What if AGI is not coming?
Skepticism is growing over whether today’s large language models and transformer-based systems can ever lead to artificial general intelligence, or whether they are just powerful but narrow statistical tools hitting compute, data, and efficiency limits. Commenters debate definitions of “intelligence” and “consciousness,” the importance of embodiment and new architectures, and whether algorithmic advances have truly kept pace with hardware scaling. Many expect useful but incremental improvements and potential market cooling, while others argue that sustained investment and gradual breakthroughs make some form of human- or superhuman-level machine intelligence ultimately inevitable.
Article framing and premise
- Several commenters say the essay doesn’t truly explore “what if AGI never comes” and instead argues AGI is unlikely, without stating it plainly.
- Some see a financial angle: no one wants to “tell the children Santa isn’t real” while AI stocks are inflating.
What is AGI / intelligence?
- Disagreement on whether AGI is already here (e.g., with transformers enabling broad transfer) versus AGI as something requiring consciousness, agency, or “soul-like” properties.
- Some argue “general intelligence” is just a bundle of capabilities; others suggest humans may be near an upper bound of useful intelligence for our environment.
- There’s frustration about shifting or vague definitions; claims that terms like AGI/ASI are used to dodge criticism.
Brains, silicon, and embodiment
- One camp: biological brains prove low‑power general intelligence is possible, so synthetic versions should be too “eventually.”
- Opposing view: we don’t understand neurons or whole organisms (e.g., C. elegans simulations), so “just scale compute” is hubristic.
- Debate over whether true intelligence requires embodiment and interaction with the physical world, versus full simulation being sufficient in principle.
- A minority claim real intelligence/life may be thermodynamically or metaphysically constrained; others find this unclear or unconvincing.
Scaling, algorithms, and limits
- Some assert no major AI breakthroughs in decades and that LLMs are just scaled-up, inefficient curve‑fitting.
- Others strongly dispute this, citing transformers, state-space models, residual connections, batch norm, and measured algorithmic efficiency gains as substantial advances.
- Concerns about limits: hardware cost, consumer-device constraints, data exhaustion and AI‑generated data “poisoning.”
- Counterclaims: “limits” have been repeatedly broken; curated synthetic data and newer web scrapes may actually help, and “model collapse” is called a myth by some.
Current LLM capabilities and gaps
- LLMs are seen as impressive but still “stochastic parrots” by many: good at language, coding assistance, and benchmarks (e.g., NYT Connections), yet lacking reliable reasoning, autonomy, and in‑place code editing.
- Some argue humans also often parrot; others stress humans can genuinely create new concepts (e.g., calculus, new genres) whereas LLMs recombine existing data.
Timelines, hype, and impact
- Several predict a near‑term plateau or bubble pop (Nvidia, AI startups), comparing to past tech hype (VR).
- Others think progress and investment will continue, but AGI will emerge gradually, if at all.
- Some say even if AGI never arrives, current AI is already useful; others worry more about a slide into “inane banality” and educational degradation than about a singularity.