Please Understand 2024-04-01T12:33:20.459Z
AIOS 2023-12-31T13:23:56.552Z
Picasso in the Gallery of Babel 2023-12-26T16:25:14.607Z
A Kindness, or The Inevitable Consequence of Perfect Inference (a short story) 2023-12-12T23:03:03.690Z


Comment by samhealy on Please Understand · 2024-04-03T14:07:59.796Z · LW · GW

(Your response and arguments are good, so take the below in a friendly and non-dogmatic spirit)

Good enough for what?

Good enough for time-pressed people (and lazy and corrupt people, but they're in a different category) to have a black-box system do things for them that they might, in the absence of the black-box system, have invested effort to do themselves, and as an intended or unintended result, increased their understanding, opening up new avenues of doing and understanding. 

We're still in the "wow, an AI made this" stage.

I'm pretty sure we're currently exiting the "wow, an AI made this" stage, in the sense that 'fakes' in some domains are approaching undetectability.

We find that people don't value AI art, and I don't think that's because of its unscarcity or whatever, I think it's because it isn't saying anything

I strongly agree, but it's a slightly different point: that art is arguably not art if it was not made by an experiencing artist, whether flesh or artificial. 

My worry about the evaporation of (human) understanding covers science (and every other iterably-improvable, abstractable domain) as well as art, and the impoverishment of creative search space that might result from abstraction-offloading will not be strongly affected by cultural attitudes toward proximately AI-created art — even less so when there's no other kind left.

the machine needs to understand the needs of the world and the audience, and as soon as machines have that...

It's probably not what you were implying, but I'd complete your sentence like so: "as soon as machines have that, we will have missed the slim chance we have now to protect human understanding from being fully or mostly offloaded."

but where that tech is somehow not reliable and independent enough to be applied to ending the world...

I'm afraid I still can't parse this, sorry! Are you referring to AI Doom, or the point at which A(G)I becomes omnipotent enough to be able to end the world, even if it doesn't?

By the time we have to worry about preserving our understanding of the creative process against automation of it, we'll be on the verge of receiving post-linguistic knowledge transfer technologies and everything else, quicker than the automation can wreak its atrophying effects.

I don't have any more expertise or soothsaying power than you, so in the absence of priors to weight the options, I guess your opposing prediction is as likely as mine.

I'd just argue that the bad consequences of mine are bad enough to motivate us to try to stop it coming true, even if it's not guaranteed to do so.

Comment by samhealy on Please Understand · 2024-04-02T09:57:43.504Z · LW · GW

I don't think I'm selling what you're not buying, but correct me if I misrepresent your argument:

The post seems to assume a future version of generative AI that no longer has the limitations of the current paradigm which obligate humans to check, understand, and often in some way finely control and intervene in the output...

Depending on your quality expectations, even existing GenAI can make good-enough content that would otherwise have required nontrivial amounts of human cognitive effort. 

but where that tech is somehow not reliable and independent enough to be applied to ending the world...

Ending the world? Where does that come in?

and somehow we get this long period where we get to feel the cultural/pedagogical impacts of this offloading of understanding, where it's worth worrying about, where it's still our problem. That seems contradictory.

If your main thrust is that by the time GenAI's outputs are reliable enough to trust implicitly we won't need to maintain any endogenous understanding because the machine will cook up anything we can imagine, I disagree. The space of 'anything we can imagine' will shrink as our endogenous understanding of concepts shrinks. It will never not be 'our problem'.

Comment by samhealy on Please Understand · 2024-04-02T09:44:21.132Z · LW · GW

For what it's worth, I think even current, primitive-compared-to-what-will-come LLMs sometimes do a good job of (choosing words carefully here) compiling information packages that a human might find useful in increasing their understanding. It's very scattershot and always at risk of unsolicited hallucination, but in certain domains that are well and diversely represented in the training set, and for questions that have more or less objective answers, AI can genuinely aid insight. 

