Gerald Monroe

Karma: 441

Gerald Monroe 31 Dec 2021 0:56 UTC
5 points
in reply to: chasmani’s comment on: We need a theory of anthropic measure binding
Biological meat doesn’t have the needed properties but this is how SpaceX and others avionics control works. Inputs are processed in discrete frames, all computers receive a frame of [last output | sensor_inputs], and implement a function where output = f(frame) : the output depends only on the frame input and all internal state is identical between the 3 computers.
Then, after processing, last_output=majority(output1, output2, output_n)
So even when one of the computers makes a transient fault, it can contribute to the next frame.
Current neural network based systems can be architected to work this way.
The computers are not identical but if your only information about them is their outputs they are indistinguishable from each other.

Gerald Monroe 31 Dec 2021 0:49 UTC
4 points
in reply to: lsusr’s comment on: The Machine that Broke My Heart
My other comment is that you probably didn’t succeed as well as you thought you did. I am taking your story at face value—that your model was spookily accurate, that it worked way better than you could reasonably expect, etc. But scale matters. Many tech prototypes work perfectly at small scales but would fail if you had built a few thousand or million hardware instances and tried them with the user wearing them across the full range of human activity and cultures.
But say you did, and your model that’s cheap enough to train on a laptop and run on an embedded CPU is perfectly accurate. How does this solve the actual problem? Obesity is caused by humans getting incredibly strong, insidious urges to eat (and maintain their weight once fat). The actual root cause is probably poisons in the food supply, explaining the delta between specific countries of genetically comparable people, but in any case, how does this gadget help?

In The Circle everyone wears cameras all the time and thus gets peer shamed into not eating, but this has issues other than the privacy ones, such as most people not being attractive enough/popular enough to have any peers to shame them.

I’ve on many occasions tried the myfitnesspal calorie counting. It does work but the longer you do it the stronger your urges to restore your missing mass. It’s like compressing a spring. Significant weight loss is very difficult. It’s not impossible, I’ve been −35 lbs for 4 years now, but I have not been able to get from being merely ‘overweight’ to ‘normal’ BMI.

Gerald Monroe 31 Dec 2021 0:41 UTC
2 points
on: The Genetics of Space Amazons
“Sci fi plot alert”: what happens if due to random chance/genetic drift the “A” version of the X chromosome becomes more common and the last male dies? This would be more probable the smaller the population is. And, I dunno, a space rock snipes the sperm storage freezer. (similar to what happened in Seveneves)

Gerald Monroe 30 Dec 2021 18:06 UTC
2 points
in reply to: gwern’s comment on: Universal counterargument against “badness of death” is wrong
Just to add to the above: even without (massive) cognitive decline in the aged, just knowing you only have a few years left likely has an effect on someone’s decisions. Most changes and improvements, in technology and institutional processes, cause initial short term problems. They only pay off long term. If you’re in the last 5 years of your career, or your life, there’s no expected payoff for learning most new things, or for seeing a major change in how the institution you work in works.
If you can reasonably expect to live for many more centuries (since with a perfect cure for aging you’d have a life expectancy of several thousand years), you might as well start adopting new things now, maybe you’ll see a net payoff in 50 years. Or maybe you’ll procrastinate first. Could go either way.

Gerald Monroe 30 Dec 2021 17:59 UTC
5 points
on: The Machine that Broke My Heart
Did you really throw away the software and not keep it in a VCS or on an extra storage device? I feel a sort of pain thinking about it—that if it really worked as well as described, it may not at all be easy to rebuild.

Gerald Monroe 27 Dec 2021 15:00 UTC
1 point
in reply to: Dagon’s comment on: Gerald Monroe’s Shortform
R1 <-> R4 are arbitrary positive floating point numbers. Units are currency units. So “human harm” is in terms of estimated actual costs + estimate reputation damage paid for injuries/wrongful deaths, “outside boundary” is an estimate of the fines for trespassing and settlements paid in lawsuits, “paperclips made” is the economic value of the paperclips, and operating cost is obvious.

