The Surprise Over Delayed AGI

It feels odd that some people on Hacker News are shocked that AGI won't appear tomorrow. A few observations you might expand on later:

• Current LLMs still hallucinate. They produce convincing but incorrect information far too often.
• Robust reasoning is missing. Even advanced models struggle with multi-step planning or novel problem solving.
• Hardware remains a bottleneck. Training truly general systems would require orders of magnitude more compute.
• Alignment is unsolved. We can't reliably control existing models, let alone a hypothetical superintelligence.

These points make it clear that AGI isn't around the corner. Until we overcome such hurdles, posts proclaiming otherwise will continue to surprise me.

In reality, progress is steady but still narrowly focused. Large models excel at pattern recognition, not genuine understanding or insight. Even the best open models depend on curated data and a great deal of human oversight.

Why Say It's Close?

Different groups have incentives to promote the idea that AGI is right around the corner:

• Startups and investors seek buzz to attract funding and top talent.
• Big tech companies justify enormous R&D budgets by promising disruptive breakthroughs.
• Media outlets favor sensational headlines that drive traffic.
• Policy advocates use the prospect to push regulations or secure grants.

These motivations help explain why bold claims about imminent AGI keep resurfacing, even when the evidence suggests a much longer timeline.

What Would Meaningful AGI Look Like?

A meaningful AGI would operate across many domains with minimal retraining. Such a system would feature:

• General learning – quickly mastering new tasks from limited examples.
• Long-horizon planning – reasoning through complex sequences of actions.
• Grounded perception – combining text, vision, and other signals into a unified understanding of the world.
• Robust alignment – reliably acting in line with human intent and values.

We're far from achieving these capabilities. Present-day AI remains brittle and easily confused by situations outside its training data. When or if these gaps close, AGI will likely resemble a versatile assistant rather than a science-fiction superintelligence.