I would like to ask whether it is not more engaging if to say, the caring drive would need to be specifically towards humans, such that there is no surrogate?

Definitely need some targeting criteria that points towards humans or in their vague general direction. Clippy does in some sense care about paperclips so targeting criteria that favors humans over paperclips is important.

The duck example is about (lack of) intelligence. Ducks will place themselves in harms way and confront big scary humans they think are a threat to their ducklings. They definitely care. They're just too stupid to prevent "fall into a sewer and die" type problems. Nature is full of things that care about their offspring. Human "caring for offspring" behavior is similarly strong but involves a lot more intelligence like everything else we do.

TLDR:If you want to do some RL/evolutionary open ended thing that finds novel strategies. It will get goodharted horribly and the novel strategies that succeed without gaming the goal may include things no human would want their caregiver AI to do.

Orthogonally to your "capability", you need to have a "goal" for it.

Game playing RL architechtures like AlphaStart and OpenAI-Five have dead simple reward functions (win the game) and all the complexity is in the reinforcement learning tricks to allow efficient learning and credit assignment at higher layers.

So child rearing motivation is plausibly rooted in cuteness preference along with re-use of empathy. Empathy plausibly has a sliding scale of caring per person which increases for friendships (reciprocal cooperation relationships) and relatives including children obviously. Similar decreases for enemy combatants in wars up to the point they no longer qualify for empathy.

I want agents that take effective actions to care about their "babies", which might not even look like caring at the first glance.

ASI will just flat out break your testing environment. Novel strategies discovered by dumb agents doing lots of exploration will be enough. Alternatively the test is "survive in competitive deathmatch mode" in which case you're aiming for brutally efficient self replicators.

The hope with a non-RL strategy or one of the many sort of RL strategies used for fine tuning is that you can find the generalised core of what you want within the already trained model and the surrounding intelligence means the core generalises well. Q&A fine tuning a LLM in english generalises to other languages.

Also, some systems are architechted in such a way that the caring is part of a value estimator and the search process can be made better up till it starts goodharting the value estimator and/or world model.

Yes they can, until they will actually make a baby, and after that, it's usually really hard to sell loving mother "deals" that will involve suffering of her child as the price, or abandon the child for the more "cute" toy, or persuade it to hotwire herself to not care about her child (if she is smart enough to realize the consequences).

Yes, once the caregiver has imprinted that's sticky. Note that care drive surrogates like pets can be just as sticky to their human caregivers. Pet organ transplants are a thing and people will spend nearly arbitrary amounts of money caring for their animals.

But our current pets aren't super-stimuli. Pets will poop on the floor, scratch up furniture and don't fulfill certain other human wants. You can't teach a dog to fish the way you can a child.

When this changes, real kids will be disappointing. Parents can have favorite children and those favorite children won't be the human ones.

Superstimuli aren't about changing your reward function but rather discovering a better way to fulfill your existing reward function. For all that ice cream is cheating from a nutrition standpoint it still tastes good and people eat it, no brain surgery required.

Also consider that humans optimise their pets (neutering/spaying) and children in ways that the pets and children do not want. I expect some of the novel strategies your AI discovers will be things we do not want.

TLDR:LLMs can simulate agents and so, in some sense, contain those goal driven agents.

An LLM learns to simulate agents because this improves prediction scores. An agent is invoked by supplying a context that indicates text would be written by an agent (EG:specify text is written by some historical figure)

Contrast with pure scaffolding type agent conversions using a Q&A finetuned model. For these, you supply questions (Generate a plan to accomplish X) and then execute the resulting steps. This implicitly uses the Q&A fine tuned "agent" that can have values which conflict with ("I'm sorry I can't do that") or augment the given goal. Here's an AutoGPT taking initiative to try and report people it found doing questionable stuff rather than just doing the original task of finding their posts.(LW source).

The base model can also be used to simulate a goal driven agent directly by supplying appropriate context so the LLM fills in its best guess for what that agent would say (or rather what internet text with that context would have that agent say). The outputs of this process can of course be fed to external systems to execute actions as with the usual scafolded agents. The values of such agents are not uniform. You can ask for simulated Hitler who will have different values than simulated Gandhi.

Not sure if that's exactly what Zvi meant.

But it seems to be much more complicated set of behaviors. You need to: correctly identify your baby, track its position, protect it from outside dangers, protect it from itself, by predicting the actions of the baby in advance to stop it from certain injury, trying to understand its needs to correctly fulfill them, since you don’t have direct access to its internal thoughts etc.

Compared to “wanting to sleep if active too long” or “wanting to eat when blood sugar level is low” I would confidently say that it’s a much more complex “wanting drive”.

Strong disagree that infant care is particularly special.

All human behavior can and usually does involve use of general intelligence or gen-int derived cached strategies. Humans apply their general intelligence to gathering and cooking food, finding or making shelters to sleep in and caring for infants. Our better other-human/animal modelling ability allows us to do better at infant wrangling than something stupider like a duck. Ducks lose ducklings to poor path planning all the time. Mama duck doesn't fall through the sewer grate but her ducklings do ... oops.

Any such drive will be always "aimed" by the global loss function, something like: our parents only care about us in a way for us to make even more babies and to increase our genetic fitness.

We're not evolution and can aim directly for the behaviors we want. Group selection on bugs for lower population size results in baby eaters. If you want bugs that have fewer kids that's easy to do as long as you select for that instead of a lossy proxy measure like population size.

Simulating an evolutionary environment filled with AI agents and hoping for caring-for-offspring strategies to win could work but it's easier just to train the AI to show caring-like behaviors. This avoids the "evolution didn't give me what I wanted" problem entirely.