Gerald Monroe 27 Dec 2021 14:59 UTC
1 point
in reply to: JBlack’s comment on: Gerald Monroe’s Shortform
This is, arguably, AGI. The reason it’s AGI is because you can solve most real world problems by licensing a collection of common subcomponents (I would predict some stuff will be open source but the need for data and cloud compute resources to build and maintain a component means nothing can be free), where you only need to define your problem.
In this specific toy example, the only thing that is written by human devs for their paperclip factory might be a small number of json files that reference paperclip specs, define the system topology, and references another system that actually handles marketing/selling paperclips and sets the quotas.
Obviously then the meta task of authoring this file could be done by another AI based system that learns from examples of humans authoring these files.
So the system as a whole has the capabilities of AGI even though it’s made of a large number of narrow AIs. This is arguably how our brains most likely work at the subsystem level so there’s prior art.
Also I can see how you might get sloppy and stop adhering to the ‘sparseness’ criterion. Make your paperclip factory of subsystems that are smarter than they need to be, increasing their flexibility in novel situations at the cost of potentially unbounded behavior.

Gerald Monroe 25 Dec 2021 21:57 UTC
1 point
in reply to: Ansel’s comment on: Gerald Monroe’s Shortform
Paperclip quality control is an agent that was trained on simulated sensor inputs (camera images and whatever else) of variations of paperclip. Paperclips that are not within a narrow range of dimensions and other measurements for correctness are rejected.

It doesn’t have any learning ability. It is literally an overgrown digital filter that takes in some dimensions of input image and outputs true or false to accept or reject. (and probably another vector specifying the checks failed)

We can describe every subagent for everything the factory needs as such limited, narrow domain machines that alignment issues are not possible. (Especially as most will have no memory and all will have learning disabled)

Gerald Monroe 25 Dec 2021 19:45 UTC
1 point
on: Gerald Monroe’s Shortform
The trivial solution to AI alignment.
Preface: I’m not expecting this solution to work, I just want to understand why the ‘ez’ solution doesn’t work.
A paperclip maximizer:
The paperclip maximizer is a paperclip factory designer/manager. It has the role to select, from a finite set of machines available on the market, machines to manufacture, inspect, and package paperclips.
It’s reward heuristic is H = ( (quota—paperclips_made) x R1 - <human harm> x R2 - <machines outside factory boundary> x R3 - <operating cost> x R4).
So the machine gets the most reward if it makes up to the quota of paperclips for the lowest cost, with the least machinery outside the factory bounds, and no humans hurt. (it has perception stack subsystems that try to detect humans inside the factory boundary, and estimates harm if any machine impinges on the space they occupy or are predicted to occupy in near future frames)
The other design aspect is functional sparseness. This means :
The machine’s neural network is the smallest network that maximizes reward from the possibilities tried during design and training
The machine exports common problems to other machines.
Common problems : paperclip quality control, paperclip production counts, operating cost calculation, machine perception, machine pathing for paperclip manufacture, and so on.
What this would mean is that this machine shares these subsystems with other AI systems that have totally different incentives, and these subsystems themselves were designed to accomplish roles without the paperclip factory planner having input into their design. Thus it can’t install a corrupt “production counter” to declare it’s met the quota without energy usage. Or redesign the paperclip plant materials porters to have onboard weapons. And it lacks the cognitive ability to consider most of these things because this doesn’t improve reward during the training phase.
Why won’t this work?

Gerald Monroe’s Shortform

Gerald Monroe25 Dec 2021 19:45 UTC

4 points

9 comments1 min readLW link

Gerald Monroe 19 Dec 2021 10:54 UTC
1 point
in reply to: Viliam’s comment on: Why Bedroom Closets?
Yes. I also sorta daydream about similar things. With real estate we’ve got this archaic model mired in all these artificial rules and barriers and intermediaries who are legally allowed to consume a chunk from every transaction.
A proper setup would be, well, once we get robotics to be more flexible and reliable (obviously using the latest breakthroughs in reinforcement learning/neural networks), we’d make everything out of cubical modules like you say.
So you assign the robots to clear a site and the first ones come and demolish anything in the way. Then the next set come and install a concrete foundation and the appropriate base supports. Then the next set install an open framework of girders with however many floors the building will be allowed to have. The framework is made from modules that bolt together in pre-manufactured sections, with obviously a utility/elevator core going up the middle that is prewired and pre plumbed.
Then the buildings functional modules are cubes of a standard size—probably using the ‘fold flat’ technology several startups are working on that makes them transportable in standard dimension trucks—and they come from a factory prefabricated and pre-furnished and designed for the purpose of their occupants. Obviously, the smart, flexible robotics in the factory can build a cube with an interior layout with tens of thousands of permutations, as the materials they need are just in time ordered from other feeder factories by robotic vehicle.
Once on site, most of these buildings have an outer set of rails and a winch system that lets the cube be autonomously raised and then moved laterally to it’s final position while the building is in use.
The cubes are inherently designed to be maintainable by other robotics—all the functional systems inside them are subdivided into modules that can be removed all in one piece, so that maintenance robots don’t have to be very smart, they just go pull the module that self reports a failure or it’s peer modules claim is not responding. (and they come back and replace peer modules and reporter modules if this doesn’t clear the fault)
And yeah obviously as the interior of a cube gets dated, or no longer to the occupant’s needs it gets removed, recycled in a factory, and a new one built. Since robots do 99.9% of the work, costs are cheap and recycling and building a new one (often with used parts from previously recycled modules) is cheap.

