Models Are Improving Across the Board — Denying This Trend Is Irresponsible

An Underrated, All-Around Evolution

Recently, a tweet sparked considerable discussion: "It's happening, and everyone should understand it — models are getting better at almost everything. Saying they aren't improving is irresponsible." Short as it is, this statement hits on a fact obscured by emotional debate: AI model progress isn't confined to a single metric — it's happening comprehensively and systematically.

Discussions about model capabilities often fall into two extremes — either exaggerating a single breakthrough or using isolated failure cases to deny overall progress. This tweeter's reminder is blunt: if even the basic judgment that "models are getting better across the board" is refused, the discussion loses its factual foundation.

Why "Better Across the Board" Matters More Than "A Single Breakthrough"

In the past, we were used to measuring models by a benchmark score or a leaderboard ranking. But the truly noteworthy signal is that models are improving simultaneously across reasoning, code, multilingual ability, instruction following, and long-context understanding. This "moving forward on all fronts" means an overall elevation of underlying capability, not surface-level results stacked up through targeted tuning.

A model that scores high on math problems but frequently errs in everyday conversation still has limited practicality. Only when a model shows stable progress across multiple unrelated tasks does it come closer to the word "general." This is the core meaning of "better across the board," and the reason this Twitter user stresses that "everyone needs to understand it."

Benchmark: A standardized evaluation set for measuring AI model capabilities, such as MMLU (multitask language understanding), HumanEval (code generation), and GSM8K (elementary school math reasoning). Each benchmark is typically designed for a specific type of capability, and scores are easy to compare across models, so they're often used as the basis for leaderboard rankings. But benchmarks have clear limitations: test questions may leak into training data (data contamination), and models may overfit to specific question types, inflating scores while actual generalization remains weak. For this reason, a high score on a single benchmark doesn't prove a model is genuinely stronger — simultaneous, multidimensional, cross-task improvement better reflects an overall elevation of underlying capability. This also explains why "moving forward on all fronts" deserves more attention than a new record on some leaderboard.

Why Denying Progress Is Irresponsible

Voices denying model progress usually come from two sources. One is based on outdated experience — judging today's models by how they performed months ago; the other treats isolated failure cases as the overall conclusion, ignoring statistically meaningful improvement.

The problem with both approaches is that they mislead others' expectations and decisions. Developers who misjudge model capabilities may miss out on usable tools; ordinary users influenced by the "AI can't do it" narrative may underestimate the value of the tools at hand. Publicly claiming that models aren't getting better amplifies cognitive bias at the level of information spread — this is exactly the weight behind the word "irresponsible."

Hallucination: Refers to a model generating content that appears plausible but is factually incorrect or fabricated out of thin air — one of the most closely watched flaws of current large language models. It stems from the mechanism by which models generate text probabilistically: the model optimizes for "the likelihood of the next word," not "the correctness of the fact," so it can still confidently output wrong information when lacking reliable grounding. The existence of hallucinations is often used to argue that "AI is unreliable," but that and "models aren't improving" are two different propositions. In recent years, improvements in instruction following, retrieval augmentation, and expressions of uncertainty have significantly reduced the frequency of hallucinations, though not eliminated them. Understanding the mechanism of hallucination helps distinguish "capability limitations" from "overall stagnation," and avoids using isolated failure cases to deny systematic progress.

Progress Is Not Perfection

To be clear, acknowledging that models are getting better does not mean ignoring their flaws. Hallucinations, broken reasoning chains, and unstable handling of edge cases still exist. But "flawed" and "not improving" are two different things. A rational attitude acknowledges both at once: capabilities are rising, and problems remain.

This Twitter user's statement is worth unpacking because it pulls the discussion back to a basic factual level — when evaluating AI, first acknowledge progress, then discuss limitations, rather than using limitations to deny progress. This attitude applies equally to the tech community and ordinary users.

Broken reasoning chains: Refers to a model in multi-step reasoning tasks where intermediate steps show logical leaps, contradictions, or omitted key steps, ultimately leading to a wrong conclusion. Such problems are especially common in mathematical proofs, complex planning, and multi-hop question answering. Causes include: the model lacking explicit verification mechanisms for intermediate states, information decay in long contexts, and training objectives favoring "looking reasonable" over "being logically rigorous." Methods like Chain-of-Thought prompting and test-time compute were introduced precisely to alleviate this problem, having the model explicitly unfold steps and self-check. Even so, broken chains still appear in edge cases — one concrete manifestation of "progress is not perfection."

Practical Takeaways for Users

For those who incorporate AI tools into their workflows, this tweet offers a pragmatic reminder: regularly re-evaluate the models you have on hand. Conclusions from a few months ago may already be outdated, and tasks that previously failed may now be doable. Rather than clinging to the old impression that "AI can't do it," use concrete tasks to verify the current boundaries of capability.

Models are getting better at almost everything. The sentence itself isn't complicated, but it asks us to stay updated and stay verifying, rather than getting stuck in emotional tribal alignment.