Hacker News AI Goalposts: Tracking AI Progress and the Moving Target of AGI

The pursuit of Artificial General Intelligence (AGI) is often marked by "goalposts"—specific tasks that skeptics believe AI will never achieve. A community project hosted at stoppels.ch/goalposts tracks these challenges extracted from Hacker News discussions, allowing users to vote on whether these milestones have been met. The results highlight a persistent tension between AI's ability to perform a task once and its ability to perform it reliably and routinely.

The Reliability Gap: One-Off Success vs. Routine Capability

While many frontier models can now perform tasks that were once considered impossible, a critical distinction exists between a stochastic success and a reliable tool.

Users note that many LLMs could accomplish a task given an infinite compute budget and infinite attempts, but they cannot routinely complete the task in a professional workflow. For example, the ability of an AI to surprise a user with an unprompted question has occurred, but it is not a routine capability that developers can rely on for application design.

Domain-Specific Failures in Specialized Knowledge

Despite the perceived "solving" of coding, AI continues to struggle with specialized, real-world physical knowledge and complex reasoning in niche domains.

Physical and Mechanical Knowledge: Users report that AI often fails in fields like agriculture, construction, and mechanics. One user cited an example where an AI provided incorrect and potentially catastrophic advice regarding the inhibition of platinum cure silicone by plaster, hallucinating sources that did not support the claim.
Visual-Spatial Reasoning: ASCII art remains a significant hurdle. While a model might generate a passable ASCII foot, other models (such as ChatGPT's free tier) may misidentify that same ASCII art as a "train/locomotive," indicating a failure in spatial understanding.
Complex Data Interpretation: AI tools still struggle with interpreting "noisy" data, such as chess scoresheets containing mistakes, where a human expert can infer intent based on the context of the level of play.

The Evolution of AI Challenges (2016–2026)

There has been a noticeable shift in the types of challenges users set for AI. Earlier challenges focused on basic functional capabilities, while modern challenges focus on higher-order cognitive functions and physical interaction.

2016 Era: Challenges centered on passing the Turing test, writing basic code, or performing simple digital actions (e.g., "order me a coffee").
2026 Era: Challenges have evolved to include opening physical doors, emulating human pettiness convincingly, and making novel scientific breakthroughs.

The Turing Test and the "Parrot" Problem

Discussion around the Turing test suggests it may be fundamentally flawed because humans are often unable to distinguish between a real person and a sophisticated "parrot." Some argue that the ability to pass a Turing test for a short period does not equate to passing an "adversarial" Turing test over an extended period, which remains a far more difficult milestone.

Mathematical and Scientific Breakthroughs

Some users believe the "last exam" for AI has been partially met through the resolution of open math problems, such as the Navier-Stokes existence and smoothness problem. However, others argue that passing a single test (e.g., 1/6 of a proposed set of challenges) is insufficient to claim a victory, emphasizing that the "goalposts" for AGI are often a moving target based on the relatifve complexity of the requirements.