Mark Worrall (@maw501) 's Twitter Profile
Mark Worrall

@maw501

Founder. Building a science-backed online learning platform.

Musings on AI, Knowledge & Education.

ID: 466806719

linkhttps://nodeledge.ai/ calendar_today17-01-2012 20:23:47

519 Tweet

279 Followers

498 Following

Mark Worrall (@maw501) 's Twitter Profile Photo

This feels like a peer-reviewed advert for what I've been building. If you're acquiring a new skill - genuinely learning something you don't yet understand - AI assistance is NOT neutral. It's actively working against you unless you fight to stay engaged. This is the

Mark Worrall (@maw501) 's Twitter Profile Photo

Should your teenager learn to code? I think the answer is an emphatic yes. Here’s why. Your kid has exactly one lever to pull in this world: their own knowledge and skills. Nothing else compounds over time. Nothing else transfers between jobs, industries, or whatever tools come

Should your teenager learn to code? I think the answer is an emphatic yes. Here’s why.

Your kid has exactly one lever to pull in this world: their own knowledge and skills. Nothing else compounds over time. Nothing else transfers between jobs, industries, or whatever tools come
Mark Worrall (@maw501) 's Twitter Profile Photo

My A-level maths teacher used to teach by saying "open the book, do the odd questions". I'm ashamed to say it took me years to realise that wasn't intellectual freedom or treating us like grown-ups. It was sink or swim dressed up as sophistication. Setting a high bar is cheap.

Mark Worrall (@maw501) 's Twitter Profile Photo

AI bike-shedding: debating whether Opus 4.6 or Codex 5.3 is the best coding model vs. actually getting good at using them. My social feed is flooded and it’s exhausting. The irony is that everyone forgets that when a better model drops, everyone gets it too. There’s no moat in

AI bike-shedding: debating whether Opus 4.6 or Codex 5.3 is the best coding model vs. actually getting good at using them.

My social feed is flooded and it’s exhausting. The irony is that everyone forgets that when a better model drops, everyone gets it too. There’s no moat in
Mark Worrall (@maw501) 's Twitter Profile Photo

AI providers are busy compressing humanity's knowledge into model weights. To some extent, I think of what I'm doing at Nodeledge as pulling the knowledge back out, and carefully structuring it. This involves making all prerequisite relationships explicit, creating clear,

Mark Worrall (@maw501) 's Twitter Profile Photo

AI doesn't shrink the software problem, it expands it. More capability means more ambition means more novel problems. It's demand-side, not supply-side. Claiming we can "solve software" misunderstands what software is. Software encodes human knowledge and intent into systems

Mark Worrall (@maw501) 's Twitter Profile Photo

The managing partner in this story gets value from AI precisely because he has the expertise to direct it and catch its mistakes. Take that away and you're generating output you can't verify. Whilst I enjoyed the post I think it gets (at least) two things fundamentally wrong:

Mark Worrall (@maw501) 's Twitter Profile Photo

You do not have to learn anything. But it may harm your future if you cannot recall, unaided, what you later claim to understand. Anything you do learn may be the only thing they can't automate. I’m only partly joking, but it’s absolutely true that the wave of disruption that's

You do not have to learn anything. But it may harm your future if you cannot recall, unaided, what you later claim to understand. Anything you do learn may be the only thing they can't automate.

I’m only partly joking, but it’s absolutely true that the wave of disruption that's
Mark Worrall (@maw501) 's Twitter Profile Photo

AI in education goes wrong when it does the thinking the student is supposed to be doing. That's the only part worth worrying about. Everything adjacent to that - setting context, adjusting difficulty, feedback, tracking progress - is fair game. This is not off-loading learning,

Mark Worrall (@maw501) 's Twitter Profile Photo

Better prompts get better results. Obviously. But prompt engineering has a ceiling, and that ceiling is set by what you already know. There's actually good reasons for this: and key to it is that experts and novices literally perceive problems differently. There's decades

Mark Worrall (@maw501) 's Twitter Profile Photo

$10M on the table. All you have to do is use AI to derive a new result in theoretical physics. You get the best model in the world and six months. Where would you even start? This isn't hypothetical, by the way. GPT-5.2 did help derive a new result in theoretical physics. So

$10M on the table. All you have to do is use AI to derive a new result in theoretical physics. You get the best model in the world and six months.

Where would you even start?

This isn't hypothetical, by the way. GPT-5.2 did help derive a new result in theoretical physics. So
Mark Worrall (@maw501) 's Twitter Profile Photo

Translation: our benchmark is not valid, and 50% success rate would get a human fired in under a day / week, but we'll still publish and feed into the hype.