#LLMs can be so weird sometimes. Claude insists on converting ’ (U+2019) to ' (U+0027) in its output, and it's almost impossible to get it to acknowledge the change (loves gas lighting!) or to not do that change.
I assume it's some random thing they've done during training, but it's quite annoying.
This is an example for another latent dysfunction:
Fewer (public) questions get asked on the internet and so knowledge is not spread, but contained, making it more individualized.
This also creates an even stronger bias towards older content, so people might take the shortcut and use a more established technology, instead of looking into new, less explored, but more innovative solutions.
I’ve just recently (2 months) started experimenting with #llms for coding related tasks. Now my social media bubble here on mastodon, which I also just rejoined recently, is madly viotriolic against LLMs.
I am increasingly leaning towards the notion that LLMs can definitely be an asset when dealing with code. Use them like a tool.
But I’m also afraid of being ostracized by this community for entertaining such thoughts. Wild times.
A new paper for our AI and Democratic Freedoms series explores whether and how #LLMs can enhance democratic participation, improve political equality, and foster meaningful deliberation. https://knightcolumbia.org/content/can-ai-mediation-improve-democratic-deliberation
First LLM scammer trying to be my friend (that I know of). Here we go.
vibe physics
Angela Collier tears into billionaires and tech bros who think they can push the boundaries of physics by prompting a sycophantic language model.
https://m.youtube.com/watch?si=-AMxQgiyNgoZjQsC&v=TMoz3gSXBcY&feature=youtu.be
How is it legal for LLMs to provide medical facts? Or any facts at all, for that matter...
LLM coding is the opposite of DRY
This is really funny. In Chomskyan times, Chomsky revised his framework every three years so that all computational linguists got frustrated since it takes about three years to develop a computational system with certain basic assumptions. (talk to Ed Stabler or read up in my Grammatical Theory textbook).
Now, linguists are fed up with AI systems, since their architecture changes frequently and they need four years with their experiments to find out how they work.
=:-)
In-context learning has been consistently shown to exceed hand-crafted neural learning algorithms across the board.
But it's limited by the length of the context. Even with neural architectures allowing context to grow to infinity, these come with high costs and scaling problems.
Is there a way to incorporate new knowledge learned in-context back into neural network weights?
Of course there is!
Let's imagine we have a lot of data, sequences of instructions and outputs where in-context learning happens.
From this data we can produce a dataset of synthetic data which presents the new knowledge learned. We can continually train the model with this dataset.
Of course this is super slow and inconvenient. But as a result we'll get a dataset with in-context learning happening, and old model weights against new model weights.
We can use this data to train a neural programmer model directly!
That model would take in the context as such, and if in-context learning has happened in those interactions, it can predict the changes to the neural network weights which would happen if the long and heavy synthetic data pipeline had been run.
Instead of the heavy pipeline, we can just use the neural programmer model to directly update the large model weights based on the in-context learning it experienced, to crystallize the learnings into its long-term memory, not unlike what hippocampus does in the human brain.
"Something that has become undeniable this month is that the best available open weight models now come from the Chinese AI labs.
I continue to have a lot of love for Mistral, Gemma and Llama but my feeling is that Qwen, Moonshot and Z.ai have positively smoked them over the course of July."
PSA: Do not use ChatGPT as a therapist. Think of what's gone into that particular sausage.
LIKE SERIOUSLY.
Disabling Autocorrect because it steals all it's spellings from the dictionary without giving it any credit. 
Writing is hard work, but don't underestimate the motivating power of annoyance.
Do you know, or are you, someone who would benefit from understanding how "AI" tools like ChatGPT really work, why we can never trust them to be correct, and what the ethical concerns are? I've finally finished writing an explainer and I'd really appreciate you sharing it if you find it useful.
VibeCleaner specializes in cleaning up vibecoded software -- "so your product can scale, your team can breathe, and the next dev won't quit."
Does VibeCleaner clean up your vibecoded software with humans or AI? I don't know. If it's AI, then the irony is, using AI to clean up the mess made by AI. Does it work? Nobody knows -- you can't try it yet, but you can join the waitlist.
"Far be it from me to accuse Anthropic of this. When they designed MCP, the idea was to quickly and easily extend chat interfaces with tool functionality (and a whole bunch of other stuff that folks ignore in the protocol!). For that context, it’s actually a good fit for the job (bar some caveats that can easily be fixed).
No, the dünnbrettbohrer of the MCP world are the implementers of the MCP servers themselves. Right now, it’s the peak of the hype cycle of inflated expectations, meaning a lot of people are selling low-code, or no-code, dressed up as MCP — but it’s still the same old shenanigans under the hood.
What I would like to achieve today is to give you simple guidance on when, how, and where to use MCP without shooting yourself in the foot (such as with Github’s latest MCP server disaster, an exploit that left private repository data vulnerable to attackers)."
https://nordicapis.com/mcp-if-you-must-then-do-it-like-this/
#Academia
#AI
#LLM
#LLMs
#AcademicChatter
I *love* #preprint servers, but biorxiv, medrxiv, and research square all said they don't take more opinion type papers.
Hat tip to @WiseWoman for suggesting #arXiv.
I can *FINALLY* say that it's officially released as a preprint.
Differentiating hype from practical applications of large language models in medicine - a primer for healthcare professionals.
https://arxiv.org/abs/2507.19567
DOI: 10.48550/arXiv.2507.19567
