it's the same pattern we saw with calculators, spell-check, and now AI

tools are only as good as the person wielding them. if you can't think critically without them, you're just amplifying mediocrity at scale

I've seen this firsthand: two people using the same AI, one creates something remarkable, the other produces generic slop. the difference is always the human's taste, judgment, and ability to push back on bad outputs

@itsreallyvivek The failure mode isn't using the model, it's outsourcing the first pass of judgment. Once you've read the primary source the model becomes a sparring partner instead of an oracle. Hard to keep that discipline when the summary is one keystroke away.

i first ask the model for the landscape and then i imagine how it should be done, then let the model write it and i made comments on the code until im happy with it. the i ask model to formally verify each path and i check its proofs. this is much slower and requires effort, but am faster in the end since it actually works the first time. not all paths require this tho, some are better left vibecoded. especially when you aren't sure if to build it or how, then obviously get it out there first and go back and rewrite when sure it's something to keep in.

@itsreallyvivek everyone nods at this and then opens the summary anyway, because the primary source is 40 pages and the summary is one tweet. knowing the rule was never the hard part, sitting with the 40 pages is

@itsreallyvivek the actual thing' is already a shrinking island. search results, docs, the news in your feed are mostly model-chewed before you reach them. the willpower to read the raw source first matters less every year than whether a raw source is still there to read.

@itsreallyvivek In other words, "Acquire domain knowledge."

And yeah, he's right.

@itsreallyvivek i like to use AI for selecting materials worth to read. then i will check the paper / blogpost myself to get fully engaged.

@itsreallyvivek this!! it might feel satisfactory to swiftly "learn" a lot of things using AI, but it mostly results in a shallow understanding of each topic

trying to maintain a consistent deep flow while doing something is what makes one virtuoso in the long run imho

@itsreallyvivek I think this advice held true even before LLMs. Distilled and compressed summary blogs/videos have always existed. If you read only those then the ceiling of your understanding is going to be that of the author of the summary, not of the primary material.

@itsreallyvivek It's the exact opposite.

The people stacking and composing agents are unlocking things that didn't exist as a category.

@itsreallyvivek When I read your post, I suddenly realized that a new education system is essentially needed. The old one was training cogs for the industrial revolution. People of this caliber in the future system will be the raw material, the resource for AI👇

@itsreallyvivek Tldr grok summarize

@itsreallyvivek Amazing advice

@itsreallyvivek 'the actual thing' is already a shrinking island. search results, docs, the news in your feed are mostly model-chewed before you reach them. the willpower to read the raw source first matters less every year than whether a raw source is still there to read.

@itsreallyvivek The real risk with AI was never replacement, but cognitive passivity. The people who’ll thrive are probably the ones using AI to extend their thinking, not outsource it

@itsreallyvivek can outsource bias so don’t do it

@itsreallyvivek Lately I feel anxious starting 20h long books but it’s worth it every time

@itsreallyvivek Begin by orienting from first principles.

Once you have, add external data.

This sequence allows you to think more clearly.

@itsreallyvivek @granawkins It’s gonna be so hard when the agent becomes the default interface for knowledge

@itsreallyvivek Solid advice that most of us will read and then immediately paste into ChatGPT asking it to summarize lol