Im not saying LLMs are not impressive or useful — Im pointing out that corporations behind commercial AI models are capitalising on our emotional response to natural language prediction. This phenomenon isnt new – Weizenbaum observed it 60 years ago, even with the simplest of algorithms like ELIZA.

Your example actually highlights this well. AI excels at language, so it’s naturally strong in teaching (especially for language learning ;)). But coding is different. It’s not just about syntax; it requires problem-solving, debugging, and system design — areas where AI struggles because it lacks true reasoning.

There’s no denying that when AI helps you achieve or learn something new, it’s a fascinating moment — proof that we’re living in 2025, not 1967. But the more commercialised it gets, the more mythical and misleading the narrative becomes

> system design — areas where AI struggles because it lacks true reasoning.

Others addressed code, but with system design specifically - this is more of an engineering field now, in that there's established patterns, a set of components at various levels of abstraction, and a fuck ton of material about how to do it, including but not limited to everything FAANG publishes as preparatory material for their System Design interviews. At this point in time, we have both a good theoretical framework and a large collection of "design patterns" solving common problems. The need for advanced reasoning is limited, and almost no one is facing unique problems here.

I've tested it recently, and suffice it to say, Claude 3.7 Sonnet can design systems just fine - in fact much better than I'd expect a random senior engineer to. Having the breadth of knowledge and being really good at fitting patterns is a big advantage it has over people.

You originally said

> They push the narrative that they’ve created something akin to human cognition

I am saying they're not doing that, they're doing sales and marketing and it's you that interprets this as possible/true. In my analogy if the company said it's a hammer that can do anything, you wouldn't use it to debug elixir. You understand what hammers are for and you realize the scope is different. Same here. It's a tool that has its uses and limits.

> Your example actually highlights this well. AI excels at language, so it’s naturally strong in teaching (especially for language learning ;)). But coding is different. It’s not just about syntax; it requires problem-solving, debugging, and system design — areas where AI struggles because it lacks true reasoning.

I disagree since I use it daily and Claude is really good at coding. It's saving me a lot of time. It's not gonna build a new Waymo but I don't expect it to. But this is besides the point. In the original tweet what Sabine is implying is that it's useless and OpenAI should be worth less than a shoe factory. When in fact this is a very poor approach to look at LLMs and their value and both sides of the spectrum are problematic (those that say it's a catch all AGI and those that say hurr it couldn't solve P versus NP it's trash).

I think one difference between a hammer and an LLM is that hammers have existed since forever, so common sense is assumed to be there as to what their purpose is. For LLMs though, people are still discovering on a daily basis to what extent they can usefully apply them, so it's much easier to take such promises made by companies out of context if you are not knowledgeable/educated on LLMs and their limitations.

>they're doing sales and marketing and it's you that interprets this as possible/true.

You've moved the goalpost from "they're not saying it" to "they're saying, but you're not supposed to believe it."

The companies are not doing it. This is what I am saying.

You admitted earlier that they are:

Person you replied to: they intentionally use suggestive language that leads people to think AI is approaching human cognition. This helps with hype, investment, and PR.

Your response: As do all companies in the world. If you want to buy a hammer, the company will sell it as the best hammer in the world. It's the norm.

As a programmer (and GOFAI buff) for 60 years who was initially highly critical of the notion of LLMs being able to write code because they have no mental states, I have been amazed by the latest incarnations being able to write complex functioning code in many cases. There are, however, specific ways that not being reasoners is evident ... e.g., they tend to overengineer because they fail to understand that many situations aren't possible. I recently had an example where one node in a tree was being merged into another, resulting in the child list of the absorbed node being added to the child list of the kept node. Without explicit guidance, the LLM didn't "understand" (that is, its response did not reflect) that a child node can only have one parent so collisions weren't possible.

> proof that we’re living in 2025, not 1967. But the more commercialised it gets, the more mythical and misleading the narrative becomes

You seem to be living in 2024, or 2023. People generally have far more pragmatic expectations these days, and the companies are doing a lot less overselling ... in part because it's harder to come up with hype that exceeds the actual performance of these systems.