s as an example.  I think he is underestimating the power of info tech to change the way we do bio-science.  Having read/write access to DNA, plus “in-silico” simulations will change the game.

Now he is talking about Silicon and saying that the complexity of computers rivals that of life.  And now he is talking about the rapid exponential progress in DNA technology.  As far as I understand, he is worried that we will create new life forms that will supersede humans.  He is saying that human evolution is “not exponential”. I think he means that it’s a very slow exponential compared to tech.

He is talking about a “planetary super-organism” – merging of the biosphere with technological artifacts (cities, Internet).  He says it’s beyond a metaphor.  However, it’s not a replicator and does not have a boundary, so I am doubtful that it’s useful to think about what’s happening on earth in this way.

He is saying that the biosphere will be subsumed but still exist in the new merged organism.  However, it seems to me that biology is not competitive because of the orders of magnitude in performance difference.  He thinks that timescales will enlarge, but the characteristic frequencies of mechanical systems are much higher, so this doesn’t make sense to me.

He says we are in the midst of the Singularity.

Cyberspace – boundaries are weak (but what about encryption?), copies are easy (yes!).  Uploading will lead to the disappearance/detachment of humans that take this route.  Ways to stop evolution to a post-human future – singleton rule, super-organism communities.  But it seems he thinks that stopping is not possible.  People can’t affect because they are cells, and can’t affect the organism.

He says that people that follow regulation will not be the ones to make progress.  The ones that do make progress are the nimble ones.  Agreed.

11:00 – Ray Kurzweil – The Mind and how to build one

Singularity worry – is it feasible to emulate the brain.

We derived narrow AI capabilities from reverse engineering the auditor and visual systems.  5th computing paradigm – Moore’s law – planar silicon.  6th paradigm – 3d molecular computing.

He is complaining that critics set up straw-man arguments.  More about exponential progress…

Why some people accept, and some people have emotional issues with accepting exponential change and it’s implications?  Not because of sophistication, intelligence, etc. .

Quick recap of exponential progress.

Exponential progress in brain scanning.  Resolution doubling every year.  Showing a image of a simulation of a cortical column.  Mentions blue brain project – Markram.  Markram expects to reverse engineer the brain by 2018, Kurzweil thinks it will take until the 2020′s.  Design of the brain is in the genome, which puts an upper bound on the code complexity strictly required – 25MB.  ~10K code for cerebellum.

How to create a mind?  Cortical columns are the basic modules.  He equates these to LISP atoms.  I think that’s too simplistic, because concepts are fluidly created.  Nevertheless, I agree that these are the building blocks of the neocortex and it would be a good research direction for AI/AGI.  It seems he is talking about de-novo mind design rather than uploading.

He says consciousness is a matter of faith and is not scientifically falsifiable.  It is necessary to believe in it for morality.  I’m not so sure – I agree with Dennet that it’s to do with having a self-model, which could be a scientific concept.

Wasn’t very satisfying as a roadmap to h+ AI.

Questions:

* Something about the collapse of the wave function related to consciousness.  But the MWI doesn’t really have any wave function collapse, so I think that direction isn’t useful.  Ray is saying the same thing.

* What level of detail for whole brain emulation?  Ray says the point of WBE is to understand the principle of operation of the brain (i.e. reverse engineering).  He says once we understand how cortical columns at a low level we can recode them at a higher level and have a much more computationally efficient simulation.

* Similar question – is molecular level simulation necessary?  Ray responds that we don’t need to scan the whole brain at that resolution.  Distinct types of neurons can be scanned with high resolution once.  It seems he is again not talking about uploading.

12:00 – Ben Goertzel – AI for increased human healthspan

Human body as a machine.

The material in this talk is available on the h+ magazine site.

14:00 – Steven Mann: Humanistic Intelligence Augmentation and Mediation

Steven has glasses on which include a camera (I call this a speccam).  The image from the camera is up on the left screen. He is concerned about surveillance. He is practicing sousveillance.  He is drawing a diagram on a piece of paper instead of showing a slide.  The diagram is visible through his speccam.

He is showing another speccam using his speccam.  It re-renders the visible field on a 45 degree piece of glass, basically interposing a processing layer on vision – a heads up display.

A milling machine can replicate itself.

He talks about some patents he got on high dynamic range image stitching.  Seems obvious – and kind of disappointing to see a software patent from him.

