The article <a href="https://blog.plover.com/2026/02/05/#micro-worlds" author="Mark Dominus" source="The Universe of Discourse">John Haugeland on the failure of micro-worlds</a> describes a very early experiment in natural-language processing from the late 1960s called <a href="https://en.wikipedia.org/wiki/SHRDLU" author="" source="Wikipedia">SHRDLU</a>. <bq>SHRDLU could handle this too, although I think its mechanism was different: it would interact with the separate blocks world subsystem and ⸢actually⸣ try to put the block on the pyramid; the simulated physics would simulate the block falling off the pyramid, and SHRDLU would discover that its stacking attempt had been unsuccessful. With Claude, something very different is happening; there is no physics simulation separate from Claude. <b>I think the answer here demonstrates that Claude's own model includes something about pyramids and something about physics</b></bq> Does it though? Where would it have acquired this model? Why would it suddenly be modeling physical laws unless some layers surrounding the text generator had been bolted on? As an engineer, I would love to know how much of what goes into these answer is actually located somewhere in calculation units that have nothing to do with a transformer-based, attention-enhanced LLM. If it's the LLM doing it, then I don't know which part of its architecture the answer comes from. I don't see the mechanism because, so far, we've managed to explain a tremendous amount of its "behavior" (responses) with statistics. Is there a reason to have stopped assuming that this is the mechanism underlying the supposedly improved "understanding"? What is the reason we're rounding up <i>now</i> as opposed to two years ago? As far as I know, we're just throwing more horsepower at these tools but haven't significantly changed the architecture that would lead us to believe that a "world model" is now governing token choice rather than "statistics". I might have missed something, though; my attention wanders. <bq>Are there are any people who are still saying “it's not artifical intelligence, it's just a Large Language Model”. I suppose probably.</bq> Well buddy, I don't spend any time talking to these things, so I admit that my thinking kind of got stuck at that stage. I think that other people starting rounding up to "this is intelligence" because they started having too much fun with it and they didn't want it to look like they were just playing video games. So, instead of talking about the mechanisms that go into these models---if they're at all different from what we presented a few years back---they talk about how it "seems intelligent" ... and how everyone who doesn't think this is a benighted fool, if not a heretic. I can't help but recognize that this mechanism is the same as the one employed by any true believer in any other faith. First, you're "born again". And then you start to castigate anyone who isn't. Classic MLM tactics. Human psychology is utterly banal. <bq>But as a “Large Language Model”, Claude <b>necessarily includes a model of the world in general</b>, something that has long been recognized as an enormous prerequisite for artificial intelligence. Five years ago a general world model was science fiction. Now we have something that can plausibly be considered an example.</bq> There it is again: the author is "rounding up" quite significantly because he doesn't have another explanation for what he thinks he's seeing. Does the LLM have a model of the world encoded within its statistical matrices? That's quite a claim, seemingly belied by the many, many times that it gets things wildly wrong. Is it that it has a model of the world but is kind of dumb sometimes, like a child? If it had a model of the world, then why is <i>context</i> so important to keep it on the primrose path? What is the theory here? Is it that the author wants it desperately to be more than it is? Would he marry it? Invite it to dinner? Watch a movie with it? Maybe. <bq>And second: maybe this isn't “artifical intelligence” (whatever that means) and maybe it is. But <b>it does the things I wanted artificial intelligence to do</b></bq> You've found a tool. You're happy with its functionality. Good for you. I have completely different expectations and quickly grow bored because there are only so many hours in a day and I am not in any way attracted to spending any of them talking to a chatbot. There is too much of actual human culture and art out there for me to bother with artificially generated content. Was there too little of everything before? Did we not have enough books or movies? Are there not enough people to talk to? What are we even talking about here? <img src="{att_link}robot_army.webp" href="{att_link}robot_army.webp" align="none" title="Robot army" scale="75%">