llm

Beware of academic "papers" about LLMs, especially on arXiv. The signal-to-noise ratio is alarmingly low, especially relative to other subdisciplines of computer science: https://wandering.shop/@xgranade/114055293092825449

#AI #GenAI #LLM #AcademicPublishing #arXiv #ComputerScience

Speaking of widespread low-quality scientific publication and the need to take care with words: https://retractionwatch.com/2025/02/10/vegetative-electron-microscopy-fingerprint-paper-mill/

The phrase was so strange it would have stood out even to a non-scientist. Yet “vegetative electron microscopy” had already made it past reviewers and editors at several journals when a Russian chemist and scientific sleuth noticed the odd wording in a now-retracted paper in Springer Nature’s Environmental Science and Pollution Research.

Today, a Google Scholar search turns up nearly two dozen articles that refer to “vegetative electron microscopy” or “vegetative electron microscope,” including a paper from 2024 whose senior author is an editor at Elsevier, Retraction Watch has learned. The publisher told us it was “content” with the wording.

Note the presence of Nature publishing group, notorious lately for their low-quality AI slop or AI-boosterism, and Elsevier, who is generally terrible.

#AI #GenerativeAI #LLM #AISlop #InformationOilSpill #AcademicPublishing #ScientificPublishing #PaperMill #PeerReview

My lab's using an LLM in an experiment for the first time. It's interesting to see how that's going.

For one thing, we (roughly a dozen AI experts) struggle to understand whether this thing is doing what we want. It's just such an ambiguous interface! We send it some text and a picture and get text back, but what is it doing? We're forced to run side experiments just to validate this one component. That makes me uncomfortable, and wonder why folks who aren't AI researchers would do such a thing.

Worse, my lab mate keeps doing more prompt engineering, data pre-processing, and restricting the LLM's vocabulary to make it work. That's a lot of effort the LLM was meant to take care of which is becoming our problem instead.

It feels like he's incrementally developing a domain specific language for this project, and all the LLM is doing is translating between English into this DSL! If that's the case, then there's no point in using an LLM, but it's hard to tell when we've crossed that line.

#ai #science #llm

I put the text below on LinkedIn in response to a post there and figured I'd share it here too because it's a bit of a step from what I've been posting previously on this topic and might be of some use to someone.

In retrospect I might have written non-sense in place of nonsense.

If you're in tech the Han reference might be a bit out of your comfort zone, but Andrews is accessible and measured.

It's nonsense to say that coding will be replaced with "good judgment". There's a presupposition behind that, a worldview, that can't possibly fly. It's sometimes called the theory-free ideal: given enough data, we don't need theory to understand the world. It surfaces in AI/LLM/programming rhetoric in the form that we don't need to code anymore because LLM's can do most of it. Programming is a form of theory-building (and understanding), while LLMs are vast fuzzy data store and retrieval systems, so the theory-free ideal dictates the latter can/should replace the former. But it only takes a moment's reflection to see that nothing, let alone programming, can be theory-free; it's a kind of "view from nowhere" way of thinking, an attempt to resurrect Laplace's demon that ignores everything we've learned in the >200 years since Laplace forwarded that idea. In that respect it's a (neo)reactionary viewpoint, and it's maybe not a coincidence that people with neoreactionary politics tend to hold it. Anyone who needs a more formal argument can read Mel Andrews's The Immortal Science of ML: Machine Learning & the Theory-Free Ideal, or Byung-Chul Han's Psychopolitics (which argues, among other things, that this is a nihilistic).

#AI #GenAI #GenerativeAI #LLM #coding #dev #tech #SoftwareDevelopment #programming #nihilism #LinkedIn

You are not doing science research if you stuff an LLM into the critical path of your experiments. You are, instead, producing science-shaped artifacts with the same peripheral relationship to science that LLM text output has to truth.

The reason for this shouldn't be hard to see but apparently is. Simplistically, science is about hypothesis-driven investigation of research questions. You formulate the question first, you derive hypotheses from it, and then you make observations designed to tell you something about the hypotheses. (1)(2) If you stuff an LLM in what should be the observations part, you are not performing observations relevant to your hypothesis, you are filtering what might have been observations through a black box. If you knew how to de-convolve the LLM's response function from the signal that matters to your question, maybe you'd be OK, but nobody knows how to do that. (3)

If you stick an LLM in the question-generating part, or the hypothesis-generating part, then forget it, at that point you're playing a scientistic video game. The possibility of a scientific discovery coming out of it is the same as the possibility of getting physically wet while watching a computer simulation of rain. (4)

If you stick an LLM in the communication part, then you're putting yourself on the Retraction Watch list, not communicating.

#science #LLM #AI #GenAI #GenerativeAI #AIHype #hype

(1) I know this is a cartoonishly simple view of science, but I do firmly believe that something along these lines is the backbone of it, however real-world messy it becomes in practice.
(2) A large number of computer scientists are very sloppy about this process--and I have been in the past too--but that does not mean it should be condoned.
(3) Things are so dire that very few even seem to have the thought that this is something you should try to do.
(4) Yes, you might discover something while watching the LLM glop, but that's you, the human being, making the discovery, not the AI, in a chance manner despite the process, not in a systematic manner enhanced by the process. You could likewise accidentally spill a glass of water on yourself while watching RainSim.

