Alexander Schmidt-Lebuhn

@anschmidtlebuhn.bsky.social

369 Followers 111 Following

Botanist, taxonomist, phylogeneticist.

Screenshot of an AI assistant for writing a personal message on electronic farewell or celebration cards, which defeats the purpose of personal messages. It asks to construct a prompt consisting of the occasion, the relationship the user has with the recipient, the tone (funny, formal, inspirational, etc.), and what style to plagiarise ("e.g., Yoda or Dr. Seuss). This egregious waste of electricity could mostly be replaced with a randomised selection of stock phrases. More importantly, if I were ever to receive such an electronic card, I would prefer somebody writing a single misspelled sentence themselves over generating this kind of slop. On the right is what looks like a chat with one alleged message reading "Hi there Welcome to... by Jess". Puzzlingly, the mystery message perennially claims to be one day old regardless of what day I logged in, and there is no context for who or what Jess is. Maybe it is just me, but the entire interface screams fraud.
Screenshot of Google Maps showing a street allegedly named "Now a car park".
Mar Hicks @histoftech.bsky.social For many years, people didn’t randomly film others in public—not because we didn’t have the technology, but because the social contract said you didn’t do that. That’s why Meta is going full court press on their PR and celebrity endorsements. They remember how people ran google glass outta town.
An AI generated image apparently shared on social media by the current president of the USA, or at least by his team under his name. It purports to show a cartoonishly compressed Kharg Island being attacked by US military aircraft. The attacking fighter planes are of a variety of types, including one that has cockpits on both ends and another that looks vaguely like an X-Wing from Star Wars. The distant half of the island and most of the empty beaches are full of explosions, while every attacker seems to have missed the closer half of the island. The airplanes are unrealistically close to each other, so that we can expect several of them to imminently crash into each other. Also, the AI doesn't understand how missiles work, so it looks as if the jets that are meant to be shooting the island are instead being shot down.
A painting of Jesus getting his feet washed by Maria Magdalene, to the amazement of his strangely geriatric followers.
Two concepts illustrated with phylogenetic species trees containing gene trees. Incomplete lineage sorting (ILS) * dataset: shallow phylogeny of very closely related species, and only visible if multiple individuals per species and/or gene copies per individual are used in the analysis * pattern: gene copies in species are not reciprocally monophyletic * cause: rapid speciation and large effective population sizes Deep coalescence * dataset: deep phylogeny of distantly related species, a kind of analysis that often uses only one individual per species and only one gene copy per individual (e.g., HybPiper) * pattern: gene tree incongruent with species tree * cause: incomplete lineage sorting many millions of years ago that is now long resolved, so the LS is not I anymore. It is not incomplete lineage sorting.
Screenshot of a sparkly button with the popup saying, "Have questions about this content? Try AI Companion." Also, AI companion is bolded, so that it screams at the reader. Also, excuse me, but "content" is an interesting choice of words for a research paper.
A ball of red fluff vaguely reminiscent of the inflorescence of a Myrtaceae, but subtended by rose leaves.
A plate reading "Schutze dieses Haus und alle die da gehen ein und aus", or in English: "Protect this house and all those who enter or leave it." However, only the word Haus is really legible because its letters are cut out, whereas the letters of all other words are written in very faded ink.
Photo of a plant looking like a yellow ball on a stick.
Photo of an analog clock without pointers and with its numerals jumbled up more or less randomly.
HANS HEITMAN ANNA MEYERS BAVWEN IST EIN FEINER LVST ABER ICH HABE ES NICHT GEWVST DAS ES HETTE SOVIEL GEKOST ANNO 1665 Translated to English: Hans Heitman. Anna Meyers. Building is great fun, but I had not realised how expensive it would be. Anno 1665
A minuscule, white hair branched into six arms on a leaf surface at high magnification.
Flower-head of Erigeron canadensis, the Canadian fleabane.
A house in the town center of Lüneburg in Germany. The lower three storeys are tall and have large windows, and above that are four much lower storeys that were likely originally used as storage space. The main door is surrounded by a beautifully painted archway two storeys tall. Fraktur letters on the building read "Die alte Raths-Apotheke", meaning The old council pharmacy.
AI-generated advertisement for "American quality pistachios", whatever that means. It shows an enormous number of pistachio-shaped, green people who have identical, utterly vapid happy expressions. They are meant to be watching a football match on a large screen but are all, for inexplicable reasons, looking away from it. Interestingly, the pistachio people in the crowd have shells, but the pistachio people football players on the screen and a single individual standing above the crowd on a table don't.
A plant press made of eight 24 cm pieces and eight 32 cm piences of wood.
