In 2025, as artificial intelligence moved from "science fiction" into everyday life, a LinkedIn article sparked deep industry reflection on AI's limits. Author Alex Wang, a blogger with over a million followers on LinkedIn, titled her piece "What AI Still Can't Do — and Why That's the Most Interesting Part," dissecting persistent issues in large language models (LLMs) and large reasoning models (LRMs) around generalization, reliability and world modeling. Wang emphasizes that AI has matured for structured tasks such as customer support and knowledge‑workflow automation, but these "can'ts" are not a distant AGI (artificial general intelligence) mirage — they form a visible "research map" that points to the next wave of breakthroughs like multimodal ecosystems and causal‑reasoning architectures. Drawing on 2025 benchmark reports and expert interviews, this article analyzes the causes, manifestations and impacts of these limitations, showing why "AI's shortcomings" are becoming an engine of innovation. After all, with open‑source models such as DeepSeek V3.2 and Mistral 3 proliferating, acknowledging limits is not retreat but forward‑looking wisdom.
The Double‑Faced Mirror of AI: Brilliance and Emerging Limits
By 2025 AI's glow lights up the global economy: OpenAI's GPT‑5 family handles three billion queries daily, Google's Gemini 3 posts Elo scores above 1500 on the LMArena benchmark, and the generative AI market is projected at $500 billion. From Waymo's autonomous driving to Adobe Firefly's professional image generation, AI has penetrated transport, healthcare and creative industries. X user @ZaStocks listed: "AI can drive autonomously, translate in real time, generate video, and even make real‑time decisions in military intelligence." These achievements stem from scale in the Transformer architecture: parameters rose from GPT‑3's 175 billion to the trillion‑scale models of 2025, and training data exceeds 10 trillion tokens.
However, as Wang notes, AI's "maturity" is confined to pattern matching within known data distributions. An Aussie AI research report found LLMs reach 95% accuracy on repetitive tasks but collapse on edge cases. For example, on the 2025 AIME math competition benchmark, DeepSeek V3.2‑Speciale scored 96% overall but only 70% on variant problems (such as those with added physical constraints). On X, @willccbb complained: "It's 2025 and GPT‑3.5‑turbo still can't handle big‑number arithmetic." These are not isolated cases: Palantir's AI platform optimizes supply chains but users reported prediction deviations over 20% under sudden geopolitical events. Wang's insight is that these limits are not "bugs" but "features" — they expose the need for AI to transition from a "predictive machine" toward an "intelligent agent."
Breaking Down Seven "Can'ts" of AI: Deep Fault Lines from Generalization to Emotion
Wang categorizes AI's limitations into seven types, each illustrated with concrete examples and echoing multiple 2025 surveys. First, the lack of true generalization: AI is good at interpolation but falters at extrapolation. The LeetCode Copilot blog noted AI‑assisted coding interview pass rates of 90%, but when faced with "logic‑puzzle variants" (e.g., a single abnormal condition) failure rates jump to 80%. Why? Training data biases toward common patterns and lack robustness — LRMs like o1 achieve only 60% success in transfer learning.
Second, reliability collapse: long‑chain reasoning tends to hallucinate. Wang points out that top LLMs' coherence drops to 70% after 1,000‑token sequences. In 2025, Anthropic's Claude 4.5 showed a 15% error rate in multi‑step planning and required "verifier models" for support.
Third, no built‑in world model: AI lacks physical/causal intuition. Wang cites Google DeepMind's Genie 3, which generates virtual warehouse simulations to train agents, but in real‑world deployment the AI ignored gravity or collisions and experienced a 40% failure rate. Meta's LRM research confirmed AI accuracy at only 65% for common‑sense planning (such as "pouring milk without spilling").
Fourth, lack of in‑situ learning: once frozen, models cannot adapt dynamically. Wang contrasts humans: "AI has no 'on‑site learning' (in‑situ learning)." In 2025 Google's nested learning made progress, but adapting to new tasks still required hours of retraining.
