Sometimes, the most valuable insights come from the unexpected places – like a 2018 AI paper that refused to claim its work was anything more than incremental improvements.
In the world of scientific publishing, researchers are often trained to emphasize the significance of their findings. Terms like “significant advance” and “expands our understanding” frequently appear in papers, sometimes with a touch of hyperbole. But occasionally, a study emerges that bucks this trend by acknowledging its modest contributions.
This is the case with YOLOv3, an object detection system developed by Joseph Redmon and Ali Farhadi. Their 2018 paper, published on the arXiv preprint server, titled “YOLOv3: An incremental improvement,” stands out for its remarkably modest claims. The authors openly acknowledged that their work was little more than a collection of small tweaks to an existing system.
The Modest Approach to Academic Achievement
YOLO (You Only Look Once) is an object detection system that can identify multiple objects within a single image. It’s powerful enough to beat CAPTCHA-style tests and has been used to spot smuggled ships. But by 2018, the creators of YOLO seemed content to let their earlier work speak for itself.
The paper’s modest tone begins with its title: “An incremental improvement.” The abstract claims “We made a bunch of little design changes to make it better.” The introduction opens with the frank admission: “Sometimes you just kinda phone it in for a year, you know? I didn’t do a whole lot of research this year. Spent a lot of time on Twitter.”
The authors credit much of their work to the contributions of others, describing their own improvements as “nothing like super interesting, just a bunch of small changes that make it better.” They even included a dedicated section titled “Things we tried that didn’t work,” a refreshing honesty that Feedback believes should become standard practice in scientific publishing.
Honesty in Science
This unusual approach to academic publishing resonates with what sociologist Per Engzell observed about limitations sections in academic papers: they represent “radical honesty for exactly one paragraph.” Data scientist Johan Ugander called YOLOv3 “the most honest paper.”
This honesty contrasts sharply with the typical academic incentive structure, where researchers often feel pressured to maximize the perceived impact of their work. The YOLO authors seemed content to simply improve upon their previous work without grand claims.
Science Communication in Crisis
While the YOLO paper offers a refreshing dose of honesty in science, other recent examples highlight the challenges of science communication. Feedback received a press release from the World Avocado Organization defending the environmental sustainability of avocados – a claim that seems almost certain to be questionable given the environmental impact of avocado farming.
This example fits into a broader pattern we’ve observed: organizations often defend their own products with scientific language, regardless of the actual evidence. As Feedback noted, these claims may be technically defensible or not, but they clearly exist within a specific incentive structure.
Then there’s the growing industry of selling sleep – mattresses, pillows, sleep apps, and more. Companies like Amerisleep send out press releases claiming that “rising youth deaths” could result from sleep deprivation, linking their products to serious health outcomes.
The Curious Case of Sleep Research
Sleep research itself isn’t necessarily flawed. Studies show that sleep improves problem-solving abilities and memory consolidation. During sleep, the brain processes memories and forms new connections, sometimes generating new insights by combining disparate ideas.
This explains why we often hear the advice to “sleep on it” when making important decisions. The prefrontal cortex, our brain’s inner critic, is less active during sleep, allowing for more creative thinking.
But the marketing of sleep research raises questions: Is the science sound, or is it being used primarily as a marketing tool? And what does this say about our society’s relationship with science?
Conclusion
The YOLOv3 paper offers a valuable lesson in scientific humility. Sometimes, the most significant contribution isn’t the discovery itself, but the honest acknowledgment of what was achieved – and what wasn’t.
As we navigate an increasingly complex relationship with scientific information, we might do well to remember that the most valuable insights often come from unexpected places – like a paper that refuses to overstate its modest contributions
