Why I don’t think AGI is right around the corner

Current LLM Capabilities and Limitations 00:00

  • There's debate about AGI timelines, with estimates ranging from 2 to 20 years
  • The speaker argues that current LLMs, while impressive, are not causing economic transformations on par with the internet
  • The real challenge is that LLMs struggle to perform humanlike labor due to fundamental capability gaps
  • Attempts to use LLMs for tasks like transcript rewriting or co-writing essays have yielded only partial success
  • LLMs lack the ability to improve over time through continual learning, unlike human workers who adapt, refine skills, and learn from feedback
  • Feedback mechanisms for LLMs (like prompt engineering or RL fine-tuning) do not produce the same learning and improvement seen in humans
  • Human usefulness comes from building context, learning from mistakes, and incremental improvement, processes not mirrored in current LLMs

The Bottleneck of Continual Learning 02:09

  • Teaching a human is an iterative, experience-based process; LLMs can only be "taught" through repeated instructions, which is ineffective for complex skills
  • Current reinforcement learning (RL) techniques are not adaptive and personalized enough to replace on-the-job learning in humans
  • Human editors improve by noticing nuances and iterating based on audience feedback; LLMs miss out on this tacit knowledge accumulation
  • The idea of LLMs developing their own RL environments to practice weaknesses is conceivable but far from reality and difficult to generalize

Context Window and Memory Challenges 04:09

  • LLMs can sometimes improve within a single session, but this learning is lost once the context window resets
  • Techniques like rolling context summaries (e.g., Cloud Code) may work for text-based fields, but are brittle elsewhere—important learnings are often omitted from summaries
  • Without the ability to retain and build on long-term experiences, LLMs can't perform as reliably as humans in evolving workflows

Economic Impact and Automation Limits 05:30

  • Arguments that current AIs will automate most white-collar jobs are overly optimistic
  • Without the ability to personalize and improve, current AIs could automate less than 25% of white-collar jobs if progress stopped today
  • Many subtasks will be automated, but the inability to accumulate context or persistent learning will prevent AIs from functioning as full employees
  • The speaker is optimistic about AI in the coming decades, especially if continual learning is solved—this would cause a step-change in model value

Potential for Future AI Progress 06:53

  • Even absent further algorithmic breakthroughs, AIs capable of on-the-job learning and knowledge sharing across instances could rapidly approach superintelligence
  • The path to continual learning is expected to be incremental, with imperfect versions appearing before humanlike learning is achieved

Skepticism Toward Near-Term AI Milestones 08:08

  • Some researchers predict highly capable autonomous agents (e.g., doing complex tax prep) by the end of next year, but the speaker is skeptical
  • The AI agent scenario described involves automating end-to-end, multi-step processes involving diverse data and human communications—a leap from current abilities
  • Challenges include: the necessity for long sequential rollouts, limited multimodal training data, and greater computation demands for images/videos
  • Most available pre-training data is insufficient for truly reliable agentic behavior; generating realistic practice data for agents remains an open issue
  • Even seemingly simple innovations (like RL approaches for math/coding) took years to implement; tackling broad computer use tasks is therefore even harder

Acknowledging Current Model Achievements 11:30

  • Despite skepticism, recent models (like Gemini 2.5) demonstrate real reasoning capabilities and problem decomposition
  • Experimenting with models in their domains of competence shows that they are intelligent, highlighting rapid progress
  • The probability of transformative AI outcomes remains wide—preparing for even unlikely scenarios is justified

Personal AI Timelines and Forecasts 13:05

  • The speaker estimates a 50/50 chance of an AI that can handle all aspects of small business tax preparation, autonomously, by 2028
  • Drawing analogy to the transition from GPT-2 to GPT-4, the timeline for full, competent computer use is seen as closer to 2028
  • Models will likely show impressive demos in 2026-2027, but reliable, end-to-end agents completing multi-day projects will take longer
  • The timeline for AI to be able to learn on the job as well as a human (e.g., a video editor learning personal preferences) is projected for 2032

Reflections on AGI Timelines and Future Progress 15:22

  • Seven years is a long time in AI development—given past leaps, it’s possible models will learn on the job by then
  • The probability of AGI arrival per year will drop after 2030 due to compute limits and diminishing algorithmic returns
  • If true AGI is delayed, the world may look relatively unchanged into the 2030s or 2040s, but dramatic change remains plausible in this decade
  • The blog post draws on discussions with AI researchers and is part of a broader practice of in-depth podcast follow-ups by the speaker
  • The speaker encourages readers to subscribe to the accompanying blog/newsletter for further discussion and insights