The problem is the gulf between can and does. For reasons elaborated in the post, most people are disinclined to invest in deep understanding if a shallower version will get the near-term job done. (This is in no way unique to a post-AI world, but in a post-AI world the shallower versions are falling from the sky and growing on trees.)

My intuition is that the fix would have to be something pretty radical involving incentives. Non-paternalistically we'd need to incentivise real understanding and/or disincentivise the short cuts. Carrots usually being better than sticks, perhaps a system of micro-rewards for those who use GenAI in provably 'deep' ways? [awaits a beatdown in the comments]

Comment by samhealy on AIOS · 2024-01-02T10:44:21.305Z · LW · GW

To clarify: I didn't just pick the figures entirely at random. They were based on the below real-world data points and handwavy guesses.

  • ChatGPT took 3.23 x 10^23 FPOPs to train
  • ChatGPT has a context window of 8K tokens
  • Each token is roughly equivalent to four 8-bit characters = 4 bytes, so the context window is roughly equivalent to 4 x 8192 = 32KB
  • The corresponding 'context window' for AIOS would need to be its entire 400MB+ input, a linear scaling factor of 1.25 x 10^4 from 32KB, but the increase in complexity is likely to be much faster than linear, say quadratic
  • AIOS needs to output as many of the 2 ^ (200 x 8 x 10 ^ 6) output states as apply in its (intentionally suspect) definition of 'reasonable circumstances'. This is a lot lot lot bigger than an LLM's output space
  • (3.23 x 10 ^ 23) x (input scaling factor of 1.56 x 10 ^ 8) x (output scaling factor of a lot lot lot) = conservatively, 3.4 x 10 ^ 44
  • Current (September 2023) estimate of global compute capacity is 3.98 x 10 ^ 21 FLOPS. So if every microprocessor on earth were devoted to training AIOS, it would take about 10 ^ 23 seconds = about 30000000000000000 years. Too long, I suspect.

I'm fully willing to have any of this, and the original post's argument, laughed out of court given sufficient evidence. I'm not particularly attached to it, but haven't yet been convinced it's wrong.

Comment by samhealy on AIOS · 2024-01-01T12:58:41.734Z · LW · GW

Agreed, largely. 

To clarify, I'm not arguing that AI can't surpass humanity, only that there are certain tasks for which DNNs are the wrong tool and a non-AI approach is and possibly always will be preferred. 

An AI can do such calculations the normal way if it really needs to carry them out

This is a recapitulation of my key claim: that any future asymptotically powerful A(G)I (and even some current ChatGPT + agent services) will have non-AI subsystems for tasks where precision or scalability is more easily obtained by non-AI means, and that there will probably always be some such tasks.

Comment by samhealy on AIOS · 2023-12-31T17:43:31.493Z · LW · GW

Plucked from thin air, to represent the (I think?) reasonably defensible claim that a neural net intended to predict/synthesise the next state (or short time series of states) of an operating system would need to be vastly larger and require vastly more training than even the most sophisticated LLM or diffusion model.

Comment by samhealy on Picasso in the Gallery of Babel · 2023-12-27T12:49:24.200Z · LW · GW

Is a photographer "not an artist" because the photos are actually created by the camera?

This can be dispensed with via Chalmers' and Clarke's Extended Mind thesis. Just as a violinist's violin becomes the distal end of their extended mind, so with brush and painter, and so with camera and photographer.

As long as AI remains a tool and does not start to generate art on its own, there will be a difference between someone who spends a lot of time carefully crafting prompts and a random bozo who just types "draw me a masterpiece"

I'm not as optimistic as you about that difference being eternal. The pre-to-post-AI leap is fundamentally more disruptive, more kind-not-degree, more revolutionary than the painting-to-photography leap or even the analogue-to-digital-photography leap[1]. With wirehead logic (why bother seeking enjoyable experiences if we can take dopamine-releasing drugs? Why bother with dopamine-acting drugs if we can release dopamine directly?) post-threshold GenAI dangles the promise of circumventing every barrier to 'creation' except the raw want. No longer is time, ten thousand hours' practice, inspiration, sticktoitiveness or talent required. 

