The alignment problem sounds abstract until you say it plainly: how do you get a system to want what we actually want, past the literal words we typed into the prompt. Every camp in AI ethics is really placing a different bet on how to solve that. Here are four bets I find genuinely interesting, plus my own two cents on each.

Yann LeCun bets on openness

He argues that locking powerful models inside a handful of labs concentrates enormous leverage in very few hands, while open code lets the wider research community audit, poke, and improve a system instead of trusting a company's word for its own safety. His rival Yoshua Bengio calls this reckless, comparing it to open-sourcing a weapon and handing out the blueprints. My take: I built RepoMind, my codebase intelligence tool, to run fully local, keeping code on the user's own machine at every step. Openness shows up as architecture, real and specific, long before it ever becomes a slogan on a slide.

Bengio's counter-bet centers on concentration of power

He worries less about a rogue AI plotting against humanity, what researchers call existential risk, or informally someone's "p(doom)," their personal probability estimate that AI ends badly, and worries more about a handful of companies quietly running the table on capital, compute, and policy influence. My take: I once watched three girls in Nairobi write their first Python loop, sharing one secondhand laptop between them. Access decides who gets to build the future, long before raw talent ever gets a say. Concentration reaches beyond a corporate problem into a doorway problem, one most people walk straight past without knowing it was there.

The technical camp bets on alignment through RLHF

Reinforcement learning from human feedback, where a model earns reward for outputs humans rate as helpful, harmless, and honest, shorthanded by researchers as the HHH principle. Anthropic pushes this further with constitutional AI, teaching the model a written set of values directly, building conscience into the foundation instead of only correcting bad outputs after launch. My take: the best fix I ever shipped at Smartant lived inside the rule itself, baked into the design from the very first draft. Good architecture scales further than good policing, every single time.

A fourth bet gets less airtime but hits closer to home for me

Scholars studying what they call data colonialism point out that the entire AI supply chain, labeling, moderating, red-teaming, runs largely on cheap labor from the Global South, while profit and credit both flow north. My take: growing up between Kenya and India taught me that data has a hometown. Every dataset traces back to real people, and right now most of them earn a paycheck where a byline would be fairer.

Where they meet

Here is where all four bets actually meet. They argue fiercely about speed and structure, yet they agree on something bigger: power shapes technology, and technology reshapes power right back. Openness guards against monopoly. Access decides who gets to invent. Values belong in the foundation, ahead of any guardrail bolted on after launch. Origin deserves credit over extraction. Socrates said wisdom starts with knowing how little you know. Building genuinely ethical AI might start the same way, with an honest count of how many rooms remain closed to you.

So here's my question, the one I keep circling back to. If the room designing our future stays this narrow, whose intelligence are we actually scaling?