This slop image is meant to illustrate the non-consensual use of music for training music generating models. The left half of the image shows a large number of what are presumably meant to be either album covers or screenshots of music videos being sucked into a funnel. The funnel sits atop a machine labelled "AI". The machine spits out wavy lines and notes to its right. As indicated in the post, a good part of the training data misses the funnel of the machine, because the generative AI used to slop out this image does not actually understand image composition.
Garbage image extruded by generative AI to illustrate an essay on the social media bans for children. The left half is taken up by a smartphone. Its screen shows a mass of very similar looking teenagers staring at their smartphones, the hand of an evil puppeteer holding them on strings, and a an assortment of like symbols and melted f symbols floating up above the crowd. In the upper left corner of the screen are are several app logos, and in the lower left corner, a gold-like substance pours out of the phone into a box labelled "PROFIT ADS CONTROL". Very subtle. The on/off and volume control buttons of the phone are on the left side, which seems unusual. The right half of the AI slop garbage image is taken up by the end of a fence. The fence has half of a lock and chain awkwardly floating in the air, implying that the fence is somehow locked, or something. On the fence is a large sign saying "under 16", with the 16 in a forbidden sign. A single sad teenager is standing next to the fence, staring at the large smartphone screen right in front of him. It is implied that he is somehow excluded, although he is in no way kept away from the large smartphone and could easily squeeze between the phone and the fence if he wanted to. He could also help himself to the gold-like substance in the box labelled PROFIT ADS CONTROL, but that is a different matter.
Screenshot of the same Bluesky post this one is a response to, but it shows that the preview text for the link to the job vacancy is "Error".
A client wanted to have a call yesterday because he thought our math was wrong. After we got on the phone he says “I’m checking it on the calculator and it works out, but I’m sure it’s wrong”. I ask him why and he says chatgpt told him. People gotta stop being stupid like this
Quoted from paper presenting Co-Scientist: It is important to emphasize L that these wet-lab experiments function as a viability check of the drug repurposing pipeline, yet they are not a replacement for the rigorous pre-clinical and clinical assessment typically C required for therapeutic validation. They provide an efficient biological reality check allowing us to rapidly evaluate AI-generated hypotheses before committing to more resource-intensive T validation studies necessary for clinical translation.
Fallen tree blocking a cycling path
Cropped from reddit post: "no doubt that US-developed AI models lead the pack by 1–2 generations. But if the average developer or startup cannot afford to use them freely, that leadership means absolutely nothing."
Macro image of the basal half of the involucral bract of an everlasting paper daisy, with an arrow pointing out the petiole or stalk.
Small red chili berries air-drying while hung from curtain rod.
A sparkly symbol indicating generative AI next to the title Reading Assistant. The window helpfully suggests some buttons to press if you want to boil a metaphorical lake instead of spending ten seconds looking at the readily available abstract of the paper. Questions you could ask: Suggested question: What impacts does the rapid expansion of A. gayanus have on protected area values? Suggested question: How does A. gayanus threaten biodiversity within management zones? Suggested question: What are the implications of increased A. gayanus cover for native ecosystems? Actions you could take: Summarize this article Summarize experiments
Not really sure if this can even be described. It is a vaguely three-dimensional attempt at drawing a bar plot by an AI that clearly has no idea how bar plots work, and without any data underlying it. There is a legend next to it that would be more appropriate to a dot plot, and all labels on the alleged figure are random gibberish reminiscent of alien glyphs in science fiction. I really wonder where you can even still access generative AI that is this bad. All the currently available models would have produced more subtly wrong images.
Screenshot of a now deleted Bluesky post making the claim that "LLMs are going to underpin every search engine that researchers use whether they are aware of it or not. Textual embeddings and retrieval augmented generation are simply just better at this than the tools we had previously", which implies that the poster has neither used the tools we had previously nor checked the results of retrieval augmented generation against sources. Let's just say the first thing you should when using one of those LLM-based research tools it to ask it for something where you already know the answer, in this case to do a literature search on something you have just worked on, then see if it understood what the various sources actually meant. I have tried that and concluded that I won't touch them with a ten foot pole.
Screen shot of two Bluesky posts (because both replies and quotes are disabled) saying: Most of the people in my life aren't in tech. And they aren't academic theory professors. It's been really interesting to ask them about where they see AI impacting their jobs and tasks. My brother in shipping logistics: "it's changing everything we are doing" My friend in architecture: "we use it" My law friends: "using it every day" My doctor family members: "using it every day" They all have critiques, too. But a lot more balanced than on this platform.
The title of a scientific publication is provided as Studies on Andropogon gay anus, instead of studies on Andropogon gayanus. I did not edit this. This is what Google Scholar showed me.
Bottom of an acorn of an oak tree seen through a 200x macro lens. Flat, with a jagged ring all around, presumably because that is where the vessels were.
