Alexander Schmidt-Lebuhn

That's what is so odd to me: I have many colleagues who use LLMs for various purposes. Nearly every time I try them, they fail me, because they don't know anything or are too unreliable (stochastic). Am I just unlucky? Or does everybody else suffer from Gell Mann Amnesia?
bsky.app/profile/ansc...

I must admit I am not trying it myself, because I am capable of writing things like "send abc to xyz" into my notebook, so I cannot speak from personal experience. But to my understanding, my colleague who said that it gets lots of things wrong was also using Copilot.

Ah, I misread your intention, sorry.

In fairness, a full recording isn't the point here, it is a summary, minutes, or 'actions', as the jargon goes. But as long as people spend that much time correcting the LLM's mistakes, what's the point? Taking notes isn't even that much work in the first place, and, as you say, a good exercise.

A colleague made that case and unironically followed it up by saying how long he spends correcting the summary because the LLM gets most people's names wrong etc. It seems easier to say, "Dan, can you take notes today?"

There are real uses, though: coding, spam, plagiarism, cheating at term papers.

The first question is what bullish means here.

The tech is useful to coders, and although OpenAI and Anthropic will go bust, the tech will persist because open source LLMs can be run locally?

Or:

In five years, LLMs will make Hollywood movies and "solve physics"?

There should be world in which one can say, "many of my colleagues say they find LLMs useful for coding" without being a shill *and* "this IP theft should be illegal, and would LLM users pay fees that reflect real costs?" without being being an irrational hater. Alas, our world is not that world.

Can you get it to explain how they will become profitable? I think Altman at some point said that was the plan, ask the very expensive LLM how to make money with it.

So, why do people like Yglesias see that as innovation? Because they are incapable of thinking further than a single step. Taxi cheaper! Good for consumer! Yay! ... without realising that nearly every consumer is also a worker, and we are all collectively losing by undermining labour protections.

Great read. Much of the innovation of the last 20 years isn't innovation but merely increased exploitation: taxis and food delivery, but without decent pay and job security; hotels, but without standards; financial securities, but without consumer protections; book sales, but as monopoly & monopsy.

I didn't know the bar in that painting was called PWUles, so that's interesting new information.

bsky.app/profile/nytp...

"Organisms" apparently means only animals, because I would be extremely surprised if plants and fungi can "choose" where they live and what they eat.

In fairness, they write "problems where it saw them", which doesn't claim that they were actual problems at any objective level. Which doesn't make this better, of course. We were just having fun! for the attention! ... until the iceberg hit. Oops.

Even if students using their free speech rights to protest is a problem (it isn't), there is an opportunity cost in public discourse. If you keep the lookout on the Titanic distracted with your rants about the wokeness of the passengers, they may not spot the iceberg looming ahead.

Yes, this or something similar was clearly going to happen. Still confused why the EU would agree with him on anything, because to a narcissistic bully, accepting 15% was saying "we are weak and can be extorted". Do they not have a theory of mind?

I think we can imagine things that a virtual representation of a politician could be tricked into saying that would be opposed to their real views or vile. These things are probabilistic and never 100% predictable. If it is the politician's own official bot, that's an accident waiting to happen.

I am not in the UK, so not my concern. But I am trying very hard to understand how this can be "useful" except as a ticking time-bomb. Why would I want a simulacrum of me that could at any moment say or do something deeply embarrassing to me? Guaranteed that some users will try to make that happen.

I need you are an expert botanist. You are a senior student at the University of Genshin (or herbaceous; Annual herb; leaves: light green and white-greenish green coloration. The user has given me some information about a plant growing in the US-Mexican poppy (Lamiacea oculata; Aster species with 18-30 days to flower, but can be found wild in Europe. The annual herb; Annual herb; I don't know what that is. No matter how I try to relax and start from scratch from seed-eating plant? No problem. The first line of the answer should have more than two years ago

Today I tried DeepSeek, and it grew increasingly incoherent over 25 pages worth of response. I didn't think "get the data from this bit of text and put it into this template" was like glimpsing the true form of Nyarlathotep, but something about JSON templates has now driven two open source LLMs mad.

Today is another day when I am confused why US-Americans continue to call former office holders president, vice president, and mayor years after the people in question ceased holding those offices.

The whole point of LLMs is that they are supposed to "understand" human language and respond in human language. If this one responds with ",",",",", I am better off just writing a deterministic script to extract the data and clean up afterwards. This doesn't seem worth the hype.

Beware of the Google AI salesman and its cronies Exposing the overly salesy AI Overviews that will push you to buy bad products and exploring the system making it possible.

This is pretty depressing to read. When I use Google these days (where I can't figure out how to make DuckDuckGo the default), I instinctively scroll down until I find the actual search results, but one should take a step back and realise what they are doing here.

housefresh.com/beware-of-th...

It's truly baffling how many people have no idea how anything works.

in a side"', m",", c",",i/",",",","", fruit", "c",arilc" ir",g," shape",": comp", ", c" t",it"", "f"",e",gates" in" fpl", [">"", ",", ",", (this" is, ", in" ", to ",", arce." inter ", ","", ",",",",",",","",","",", in", "", in"",in"":", in", in c in them"",", "fc" " if", [",", in"",", in",", ", fruit"",",", ",it"",",".", in",",",",""," ","",", in", "\item",",",",", (","",",",","",",",",",", in ain","," ", "", seed",", (",",","","",",", ","," "",", gif c",", (","cane, ",",","",", ar [",", ",",",", t", ",", (sube", ", in ", inter a sp","", scar\t", "f", [",",", in string" hb.

Trying out another LLM in Ollama, this one allegedly optimised to extract structured information from text, and finding that it its outputs are deranged nonsense. Why does this stuff never work reliably for me when so many other people use them for coding, meeting notes, etc.?

It is strange how little it seems to have worried intellectual frauds so far that while LLMs cannot automate nurses, landscapers, lab technicians, or janitors, they can easily automate intellectual fraud.

Aren't these two different issues, though? A hypothetical world in which LLMs are profitable and this wasn't a bubble at all would still be a world in which the business of "write report and create slick slideshow telling a CEO what they wanted to hear" could be replaced with LLMs doing just that.

PhyloControl, a phylogeny visualisation platform for risk analysis in weed biological control, is now available. Designed in consultation with biocontrol researchers, and result of a collaboration led by Stephanie Chen, a CSIRO postdoc in my team.

doi.org/10.1016/j.bi...

That's a problem, then, because the only way it will ever get overturned is if an effective labour movement already exists and campaigns for it to be overturned.

It has been fairly clear for some time what will cause the next large economic crisis, although admittedly, deregulation of cryptocurrencies is another strong contender.

New paper on our software application developed for extracting data from biological specimen images in natural history collections:

"SpeciMate: Improving metadata extraction from digitised biological specimens", published in Biodiversity Data Journal, bdj.pensoft.net/article/1605...