Interview, Podcast, Other
Why I don’t think AGI is right around the corner
- AI progress timelines among experts vary widely from two to twenty years, with personal views updated as of July 2025 suggesting a range of outcomes rather than a single fixed date.
- A fundamental bottleneck for Fortune 500 transformation is identified as the difficulty of extracting human-like labor from LLMs due to a current lack of continual learning capabilities.
- Future timelines for useful LLM tools are expected to extend as current models fail to improve over time or build context comparable to human learning.
- Current LLMs possess a higher baseline than average humans but cannot incorporate high-level feedback for improvement, a limitation expected to persist until "smarter models" develop self-directed reinforcement learning loops.
- Progress in continual learning is predicted to trigger a significant value discontinuity and potentially an intelligence explosion, though a "software-only singularity" is considered unlikely.
- A broken early version of continual learning or test-time training is anticipated to appear before systems achieve true human-like learning, likely occurring within the next few years.
- Without advancements in continual learning, less than 25% of white-collar employment is expected to disappear even if current AI progress halts.
- Reliable computer use agents capable of end-to-end tax preparation are forecasted by the end of next year, though this specific timeline is viewed with skepticism.
- Increasing horizon lengths for agentic tasks are expected to slow progress due to high compute intensity associated with processing images and video, compounded by insufficient multimodal pre-training corpora.
- Algorithmic innovations to solve computer use are expected to face significant delays, referencing the two-year gap between GPT-4 and O1 as a precedent for difficulty.
- An AI capable of handling end-to-end tax tasks with the competence of a general manager is predicted to arrive around 2028.
- AI is expected to learn on the job with human-like depth for white-collar work, such as video editing, by 2032.
- Post-2030 AI progress is forecasted to rely primarily on algorithmic advances rather than scaling training compute, as the latter is expected to plateau before the end of the decade.
- AGI arrival in the 2030s or 2040s is considered a possibility, with scenarios ranging from a relatively normal world until those decades to outcomes described as "truly crazy."
- Current pre-training data lacks a large corpus of multimodal computer-used data, raising doubts about the sufficiency of current resources for building reliable agents.
- The speaker notes a disagreement with the view that today's systems would be more economically transformative than the internet even if AI progress ceased.