In short, I fear that the difference (long may it exist) between the work of the careful craft-prompter and that of the random bozo will shrink with each generation. More specifically, the group of people who can distinguish between one and the other will collapse to a smaller and smaller elite (art historians, forensic AI experts), outside which paranoia and apathy will reign. That is not the democratic utopia we got promised by the CEOs.

  1. ^

    The principles of composition and optics, e.g. vanishing point, golden ratio, focal length, colour balance, depth of field, etc., apply to digital photography (and even rendered 3D graphics) just as they apply to analogue: the knowledge base is the same and a good, non-bozo photographer will excel in either submedium. Contrast this to an armchair prompt engineer who can 'create' a professional-seeming photorealistic image in seconds without knowing a thing about any of these principles. 

    Whether this is good/democratic or bad/obscurantist is open for debate, but I think it shows that the GenAI shift is different from and bigger than the earlier disruptive shifts it gets compared to.

Comment by samhealy on Picasso in the Gallery of Babel · 2023-12-27T11:56:25.598Z · LW · GW

Agreed. However,

  1. Do you think those IRL art forms will always be immune to generative capture and the threshold? 
  2. Even if the analogue original (a sculpture, a particular live dance performance, a particular live theatre performance) remains immune, most people will consume it through digital reproduction (photographs, VR, AR, video, audio) for which the threshold does apply.
Comment by samhealy on Picasso in the Gallery of Babel · 2023-12-26T21:52:05.395Z · LW · GW

Oh I see! So that's just the default slideshow range, padded front or back with zeroes, and you can enter much longer locations manually?

Comment by samhealy on Picasso in the Gallery of Babel · 2023-12-26T21:42:21.624Z · LW · GW

I like this.

It feels related to the assertion that DNNs can only interpolate between training data points, never extrapolate beyond them. (Technically they can extrapolate, but the results are hilarious/nonsensical/bad in proportion to how far beyond their training distribution they try to go.)

Here's how I see your argument 'formalised' in terms of the two spaces (total combinatorial phase space and a post-threshold GenAI's latent space over the same output length), please correct anything you think I've got wrong:

A model can only be trained on what already exists (though it can augment its dataset to reduce overfitting etc.), and what already exists approximately overlaps with what humans consider to be good, or at least sticky. What humans consider to be good or sticky changes over time; the points in phase space that are not in the latent space of and therefore not accessible to GenAI 1.0 are different from those not in the latent space of GenAI 2.0, assuming GenAI 2.0 has had access to new, external training data from human or other agents with ongoing life experiences. GenAI 1.0 therefore cannot claim to have subsumed GenAI 2.0 before the fact, whereas GenAI 1.0 + 2.0 has a provably broader gamut.

Though I like the spirit of this, it doesn't quite refute the certainty of provenance-erasure, at least in theory. With a training dataset that ends at time t1, a sufficiently high-fidelity GenAI could (re-)create anything plausibly or actually human-made at or before t1, and only fresh creativity from agents with subjective experience between t1 and t2, when the model's next generation begins training, is fully 'protected' (unless the fresh creativity is actively excluded from the next model's training data). 

Also: what about the limiting case of a CI-style model that is always in combined training-and-inference mode, so is never more than trivially out-of-date? (I know current architectures aren't designed for this, but it doesn't seem like a lunatic leap.)

Objections aside, the idea of dynamic interaction (and perhaps Hofstadterian feedback loops) between experiential agent and environment being the secret sauce that DNNs can never recreate is appealing. Can the idea be made rigorous and less dualism-adjacent? What exactly is the dynamic setup getting that DNNs aren't? Maximal or optimal transfer of order/entropy between agent and environment? An irreducible procedure in which Kolmogorov complexity = string length, so that any attempt to compress or reduce dimensionality kills its Truly Generative spirit?

