Maybe for context about what an organoid is, and why this very much matters:
Drug trials on a close-enough vat-grown copy of an human organ gives us the chance to better study drug's potency, effects and side-effects, without harming primates, cheaply, at scale, and closer to in-vivo humans.
In our research, our collaborators coax cells extracted from blood draws all the way into neural organoids with synaptic activity. Absolutely amazing stuff.
IMO, I want to know why working with fetal tissue is needed anymore, especially when it evokes such strong feelings in a large segment of our society e.g. anti-abortion advocates. Right or wrong, it creates LESS trust in science when when we need more trust in science.
Am I reading it correctly that the authors say it's possible to build organic cellular tissues that mimic parts of the human brain (developing typical markers in the different regions of the developing human fetal brain) from human fetal cells of the brain or spinal cord? And these organoids show signs of electrical activity?
"We thus find that the human developing brain is capable of robustly self-organizing in vitro over multiple passages into a 3D, layered structure."
This only works if the article isn't covering anything very novel. If there isn't a fairly wide base of explanations on the topics within, it struggles and hallucinates.
ChatGPT is the ultimate non-human bullshit artist so far developed. It's scary to think LLMs will improve and will be used by used unethically by students, professionals, state actor boiler rooms, and targeted harassment and manipulation of individuals. For the case of proving original work, it's worth considering recording chain-of-custody with an "anti-cheat" app for data capture and written works will be necessary to prove it was generated by or directed by a human being by recording specific data of creation to establish zero-knowledge, zero-reputation provenance.
The next question becomes: When automation and AI are used in drug discovery and STEM ever more autonomously (they are mostly narrow AI initiated by a human with specific goals today), what's the protocol for distributing credit? It seems plausible in the near future with large piles of computing power to train LLMs against journals, give it access to data, drug/protein/ligand databases, and let it find and screen beneficial candidate molecules for critical need medications such as new classes of antibiotics and antimycotics, and rare degenerative diseases, perhaps including gene edits. Efficient, progressive layers of screening methodology seem important. IIRC, 15 years ago (c. 2007) biomedical informatics people at Stanford (SAIL -> SMI -> BMIR) were using NLP heuristics against the literature for meta-analyses. I assume things progressed by several lightyears hence.
Linguistic transformations are the feature that I trust LLMs the most with. It requires relatively little encoding of knowledge outside of language. Parsing jargon to convey novel ideas is absolutely a doable task for a language model.
Yes. If you can necessarily and sufficiently prompt an LLM with authoritative facts, it can usually generate reasonable copy around it. For example, if you were a PwC or McKinsey consultant tasked with rapidly generating a voluminous, glossy deliverable in need of a consistent, formal style, LLMs offer an initial starting point Easy Button for content generation prior to human editing (perhaps assisted by an on-prem "grammarly" app) and pasting into ye olde templates.
That's not going to help you understand the subject material better. It's just going to assemble a facsimile of the lowest common denominator of other people on the internet talking about it. If anything, it will hurt your understanding and give you a false sense of knowledge.
Given how well chatGPT can explain code, I'm reasonably confident it can do much better than the lowest common denominator of people on the internet talking about any given expert document.
May well still give a false sense of knowledge, but that's a very common problem for all of us.
An LLM can reasonably explain code because code is both a closed, finite set of instructions and heavily explained in the training data that contains a vast number of learning materials.
A novel scientific paper is much, much different than code.
> A novel scientific paper is much, much different than code.
Sure, just not in a way that matters.
Human language is a "closed, finite set" of symbols (to the same extent as code, at least), all of which are heavily explained in the training data that contains a vast number of learning materials.
Science is about the novel things; but when the things are so novel the research has to invent new words to discuss these things, those new words come with explanations of what they mean.
It is clear that you do not understand how an LLM works if you believe it can explain a novel scientific paper to you with any reasonable degree of accuracy.
It doesn't look up or use existing explanations of papers like some kind of internet or database search. You have a fundamental misunderstanding of how it all works.
The fact that you don’t know that GPT was primarily trained on public crawler data and feel inclined to make authoritative comments on it, refuting people who do understand how an LLM works is exactly what I mean by “false sense of knowledge”.
BTW the idea was you paste the article into ChatGPT and ask it to explain the text in laymans terms. Maybe that was the point which you misunderstood. Of course it won't know about papers that it hasn't been trained on.
You don't know how it works. The fact that it is a "novel" scientific paper is completely irrelevant. Try it on a new paper just released today, that is definitely not in the training set.
