# alignment through structure we cannot achieve ai alignment just by telling ai to be good, because good is a subjective notion we cannot align it to our own reward functions, because we cannot formulate them in any computable way but with vague and broad terms we cannot rely on governance, because companies and countries have strong incentives to accumulate power, and do that first - as we are witnessing in real time, government is only democratic or friendly as long as the people who lead it we also cannot rely only on technical methods, because we have little control of what comes in the training data and fine-tuning can easily be bypassed, mechanistic interpretability is vague, and a superintelligence that teaches itself would be able to retrain without any human-imposed technical barriers anyway it is well said > we don't rise to the level of our expectations, we fall to the level of our training if it will be at all possible to misuse ai, it will be misused if it at all will be possible to start and win a war with low risk by using ai, people will do that if it will be possible to get absurdly rich at the cost of the whole society, people will do that checks and balances do not work when they are based on someone's decision. they work when they are deeply embedded into a resilient system therefore, the only way to align not just ai with human preference, but also, more importantly, humans among each other, is to create ai in a societal, political, technological structure, in which alignment will be an inevitable byproduct as we have learned through history, those system are such that limit the accumulation of power and distribute the future among people. it may not be perfectly even distribution, and while equality and fairness are in many respects desirable, but for any practical purposes, unachievable. even more so, as dozens of socialist and communist experiments across history have proven, a desire for evenness can get more deadly than a direct goal of extermination, as the numbers of murdered in the soviet union and china are greater than those by the nazi germany and japan in context of ai alignment, 1. ownership of data, models, and compute 2. distribution of resources, capabilities, and usage 3. decentralization of control, governance, and power are necessary prerequisites such that humanity doesn't kill itself and does not end up under control of one super powerful ai those features define the structure of the worldwide ai computer that is being built on top of the internet by the likes of microsoft, google, facebook, nvidia, alibaba, tencent, baidu, and other, collectively spending hundreds of billions of dollars to shape the ai industry - and the future of civilization - to fit their interests, which have nothing to do with the interests of the global society the question of alignment is the question of power, wealth, and freedom - if survival wasn't already enough we only have until 2030 to decide what world we want to live in --- // 13 apr 2025 #alignment