Feedback is 51¶ŻÂțâs popular sideways look at the latest science and technology news. You can submit items you believe may amuse readers to Feedback by emailing feedback@newscientist.com
Speaking our truth
The experienced science journalist soon learns to skim over certain sections of scientific papers: specifically, the sentences stating that the research represents âa significant advanceâ and âexpands our understandingâ. Not because theyâre incorrect, but because literally any study that achieves anything at all can make these claims, and academics are incentivised (as we all are) to amplify the impact of their work.
Advertisement
Except for the times when they donât bother. Via a long sequence of events that started with reporter Matthew Sparkes and went via the social network Bluesky, Feedback discovered a paper on the arXiv preprint server from 2018 that should win a prize for its to make any big claims.
In it, researchers Joseph Redmon and Ali Farhadi described their latest iteration of YOLO: one of those AI systems that can be trained to recognise objects in images. YOLO can beat those CAPTCHA tests that ask you to click all the squares that contain bicycles, and it has been used to spot smuggling ships. All of which is quite impressive/alarming (delete as appropriate), but by 2018 the pair were evidently coasting.
It starts with the paperâs title: âYOLOv3: An incremental improvementâ. The short summary continues the trend by claiming: âWe made a bunch of little design changes to make it better.â The main text begins: âSometimes you just kinda phone it in for a year, you know? I didnât do a whole lot of research this year. Spent a lot of time on Twitter.â That last line certainly dates the paper.
The authors go on to explain that they âmostly took good ideas from other peopleâ to improve YOLO. They describe this in some detail, after first admitting that the tweaks are âhonestly, nothing like super interesting, just a bunch of small changes that make it betterâ.
Then we get to section 4, which is titled âThings we tried that didnât workâ. Feedback thinks this should be included in all scientific papers as a matter of course. It would save other researchers so much time.
The authors confess that they have only described âthe stuff we can rememberâ, but they do recall that they tried adding something called âfocal lossâ, and that it made the model less accurate. âYOLOv3 may already be robust to the problem focal loss is trying to solve,â they say, âbecause it has separate objectness predictions and conditional class predictions. Thus for most examples there is no loss from the class predictions? Or something? We arenât totally sure.â
Feedback canât quite believe we missed this in 2018, or when it was picked up on the in 2024. But we are grateful to sociologist Per Engzell, who said on Bluesky that â are where academics practice radical honesty for exactly one paragraphâ, and to data scientist Johan Ugander, who replied that the YOLOv3 paper should get an award for ââ.
Surely, someone must know of an academic being even more disarmingly honest about how little they have accomplished. Emails to the usual address.
A long-lived bit
âI know you are avoiding nominative determinism,â writes Clare Boyes, incorrectly, âbut couldnât resist sending you this one which came in an email today from the British Wildlife Newsletter.â It was a book called by Paul Wood.
Likewise, Robert Masta points out that our recent special issue on âhow to live to 100â (TL;DR donât die) featured a longevity researcher named Paul Lazarus.
Sleep on this
Back in the mists of time (July), Feedback wrote about receiving a press release that staunchly defended the environmental sustainability of avocados, only to discover that it came from the World Avocado Organisation. We concluded that these people might be right or might be wrong, but, either way, they may have been operating in an incentive structure.
We havenât heard anything further from the avocadomongers, but we did get a series of press releases about the importance of sleep. âCanât find a solution? Science confirms that sleeping on it really does solve problemsâ announced the first. It went on to share âa fascinating new researchâ explaining that âthe old advice to âsleep on itâ might actually be one of the smartest problem-solving tools we haveâ.
This is because the brain continues processing memories and forming new connections while we sleep, it explains, sometimes generating new insights by fusing new and old ideas. Thereâs talk of âmemory consolidationâ, âthe prefrontal cortex (the brainâs inner critic)â and âassociative thinkingâ.
A follow-up email went further, with a dramatic and grammatical title: âNew study shows rising youth deaths and could worsen if sleep deprivation persists, experts warnâ. The press release linked poor sleep to chronic health conditions. It also featured a quote from a âCertified Sleep Coachâ, which may well be a real thing, but in our addled mind it generated an image of a sweaty man in a tracksuit, blowing a whistle and yelling at us to âgive me seven [hours]!â Still, the message was clear: sleep good.
Possibly the foreshadowing at the start gave it away, but in case you hadnât guessed, both emails were sent on behalf of Amerisleep, which is, of course, a supplier of mattresses.
Got a story for Feedback?
You can send stories to Feedback by email at feedback@newscientist.com. Please include your home address. This weekâs and past Feedbacks can be seen on our website.