Gerald Monroe 1 Dec 2021 8:42 UTC
1 point
in reply to: Viliam’s comment on: Why Bedroom Closets?
Modular housing would let you do that, it hasn’t taken off in the USA in favor of finance scam built on site places that are extremely expensive to modify. So you need a design that meets most needs no matter what they are, which is part of the reason US houses are so big with extra rooms.

Another factor is that it’s expensive to switch houses. In theory you should be able to just trade up and down with minimal hassle. In practice you pay approximately 5% of the value of the property to various people as fees! Not to mention the time wasting viewing and bidding process, which would be much more efficient if houses were more, well, identical to each other so that prices can be easily discovered.

Gerald Monroe 29 Nov 2021 11:01 UTC
5 points
on: Why Bedroom Closets?
Dining rooms. Foyers with double height ceilings. Sun rooms. upstairs kitchens. Commercial grade kitchens in houses meant for 4-5 total occupants.
There’s a lot of ways to waste space in housing, and the other factor as you figured out is that there’s not a whole lot of engineering effort put in. A methodical way to design a house would be to sample the movement and activities of the occupants, over a decent sample size, over a period of years. Find out in the data where people go and what they do, how long they spend on a task, where do they get less task performance because a space is too cramped.
And then develop a model and iterate many designs and converge on good layouts and house designs. Which you then build in factories and install on-site as prefab modules.
But somehow people have gotten convinced that this would be too cheap, and this means their neighbors might be poor, and maybe those neighbors will do bad things to them. Or something. It’s complex.
For offices this has mostly been done. The buildings are poured on site for various economic reasons but a lot of the materials are prefab and obviously the cube farms themselves are prefab. The exact layout has been carefully iterated on.

Gerald Monroe 14 Nov 2021 17:36 UTC
0 points
in reply to: TLW’s comment on: Postmodern Warfare
So TLW, at the end of the day, all your objections are in the form of “this method isn’t perfect” or “this method will have issues that are fundamental theorems”.
And you’re right. I’m taking the perspective of, having built smaller scale versions of networked control systems, using a slightly lossy interface and an atomic state update mechanism, “we can make this work”.
I guess that’s the delta here. Everything you say as an objection is correct. It’s just not sufficient.
At the end of the day, we’re talking about a collection of vehicles. Each is somewhere between the size of a main battle tank and a human sized machine that can open doors and climb stairs. All likely use fuel cells for power. All have racks of compute boards, likely arm based SOCs, likely using TPUs or on-die coprocessors. Hosted on these boards is a software stack. It is very complex but at a simple level it does :
perception → state state representation → potential action set → H(potential action set) → max(H) → actuators.
That H, how it evaluates a potential action, takes into account (estimates of loss, isActionAllowed, gain_estimate(mission_parameters), gain_estimate(orders)).
It will not take an action if not allowed. (example, if weapons disabled it will not plan to use them). It will avoid actions with predicted loss unless the gain is high enough. (example it won’t normally jump out a window but if $HIGH_VALUE_TARGET is escaping around the corner, the machine should and will jump out a window, firing in midair before it is mission killed on impact, when the heuristic is tuned right)
So each machine is fighting on it’s own, able to kill enemy fighters on it’s own, assassinate VIPs, avoid firing on civilians, unless the reward is high enough [it will fire through civilians if the predicted gain is set high enough. These machines are of course amoral and human operators setting “accomplish at all costs” for a goal’s priority will cause many casualties].
The coordination layer is small in data, except for maybe map updates. Basically the “commanders” are nodes that run in every machine, they all share software components where the actual functional block is ‘stateless’ as mentioned. Just because there is a database with cached state and you send (delta, hash) each frame in no way invalidates this design. What stateless means is that the “commander” gets (data from last frame, new information) and will make a decision based only on the arguments. At an OS level this is just a binary running in it’s own process space that after each frame, it’s own memory is in the same state it started in. [it wrote the outputs to shared memory, having read the inputs from read only memory]
This is necessary if you want to have multiple computer redundancy, or software you can even debug. FYI I actually do this, this part’s present day.
Anyways in situations where the “commander” doesn’t work for any of the reasons you mention...this doesn’t change a whole lot. Each machine is now just fighting on it’s own or in a smaller group for a while. They still have their last orders.
If comm losses are common and you have a much larger network, the form you issue orders in—that limits the autonomy of ever smaller subunits—might be a little interesting.
I think I have updated a little bit. From thinking about this problem, I do agree that you need the software stacks to be highly robust to network link losses, breaking into smaller units, momentary rejoins not sufficient to send map updates, and so on. This would be a lot of effort and would take years of architecture iteration and testing. There are some amusing bugs you might get, such as one small subunit having seen an enemy fighter sneak by in the past, then when the units resync with each other, fail to report this because the sync algorithm flushes anything not relevant to the immediate present state and objectives.

