Wait, what?!? One of the biggest AI companies wants to help write the laws regulating their own products. What could possibly go wrong? So I reviewed the reporting on Anthropic's proposed AI safety legislation. While I don't have the underlying legislative text or supporting data (which is behind a paywall), there are some obvious questions we should all be asking: 1. Evidence: What measurable evidence supports the proposed risks and regulatory thresholds? Are we regulating demonstrated dangers or hypothetical catastrophes? 2. Accountability: Could these rules reduce Anthropic's liability when its AI causes harm to customers or other companies? Who ultimately remains responsible? 3. Competition: Will compliance costs disadvantage AI startups while protecting established players? How convenient. 4. Foreign open-weight models: Could these regulations restrict our ability to download, modify, or deploy models developed outside the US? 5. Truly Open source models: What happens to genuinely open-source, community-developed models like OpenWALDO? Who bears compliance costs and legal responsibility when there's no single corporate developer? 6. Enforcement: Who decides what's dangerous, and what objective, reproducible evidence is required? Will independent evaluators have meaningful authority, or will the AI companies effectively define their own standards? 7. Responsibility: Does the legislation distinguish model behavior from failures in agents, infrastructure, permissions, and human oversight? I'm supportive of reasonable, evidence-based AI safety regulation. But when an industry leader helps draft the rules governing its own industry, yellow flags should go up and scrutiny is warranted. Are we protecting the public, or potentially protecting the companies writing the rules? Maybe both. That's why we need transparency before drawing conclusions.

This is fantastic. @BenAffleck understands and can reasonably explain something that many of the so-called AI "experts" seem determined to over complicate and mislead people on. In fact, many of these experts are rapidly losing credibility with people who actually understand how this technology works. Unfortunately, they've already convinced a lot of people that AI is something it isn't. And as the saying goes, it's easier to fool someone than to convince them they've been fooled. An LLM is a computer program. We designed it. We know the architecture, the mathematics, and how inference works. Much of the underlying code and research is open for anyone to inspect. All LLMs processes tokens through mathematical operations using learned weights and context to predict what comes next. The results can be astonishingly accurate, useful, and convincing. But the LLM isn't alive. It doesn't love you, hate you, care about you, or sit around contemplating its existence or understanding. And yet people insist that somewhere inside those matrix operations, consciousness has magically materialized. Seriously? This isn't biological life evolving in nature. It's software executing computations on hardware we designed and built. If you're claiming those computations somehow produce subjective experience, explain the mechanism (I'll wait...). The remarkable achievement here is that we've created software capable of convincingly emulating aspects of human intelligence and behavior. That's freakin' amazing and cause to celebrate the technology! But please stop confusing a computer program with a living, conscious being. It's absolutely not. The fact that so many people have fallen for that illusion is a testament to how convincing the technology has become. But here's the clincher: when we misunderstand this, and we start treating computer programs as conscious, independent beings, the legal implications could be disastrous. Suddenly, "the AI did it" becomes an excuse to shift blame away from the people and companies responsible for building, deploying, and operating these systems. AI cannot be held accountable. But people and companies can be, and must be held accountable. Ben gets it. More people should.