A preview of a journal article shown by BlueSky together with the link to that article, only instead of showing the title of the article, it reads, "Please wait whilst we redirect you". Also, just lookest atst thatst pretentioust littlest "whilst".
LLM tool on the Outlook calendar suggesting what it could do: Identify my top three priorities. Which of my collaborators should I meet with? Review pending invites. Find my next one-on-one with somebody.
Updated Computer Vision Model The new computer vision model has over 1,400 new taxa! Check them out here
A cluster of tubular, red flowers with very short lobes and long styles.
Two flowers of the mangrove Avicennia marina.
Lots of white fluff before a green background. Seriously, these plants are mostly the seeds' parachuting hairs at this stage.
Two books with green covers, but the spines have turned blueish over the years.
Lower side of a fern frond. The margins are slightly revolute, and there are two rows of brown sporangia on each pinnule.
Close-up of flower of native gentian (Gentianella) at Tongariro National Park, New Zealand.
Close-up photograph of a heather plant in flower
Diagram showing that New Zealand is the island worst affected by invasive weeds in the world.
OpenAI and Anthropic are losing billions of dollars per year, even on $200 per month subscriptions. In the long term, LLMs will have to be priced at much more than most people will be ready to pay. When that happens, users can simply switch to running freely available models on-device, so LLMs are here to stay even if OpenAI and Anthropic go bankrupt. I have tried both the free versions of LLMs online and running freely available models on my laptop, and none of them perform very well. You have to sign up for a pro subscription and try one of the cutting-edge models, otherwise you don't get to criticise LLM performance. (Now start over from the top.)
The meme or the dog refusing to hand the ball back to the human but still expecting the ball to be thrown, but edited so that the dog says "please profit", "no paying for labor", and "only profit" instead.
circular reasoning works because circular reasoning works because etc.
A cat resting on a mouse pad and blocking access to a keyboard.
Sichuan pepper fruits on the plant
A cartoon in three panels, with the following speech, all delivered by the same phone operator. Nine one one what is your emergency? What do you mean you are being murdered? That's illegal, people can't do that.
Chat with Copilot in PowerPoint. The available responses are "not now" and "chat now". Notably missing is a tick box labelled "don't ask again".
This is the usual graph of CO2 emissions per year from https://ourworldindata.org/co2-emissions, illustrating the primary risk of catastrophe faced by humanity. That risk is also in second to tenth place, after which we maybe have resource overuse, and then global warming again in places twelve to fifty. CO2 emissions as depicted in the graph are rising continuously except during years of global economic crisis, only to resume their increase year by year immediately after. This means that despite all the hopeful stories about solar power, we have not even started to take the foot off the gas while driving towards a concrete wall, we are instead still accelerating. Asmodei's company uses an enormous amount of energy to run its plagiarism bot, contributing to the problem.
Spring-flowering in the arid zone of Australia, here New South Wales in 2013. The entire landscape is white to the horizon except for a few dots of yellow, but I assume the problem is that these are not fleshy-fruited taxa but mostly Asteraceae (daisy family) and Goodeniaceae, which apparently were not part of the data set.
Detail of a shrubby everlasting daisy showing several flower-heads. In the lower right corner is a blackish ant with worryingly large, serrated mandibles.
A post arguing that Meta glasses should be tolerated despite their use by, as they wrote in a previous post that is now blocked, "creeps". It says: They'll always find a way (the "creeps", that is). I love my glasses. Super useful for handsfree video/pics while doing outdoor sport activities as well as music and driving directions. Looking forward to also trying the live translate next time I travel.
A cartoon of three frames comparing transit with fossil fuel vehicles, electric vehicles, and "walking, biking, and public transit". The first two frames are identical masses of asphalt and numerous cars stuck in traffic jams. The third is a livable city in which humans are centered. Although cars with automated driving were not the topic of this cartoon, the same problem applies to them.
Macro photo of a flower of Thai Basil.
Google Scholar search results to the query "taxonomic revision euchiton" showing the first six results. All are taxonomic papers that in some way deal with the cudweed genus Euchiton.
AI-generated alleged art showing two elephants made of watermelons. The larger elephant spouts a stream of melon juice from its trunk, which in this context logically would mean that it has a severe nosebleed. The shadows cast by the elephants on the ground go in the right direction, but the shadows on their bodies are wrong, as if a second sun is behind and to the right of the viewer. The background is a perfectly flat landscape of short-cropped grass. The sky is lightly cloudy. The overall impression is of extremely ugly AI-generated watermelon elephants having been imposed on one of the classic default desktop wallpaper images of Windows except, as mentioned, the ground is too unnaturally flat. The text under the image reads, Turn your words into high-quality videos on Gemini. Share, save, or spark inspiration. The text on the image reads, Let your imagination run wild with Gemini.