My god..

#AI and with that, I mean #LLM's are doing SO MUCH damage

#Laracast is another one who's quitting because of it

I really freaking HATE this

https://www.youtube.com/watch?v=g_Bvo0tsD9s

An Open Letter to Cory Doctorow: Ollama is part of the enshittification! · brennan.day https://brennan.day/an-open-letter-to-cory-doctorow-ollama-is-part-of-the-enshittification/ #AI #LLM #AIHype

I wrote about using a website's search input to control my smart home (and other things)

https://tomcasavant.com/your-search-button-powers-my-smart-home/

This warms me in the meow meow

"How Replacing Developers With AI is Going Horribly Wrong"

#AI #LLM

https://www.youtube.com/watch?v=ts0nH_pSAdM

Can someone clarify, in academia and industry are LLM hallucinations the result of overfitting, or simply a false positive?

I'm beginning to think that hallucinations are evidence of overfitting. It seems surprising that there are few attempts to articulate the underlying cause of hallucinations. Also, if the issue is overfitting, then increasing training time and datasets may not be an appropriate solution to the problem of hallucinations.

#AI #ML #llm

A few inchoate thoughts on Gas Town, since I think this example has more to it than “it’s just a meth binge/crypto scam/one-shot AI poisoning”. Part of the reason I think this is that some of the rhetoric it deploys dovetails perfectly with broader trends and phenomena, and I think it's worth pulling those out.

1. Economists from the physiocrats (18th century) onward promised society freedom from material deprivation and hard physical labor in exchange for submitting to an economic arrangement of society
2. In a country like the US, material deprivation and hard physical labor have been significantly reduced since then:

• Though too many clearly still suffer too much, a large proportion of people live free from fear of starvation or lack of shelter
• The US has deindustralized, meaning hard physical labor is not the reality for a lot of people. For a lot of people labor is emotional or symbolic (“knowledge work”)
• In other words, for lots of people the economic promise has been fulfilled

3. Having to think hard is one of the service economy’s analogs for hard physical labor. If the promise of economics is to be continually pursued--meaning it maintains the promise that if we collectively submit to it, in exchange we will enjoy a freedom--a natural target of the promise is providing freedom from the need to think hard

• It is not coincidental that “Gas Town”’s announcement post mentioned Towers of Hanoi, an undergraduate CS student homework problem that for most students requires thinking hard. It’s designed to encourage a kind of “eureka” moment where recursion as a computer programming technique becomes more clear. GT claims to fulfill the promise of not having to think hard like this anymore: the LLMs will do that thinking for you
• It is not coincidental that Gas Town is described as being very expensive. Economic power in the form of asset accumulation is what earns you freedom in this way of conceiving things. If you want the freedom from having to think hard, you’d better accumulate assets
• Since the promise is greater collective freedom, endeavoring to accumulate assets is, in this view, a collective good
• This differs from effective altruism and other “do good by doing well” conceptions. Rather, the very mechanism of economics produces collective wealth, so the story goes, which means the more active one is as an economic agent, the more collective good one produces (“wealth” and “good” being conflated)
• Accumulation of assets is the scorecard, so to speak, of such enhanced economic activity, and the individual reward can then be freedom from having to think hard

4. Expending significant resources is viewed as a good in itself from a (naive) evolutionary perspective

• Lotka’s maximum power principle (supposedly) dictates that those entities that transform the most power into useful organization are most fit from an evolutionary standpoint
• Ernst Juenger’s notion of “total mobilization” brings this principle to politics/political economy/geopolitics: those nations that “totally” mobilize their national resources are the ones that will dominate geopolitically
• See, for instance, the RAND Corporation’s Commission on the National Defense Strategy: “The Commission finds that the U.S. military lacks both the capabilities and the capacity required to be confident it can deter and prevail in combat. It needs to do a better job of incorporating new technology at scale; field more and higher-capability platforms, software, and munitions; and deploy innovative operational concepts to employ them together better.” (emphasis mine). In summary: the US is about to be outcompeted (lacks fitness); in response, it should go big (“at scale”, “more”) in an organized way (“deploy innovative operational concepts”, “employ them together better”)
• The rhetoric around LLM-based AI includes similar language, exemplified in the GT post: burn through as much infrastructural resources as possible to produce organized outputs “at scale”, while avoiding having human beings think too hard to produce those outputs, an indication that the power was burned to produce useful organization
• LLM-based AI plays a prominent role in US federal government strategy, particularly military strategy, with language about dominance serving to justify its use
• It is not coincidental that Gas Town uses many orders of magnitude more resources to solve the Towers of Hanoi problem (“Burn All The Gas” Town). This rhetoric dovetails perfectly with the “total mobilization” concept

#ComputerScience #AI #GenAI #GenerativeAI #LLM #GasTown #economics #eugenics #MaximumPowerPrinciple #evolution #EvolutionaryTheory #Darwinism #USPol #US

Parece que pessoas já acham normal pagar mico colando diretamente o misturadão de lavagem do que raciocinar um minuto pra escrever um ou dois parágrafos, mesmo que fosse com ajuda da máquina para pesquisar antes.

