back to writing

the white-collar bloodbath

I love the dwarkesh pod. it's consistently one of the best sources of calibrated thinking across ai, technology, economics, history and future - it really is only dwarkesh and lex who do it this well.

the 'continual learning bottleneck'

recently, dwarkesh has been sharing his case for why he thinks 'agi isn't right around the corner'.

his case is based largely on the lack of 'continual learning' in current models - ie. they need a continual learning loop (or, as gwern puts it, a 'day-dreaming' loop) - basically, the ability to learn from experience without explicit retraining.

it's a real limitation.

some people think we'll solve it quickly and continue scaling; others think we'll need fundamentally new ideas and breakthroughs. it's an open and interesting debate.

the 25% claim

this 'continual learning deficiency' also suggests that economic impact - specifically 'white-collar-work' displacement, might not be as severe as once expected.

when dwarkesh interviewed sholto and trenton from anthropic, they suggested that current ai is already sufficient to automate significant white collar work.

dwarkesh has pushed back against this claim, estimating it would be more like 25%.

he shares his own experiences attempting to apply ai within his podcast production - explaining he still needs humans and human judgment throughout the entire process, from post production to transcripts.

separately, as he puts it - "ai can't even do my taxes".

incorrect framing

i think this framing fails to capture the full picture and drastically underestimates real white collar job displacement.

if we're talking about completely automating white-collar jobs, one-to-one, with current systems, then yes - 25% seems a reasonable estimation.

but if we also consider augmentation and scaffolding, the estimate should be much higher.

augmentation and the multiplication effect

it's becoming increasingly clear that with the right context, humans can do exponentially more work with ai.

mit research proves workers complete tasks 37-40% faster with 18% higher quality using chatgpt.

it's obviously a jagged and messy frontier - with plenty of examples suggesting the augmentation effect cannot be applied uniformly. recent studies on developer productivity show mixed results, but this is more a matter of misuse than model efficiency.

let's apply this to white-collar work.

i don't think it's outlandish to claim that a single ai-native 'generalist' employee or contractor, with the right context and data, could hypothetically do the same job as an entire department or team.

if you've used the most recent frontier models and their integrations, and ever worked with a marketing agency or software company, you know that way more than 25% of positions are vulnerable.

marketing coordinators, content creators and copy writers, administrative assistants.

these roles certainly require degrees of continual learning and nuance and can't be replaced one-to-one by ai. but a single ai-native employee or contractor, with the right context and data, could now handle a significant chunk of this work - maybe the work is subpar, but sufficient and far more scalable.

scaffolding: pulling capabilities forward

in the conversation, dwarkesh also uses the example, "ai can't do my taxes".

it's true, we're still aways from 'personal assistant escape velocity'. you wouldn't hand over end-to-end tax submission to chatgpt, yet.

but this is where 'scaffolding' comes in.

if there's enough economic incentive, companies will build scaffolding around existing models enabling users to complete tasks - like ai-powered tax preparation.

there are currently tens of thousands of engineers building this scaffolding - 'pulling model capabilities forward' through customized solutions whether it's thin wrappers over existing tools, fine-tuned models on custom datasets, or complex agentic workflows.

often, these solutions aren't 'smarter ai replacing dumber humans', it's just a data and scale game.

if you have a successful podcast and a small, efficient team - you're better off keeping those humans and ensuring they're using ai.

but these is not the reality - for most businesses powered by white-collar work, ai and automation will soon become non-negotiable.

the extreme example: klarna's support bot handling 2.3 million conversations, doing the work of 700 full-time staff members, resolving issues in under two minutes.

ibm is pausing hiring for 7,800 back-office positions they expect ai and automation to handle. that's 30% of their back-office workforce over five years.

the harsh reality, is most organizations operate under the "garbage can model"—chaotic, inefficient systems where problems, solutions, and decision-makers collide randomly. most white-collar workers spend their days navigating bureaucratic inefficiencies, redundant processes, and information silos.

as ethan mollick notes,

even if we substract the bitter lesson, and assume model progress plateu, there is still massive opportunity for a single ai-native entity to come in and completely overhaul things.

the white-collar bloodbath

white collar work makes up basically half of the world's workforce - billions of people.

if we're talking about completely automating white-collar jobs with current systems, then yes - 25% seems reasonable.

but considering augmentation and scaffolding, the estimate should be much higher.

immediately susceptible to augmentation (40-50% of white-collar roles):

  • administrative and office support: 46% of tasks automatable
  • marketing coordinators and content creators
  • customer service representatives
  • data analysts and researchers
  • basic legal and financial tasks

immediately susceptible to scaffolding (30-40% of white-collar roles):

  • bookkeeping and accounting: custom software already handles 80%+ of tax preparation
  • hr and recruitment: automated screening, scheduling, basic employee queries
  • sales coordination and crm management
  • basic project management and scheduling

combined, we're looking at 60-70% of white-collar work being significantly impacted.

obviously, it's hard to make claims about which specific roles are fully automatable versus susceptible to augmentation or scaffolding; there's a world of nuance.

not to mention, this is the 'if ai progress completely plateaus' scenario (there's a strong case we continue improving), and we're not even considering potential economic incentives or geopolitical pressures.

i think it's fair to claim that far more than 25% of white-collar work is in danger within the next 3-5 years.

if it turns out that less than 25% is displaced, it likely had more to do with resistance, political pressure and slow adoption than any limitations posed by the 'continual learning bottleneck'.

References