We got @OncourseAI to 80k+ signups as a small bootstrapped team, and I built @caddie_app in a week on the side.
So we wrote down the 15 principles we actually work by with agents
My favourite: keep at least five agents running, day and night.
usecaddie.app/blog/agents
Loved this write up about my all time favorite movie substack.com/@pranatik/note/…
1
57
The point was never to study more
getoncourse.ai/blog/how-can-…
70
No question to go unanswered ... ... ... getoncourse.ai/blog/how-shou…
62
The notion of *unearned wisdom* 👌 good to find a name for something we see often but were unable to call it …
Ever wonder why ChatGPT can write a sonnet about quantum physics but sometimes fails at counting letters? There's a fascinating psychological explanation – and it starts with a warning Carl Jung wrote decades ago.
You know how you can't just read about riding a bike and suddenly be good at it? You have to fall a few times. Build muscle memory. Go from wobbly to confident.
Jung called knowledge without that struggle "unearned wisdom." And he warned: beware of it.
Here's what's wild: this 20th-century insight might explain one of AI's biggest unsolved problems.
(Stick with me – this gets interesting.)
Think about how you actually learned math. First addition, then multiplication, then algebra. Each builds on the last. You can't skip steps.
Your brain literally organized itself differently because of that sequence. It created what researchers call "Unified Factored Representations" – modular, adaptable knowledge.
Now imagine learning addition, calculus, and differential equations all at the same exact moment.
Sounds ridiculous, right?
That's basically how we train AI.
Modern language models are "epistemological vacuum cleaners" (love that phrase). They ingest millions of documents simultaneously – nursery rhymes alongside PhD dissertations, basic arithmetic alongside advanced proofs.
Everything flattened into one massive parallel gulp.
Here's where it gets problematic:
When you learn naturally, your knowledge is modular. You can update your understanding of, say, cooking techniques without forgetting how to ride a bike.
But when AI learns everything at once? It creates what researchers call "Fractured Entangled Representations."
Imagine dropping all the buildings of a city from the sky at once instead of building them street by street.
Sure, you have a city. But try to renovate one building? The whole thing might collapse. Everything's tangled with everything else.
That's the AI's knowledge structure.
This explains SO much:
Why AI hallucinates facts it should "know"
Why updating one capability can break another (catastrophic forgetting)
Why these systems can't do genuine compositional reasoning
Why they're bad at truly novel problems
The knowledge was never earned. It was dumped.
The hidden cost is what researchers call "destroyed evolvability."
You can't easily improve these systems without retraining from scratch. The representations aren't modular enough. Touch one thing, ripple effects everywhere.
It's intellectual spaghetti code.
Think about a master carpenter. She didn't read every woodworking book at once. She:
Learned basic cuts
Struggled with joints
Felt wood grain
Made mistakes
Developed intuition
Her knowledge compresses elegantly. Yours does too.
AI's knowledge? Bloated and brittle.
Here's the part that keeps me up at night:
We're building systems with vast knowledge but no foundation. They're like buildings with impressive facades but no load-bearing walls.
Jung intuited this – knowledge without the struggle of earning it lacks depth.
But wait – this isn't just philosophical musing.
Some researchers are exploring solutions:
Curriculum learning (carefully sequenced training)
Adversarial methods (creating productive struggle)
Meta-learning (optimizing for future adaptability)
Teaching AI like we'd teach a child, not filling a database.
Next time you interact with AI, notice:
When does it feel shallow vs. insightful? When does it confidently claim something false? When can't it adapt a concept to a new context?
You're probably seeing the fingerprints of unearned knowledge.
The deeper question: Can we create artificial wisdom, or just artificial knowledge?
Wisdom requires synthesis, struggle, developmental stages. You can't speedrun it.
Makes you wonder if we're optimizing for the wrong thing entirely.
The price of unearned knowledge isn't what the AI fails to learn.
It's what it becomes incapable of learning in the future.
Jung saw this in human psychology. We're rediscovering it in machine learning.
Some lessons, it turns out, are timeless.
85
prasad kompalli retweeted
Why are we pretending that the pinnacle of AI interaction is typing questions and getting text back?
We are pattern-matching, visual-processing, spatial-reasoning creatures who happened to develop language as a useful hack. We think in relationships and patterns, yet currently we're forced to translate everything into text requests and then back into understanding.
Read more on my latest blog: Re-imagining AI Interfaces
good read on UI for AI world : open.substack.com/pub/somest…
@harshaog is hinting at whats coming in @OncourseAI (-:
3
82
great to see the nextgen win the match. Finally a team thats grounded, hungry for learning and focussed to perform and with no one who feels they are above the game itself... #INDvsENG2025
1
1
143
Nice to have watched this live yesterday w/ kids together ( and couldn’t resist following up watching the iconic docking scene - ) #docking #indiainapace #isro #AxiomMission4
192
Tech head tells me “last week Devin gave max PRs and the feature change you asked was completely done by Devin”
Amazed at how fast and far AI for coding has come and what it lies ahead for software dev.. @getOncourseAI
#AI #AIforcoding #genAI
2
83
A small observation : more than solving HL math/physics/coding problems, I find getting LLMs to 'formulate' good set of solvable problems in a given topic ( algebra, geometry ... ) is a challenge. LLMs should be benchmarked in this. #GenAI #LLMbenchmarks
2
3
160
Its been super exciting , ...building cool new learning experiences ..!
This quoted post is unavailable.
prasad kompalli retweeted
Are you a fresher who wants to help sculpt the future of Learning with AI? Let's chat. We're hiring a Software Engineering & a Product Intern.
prasad kompalli retweeted
Want to help sculpt the future of Learning with AI? Let's Chat.
We're hiring for Full-Stack/Front-end/Back-End Engineers.
Watched live #Chandrayaan3 along with the kids. - all feeling super proud and inspired
2
390
One more publication and one more reason to be proud of Team @MFineCare @IndanaRaja wjarr.com/content/efficacy-d…