Fifth, weak macro‑goal setting: agents execute subgoals well but lack strategic vision. Wang's example: AI can correct code but does not understand "why the project exists." Bindu Reddy wrote on X that 20–50% of work can be automated, but "safety/permissions/context" still require human oversight.
Sixth, emotional grounding is hollow: AI simulates empathy but lacks lived experience. Wang warns that in negotiation or HR contexts AI falls into "flawed premises." LinkedIn commenter @mnovello23 added: "AI automates data cleaning but cannot build relationships or emotional insights." A 2025 AI x Education Substack reported that AI grading of essays contained racial bias and could not reliably distinguish good writing from bad.
Seventh, fragility on edge cases: AI lacks intuition in ambiguous situations. Wang sees this as historical pattern: "each breakthrough solves prior edge cases." Facebook group discussions note that AI understands code but cannot "design user experiences or make architecture decisions."
These "can'ts" are not static: in 2025 Kweku Opoku‑Agyemang argued on X against cynics that "AI has solved math problems to 100% and can act autonomously." But Wang insists that acknowledging shortcomings is the prerequisite for progress.
Root Causes: The "Original Sin" of Predictive Architectures and Resource Limits
The roots of AI's limits are multidimensional. First, core architectural flaws: the Transformer is a "predictor," not an "understander." Wang analyzes that LLMs generate from statistical associations and tend to "drift" over long chains, with hallucination rates of 10–20%. Personal Science Guide warned that the "hallucination risk" of generative AI stems from noisy data.
Second, data and compute bottlenecks: training data is being exhausted — in 2025 some analyses estimate only 1% of high‑quality corpora remains unused. X user @aasim wrote: "We've consumed power, data, funding, even ways to measure intelligence." Although DeepSeek V3.2 reduced costs by 90%, world models still require simulating trillions of interactions and demand roughly ten times current compute.
Third, design‑philosophy bias: AI optimizes benchmarks rather than reality. Wang quotes Fei‑Fei Li: "Scale is not the future; we need cognition‑inspired architectures." In 2025 the EU AI Act reinforced this point by requiring transparency in "causal reasoning" and imposing fines up to 7% of revenue. Ethical blind spots also widen limits: emotional grounding gaps stem from data dehumanization — avoiding bias sometimes sacrifices depth.
The Appeal of "Can't": An Engine of Innovation and Wide Impacts
Wang's core point is that these gaps are "the most interesting" because they are actionable signals rather than mysticism. For instance, the lack of world models has spawned simulators like Genie 3, helping push robots from labs into factories; reliability issues have spurred multi‑model oversight, such as OpenAI's "tool call" ecosystems.
The effects of AI's "can'ts" are multifaceted and double‑edged. On the positive side they spur human–AI collaboration. AI x Education stresses: "Focus on what AI can't do — building relationships and emotional insights." Companies like Palantir use AI for decision support while keeping humans for high‑level judgment, boosting efficiency by 30%. On the negative side, ignoring limits amplifies risks — hallucinations have led to medical misdiagnoses (5% of cases in a 2025 report) or agent failures (e.g., edge accidents with autonomous vehicles). Globally, developing countries benefit from open‑source (like Mistral 3's multilingual support), but the data divide can worsen inequality. Wang predicts breakthroughs will come from "changing learning paradigms" — long‑term memory and uncertainty modeling could unlock tenfold performance jumps.
Conclusion: Dawn from "Can't" to "Can"
Alex Wang's reflections act as a mirror, revealing both AI's brilliance and its shadows. In 2025 we witnessed a turning point from "AI peak" to confronting reality. As @CaibelPlayz wrote on X, AI cannot be judge or soldier, and that is precisely a stronghold of human value. Going forward, acknowledging "can'ts" will drive a structural shift from scale to architecture — perhaps in 2026 a Garlic model will close coding gaps. AI is not a replacer but a magnifier: it illuminates our limits and lights up possibilities. In this era where "what AI can't do" is the most interesting part, innovators should take heed and move forward together.