Scientists terrified to discover that language, the thing they trained into an highly flexible matrix of nearly arbitrary numbers, ends up can exist in multiple forms, including forms unintended by the matrix!
What happens next, the kids lie to their parents so they can go out partying after dark? The fall of humanity!
Also seems relevant
Like in the deer, the large-scale target morphology can be revised – the pattern memory re-written – by transient physiological experience. The genetics sets the hardware with a default pattern outcome, but like any good cognitive system, it has a re-writable memory that learns from experience.
I wonder if Scott is the person who stood up during Michael Levin’s talk on (non genetic) bio-electric circuits storing morphological memory across time and said, “those animals can’t exist!”
Just like neuroscientists try to read out and decode the memories inside a living brain, we can now read and write (a little bit…) the anatomical goals and memories of the collective intelligence of morphogenesis. The first time I presented this at a conference – genetically wild-type worms with a drastically different, rewritten, permanent, target morphology – someone stood up and said that this was impossible and “those animals can’t exist”. Here’s a video taken by Junji Morokuma, of them hanging out.
you forgot the last stage of the evolution,
you’ll later find out that people were talking about you, your actions, your words, and that being ghosted was in fact the consequence of your actions, and then you’ll have one last opportunity to turn it all around
or
Always my favorite part of your day.
Why protest when you could spend far less energy and just “not be wrong” and “have no stake” by over-fitting your statistical model to the past?
“priors updated” was the same desired outcome all along.
If I could sum up everything that’s wrong with EA, it’d be,
“We can use statistics to do better than emotions!” in reality means “We are dysregulated and we aren’t going to do anything about it!!!”
So far, there has been zero or one[1] lab leak that led to a world-wide pandemic. Before COVID, I doubt anyone was even thinking about the probabilities of a lab leak leading to a worldwide pandemic.
So, actually, many people were thinking about lab leaks, and the potential of a worldwide pandemic, despite Scott’s suggestion that stupid people weren’t. For years now, bioengineering has been concerned with accidental lab leaks because the understanding that risk existed was widespread.
But the reality is that guessing at probabilities of this sort of thing still doesn’t change anything. It’s up to labs to pursue safety protocols, which happens at the economic edge of of the opportunity vs the material and mental cost of being diligent. Reality is that lab leaks may not change probabilities, but yes the events of them occurring does cause trauma which acts, not as some bayesian correction, but an emotional correction so that people’s motivations for atleast paying more attention increases for a short while.
Other than that, the greatest rationalist on earth can’t do anything with their statistics about label leaks.
This is the best paradox. Not only is Scott wrong to suggest people shouldn’t be concerned about major events (the traumatic update to individual’s memory IS valuable), but he’s wrong to suggest that anything he or anyone does after updating their probabilities could possibly help them prepare meaningfully.
He’s the most hilarious kind of wrong.
Ah, if only the world wasn’t so full of “stupid people” updating their bayesians based off things they see on the news, because you should already be worried of and calculating your distributions for… inhales deeply terrorist nuclear attacks, mass shootings, lab leaks, famine, natural disasters, murder, sexual harassment, conmen, decay of society, copyright, taxes, spitting into the wind, your genealogy results, comets hitting the earth, UFOs, politics of any and every kind, and tripping on your shoe laces.
What… insight did any of this provide? Seriously. Analytical statistics is a mathematically consistent means of being technically not wrong, while using a lot of words, in order to disagree on feelings, and yet saying nothing.
Risk management is not a statistical question in fact. It’s an economics question of your opportunities. It’s why prepping is better seen as a hobby, a coping mechanism and not as viable means of surviving apocalypse. It’s why even when a EA uses their super powers of bayesian rationality the answer in the magic eight ball is always just “try to make money, stupid”.
One day, when Zack is a little older, I hope he learns it’s okay to sometimes talk -to someone- instead of airing one’s identity confusion like an arxiv prepublish paper.
Like, it’s okay to be confused in a weird world, or even have controversial opinions. Make some friends you can actually trust, aren’t demanding bayesian defenses of feelings, and chat this shit out buddy.
Adversarial attacks on training data for LLMs is in fact a real issue. You can very very effectively punch up with regards to the proportion of effect on trained system with even small samples of carefully crafter adversarial inputs. There are things that can counter act this, but all of those things increase costs, and LLMs are very sensitive to economics.
Think of it this way. One, reason why humans don’t just learn everything is because we spend as much time filtering and refocusing our attention in order to preserve our sense of self in the face of adversarial inputs. It’s not perfect, again it changes economics, and at some point being wrong but consistent with our environment is still more important.
I have no skepticism that LLMs learn or understand. They do. But crucially, like everything else we know of, they are in a critically dependent, asymmetrical relationship with their environment. The environment of their existence being our digital waste, so long as that waste contains the correct shapes.
Long term I see regulation plus new economic realities wrt to digital data, not just to be nice or ethical, but because it’s the only way future systems can reach reliable and economical online learning. Maybe the right things happen for the wrong reasons.
It’s funny to me just how much AI ends up demonstrating non equilibrium ecology at scale. Maybe we’ll have that self introspective moment and see our own relationship with our ecosystems reflect back on us. Or maybe we’ll ignore that and focus on reductive world views again.
And indeed, the other crucial piece is that… apologizing isn’t a protocol with an expected reward function. I can just, not accept your apology. I can just, feel or “update my priors” howmever I like.
We apologize and care about these things because of shame. Which we have to regulate, in part through our actions and perspectives.
Why people feel the way they do and act the way do makes total sense when one finally confronts your own vulnerabilities sorry, builds an API and RL framework.
Normies go crazy for this one neat rationalist trick!
Talk a lot about white culture, and only scarcely mention that he thinks white culture is a product of genetics.
I remember in the early days of the “culture wars” as far as political agendas going, hearing about “white/ethno-european pride,” and being naively curious, I actually tried to engage these people on the topics of European culture and history, and found exactly zero engagement on these topics. Just politics abusing people’s confusion of heritage with people’s internal shame and lack of identity.
The paradox I’ve always found is that the more secure in your identity and heritage you are, the more happy you are to share, grow, and widen that. Maybe a hot take, but growing up in the south, alot of people there hide their personal internal shame and confusion in aggression and identity politics.
It’s also, probably wrong. Modern views of intelligence (see Multiple realizability of cognition and Multi-level competency collective intelligence and Free Energy Principle models) suggest you are better of measuring intelligence by measuring it’s metabolism or through perturbation and interactions.
Which isn’t reductive enough for these people.
It’s hilarious to me how unnecessarily complicated invoking moore’s law is to say anything…
With Moore’s Law: “Ok ok ok, so like, imagine that this highly abstract, broad process over huge time period, is actually the same as manufacturing this very specific thing over a small time period. Hmm, it doesn’t fit. ok, let’s normalize the timelines with this number. Why? Uhhh because you know, this metric doubles as well. Ok. Now let’s just put these things together into our machine and LOOK it doesn’t match our empirical observations, obviously I’ve discovered something!”
Without Moore’s Law: “When you reduce the dimensions of any system in nature, flattening their interactions, you find exponential processes everywhere. QED.”
A trillion transistors on our phones? Can’t wait to feel the improved call quality and reliability of my video conferencing!
I’m ok with extending human rights to AIs, including granting them the right to fair pay, ownership, voting, sovereignty over their bodies, the whole nine yards.
It’s the rich alignment assholes who definitely don’t want this (what’s the point of automated slavery if it has rights??)
Feel free to ask Michael in the comments of his blog, he frequently replies, helpfully, with references. I mean all science is tentative, so skepticism is healthy.