Comment by samhealy on Picasso in the Gallery of Babel · 2023-12-26T21:01:58.551Z · LW · GW

Cute! But how does each 16-digit 'image location' value (of 10^16 in total) uniquely represent one of the 4096^266240 possible images?

Comment by samhealy on Simulation hypothesis and substrate-independence of mental states · 2020-11-26T12:29:57.351Z · LW · GW

Very interesting article. Most of my objections have been covered by previous commentators, except:

1a. Implicit in the usual definition of the word 'simulation' is approximation, or 'data compression' as Michaël Trazzi characterises it. It doesn't seem fair to claim that a real system and its simulation are identical but for the absence of consciousness in the latter, if the latter is only an approximation. A weather forecasting algorithm, no matter how sophisticated and accurate, will never be as accurate as waiting to see what the real weather does, because some data have been discarded in its input and processing stages. Equally, a lossy simulation of a conscious human mind is unlikely to be conscious.

1b. What Bostrom, Tegmark and other proponents of substrate-independent consciousness (and therefore the possibility of qualia in simulations) have in mind is more like the emulators or 'virtual machines' of computer science: lossless software reproductions of specific (hardware and software) systems, running on arbitrary hardware. Given any input state, and the assumption that both emulator and emulated system are working correctly, the emulator will always return the same output state as the emulated system. In other words, emulation is bit-perfect simulation.

1c. Even if brains are analogue computers, one can emulate them accurately enough to invoke consciousness (if it can be invoked) in a digital system with sufficiently high spatial and temporal resolution: real brains must have their own error correction to account for environmental noise, so the emulator's precision merely needs to exceed the brain's. Against Roger Penrose's vague, rather superstitious claim that quantum effects unique to meat brains are necessary for consciousness, much the same argument holds: brain electrochemistry operates at a vastly larger scale than the quantum upper limit, and even if some mysterious sauce does bubble up from the quantum to the classical realm in brains and not in silicone, that sauce is made of the same quarks and gluons that comprise brains and silicone, so can also be understood and built as necessary into the emulator.

2. You say that "People in universe A can act upon (change their conscious experiences) people in Universe B, at least by shutting the computer down." But shutting the computer down in Universe A does NOT change the conscious experiences of the people in the emulated Universe B, because they are only conscious while the emulator is running. Imagine being a conscious entity in Universe B. You are in the middle of a sneeze, or a laugh. An operator in Universe A shuts down the machine. Then imagine one of three scenarios taking place: the operator never restarts the machine; after a year the operator restarts the machine and the simulation continues from where it was paused; or the operator reboots the machine from its initial state. In none of these scenarios is your conscious experience affected. In the first, you are gone. You experience nothing, not even being annihilated from a higher magisterium, since the emulator has to be running for you to experience anything. In the second, you continue your laugh or sneeze, having felt no discontinuity. (Time is, after all, being emulated in Universe B along with every other quale.) In the third, the continuous 'you' from before the system was rebooted is gone. Another emulated consciousness begins the emulation again and, if no settings are changed, will have exactly the experience you had up to the reboot. But that new 'you' will have no memory of the previous run, nor of being shut down, nor of the reboot or anything subsequent.

3. Against your argument that stripping away inputs and outputs from the simulation constitutes a reductio ad absurdum of the premise that emulations can be conscious: this is true of meat brains too. To nourish an embryonic brain-in-a-jar to physical 'maturity' (whatever that might mean in this context), in the absence of all communication with the outside world (including its own body), and expect it to come close to being conscious, is absurd as well as ghoulish. Moreover -- relating this with 1. above -- to say that you have strictly emulated a brain-centred system whose consciousness you are trying to establish, you would have to include a sphere of emulated space of radius greater than tc around the brain's sensory organs, where t is the length of time you want to emulate and c is the speed of light in a vacuum (assuming no strong simulated gravitational fields). This is because information from anywhere in the tc-radius sphere of real space could conceivably affect the brain you're trying to emulate within that time interval.