Greenland and the limits of Artificial General Intelligence
ChatGPT thinks this Greenland crisis is fake, giving lie to Sequoia Capital’s claim that AGI is here, says Lewis Liu
“AGI is here, now.”
That’s Sequoia Capital this week, one of Silicon Valley’s most legendary venture firms and a major OpenAI investor, declaring we’ve crossed the threshold into artificial general intelligence.
Their post also proclaims, in big bold letters, that they are “Blissfully Unencumbered by the Details”. When Sequoia speaks, the tech world listens. This post has dominated conversations across the AI builder community for days.
As a builder, venture investor, and AI scientist myself, I find their proclamation both deeply useful and profoundly dangerous.
Here’s what’s useful
Sequoia offers a functional definition of AGI: “the ability to figure things out. That’s it”. AI can now crawl through information, determine a path forward, and execute. The key shift, as they frame it, is AI moving from “talking” to “doing”. Harvey and Legora “function as associates,” Juicebox “functions as a recruiter,” OpenEvidence’s Deep Consult “functions as a specialist.” Those are their exact phrases, and I’m skeptical of the framing, but we’ll get to that.
They’re throwing down the gauntlet for builders, and this matters. AI systems can actually redline contracts now, genuinely reach out to prospects today. It’s a reminder to think bigger about what’s possible and that the horizon has expanded dramatically in just the last year.
I sent this post to my own co-founders, not to debate the philosophy, but I wanted to push us on Sequoia’s “doing” versus “talking” framework. We need to rise to this challenge ourselves.
However, calling these systems “AGI” is profoundly damaging, both to the credibility of the AI revolution and to its safe deployment. It obscures what AI agents can actually do today (hint: not general superintelligence) and offers zero guidance on how humans should interact with them (hint: don’t trust them blindly).
Let me give three examples that expose these systems’ real limitations.
AI systems cannot handle out-of-distribution situations
I wrote about this in my last column, but the Greenland crisis offers a live, unfolding example. I tested how GenAI tools, such as ChatGPT 5.2 with maximum “thinking and research” mode, would analyze the situation. Could these supposedly AGI systems help me understand a fast-moving geopolitical event?
They couldn’t even conceive of it happening.
I shared a screenshot of the Wikipedia page documenting the crisis. Every model told me it was fabricated, “bullshit”, impossible. When I pushed further, citing real news sources, ChatGPT kept telling me to “relax” – insisting “this is not a real crisis”. The models are so anchored in the traditional norms of the Western Alliance, the system simply cannot output context that contradicts its training data, even when confronted with primary sources.
AI “thinking” fails when the situation sits outside its training distribution: gaslighting users instead of escalating uncertainty, refuses humility and keeps “reasoning” even when it’s wrong. If policymakers or politicians are using these tools to understand Greenland right now, that’s genuinely dangerous.
AI systems reflect the beliefs of their builders
A Nature article published two weeks ago showed exactly this. Researchers found that LLMs reflected the ideology of their creators. Mainland Chinese models were extremely favorable toward the PRC; Western models were deeply unfavorable.
Even within Western LLMs, the bias is stark. Grok (Elon Musk’s XAI model) showed negative bias toward the EU and multiculturalism, reflecting a right-wing agenda. Google’s Gemini, perceived as more liberal, was more favorable toward both.
The AI community now accepts this as fact: LLMs mirror the ideology of their labs. So how can we trust an AI agent has a “blank slate” to “figure something out”? Especially when it must trawl through vast, nuanced data?
Claims of AGI assume neutrality; or at least it should. The evidence says otherwise.
Deterministic versus non-deterministic systems
GenAI is deeply non-deterministic. The same prompt can produce slightly different outputs or vastly different ones.
Humans intuitively understand what should be fixed versus creative. My shirt size when ordering online? Deterministic. The pattern I choose today? Creative preference. Even the latest, most powerful models blur this line constantly. You’ve seen it: GenAI treating deterministic facts as speculative, creative inputs.
This reveals a critical gap in meta-cognition – awareness of the thinking process itself. Without distinguishing what’s fixed from what’s generative, AI cannot reliably “figure things out”.
So what do we do?
We have tools at our disposal. First, choose narrower use cases where bias and out-of-distribution events are less likely. Second, ensure AI has full, personalised context anchored in reality – not agents running wild without grounding. As I’ve written before, context is king when it comes to AI agents. This also clarifies what should be deterministic versus generative. Third, deploy rules-based filters and observer agents that trigger human review when needed.
Finally, accept a fundamental truth: LLMs will always reflect their training data and their creators’ ideology. LLMs (and their creators) are political actors, whether they want to be or not. As such, AI should be controlled by individual human users, not a system done to humans. Provenance matters – tracing every decision back to a human, no matter how many steps removed, is essential for governance and safety.
At the end of the day, I don’t care what we call it – just not AGI. What we have right now is generationally powerful: AI that can talk and act efficiently within narrow, well-defined frameworks. Guided by critical guardrails, deterministic filters, and human-in-the-loop processes, these systems will add trillions of dollars of value to our global economy.
Call it Artificial Narrow Intelligence. That’s the trillion-dollar opportunity available today.