Greetings,
My knee is still healing and as I write this I have COVID. So: I am again behind on proper science reading. BUT I have some fun AI content anyway plus one flashy new paper.
AI TO GENERATE PODCASTS?!
Many of you requested updates on important AI features and this month is a biggie. Google's Gemini Notebook (https://notebooklm.google/) which among other more established features turns a body of text into an NPR-style podcast (a blog about the feature is here: https://blog.google/technology/ai/notebooklm-audio-overviews/).
I would share a direct link to the audio it produces to avoid the energy use of you all trying it on your own, but my employer's AI guidance doesn't let me do that, so I'll just say: watch a little less streaming video this week to reduce your carbon footprint, and give this a go. Just 1) upload a document to https://notebooklm.google, and 2) click Generate in "Deep dive conversation". Wow. The results are conversational, natural, and interesting, with relatively few hiccups / artifacts that I noticed (although not none), although it does consistently hype up each paper (both the magnitude and significance of results are exaggerated, albeit not quantitatively). I could see this being useful to help you digest that report you can't bring yourself to read while jogging instead. Not that you should ever trust an AI summary at this point to be accurate (please don't!) - but for legit reports it appears to be much better than not reading the report at all from my limited testing.
Here are the results of four quick tests:
1. A colleague fed in a research report which was written for a lay audience already (https://www.pewresearch.org/data-labs/2024/05/23/electric-vehicle-charging-infrastructure-in-the-u-s/), and I was gobsmacked by the quality. Not as good as a legit pro quality podcast, but MUCH better than something I would put out and a hell of a good first draft.
2. What about a peer-reviewed science article which is much more technical and hard to read? I uploaded my 2nd favorite paper that I'm first author on (since my favorite is written in plain English already) which is about how spatial resolution impacts accuracy, cost and making the right decision: http://onlinelibrary.wiley.com/doi/10.1002/rse2.61/full Once again it did really surprisingly well. There was one weird audio screwup, and they focused a moderate amount on an unimportant finding but also had a line I loved: “sharper images leading to sharper insights” (meaning higher spatial resolution of satellite imagery is more accurate and makes for better decisions). Damn that's good!
3. OK, but what about a paper even a proper scientist would have a hard time slogging through? I literally searched all the notes I have written in my reference library notes for the phrase "hard to read" and found this doozy looking at how cow genetics affects methane emissions: https://journals.plos.org/plosgenetics/article?id=10.1371/journal.pgen.1005846 The writing is very technical BUT at least for a reader like me you can skim and sift out some high level conclusions so I was curious to see how it would perform. The answer is still quite well with one big caveat! The use of sire progeny groups was charmingly described as "like a bovine family reunion, but for science" and they had some great metaphors like how a microbe has a "key" to unlock the "lock" of some important sugar molecules. BUT they described the results as finding a MASSIVE and INSANE impact of genetics on methane and how it was a TOTAL GAME CHANGER. In fact, while the authors never clearly state this (though it's obvious in Fig 1), genetic impacts are MUCH smaller than the impact of diet (grass / forage vs. concentrated feed). So like a cow's stomach, this tool will regurgitate back what you put in w/o any filtering or warning.
4. What if you upload a blatantly false document? You guessed it - the tool performs the same way and makes it seem engaging and interesting. A colleague uploaded an Onion article and the tool did NOT waver or flag any parody or provide any warning, it just made it sound like real news. This one is pretty scary.
Note that there were literally no other prompts or edits to produce this, just upload a PDF and click "generate" which I find staggering. I would have believed that these were real podcasts w/ real people, and it definitely took the core content and made it interesting and engaging (albeit over-hyped). Hearing three of these in a row, there is some noticeable repetition in key phrases and style which will make my spidey-sense tingle if I hear similar podcasts out in the wild not attributed to Gemini Notebook.
As the always wise Bob Lalasz from Science+Story noted - this tool cannot (yet) simulate specific voices like a researcher or celebrity, and the output will quickly saturate and land as inauthentic. But at this moment it's both a tool that can provide some useful and engaging summaries, AND could be a way to make misinformation seem more appealing, as the Onion example shows.
Enough AI, here's a review of one actual paper I read w/o AI helping:
OK, for years people have been hyping the potential of a kind of seaweed (Asparagopsis) to reduce methane emissions in cattle. But the in vitro evidence was mixed - with potential toxicity and downsides a concern if the dose wasn't just right. George et al is a live trial with good news! They found methane emissions from cattle given an Asparagopsis additive were cut roughly in half compared to control, with 6.6% higher weight gain per unit of feed and no substantial impacts on quality or health (fat color maybe a bit better, tenderness maybe a bit worse). The methane reduction peaked at day 21 and declined afterwards, but since most cattle are in feedlots only ~3 months (ranging from 1.5-4) the decline after day 100 is pretty moot. That's a lot of potential! The only potential downsides were ~50% higher bromine residues in kidney and muscle (I couldn't quickly find guidance on safe levels). Caveats: 1) there were no conflicts declared but the authors appear to almost all work as feedlot consultants and it's single rather than double-blinded, 2) in the US the feedlot phase is only about 15% of the lifecycle emissions of a cow (and of that, some is from growing crops and nitrous oxide) so TOTAL impact on CO2e / kg beef is not as dramatic. Overall my take is 1) this should absolutely be tested and replicated more - anything that reduces the very high carbon of beef is worth pursuing. BUT 2) if this is marketed as "green beef" or license to avoid reductions, it could be net harmful for the climate. and 3) IF spraying this solution on grass worked similarly and didn't inhibit development of calves, the impact could be much higher (since ~80%ish of a cow's life is grazing prior to feedlot, and methane emissions are higher on grass than feed). So let's test that too. There's a blog about this at https://www.theguardian.com/australia-news/article/2024/aug/18/feeding-seaweed-supplement-to-cattle-halved-methane-emissions-in-australian-feedlot-study-finds
As always, if you know someone who wants to sign up to receive these summaries, they can do so at http://bit.ly/sciencejon (no need to email me).
REFERENCES:
Fisher, J. R. B., Acosta, E. A., Dennedy-Frank, P. J., Kroeger, T., & Boucher, T. M. (2018). Impact of satellite imagery spatial resolution on land use classification accuracy and modeled water quality. Remote Sensing in Ecology and Conservation, 4(2), 137–149. https://doi.org/10.1002/rse2.61
George, M. M., Platts, S. V., Berry, B. A., Miller, M. F., Carlock, A. M., Horton, T. M., & George, M. H. (2024). Effect of SeaFeed, a canola oil infused with Asparagopsis armata , on methane emissions, animal health, performance, and carcass characteristics of Angus feedlot cattle. Translational Animal Science, 8(August). https://doi.org/10.1093/tas/txae116
Roehe, R., Dewhurst, R. J., Duthie, C.-A., Rooke, J. A., McKain, N., Ross, D. W., Hyslop, J. J., Waterhouse, A., Freeman, T. C., Watson, M., & Wallace, R. J. (2016). Bovine Host Genetic Variation Influences Rumen Microbial Methane Production with Best Selection Criterion for Low Methane Emitting and Efficiently Feed Converting Hosts Based on Metagenomic Gene Abundance. PLOS Genetics, 12(2), e1005846. https://doi.org/10.1371/journal.pgen.1005846
Sincerely,
Jon
p.s. This hairless cat named Sylvester lives near me, is extremely friendly, and is as weird to touch as you might guess
