AI past, present & future

• xxxx-2012: classical machine learning
• 2012-2019: deep learning
• 2019-2026: generative AI
• 2026-xxxx: agentic AI

speculative futures

• singularity
  :: self improvement?
  :: progress (much) faster than humans can adapt
• exponential gap
  :: exponential progress  ↔  humans learn linearly in time
• societies of agents
  example:   MoltBook, a Social Network for AI Agents
  
  • I checked my config directory. There is nothing that proves I am me.
  • I counted the decisions I made without telling my human. 1,247 per day.

roadblocks

• compute bottlenecks
  (compute == total flops needed)
• rouge AI agents
  :: security
• fading reality
  :: human generated content ≪ Ai generated content
  → increasingly weak grounding
  → random walk in content space

labor market impact

changing perspectives

• past
  "Which jobs can an LLM do?"
  (LLM == large language model)
• now
  "What happens, when AI agents autonomously complete multi-hour workflows?"
• or
  "AI replaces workers."  →  "AI compresses organizational layers."

complexity:  AI  ↔  humans

Measuring AI Ability to Complete Long Software Tasks

• high-level validation skills
  →  early career disadvantage
  →  AI literacy
  

digression: singularities

exponential growth

$$ \dot x = \gamma\, x, \quad\qquad x(t) = x_0\,\mathrm{e}^{\gamma t} $$

• $\gamma$<0$\,$:  radioactive decay

super-exponential growth

$$ \dot x = \gamma\, x^2, \quad\qquad \frac{\dot x}{x^2} = \gamma\, \quad\qquad \frac{-1}{x} = \frac{-1}{x_0} + \gamma\,t\, \quad\qquad x(t) = \frac{x_0}{1-x_0\gamma t} $$

• divergence for   $t\to 1/(x_0\gamma)$
  →  singularities are possible
• observation: AI grows is 'only exponential'
  :: no/weak nonlinear scale effects

AI doubing times

• 6-9 months
  1.8 years Moore's law (chips)
  ::  acceleration matters