There's still a problem though.

It continues to work reliably even with our current technologies

Goal misgeneralisation is the problem that's left. Humans can meet caring-for-small-creature desires using pets rather than actual babies. It's cheaper and the pets remain in the infant-like state longer (see:criticism of pets as "fur babies"). Better technology allows for creating better caring-for-small creature surrogates. Selective breeding of dogs and cats is one small step humanity has taken in that direction.

Outside of "alignment by default" scenarios where capabilities improvements preserve the true intended spirit of a trained in drive, we've created a paperclip maximizer that kills us and replaces us with something outside the training distribution that fulfills its "care drive" utility function more efficiently.

Many of the points you make are technically correct but aren't binding constraints. As an example, diffusion is slow over small distances but biology tends to work on µm scales where it is more than fast enough and gives quite high power densities. Tiny fractal-like microstructure is nature's secret weapon.

The points about delay (synapse delay and conduction velocity) are valid though phrasing everything in terms of diffusion speed is not ideal. In the long run, 3d silicon+ devices should beat the brain on processing latency and possibly on energy efficiency

Still, pointing at diffusion as the underlying problem seems a little odd.

You're ignoring things like:

• ability to separate training and running of a model
  • spending much more on training to improve model efficiency is worthwhile since training costs are shared across all running instances
• ability to train in parallel using a lot of compute
  • current models are fully trained in <0.5 years
• ability to keep going past current human tradeoffs and do rapid iteration
  • Human brain development operates on evolutionary time scales
  • increasing human brain size by 10x won't happen anytime soon but can be done for AI models.

People like Hinton Typically point to those as advantages and that's mostly down to the nature of digital models as copy-able data, not anything related to diffusion.

Energy processing

Lungs are support equipment. Their size isn't that interesting. Normal computers, once you get off chip, have large structures for heat dissipation. Data centers can spend quite a lot of energy/equipment-mass getting rid of heat.

Highest biological power to weight ratio is bird muscle which produces around 1 w/cm³ (mechanical power). Mitochondria in this tissue produces more than 3w/cm³ of chemical ATP power. Brain power density is a lot lower. A typical human brain is 80 watts/1200cm³ = 0.067W/cm³.

synapse delay

This is a legitimate concern. Biology had to make some tradeoffs here. There are a lot of places where direct mechanical connections would be great but biology uses diffusing chemicals.

Electrical synapses exist and have negligible delay. though they are much less flexible (can't do inhibitory connections && signals can pass both ways through connection)

conduction velocity

Slow diffiusion speed of charge carriers is a valid point and is related to the 10^8 factor difference in electrical conductivity between neuron saltwater and copper. Conduction speed is an electrical problem. There's a 300x difference in conduction speed between myelinated(300m/s) and un-myelinated neurons(1m/s).

compensating disadvantages to current digital logic

The brain runs at 100-1000 Hz vs 1GHz for computers (10^6 - 10^7 x slower). It would seem at first glance that digital logic is much better.

The brain has the advantage of being 3D compared to 2D chips which means less need to move data long distances. Modern deep learning systems need to move all their synapse-weight-like data from memory into the chip during each inference cycle. You can do better by running a model across a lot of chips but this is expensive and may be inneficient.

In the long run, silicon (or something else) will beat brains in speed and perhaps a little in energy efficiency. If this fellow is right about lower loss interconnects then you get another + 3OOM in energy efficiency.

But again, that's not what's making current models work. It's their nature as copy-able digital data that matters much more.

Yeah, my bad. Missed the:

If you think this is a problem for Linda's utility function, it's a problem for Logan's too.

IMO neither is making a mistake

With respect to betting Kelly:

According to my usage of the term, one bets Kelly when one wants to "rank-optimize" one's wealth, i.e. to become richer with probability 1 than anyone who doesn't bet Kelly, over a long enough time period.

It's impossible to (starting with a finite number of indivisible currency units) have zero chance of ruin or loss relative to just not playing.

• most cautious betting strategy bets a penny during each round and has slowest growth
• most cautious possible strategy is not to bet at all

Betting at all risks losing the bet. if the odds are 60:40 with equal payout to the stake and we start with N pennies there's a 0.4^N chance of losing N bets in a row. Total risk of ruin is obviously greater than this accounting for probability of hitting 0 pennies during the biased random walk. The only move that guarantees no loss is not to play at all.

Goal misgeneralisation could lead to a generalised preference for switches to be in the "OFF" position.

The AI could for example want to prevent future activations of modified successor systems. The intelligent self-turning-off "useless box" doesn't just flip the switch, it destroys itself, and destroys anything that could re-create itself.

Until we solve goal misgeneralisation and alignment in general, I think any ASI will be unsafe.

If we wanted to kill the ants or almost any other organism in nature we mostly have good enough biotech. For anything biotech can't kill, manipulate the environment to kill them all.

Why haven't we? Humans are not sufficiently unified+motivated+advanced to do all these things to ants or other bio life. Some of them are even useful to us. If we sterilized the planet we wouldn't have trees to cut down for wood.

Ants specifically are easy.

Gene drives allow for targeted elimination of a species. Carpet bomb their gene pool with replicating selfish genes. That's if an engineered pathogen isn't enough. Biotech will only get better.

What about bacteria living deep underground? We haven't exterminated all the bacteria in hard to reach places so humans are safe. That's a tenuous but logical extension to your argument.

If biotech is not enough, shape the environment so they can't survive in it. Trees don't do well in a desert. If we spent the next hundred years adapting current industry to space and building enormous mirrors we can barbecue the planet. It would take time, but that would be the end of all earth based biological life.