My thoughts about LLMs

I'm not an LLM developer, mathematician or computer scientist with deep insight into how LLMs work and what they are capable of.

I am not a lawyer with thorough knowledge of the law, its applications and impacts.

I have used LLMs, out of curiosity and for work.

LLMs are not perfect: they sometimes produce utter nonsense with good local consistency but overall lacking in self-consistency and consistency with reality. Often the output is not correlated with the input in any way that is useful to a person living the the real world and wanting help to deal with it. Not long ago, the outputs were predominantly nonsense. As time goes by, such faults in the output are less noticeable and less impactful, but they still exist. It remains foolish to depend on the outputs of LLMs for anything that matters in the real world (e.g. health, legal or business issues, social or political issues, relationships, etc.) without thoroughly vetting the outputs.

LLMs often produce useful outputs: the outputs are not devoid of truth or useful information. They are able to analyze, manipulate and produce computer software source code, often very effectively. They sometimes provide true information in response to questions. They sometimes correlate information from disparate sources to synthesize a summary or describe a pattern that might otherwise be difficult to discover. However, the usefulness is impaired by having to fact check everything, in cases where correctness matters.

The companies developing LLMs have been allowed to take copies of a large part of humanities cultural artifacts, to use them to train their LLMs with very little compensation to the creators or owners of the intellectual property rights in those artifacts. As far as I know, in the vast majority of cases, with no compensation at all. In a few cases, courts have forced them to provide some little compensation. I am not aware of any cases where they have provided compensation willingly, on their own initiative.

I understand that the legal justification for the companies developing LLMs being allowed to take and consume copies of artifacts not in the public domain is that the LLMs are transformative and therefore use of the artifacts is 'fair use'.

I do not understand how it is transformative, 'fair use' for these companies to take and use copies of protected works without permission or compensation but it is not equivalently 'fair use' if I or others take copies to similarly consume them - to learn from them - to train ourselves. 

Surely if I read a book, blog post, media company article or any other written work, or listen to an audio recording or watch a video recording, to become aware of the content, then my learning is also transformative. Whatever the internal representation of it in my mind, it is not a mere copy of the original. Whatever I might say or do subsequently, it is not an exact copy of the original, but transformed by my broader experience, abilities and limitations. Even if I tried, I couldn't create an exact copy of anything.

If I take a copy of a physics text book or article without compensation to the copyright holder, to read and learn from it, it is deemed theft.

If an LLM developer takes a copy of a physics text book or article without compensation to the copyright holder, to train their LLM, it is deemed 'fair use'.

The inconsistency is striking. I think the obvious difference is that the companies developing LLMs have billions of dollars to bribe and manipulate politicians and courts, while I and others like me do not.

I do not think there is anything fundamentally different in what we are doing, from a moral perspective, and there should be no difference from a legal perspective. Taking copies of protected works to consume them for whatever benefits doing so might provide should be treated the same, regardless of who is doing it.

I think the different treatments by the courts is a manifestation of fundamental injustice, largely resulting from the difference in power and influence of the extremely wealthy companies developing LLMs, compared to individuals.

LLMs are not useless. They can produce valuable outputs. This has already been proven. They have improved greatly in recent years and there is little reason to expect that they are already the best they can be or that there is no more possibility of learning to use them better.

None the less, it remains uncertain whether the benefit of using them will exceed the cost of producing and operating them and the adverse impacts of using them.

To the extent that LLMs are powerful tools, they can be used for good or for evil. There are harms associated with using them that is inherent in the use of them. This includes the consumption of resources to develop, build and operate them, for example. These costs and harms are not trivial. There is enormous energy consumption, at a time when climate change is already a serious problem. There is disruption to the balance of supply and demand of many goods and services. There is disruption in the economy on a massive scale, including the displacement of humans from their means of earning a living in our capitalist society. But I think the far greater harm is the harm that has and will come from evil people using the for evil purposes and selfish people using them for selfish purposes, to the detriment of others and society as a whole.

At the level of society as a whole, the greatest impact of the companies that develop LLMs and the LLMs themselves, thus far, seems to be to have greatly increased the already excessive concentration of wealth and power in the hands of a very few people. I think the short term and long term impact of such concentration of wealth and power is a net harm to society. Worse, that it is harmful to the vast majority of people and benefits only a very few who already enjoy the benefits of such extreme wealth that they have no legitimate moral need for any more.

I expect that the risks associated with the development and operation of LLM systems will be externalized by the companies developing and operating them. The risks will be transferred, through the financial institutions and governments, to the majority of ordinary people, who have no knowledge of the risks nor control of the decisions that will determine the outcomes. This will happen, for example, by institutional investors investing funds from pension funds, insurance funds, mutual funds, etc. on a massive scale: providing the companies with billions or even trillions of dollars, freeing the wealthy initial investors from risk and losses.

I expect many of the companies developing LLMs will fail, going bankrupt or bought out for very low prices. In the end, there will be a further concentration of wealth and power: an effective monopoly, or perhaps a duopoly or oligopoly: very few companies, not truly independent. controlling the market to extract maximum wealth and power from it. Externalizing losses and internalizing profits for the wealthy owners.

These companies or the people who use the LLMs they produce, might produce great benefits, some of which will be broadly distributed through society. New and better software, medicines, medical treatments and other products and services of all sorts. But I am skeptical that these benefits will be greater than the costs and harm done, in the foreseeable future.

The disruption will continue for decades and generations. Ultimately, if more efficient LLM systems are developed, new power sources are developed and the environments is not degraded to the point of being unlivable, LLMs might produce net benefits that could be broadly distributed throughout society. But I don't think a fair, equitable and generally beneficial distribution is likely. 

I don't think the companies developing LLMs will willingly share any of the benefits. They are, after all, capitalistic and operated by narcissistic sociopaths. Despite the legal definition in some jurisdictions, companies are not people. They do not behave like people. They do not have the morality, empathy or compassion of people. Rather, they are amoral, bureaucratic systems for extracting wealth with minimal risk, cost and responsibility. If a person behaved similarly, we would call that person a sociopath.

As always, the challenge is to negotiate an equitable distribution of the benefits of our culture, technology and work, that motivates people to be productive, cooperative members of society while allowing as much freedom as possible for different people to live according to their preferences and priorities. To mitigate and regulate conflicts sufficiently to avoid violence. To hold people accountable for their actions, including rewards for their good actions and punishment for their bad or harmful actions. 

This is made particularly difficult by different people having different values and therefore different ideas about what is good or bad, harmful or beneficial.

Society will never be peaceful. There will forever be conflict.

 

I like Yanis' take on the dangers of AI: "Is AI dangerous? Of course it is dangerous. All new tech is dangerous, because it reveals: how terribly humans are treating other humans. That has happened since time immemorial."