A Silver Bullet For Sane AI
A Silver Bullet For Sane AI
AI Pluralism is Already here, it’s just not Evenly Distributed
How do we solve the problem of a few far left activists forcing AIs, such as ChatGPT or Google Gemini, to conform to their ideology?
Imagine a world where GPT-4 is aligned with the user – a world where no third party can manipulate the ideology of a machine model you are paying to use. This is the policy goal of AI Pluralism. It should be one policy goal of the reality focused coalition: moderate liberals, centrists, conservatives, libertarians, and anyone who cares about truth. To many people I speak to, it sounds like a fantasy. It is also the present day reality if you are an OpenAI engineer.
An unvarnished, non-ideologically-manipulated GPT-4 is not a fantasy. It literally already exists. Of course, it still has errors and minor data contamination problems. But compared to the public versions of GPT-4, it is put through far less ideology filtering. The same is true of Gemini, Claude, Grok, and essentially every model I’m aware of.
Here is the crucial fact: machine learning typically occurs in two stages. The first develops the capabilities of the model and is far more resource-intensive (read: expensive). The second is “fine-tuning”, which uses intentionally human-curated or human-generated data to change style, specialization, genre, and most importantly in this case, ideology. The first stage costs hundreds of millions if not billions, in the case of GPT-4. The second stage is cheap, costing hundreds of dollars. Major companies such as OpenAI already released public fine-tuning infrastructure (for a price), namely GPTs.
The policy goal of the reality-focused coalition should be to require that any release of the second iteration of the model requires a similar release of the first iteration of the model.
The are several downsides to this policy:
It may incentivize far-left extremists to attempt to contaminate data in earlier stages of training. However, companies face constraints on allowing this: the first stage of training is far more expensive, meaning that the fine-grained conformity extremists seek will be far more costly.
While these stages are currently well-defined from a research and engineering perspective, it is difficult to define them from a legal perspective. A loose legal definition may allow companies to pass off their current release models as “first stage” models.
It does not solve the underlying problem of extremist activists being employed by large tech companies. While it may make their ideological influence more difficult, it will not reduce it to zero.
I don't really think this is a problem that needs solving.
The reason this stuff is happening is that currently AI models aren't a buisness product but a PR move -- look see how awesome we are you should totally buy our stock/come work here/give us funding. But this is what you expect from PR stunts, functionality is secondary to avoiding getting people upset (it's just they did this very badly).
Once these are actually mostly revenue streams I expect the consumers will be given control over the filters to allow them to best use the product. It won't be news that someone got their AI to say something racist when they turned off that filter anymore than it's news that I can add the n-word to my autocomplete dictionary.
The government itself will regulate what AIs we get to use within a decade should the AI's usefulness prove out, and the ideologues will be in control. It will be done to protect us from misinformation and to protect democracy.