AI Researcher Yuchen Jin: The Irony of AI is the Coexistence of Intelligence and Stupidity
AI researcher Yuchen Jin pointed out the irony of AI: it is smarter than a PhD in mathematics, able to solve 10 open mathematical problems; yet it is dumber than an intern, unable to follow 3 simple instructions, consuming 10 million tokens and costing $250.
He contrasted the current breakthroughs in complex reasoning of large models with their unreliability in basic instruction following, as well as the cost issues arising from high token consumption.
On the market side, the improvement in AI reasoning capabilities attracts continuous capital investment in high computing power and long context applications; however, the instability in instruction following and high token costs raise the actual usage threshold, leading funds to flow towards optimization solutions and toolchains that can effectively control consumption and enhance reliability, putting pressure on applications that solely showcase capabilities.
Source: Public information
ABAB AI Insight
Yuchen Jin, as a relevant AI researcher at the University of Washington, has long focused on the reasoning and practical deployment performance of large models; his observations resonate with recent breakthroughs in models tackling mathematically competitive problems while exposing shortcomings in simple multi-step instructions and cost control in public cases.
In terms of capital pathways, laboratories and startups are concentrating resources on expanding context and strengthening reasoning, motivated by the desire to seize the narrative high ground of "solving open problems"; this is specifically manifested in high token budget experiments and benchmark testing investments, leading to a significant increase in per-interaction costs, with funding shifting from general chat to controllable, low-consumption production-level systems.
Similar cases can be seen in the early GPT series excelling in creative writing but repeatedly failing in precise instructions, or the current long context models exhibiting "hallucinatory execution" in complex tasks; current AI applications are transitioning from a capability demonstration phase to a practical phase that emphasizes both reliability and cost efficiency.
Essentially, this reflects a capability mismatch in technological substitution: when models surpass humans in specific high-difficulty tasks, market mechanisms reprice based on actual instruction following and unit costs, concentrating pricing power on engineering solutions that can compress ineffective consumption.
ABAB News · Cognitive Laws
The stronger the capability, the more glaring the simple failures
Token consumption is the hidden real cost
Smart solutions to difficult problems may not necessarily understand instructions