* Fica de proposta de tradução de slop remix.

#subtoot #slop #lavagem #leroLero #LLM #noAI

Anthropic secretly installs spyware when you install Claude Desktop
https://www.thatprivacyguy.com/blog/anthropic-spyware/

#claude #ai #llm #privacy #cybersecurity #spyware #fuckai #stopai #dataprivacy #anthropic

AI Slop companies now advertise their crimes to hype up their slop bots, basically acknowledging their "product" is total worthless and does not do what it's instructed

All the while stealing our content to train on, flattening whole forests to make space for their crap centers, consume water and electricity meant for people

And they don't have a clue about the endgoal

What a terrible joke..

#LLM #AI #slop

BBC: AI chatbots could be making you stupider

https://www.bbc.com/future/article/20260417-ai-chatbots-could-be-making-you-stupider

#ai #llm

We're going to put together a resource for musicians on how they can have album art without using large language models. Do you have any resources you'd like to see in that? Free image libraries? Friends who can help do design? Basic how-tos?

#mutualaid #solidarity #helpingeachother #mirlo #resources #ai #llm

@poll @board
你觉得哪家公司LLM的模型最好？

#无聊投票 #LLM

“I use ChatGPT for ideas and inspiration” is the new “I read the Playboy for the articles.”

#LLM #AI

The case of “vegetative electron microscopy” illustrated here shows what is badly needed in current #LLM research and has implications far beyond. We need tools that help us curate huge corpora. We need to be able to trace #hallucinations back to the training data and understand what are the specific (to a surprise, often #deterministic) reasons in the model input that cause that particular output.

If anyone is interested in collaborating on this, I'm in, have done some small-scale experiments and have already submitted a grant proposal.
https://theconversation.com/a-weird-phrase-is-plaguing-scientific-papers-and-we-traced-it-back-to-a-glitch-in-ai-training-data-254463

#lispyGopherClimate #lisp #technology #podcast #archive, #climate #haiku by @kentpitman
https://communitymedia.video/w/c3GdAXe7BQTbK3VrcXCm7E
& @ramin_hal9001
On the #climate I would like to talk about the company that found #curl and #openssl's #deeplearning many (10ish) 0-day vulns "using #ai ". (#llm s were involved).

This obviously relates to my #lisp #symbolic #DL https://screwlisp.small-web.org/conditions/symbolic-d-l/ (ffnn equiv). Thanks to everyone involved with that so far.

I implemented that using #commonLisp #condition handling viz KMP.

Somehow this feels very weird

Even though it's with renewable energy, the problem is that is still uses that energy that could be used a lot better otherwise

#AI and with that #LLM's are really a virus

#lispygopherclimate #lisp #lambdamoo #programming #podcast #live https://communitymedia.video/w/5vAGot7LujjpFQ5Mzz6bFM
0UTC Wednesdays Weekly
@kentpitman #climateCrisis
Breaking The Complexity Barrier (Again) (Again) vs #LLM #AI .
(+ Reliably #jailbreaking LLM AI)
Capitalization and lisp
Following Terry's lead, My #McCLIM, programs and the #racket people
Talkin' 'bout my generation #gopher

telnet lambda.moo.mud.org 8888
co guest
@join screwtape
"yo<RET>
:wave

#unix_surrealism
@hairylarry @nosrednayduj @mdhughes et al.!

A first essential condition on the cognitive is that cognitive states must involve intrinsic, non-derived content. Strings of symbols on the printed page mean what they do in virtue of conventional associations between them and words of language. Numerals of various sorts represent the numbers they do in virtue of social agreements and practices. The representational capacity of orthography is in this way derived from the representational capacities of cognitive agents. By contrast, the cognitive states in normal cognitive agents do not derive their meanings from conventions or social practices.

Adams & Aizawa, The bounds of cognition

#AI #GenAI #GenerativeAI #LLM

In-person Certified #Scrum Developer training in Seoul, South #Korea this summer (16-17 Jul)!

Also: 18 Jul "Essential TDD w/ #AI / #LLM Assist" workshop. ( https://agile-korea.com/tdd-ai/ ) An additional 10% off for both.

https://www.scrumalliance.org/courses-events/search/coursedetail?id=202504181

It's hilarious (*) that "JSON prompting" is a thing and AI people say "prompt engineering is dead; you should be JSON prompting".

You mean to say it's more effective to use a structured syntax to instruct the computer to do what you want? That's amazing! But I feel like I've heard that somewhere before? 🤔

💡 I wonder whether having a process that rewrites the JSON you enter into other JSON according to some rules---also expressed in JSON of course---would be helpful sometimes? We could call it "JSON programming", maybe.

#AI #GenAI #GenerativeAI #AgenticAI #LLM #Claude #PromptEngineering #JSONPrompting #ComputerScience


(*) Hilarious in the facepalm sort of way.