Gerald Monroe 8 Nov 2021 0:44 UTC
1 point
in reply to: Stuart_Armstrong’s comment on: Is Molecular Nanotechnology “Scientific”?
(Unfortunately, I’d put AI down with flight in 1200 - intelligence exists, but we don’t understand it to any real extent, and current technologies are not approaching proper intelligence; they need new conceptual ideas)

What’s your measuring stick here? “Artificial general intelligence” doesn’t require the intelligent system to be able to have emotions, or even organism level goals, arguably. Arguably, a software stack where you can define what a robotic work system must accomplish in some heuristic language, and then autonomously generate the neural network architecture, models, and choose the robotic hardware platform to use meets the criterion of being “AGI”.
So a small number of AI engineers log in to some cloud hosted system, define some new task (‘cooking: grilled cheese sandwiches’), do a person-month or so of labor, and now worldwide anyone can get their grilled cheese cooked by a robot that is a little bit better on average than a human. (assuming they pay the license fees, which are priced to be cheaper than having a human do the task for that local labor market)
This to me seems near term, do you disagree with it’s feasibility?
Would this be “not AGI” even though you can automate entire classes of real world tasks?
(the limit is that the task needs to be modelable for your success heuristic, and for the inputs the robotic system will see. So you can automate almost any kind of physical manipulation task. While “teaching: second grade math” can’t be automated because you can’t model the facial expressions a small child will generate accurately or a set of queries a small child might ask over the entire state space of questions a child will likely ask, or model whether or not a given policy results in the child learning the math. At least, not with current simulators, obviously there is significant progress being made)

Gerald Monroe 8 Nov 2021 0:39 UTC
1 point
in reply to: NE1’s comment on: Is Molecular Nanotechnology “Scientific”?
What about “microtechnology”? These would be self assembly machines made of bigger parts, similar to the actuators in a DLP chip. Or “hybrid microtechnology”, where some other process using a catalyst has made subunits large enough to be manipulated by these “micro-scale” robotics.
Wouldn’t “hybrid microtechnology” have precisely the same real world consequences (self replication) that the nanoscale machinery you object to?

Gerald Monroe 7 Nov 2021 5:26 UTC
1 point
in reply to: TLW’s comment on: Postmodern Warfare
It might be interesting to discuss this in a more interactive format, such as on https://discord.gg/GVkQF2Wn . You do know some stuff, I know some stuff, and we seem to be talking past each other. Fundamentally I think these problems are solvable.
(1) merger of conflicting world spaces is possible. or if this turns out to be too complex to implement, you deterministically pick one network to be the primary one, and have it load from the subordinate network the current observations.
(2) If commanders need more memory than the communications channel, they must exchange deltas. These deltas are the (common observations, made as input from the subordinate platforms, and state metadata). This is how a complex simulation like age of empires worked on a modem link.https://www.gamedeveloper.com/programming/1500-archers-on-a-28-8-network-programming-in-age-of-empires-and-beyond
(3) Free air laser links is one technology that would at least somewhat obscure the source of the signaling (laser light will probably reflect in a detectable way around corners but it won’t go through solid objects) and is capable of tends of gigabits per second of bandwidth, enough to satisfy some of your concerns.

