AI and Machine Learning: What Experts Are Saying

At NeurIPS 2025, researchers kept returning to one complaint: models that ace standardized tests still stumble on edge cases, distribution shifts, and adversarial inputs. The conversation about AI has moved past hype toward a harder, more useful argument about what these systems can actually do—and where they fail.

The Foundation Model Split

Experts divide sharply on large pretrained models. Supporters point to economics: training one big model and adapting it to many tasks costs less than building task-specific systems from scratch. Critics, including Yann LeCun, argue current models lack genuine understanding and confabulate confidently. Training a frontier model now consumes over $200 million in compute alone, and the environmental bill has drawn scrutiny even from labs that benefit from the work.

The industry compromise is a tiered architecture. Large models handle complex reasoning; smaller specialized models manage high-volume, lower-complexity jobs. Anthropic and Alibaba Cloud both use this pattern. Most enterprise applications do not need frontier intelligence—they need something reliable, fast, and affordable for a specific workload.

Reliability Is the Real Problem

A language model wrong 5% of the time is a curiosity. The same error rate in clinical decision support is a liability. Google DeepMind has published on "calibrated uncertainty," where models estimate their own confidence and defer to humans when unsure. Mayo Clinic and Johns Hopkins studies show this human-in-the-loop approach improves diagnostic accuracy by 15–20% over either AI or physicians alone.

Mechanistic interpretability—mapping a model's internal representations to its outputs—may reduce hallucinations structurally. Papers from 2025 report targeted interventions cutting hallucination rates by 30–40% in controlled settings. The work is early, but it suggests failure modes can be engineered around rather than merely patched.

Energy use remains a live issue. Stanford's HAI estimated training energy for top models doubled every eight months through 2024. Mixture-of-experts architectures, which activate only a fraction of parameters per query, have cut training costs 3–5x. Most major labs now buy renewable energy, though critics note timing mismatches between generation and compute make the offsets imperfect. Some researchers, including Melanie Mitchell of the Santa Fe Institute, argue that scaling may hit diminishing returns as performance gains slow despite more compute and data, renewing interest in alternative architectures.

Policy Without Soundbites

The EU AI Act draws mixed reviews: some experts praise its risk-based tiers; others warn compliance costs will squeeze smaller firms out of the market. The U.S. has stuck with sector-by-sector guidance from the FDA, FTC, and SEC—messier, but more responsive to domain-specific risk. China has targeted rules on recommendation algorithms, deepfakes, and generative AI. Companies deploying AI across borders face a genuine patchwork, and multi-jurisdiction compliance is becoming a barrier to entry for startups.

The experts disagree openly on bigger questions: whether to accelerate or pause development, how much compute to monitor, and how seriously to take timelines for artificial general intelligence, which range from "within a decade" to "never." What unites the most useful voices is a refusal to take binary positions. They acknowledge real capabilities and real limits simultaneously, back responsible deployment without opposing progress, and engage with policy without flattening technical nuance into slogans. That stance is less exciting than the hype wars, but it is the one producing durable results—and the one the field needs more of as AI embeds itself deeper into healthcare, finance, and public infrastructure.