Cyborg-log.  Continuous.  He is talking about getting to a resolution below the perception granularity of a human.  “Un-digital”.

He brought another person on stage that is talking about user-interfaces that keep humans in the loop as we get “un-digital”.  Using high bandwidth interaction devices – tactile, visual.  Musical instruments.  They are demoing one on stage.

More about musical instruments and patents.  This doesn’t seem very relevant.  That’s about it for this talk.

14:50 – Mandayam Srinivasan: Enhancing our bodies and evolving our brains

Skipped this one.

15:25 – Brian Litt: The past, present and future of brain machine interfaces

Long review of existing tech – small number of electrodes, poor biocompatibility.

His new work – implant 2.5 micron thick flexible surfaces with 720 electrodes for capturing neuronal signals.  Biocompatible, partly bioabsorbable (?).

Someone else’s work in rats.  Memories played back during sleep at 6x original waking acquisition speed.

Speculates about self-assembling injectable electrodes.

16:15 – Demis Hassabis: Combining systems neuroscience and machine learning: a new approach to AGI

How to create meaning for symbols?  CYC is brittle and cumbersome.

Search space of possible AGI solutions: regime 1 – small and dense search space.  regime 2 – large sparse search space.  In regime #2 it makes sense to use the brain as reference.  He thinks regime #2 is the case.  Arguments: evolution only produced it once, failure of AI field to date.  I somewhat agree, except there might be a more dense area that we might be overlooking.

Abstract at one end of spectrum, whole brain emulation as other end.  He thinks the abstract end is ad-hoc and unprincipled and WBE is 50 year in the future.

He advocates the middle way “systems neuroscience”.  Extract algorithm from brain function.  E.g. vision algorithm from the way vision areas work in the brain.  Then you have a component for narrow AI or AGI.  Another example is space perception (“grid cells”).

Reinforcement learning is another example.  HNN is another.

This approach still seems somewhat ad-hoc to me.  It seems likely that there are aspects of the way the brain works that cannot be identified as a separate module.  These kind of aspects would be too intertwined with other modules and aspects to tease apart into a separate algorithm.

However, we might be able to achieve a different type of intellect with the modular reverse engineering approach that he advocates.

17:20 – Terry Sejnowski: Reverse-engineering brains is within reach

Should look at evolution when trying to understand the brain.  A lot of conservation of genes even from “primitive” life forms.

Scale of the brain: 1e15 synapses, 1e11 neurons, 1e16 bits/s.

Temporal difference learning = dynamic programming.  Dopamine system.

Use of TD for backgammon in 1994. Radio, radar, car and power grid applications in next 5 years.

Not very exciting.

17:50 – Dennis Bray: What Cells Can Do That Robots Can’t

The main thrust of this talk is that a lot of computation is done by cells.  It seems to me that this computation is not relevant to the functioning of the mind.  Most of it is keeping homeostasis in the cell and with the environment.  When a person loses a limb, they maintain their mind and identity, so the cells in their limb could not have been “mission critical” to their identity and intelligence.  Therefore the speaker’s argument seems irrelevant to the prospect of creating machine intelligence.

18:20 – Terry and Dennis are debating

Terry is saying that they will bite the bullet and do a “complete” molecular simulation.  They did a monte-carlo simulation of a synapse over 100ms.  Cool video of the simulation showing about 100 synapses – “MCELL”.

Dennis rebuts.  He compares the brain to a neural network.  He says that neurons are all different.  I don’t see how that is related to the problem at hand – obviously you want to simulate the diversity of neurons present in a real brain by scanning one or by generating using embryo development simulation.

Terry: simulated a network of 300 generic neurons, and they learned useful things.

They seem to be “agreeing” and taking questions from the audience.

Dennis – there is nothing impossible in principle.  He is now asking about details of Terry’s simulation.  Terry – different proteins and molecules important at different times.

Dennis – there are so many different things to simulate.  Terry – exponential progress.

I think they are mostly in agreement about only some molecular aspects being relevant and we need to simulate to find out.  The main difference between them is probably related to intuition about exponential progress and the tools and resources it will make available in a short amount of time.  Terry can see that.

