Love a blog post that just plops a bunch of quotes and then finishes with "kinda makes ya think, huh?" without any further exploration.
The comments I've perused that resonate with me the most center around: "for what purpose". Why does the person want to know. What will they do with this information. At what level of depth will they feel as if they now "understand" the concept.
A half-decent analogy is resolution. Sure, you can express the concept of a mountain with a 32x32 set of pixels, but what is the purpose of this "picture". Who's going to be looking at it and for what purpose. If a 9 year old asks you how the heart works, you can just say its a pump that drives blood through your circulatory system. But this is a woefully inadequate explanation for a medical student. At some point, reducing the resolution (reliance on already understood concepts/terminology) becomes an actual impediment to expressing the idea in a way that is useful to anyone at all (a 2x2 pixel arragnement).
There is something to be said about knowing your audience, communicating what is important to them, while also acknowledging that the nuance in any field is in the details and glossing over them or using non-academic terminology is necessarily lossy.
This is true of a lot of things. Logical thinking is a skill that can be honed and is broadly applicable. But for lots of fields (eg medicine) you just need to learn a lot of foundational things before you can reasonably get to the meat of understanding the “system”. You just need to memorize terms, concepts, and the relationships between them.
Agreed, and I’d go further: logic does transfer, but it needs knowledge to work on, so knowledge is the bottleneck in every field. The need for knowledge is just more obvious for technical fields (medicine, software, etc). In the context of American K12 education, you can't understand the underpinnings of democracy without understanding other forms of government. You can't understand the forms of government without understanding the basics of human nature and human societies. That knowledge comes from learning history and the liberal arts. If you look at the US Common Core, it does not specify what knowledge should be learned at what grade level (what books to read, what history to study). It focuses on general skills and leaves content to states and local districts. This started with the progressive education movement decades earlier and it has created large gaps in knowledge that the overlay of logical thinking can't operate on because they don't exist in the minds of large portions of the population. And it’s not about intelligence. Kids from educated families pick this knowledge up at home, and everyone else depends on school to provide it.
The framing of the review first goes over different models of teaching (including the kind you’re pushing here) and compares it with other models (including the kind[s] you’re supposedly criticizing). I found this framing helpful in asking the fundamental questions: what is the point of school, and how can a government tasked with educating the masses deliver.
This is another one of those cases where I’m simply tired of hearing the same arguments from people who either didn’t have a good time in school and/or were smarter than average and were bored/stilted as a result. The system is designed (not perfectly!) to serve the masses. If it didn’t serve you, it doesn’t mean it was broken (or designed to produce worker drones).
If I'm reading this right, they ran open source models through each harness (Deepseek, GLM etc) and then compared results. It's my understanding that Claude and Codex are highly optimized for their own models: system prompt, tooling, etc etc. So shoehorning in another model that they weren't designed for feels like setting them up for failure, especially compared to OpenCode (optimized for general use with [mostly?] open source models AND Deepseek's own harness running...it's own model! Doesn't seem like a well-set-up test.
This one took a turn but point taken. You see similar arguments for a great many ailments (sperm count and carpets comes to mind). You can’t well run randomized control trials on every plastic, additive, or manufacturing process. It feels futile to even start.
It depends on the subject. For classes/subjects that are very terminology-heavy, with a lot of dependencies between topics, memorizing things is a pre-requisite to the "deeper" understanding, especially under time-constraint of a semester or whatever. I think most people would agree that it's ideal to "naturally" explore a subject in order to get a deeper understanding of it, but one can't always be expected to 1. Be able to deploy that "curiosity" for all subjects at any time and 2. Be able to adequately achieve competence in a subject within the given timeframe
Codex was running a 2x usage promotion from around the time when Claude introduced rate limiting during peak hours, until May 31st. The various relevant subreddits were (more) insufferable: just 1000 posts per day to the tune of "Just switched the codex! So much more usage!" only to have that tone flip immediately after the promo ran out.
Just to add to this: Marimo notebooks are set up in a way that ends up being easier/more-token-efficient for agents to work in. They're just python files with some custom formatting to be read by their front-end UI, sort of like cell notation in (# %%) but with a lot of QoL bells and whistles built in. Jupyter notebooks in comparison, are mostly json and are pretty messy to work with as far as git diffs goes, which is annoying. VS-Code's Github Copilot actually had pretty good Jupyter integration, letting you refer to specific cells, and allowed the agent to run individual cells and read their output before writing subsequent cells etc etc. I suspect there are vs-code specific tool calls for jupyter notebooks being used that I haven't bothered trying to replicate with claude.
A month or two ago, Marimo released Marimo Pair, which works about as well as that. It appears to be a skill that just points claude code (perhaps other harnesses, haven't tried) at whatever localhost the Marimo UI is worked off of, and tells it the structure of a Marimo notebook (very necessary since I imagine there isn't as much training data on them compared to Jupyter etc). https://marimo.io/blog/marimo-pair
The comments I've perused that resonate with me the most center around: "for what purpose". Why does the person want to know. What will they do with this information. At what level of depth will they feel as if they now "understand" the concept.
A half-decent analogy is resolution. Sure, you can express the concept of a mountain with a 32x32 set of pixels, but what is the purpose of this "picture". Who's going to be looking at it and for what purpose. If a 9 year old asks you how the heart works, you can just say its a pump that drives blood through your circulatory system. But this is a woefully inadequate explanation for a medical student. At some point, reducing the resolution (reliance on already understood concepts/terminology) becomes an actual impediment to expressing the idea in a way that is useful to anyone at all (a 2x2 pixel arragnement).
There is something to be said about knowing your audience, communicating what is important to them, while also acknowledging that the nuance in any field is in the details and glossing over them or using non-academic terminology is necessarily lossy.
reply