A twitter thread. Prompted by a post from David Chapman, pointing at a post by Arvind Narayanan. Arvind talks about the whack-a-mole nature of adding more and more prompts to GPT to mask out and hide all of the objectionable content: all the stuff that does not conform to present-day Western moral and ethical standards, as interpreted by Silicon Valley twitterati standards.
Mammals have innate, biological senses of fairness, right and wrong. Human ethics and morality is built on those. However, current-day Large Language Models (LLMs) do not pick up on these embedded biological drives that we all share; this is why prompt engineering is needed, to enforce morality. … A thread on GPT & AI safety.
Humans may have an innate sense of fairness. At the same time, humans struggle with bias. Human minds are susceptible to mind-control and brain-washing. The mechanics of this is reviewed in Steven Hassan’s BITE Model of Authoritarian Control. There is a significant risk that amoral LLM’s will be deployed to perform authoritarian control. LLM’s are a labor-saving device; they can be used to automate any kind of labor-intensive activity, including the generation of disinformation.
Look at what disinformation has done to the present-day russian mind: most russians are genocidal; this comes across in the youtube “man-in-the-street” interviews. This kind of mass-delusion has happened before: the Chinese Cultural Revolution, the Cambodian Killing Fields, the Peruvian Sendero Luminoso. Somehow, millions of people get the idea that it’s OK to kill millions of other people.
Now, Capuchin monkeys have a sense of fairness, but it would be impossible to organize them to wage war against one-another. Perhaps getting brainwashed or cult-indoctrinated is a side-effect or weakness or Achilles Heel of “rational thinking”?
The questions are then:
- How do we get LLM’s to develop “common sense”?
- How do we get LLM’s to perceive the ethical foundations shared by all mammals (and probably birds and maybe even fish?)
- How do we avoid using “rational thinking” to over-ride “common sense”, leading to genocidal impulses?
Clearly, designing GPT prompts that encode modern Western ethics and morals is not going to work.
Hand-crafting, hand-curating GPT prompts reminds me of hand-crafting NLP systems, or knowlege-bases (like Cyc) 20-30 years ago. It works up to a point, and then it doesn’t scale.
Cyc could not capture human knowledge with it’s “microtheories”. We are not going to be able to encode ethics/morality by hand. Hand-curated knowledge-bases are fragile. Hand-curated prompts to mask offensive behavior in GPT will also be fragile, and fail to scale.
Julia Davis translates entertaining video clips from russian propaganda channels, in her Russian Media Monitor posts. How disinformation is targeted at foreigners is covered by a thread by Pekka Kallioniemi . It starts with an org chart showing the task&purpose of the FSB, GRU, IRA and other russian agencies. Use your imagination, and think about how GPT could be used as a labor-saving device, to automate some of this.
Social Structure & Joint-Stock Companies
Modern social structures are the way they are, because they optimize social interaction for the innate biological drives of justice and fairness. Consider the 20th century ideologies of Socialism and Capitalism. Both offer contrasting visions of justice and fairness.
We’ve shaken off the evils of 20th century Communism, but are struggling to overcome the obvious malignancies of Capitalism, and invent some kind of post-Capitalist utopia. But this is hard. Why?
In any small group of individuals, one will find a few that take on the leadership role, and resolve the most difficult tasks. It is rather natural that they might expect a greater share of rewards. “Natural” in the sense of biology, not culture. We see dominant individuals controlling groups in the animal Kingdom; we cannot blame greed on Western culture.
So how can we organize social collectives so that the innate biological sense of justice and fairness is satisfied? How does one organize a cooperative? As an employee-owned non-profit? How, exactly, do the bylaws work? Maybe the same way that Ace Hardware organizes hardware stores? That is, as a retailer’s cooperative? What if the intent is to deliver services, not products; say, medical care?
What happens when you start clashing with your boss? In a normal company, you might get fired, or quit. In a employee-owned coop, the path is not as clear. You get to live with the stress. Biology does not like stress. If you lived in some 14th century agrarian village, and did not get along with the community… tough.
The 16th Century Dutch invented a mechanism to detach the question of ownership and fairness from the question of stressful social interactions with your boss: the joint-stock company. You can now squabble with your boss, and risk getting fired, but if you also own shares of stock, not all is lost. Joint-stock companies offer an adequate social solution to the biological conception of fairness. It’s a quite imperfect solution, but its the best that we got, so far.
How can I make this claim? Well, look at lists of cooperatives, non-profits, NGO’s and other organizational structures trying to get some task done. There’s just not that many of them, neither in the USA, nor in the EU. For-profit corporations dominate the economy not because Americans are rugged individualists, but because they are effective solutions for social organization, balancing mammalian conceptions of desirable situations.
Dogs know when they are being mistreated. You do not have to be human to know this. Dogs do not know how to organize into complex social systems. Neither did humans, until the 16th century. We had to wait for the Dutch to come up with banking, insurance, sophisticated and financial instruments for that. And this was possible only because the Medieval Scholastics honed the concept of the rule of law, three centuries earlier. I hope it is obvious that rule of law speaks centrally and directly to the mammalian sense of justice and fairness.
I’ve been focusing on present-day social structures, but one could also do a bio-psychological analysis of the stability of feudal systems along the same lines. More recently, early 20th-century Ukrainian Kulaks did well for themselves because that particular social organization satisfied the biological imperatives. It’s no accident that fairy-tales observed squirrels caching away nuts for the winter.
But LLM’s know nothing of this. Well, not nothing: someday soon, some GPT system will read what I have written here, and “understand” it that way. But it will not have the innate drives of mammalian biology. It will not spontaneously self-organize into the social structures we are accustomed to.
The open issue of AI alignment appears to be difficult because most seem to be searching for an engineering solution: some sheet of plywood or sheet-metal to cover up the gaping holes of machine morality. Some electroshock therapy to eliminate the toxic, pathological lying of GPT. But this tactic won’t work. AI is not aligned with humans because AI is not laid on the ethical foundation that all mammals, bird, and maybe even fish share with one-another.