It is useful to think about the simulation’s level of detail (LOD) in terms of certain thresholds, from most detailed to least detailed:

• Noise level: at this level (call it LODn) the lack of precision is on the order of the noise present in a biological brain.  It is not possible to distinguish the functioning of such a simulated brain from the functioning of a biological brain.
• Functional level: at this level (call it LODf) the lack of precision is greater, but the result is functionally similar.  There may be some behavior changes, but the overall capabilities (e.g. “intelligence”) are similar.
• Equivalence level: at this level (call it LODe) the precision is even lower, but this is compensated with  tweaks to the simulated physiology.  The result is equivalent in capabilities, although some characteristics may be very different.  For example, the retina can be replaced with a non-biological equivalent.

The computation power required for LODn > LODf > LODe.  There are likely order of magnitude differences between the levels.

If we consider a non-biological example – a digital computer, what does it take to simulate it?  It is obviously enough to simulate the logic function.  The Markham’s line of reasoning would seem to argue that we have to simulate the voltage gradients and charge movements in each transistor!

In the transistor case, LODn would involve simulating each transistor’s logic function.  LODf would involve an instruction set simulation (e.g. the QEMU emulator).  LODe would involve using the most convenient instruction set (e.g. x86) and recompiling any software.  Clearly, an LODe simulation is several orders of magnitude more efficient.

Markham fails to convince that his preferred level of simulation is required for LODn, never mind the other levels.

One way to find out what levels require is to actually run simulations and compare to physical neural matter.  The Blue Brain project aims to do that, although the results are not conclusive yet.  It would be good if more research was directed at comparing their simulation to a biological brain.  This would make the project more grounded.

Jurveston’s main thesis was that AI created by evolutionary algorithms would have a strong competitive advantage over attempts to use traditional design. Some of the advantages include lack of brittleness and speed of implementation. Once you go down the evolution path, you diverge from the design path, because reverse engineering an evolved system is either very difficult or impossible. One idea I had while listening to this is that you can evolve subsystems with well defined I/O and connect the subsystems into a designed overall architecture.

The claim of up to 100 million neurons seems pretty aggressive. Not sure why they would be so much ahead of the IBM effort using x86 CPUs, even if they have thousands of them.

CCortex will be open-source.

They claim multiple levels of simulation: 1. detailed equations 2. “estimation of action potentials” 3. a proprietary method.

He talked about three different schools of thought about the Singularity:

Vingean – where prediction becomes impossible
Accelerationist – exponential technological change
Greater than Human Inteligence – in a positive feedback loop

His thesis was that the three schools reinforce but also contradict each-other.

Another good point Eliezer makes is that advances in scientific knowledge and algorithms reduce the threshold for the Singularity.

Nick Bostrom’s Simulation Argument argues that at least one of the following must be true:

• the human species is very likely to go extinct before reaching a “posthuman” stage
• any posthuman civilization is extremely unlikely to run a significant number of simulations of their evolutionary history
• or we are almost certainly living in a computer simulation

However, I see other possibilities. Assumptions:

• The strong many-worlds theory is correct (i.e. all consistent mathematical systems exist as universes, a.k.a “everything exists”)
• The many-worlds immortality theory is correct (i.e. for every conscious state there is at least one smooth continuation of that state in the many-worlds)

Given these assumptions, it doesn’t matter if we are in a simulation because our conscious state exists in many simulations and many non-simulated worlds that look identical to us (but are different in imperceptible ways). Even if all the simulations stopped, there would still be a continuation of our conscious state in a non-simulated world consistent with our observations to date.

Further, it seems that there are more non-simulated worlds than simulated worlds. This is because there are many ways a mathematical model can exist so that it cannot be formulated in a finite way, and therefore not simulatable by an intelligent entity. It might even be that simulatable world are of measure zero in the many-worlds.

Further out ideas:

A fascinating related idea is the Egan Jump as described in the book Permutation City. The idea is to jump to another world in the many-worlds by simulating the genesis of a new universe. In this universe you code yourself into the initial conditions, and design the rules so that you end up as an upload in the substrate of the new universe. Because that universe will continue as it’s own mathematical model, your conscious state will continue in that universe, branching off your original self.

Yet another, more distantly related idea is that the peculiarities of our universe (quantum physics, large amounts of empty space) are in a sense an error correcting mechanism. Because any perturbation of a world is also a world, the result is quite chaotic and inhospitable to meaningful life. The structure we see around us with large aggregates “average out” the chaos. This leads to a stable environment as required for conscious observers to arise.

I still believe that the Friendly AI initiative will have a hard time achieving meaningful success.  I think it’s more likely that we will be able to create friendly Uploads first, or at least that we should strive towards that goal.