I don't think AGI is right around the corner

State of AGI Timelines

• Many commenters are more skeptical than a year or two ago: LLM hype feels past its peak, AGI is seen as “not right around the corner,” with estimates often in the 2030s–2040s or “maybe never with current hardware.”
• Others point out the article’s author is actually bullish: ~50% for AGI by early 2030s, and even misaligned ASI by 2028 considered “plausible.”
• A minority think timelines are much shorter, driven by autonomous money-making agents and huge economic incentives once such systems exist.

What Counts as AGI?

• No shared definition. Competing views:
  • “Median human” or “bottom X% of humans” across all cognitive tasks.
  • Ability to replace unsupervised white‑collar workers or run companies.
  • Systems that can learn and self‑improve (e.g., design, train, and deploy better models).
  • Stronger notions: simulate reality faster than real time, or match humans in “virtually every task.”
• Some argue current LLMs are already “AGI by some definition” (better than many humans on many intellectual tasks); others insist that’s redefining the term down.

Limits of Current LLMs

• Core critiques:
  • No genuine continual learning: you can’t reliably update preferences or skills through feedback; models are mostly “frozen.”
  • Brittle reasoning and character‑level failures (e.g., counting letters, “states with W”) despite fluent chains of thought, often blamed on tokenization but seen as deeper evidence of shallow understanding.
  • Hallucinations and confident nonsense; poor at robust planning, physical reasoning, and open‑ended tasks like operating a vending machine in the real world.
  • No obvious track record of novel scientific discoveries purely from “having all human knowledge.”
• Supporters counter that:
  • Emergent capabilities across domains are a form of intelligence.
  • Speed and scalable parallelism alone could be enough to reach “superhuman” performance once baseline reasoning is good enough.

Continual Learning, Architectures, and Paths to AGI

• Many see lack of adaptive, test‑time learning as the main blocker; human‑like skill acquisition (e.g., learning saxophone, doing taxes across years) doesn’t match today’s train‑once paradigm.
• Proposed directions:
  • Agent systems with tools, search, RL, simulations, and explicit world models.
  • Hybrid systems (logic, Prolog, heuristics, specialized sub‑agents) orchestrated by LLMs.
  • Hardware advances (memristors, analog or neuromorphic chips; or even biological “vat brains”).
• Others argue scaling diverse data and compute has historically beaten hand‑crafted continual‑learning schemes and may keep doing so.

Intelligence, Consciousness, and “Souls”

• Ongoing philosophical dispute:
  • Materialist view: brain is computation; in principle reproducible on silicon (or in wetware with a Python API).
  • Non‑material or “cardinality barrier” arguments: biological systems might use effectively uncountable analog state spaces or unknown physics, putting true sentience beyond current digital machines.
• Distinction between intelligence vs consciousness vs wisdom:
  • LLMs may be “intelligent” at language but lack embodiment, caring, participatory knowing, and self‑correction in the human sense.
  • Some emphasize that current systems are “word fountains” without shared lived reality or genuine understanding.

Economic and Social Impact Without AGI

• Broad agreement that sub‑AGI systems can still be highly disruptive:
  • Many repetitive, low‑skill, or text-heavy jobs could be automated or radically changed.
  • Risk of “good enough and cheaper” systems replacing humans even when quality is worse (IKEA vs master carpenter analogy).
  • Concern about layoffs justified by “AI” more than caused by true capability, and about concentration of power if very capable systems remain expensive and centralized.
• Some hope this leads to more leisure and UBI; others worry about mass unemployment, political instability, and declining well‑being even as GDP rises.

Hype, Incentives, and Trust

• Strong suspicion that near‑term AGI claims are driven by:
  • Fundraising, valuations, and subsidies for AI labs and startups.
  • Media and influencer incentives to declare “AGI is here” with each new model.
• Academics and some practitioners often report more conservative views but feel ignored.
• Several see AGI as a “moving goalpost” or even a grift: definitions shift so it’s always a few years away, yet always just close enough to sell.

Safety, Doom, and “Intelligence Explosion”

• Some think current LLMs are already enough to start an “intelligence explosion” once they reliably help improve their own successors; others think their lack of real discovery and self‑direction undermines that narrative.
• A fraction worry about misaligned ASI in the 2020s and discuss prepping (off‑grid power, bunkers), though others call this speculative and unfalsifiable.
• Many place more weight on nearer‑term harms: economic shocks, surveillance, manipulation, military uses, and ethically fraught paths like human‑cell “computers.”