Over the past 12–24 months, the most significant change in AI applications may be that models have been embedded in more complete workflows, rather than the models themselves suddenly becoming much smarter. A software engineer and AI product professional offered this assessment in the community: models are certainly more useful now, but the increase in their “raw intelligence” may not be proportional to the improvement in their overall performance.
Here, “workflow” refers to a system of context retrieval, prompt design, tool calling, and task orchestration built around a model. Its value lies in giving models access to information, computing resources, and execution steps, making it less difficult for a model to solve a problem through a single response alone.
What this is
The poster’s experience comes from three areas. In production, once an existing model and workflow can already meet the required standard consistently, switching to a newer frontier model often produces only limited gains; the main bottleneck shifts to whether the context is accurate and the necessary tools are available. In software engineering, new models generally deliver more improvements, but the performance of the same model can still vary significantly across different workflows. At the same time, small local models can complete many efficient tasks once they have acquired the necessary knowledge.
This does not mean models have not improved. More reliable tool calling, fewer incorrect responses, and longer context windows are genuine improvements in capability; nor does smaller models handling certain tasks well prove that intelligence requires only a small number of parameters. The key change is that model upgrades and system optimization are jointly creating the perception of models becoming “stronger.”
Industry view
What deserves recognition is that this serves as a reminder for enterprises not to purchase models solely on the basis of leaderboard rankings. Benchmarks usually restrict the form of questions, while production tasks involve data quality, tool permissions, response time, and review costs. Improvements in the latter areas may determine whether a project is usable even more strongly than growth in model parameters or individual benchmark scores.
But this perspective also carries clear risks: the poster’s experience is concentrated in software engineering and a particular product environment, so it may not apply to knowledge-intensive fields such as research, medicine, and complex decision-making. New models can sometimes find more ingenious ways around obstacles, appearing only too slow or unstable under strict production constraints. Underestimating frontier models on this basis could also mean missing their advantages in knowledge coverage, nuanced judgment, and cross-domain generalization. It is true that AI companies’ expectations for public listings may amplify market narratives about capability leaps, but underestimating model value would likewise be detrimental to investment in infrastructure and toolchains.
Impact on regular people
For enterprise IT: When evaluating AI projects, teams should separately record the contributions of model upgrades and workflow redesign; procurement contracts should also avoid equating leaderboard performance directly with business outcomes after deployment.
For individual careers: Asking the right questions still matters, but providing clear materials, breaking down steps, and establishing review mechanisms may improve output quality and stability more than constantly chasing the latest model.
For the consumer market: Intelligent applications will continue to spread, but a given function does not necessarily need to rely on the most expensive large model. Smaller, cheaper, or locally run solutions may enable lower prices and better privacy.