Gerald Monroe 2 Nov 2021 9:01 UTC
1 point
in reply to: Sune’s comment on: Secure homes for digital people
Only the writes are blocked.

Gerald Monroe 2 Nov 2021 5:28 UTC
1 point
on: Secure homes for digital people
I mean a simpler version would just be hardware that has limits baked into the substrate so certain operations are not possible.

The most trivial is you let core personality has copies in secure enclaves. These are mind emulator chips that is loaded with the target personality and then a fuse is blown. This fuse prevents any outside source from writing to the memory of the target personality.

It can still grow and learn but only by internal update rules to internal memory.

Gerald Monroe 31 Oct 2021 5:45 UTC
0 points
in reply to: TLW’s comment on: Postmodern Warfare
So note I do work on embedded systems IRL, and have implemented many, many variations of messaging pipeline. It is true I have not implemented one this complex, but I don’t see any showstoppers.
1. This is how SpaceX does it right now. In summary, it’s fine to have some of the “commanders” miss entire frames as “commanders” are stateless. Their algorithm is f([observations_this_frame|consensus_calculated_values_last_frame]). Resynchronizing when entire subnets get cut off for multiple frames and then reconnected is tricky, but straightforward. (a variation of the iterative algorithms you use for sensor fusion can fuse 2 belief spaces, aka 2 maps where each subnet has a different consensus view of a shared area of the state space)
2. It does, there is not a way to broadcast information that doesn’t reveal your position.
3. Please define TPM. Message payloads are fixed length and encrypted with a shared key. I don’t really see an issue with the enemy gaining some information because ultimately they need to have more vehicles armed with guns or they are going to lose, information does not provide much advantage.
4. Thermite, battery backed keystores. And the vehicles don’t have the actual source for the algorithms used to develop it, just binaries and neural network files. Assuming that the enemy can bypass the self destruct and exploit a weakness in the chip to access the key in the battery backed keystore, they just have binaries. This doesn’t give them the ability to use the technology themselves*. Moreover, the agents are using near optimal policy. A near optimal policy agent has nothing to exploit—they are not going to make any significant mistakes in battle you can learn about.
5. Nothing like this. The “commander” agent’s guessing from prior experience optimal configurations to put it’s troops. The “subordinate” agent it is querying runs on the same hardware node. So these requests are obviously IPC using shared memory. And the commander makes a finite number of “informed” guesses, gets from the subordinate which “plans” are impossible, and selects the best plan from the remaining (with possibly some optimizing searches in nearby state space to the current best plans). This will select a plan chosen from the set of { winning battle configurations in the current situation | possible according to subordinate } that is the best of a finite number of local maxima.
I am not sure your “glass jaw” point. OpenAI is a research startup with a prototype agent. It can’t be expected to be flawless just because it uses AI techniques. Nor do I expect these military drones to not have entire real life battles where they lose to a software bug. The difference is that the bug can be patched, records of the battle can be reviewed and learned from, and the next set of drones can learn from the fallen as directly as if any experience that was uploaded happened directly to them.
At the end there I am assuming you end up with varying sizes of land vehicle because they can carry hundreds of kilograms of armor/weapons. Flying drones do not have even in the same order of magnitude the payload capacity. So you end up with what are basically battles of attrition between land vehicles of various scales, where flying drones are instantly shot out of the sky and are used for information gathering. (a legged robot that can climb open doors and climb stairs I am classifying as a land vehicle).
Maybe it would go differently and end up being a war between artillery units at maximum range with swarms of flying drones used as spotters.
*I think this is obvious, but for every piece of binary or neural network deployed in the actual machine in the field, there is a vast set of simulators and testing tools and visualization interfaces that were needed to meaningfully work on such technology. This ‘backend’ is 99% of what you need to build these systems. If you don’t already have your own backend you can’t develop and deploy your own drones.

Gerald Monroe

Ger­ald Mon­roe’s Shortform

Gerald Monroe’s Shortform