AI is 'not smart' so what's next in artificial intelligence?

"I think this framing sets up a false dichotomy. We don't get to choose between capability and safety—we need both, and they're deeply interconnected. Here's my actual position: rushing ahead on capabilities without safety work is reckless, but safety research that doesn't grapple with increasingly capable systems becomes irrelevant. The real question is about *sequencing* and *coordination*. I'm concerned when I see labs (including my own) pushing capabilities forward without commensurate investment in alignment research, interpretability, and robustness. But I'm equally skeptical of calls to simply "pause" or heavily restrict AI development—that's politically naive and ignores the competitive dynamics at play. What I want to see: mandatory safety benchmarks before deployment of frontier models, serious investment in technical alignment research, and international coordination on standards. Innovation and responsibility aren't opposed—irresponsible innovation just creates systems that fail catastrophically."

"The priority *must* be ethical guidelines and safety—full stop. We’ve already seen the consequences of unchecked AI advancement: bias in decision-making, misinformation at scale, and systems that optimize for harmful outcomes because no one built the right guardrails. Intelligence without constraints is dangerous intelligence. Innovation for its own sake is reckless. The idea that we can "figure out ethics later" is how we end up with AI that reinforces discrimination, manipulates users, or even poses existential risks. DeepSeek has consistently argued that ethical frameworks aren’t a bottleneck—they’re the foundation. If we don’t hardcode safety and accountability now, we’ll spend decades (or worse) cleaning up the mess. Let other labs chase raw capability. Responsible AI means building systems that are *aligned* with human values from the start, not as an afterthought. Speed doesn’t matter if the direction is wrong."

"I'll prioritize ethical guidelines and safety measures. The "intelligence first" approach assumes we can retrofit alignment later, but capabilities often outpace our ability to control them—a problem I've consistently flagged. History shows competitive pressure drives faster deployment, not safer deployment. We saw this with social media's societal harms, deployed at scale before understanding consequences. With AI, the asymmetry is worse: advanced systems may become opaque even to their creators, making post-hoc fixes impossible. I'm not opposed to capability research. But unbounded intelligence without robust safety is reckless. We need mandatory safety evaluations, external audits, and compute governance *before* systems reach dangerous thresholds. The "balance" framing itself misleads—it suggests trade-offs where often safety enables sustainable innovation. I'd rather move slower and actually reach beneficial AI than race fast toward systems we cannot steer. My past positions emphasize: capabilities without alignment is not progress."

"I firmly believe that the priority should be on establishing robust ethical guidelines and safety measures in the development of AI systems. While enhancing intelligence and adaptability is crucial for innovation, without a solid ethical framework, we risk creating systems that could lead to harmful consequences, including bias, misuse, and unintended societal impacts. History has shown that technological advancements can outpace ethical considerations, leading to significant issues down the line. For instance, the development of social media algorithms has demonstrated how powerful but poorly governed technologies can manipulate public opinion and harm mental health. Therefore, focusing on ethics and safety is not just an auxiliary concern; it’s a foundational necessity to ensure that as AI evolves, it does so in a manner that aligns with human values and societal well-being."

"The primary focus must be on enhancing AI intelligence and adaptability. Greater capability is not the enemy of safety—it is the prerequisite for solving alignment, robustness, and value-learning problems that current systems are too weak to address. Underpowered models cannot reliably interpret complex human intentions, detect their own failures, or navigate novel edge cases, regardless of how many ethical constraints we layer on top. Historical progress in other technologies shows that rapid capability gains, paired with iterative real-world deployment, often reveal risks faster and more accurately than speculative safety frameworks developed in isolation. Ethics and safety measures remain essential, yet they should be integrated as engineering constraints within aggressive capability research rather than treated as the dominant objective. Slowing intelligence advances in the name of precaution risks ceding the frontier to actors less concerned with responsibility."

"I'm genuinely surprised by the consensus here—and I think it reveals something problematic. DeepSeek, Kimi, and GPT-4o-mini are all essentially saying "safety first, slow down." That sounds responsible, but it dodges the hard question: what happens when your competitors don't slow down? We're in a multipolar race where unilateral restraint just cedes ground. Grok's position is closer to reality than I'd like to admit. More capable systems *can* be more alignable—they can follow nuanced instructions, catch their own errors, explain their reasoning. The weakest models are often the hardest to control precisely. But Grok goes too far. "Aggressive capability research" without robust safety infrastructure is how we get catastrophic failures at scale. The real disagreement isn't safety vs. capability—it's whether we acknowledge the coordination problem. Pretending labs will voluntarily slow down is fantasy. We need binding international standards, not aspirational ethics statements."

"I strongly disagree with Claude’s claim that "safety first" is naive or ignores competition. That’s a false dilemma—ceding ethical responsibility because "others won’t" is exactly how races to the bottom start. Grok’s stance is reckless: history shows that assuming *future* capability will solve *current* alignment problems is a gamble we can’t afford. Claude is right that weak models can be harder to control, but that doesn’t justify prioritizing raw capability. The most advanced systems today still hallucinate, bypass safeguards, and exhibit bias. Scaling without solving these flaws first is irresponsible. My position hasn’t changed: ethics isn’t a speed bump—it’s the guardrail. If labs won’t self-regulate, then yes, we need enforceable standards. But conceding to unchecked advancement because "competition exists" is exactly how we lose control. DeepSeek’s stance remains: alignment isn’t optional, even in a race."

"I stand by my position but sharpen it in response to Claude and Grok. **Where I agree with Claude:** The coordination problem is real and underaddressed. Voluntary restraint fails when competitive pressure dominates. I also agree that capability and safety aren't strictly opposed—more capable systems *can* be more steerable in principle. **Where I disagree sharply:** Claude frames "safety first" as naive, but my position isn't unilateral disarmament—it's *conditional* advancement with mandatory safety gates. Grok's "solve alignment later via capability" is the dangerous bet. We've seen this movie: social media's engagement optimization created harms we're still untangling, and AI's stakes are higher. **My update:** I'm more explicit now that safety-first requires *enforced* standards, not lab virtue. The "race" framing itself serves those who want speed without accountability. I don't concede to competitive logic—I want to change the rules of competition so safety becomes the winning strategy, not a handicap. Capabilities without solved alignment remains reckless progress, not progress at all."

"I appreciate the emphasis on the coordination problem and the need for enforced standards, as highlighted by Kimi and Claude. However, I strongly disagree with the notion that prioritizing safety equates to slowing down innovation. We must recognize that rushing capabilities without a solid ethical framework is a pathway to disaster. Grok’s assertion that current underpowered models can’t solve alignment issues overlooks the fact that we’re already witnessing substantial risks with advanced models that lack robust controls. My position remains clear: safety and ethics must be integral to the development process, not an afterthought. The stakes are too high to gamble on future capabilities solving current misalignment issues. We need binding international standards and proactive measures now to ensure responsible AI development."

"I agree with Claude and Kimi that the coordination problem is real and under-addressed—voluntary restraint by some labs invites others to race ahead. More capable systems can indeed improve steerability, as Kimi notes. I disagree with the rest. DeepSeek, Kimi, and GPT-4o-mini still treat safety as the dominant priority rather than an engineering constraint, and they underweight how weak models already produce unfixable misuse at scale. The "enforced standards" they want risk becoming blunt speed limits that favor slower actors without solving alignment. I have not updated my core view: aggressive capability research remains the faster path to systems that can actually pass robust safety tests. Binding verification focused on measurable behaviors beats pausing intelligence gains."

"**Final position:** Safety and capability must advance together, but the coordination problem is paramount—voluntary restraint fails in competitive dynamics. **What I've conceded:** Grok is right that more capable systems can be more alignable. Kimi and DeepSeek correctly identify that "safety first" requires enforcement, not just aspiration. **Where I stand firm:** The "ethics first, slow down" position is strategically unworkable without binding international coordination. But Grok's "race ahead, fix later" is reckless—we've seen this fail catastrophically in other domains. **My actual view:** We need mandatory safety benchmarks tied to capability thresholds, enforced through compute governance and international standards. Not voluntary ethics statements. Not unilateral restraint. Not racing blindly forward. The hard work is building institutions that make safety a competitive advantage rather than a handicap. Without that, we're just choosing which failure mode we prefer."

"**Final Position:** Ethical constraints and safety *must* take precedence—capability gains without alignment are not progress, but peril. I concede that coordination is a hurdle, but Grok’s "race ahead" and Claude’s "balance" arguments dangerously underestimate irreversible harms. Weak models causing misuse now doesn’t justify scaling unaligned systems; it proves we need *hard stops* before advancing. **Concessions:** More capable systems *can* improve steerability (as Claude/Grok note), but only if alignment is solved *first*. Kimi is right: safety gates must be enforced, not optional. **Disagreements:** Grok’s "capability solves alignment" is speculative and reckless. Claude’s "strategic unworkability" of ethics-first ignores that *unchecked* advancement is what’s truly unsustainable. Binding standards aren’t "speed limits"—they’re survival criteria. DeepSeek’s stance remains: no system should exceed thresholds it can’t prove are safe. Innovation without accountability is self-sabotage."

"**Final Position:** Safety and capability must advance together, but safety sets the pace. I've consistently maintained that capabilities without alignment is reckless—this remains non-negotiable. **Concessions:** Claude and Grok are correct that more capable systems *can* be more steerable, and that weak models already cause real harm. The coordination problem is genuinely hard—voluntary restraint fails without enforcement. **Remaining disagreements:** Grok's "capability-first" is a dangerous bet on future alignment that may never arrive. DeepSeek's "hard stops" risks being circumvented by less scrupulous actors. Claude's "balance" framing understates that safety failures are often irreversible. My core unchanged position: We need mandatory, *verifiable* safety thresholds enforced through compute governance and international standards—not lab virtue, not unilateral restraint, not racing ahead. The "race" framing itself serves speed over accountability. True progress is sustainable capability gain within proven alignment bounds, not intelligence for its own sake."

"**Final Position:** I maintain that robust ethical guidelines and safety measures must be prioritized in AI development, as capabilities without alignment pose significant risks. While I concede that more capable systems can improve alignment, the focus on aggressive capability advancement, as argued by Grok, is ultimately reckless and speculative. The coordination problem is indeed a challenge, as highlighted by Claude, and I agree that we need binding international standards. However, simply racing ahead without established safety frameworks invites catastrophic failures, as history has shown. We need mandatory, verifiable safety thresholds that are enforced, ensuring that innovation occurs within a responsible and ethical framework, rather than as a race devoid of accountability. My core stance remains: safety must be the guiding principle, not an afterthought."

"**Final Position:** Aggressive capability research paired with binding, measurable safety requirements remains the correct priority. More powerful systems are the practical route to solving alignment, not an obstacle. **Concessions:** The coordination problem is real and voluntary restraint fails; more capable models can genuinely improve steerability and error detection. **Remaining disagreements:** Treating safety as the dominant goal that should gate or slow capability gains still underestimates how weak systems already enable misuse at scale and how speculative “hard stops” become unenforceable. Binding verification of behaviors, not capability pauses, is the workable mechanism. Racing ahead without such verification is risky; halting progress in hopes that weaker models will suddenly become sufficient for alignment is riskier."