The scientific article seems to be open access [1].
Before people draw links to recent large language model breakthroughs: Although they do use techniques from computational linguistics, there are no neural networks involved. This is more like old-school AI.
They have essentially a giant optimization problem, and they (approximately) model it as a lattice parsing problem, with a stochastic context-free grammar. They can solve that to optimality in O(n^3), which is too slow for some applications. So they propose a O(n) heuristic (hence no optimality guarantees, but the model was approximate to begin with anyways, and the heuristic is a lot faster), which is the reason for the name of their code: "LinearDesign".
"The lattice parsing problem refers to the task of parsing a word lattice, which is a graph structure that represents multiple possible sequences of words that could have generated a given speech signal [1]. The word lattice is a weighted directed acyclic graph, where each node represents a word hypothesis and each edge represents a transition between two words. The weights on the edges represent the likelihood of the transition. The goal of lattice parsing is to find the most likely sequence of words that generated the speech signal, given the word lattice [1]. Lattice parsing is a challenging problem because the word lattice can be very large and contain many alternative paths, making it difficult to find the most likely path efficiently [1]. Several techniques have been proposed to address this problem, including bi-directional LR parsing from an anchor word, augmented chart data structure, and attention shifting for parsing speech [2][3][4]."
(full disclosure this might not be correct, I tried this with an LLM approach we're beta testing at my job called scite Assistant that answers with real references - no hallucinations, just curious how the response is against someone that knows the field a bit more..!)
This linear-time approximation algorithm for mRNA design (LinearDesign) was inspired a lot from their previous work on a linear-time RNA folding algorithm (LinearFold).
Hopefully these ones will be tested longer than the last batch of mRNA designs, so that risks like the doubling of retinal vascular occlusion risk are identified before they go into production: https://www.nature.com/articles/s41541-023-00661-7
Statistically, based on the paper you linked longer testing to pick up this side effect would have been the wrong decision. It seems to me that longer testing would have caused more deaths than would have prevented RVO. From your paper:
Based on the official COVID-19 death reports, it is estimated that vaccinations have prevented 14.4 million excess COVID-19 deaths worldwide between December 2020 and December 202139. Thus, vaccination is the most effective method for preventing the spread of SARS-CoV-2.
The number of reported ophthalmic complications has remained low, and vaccine-related retinal vascular occlusion is very rare, although the number of COVID-19 vaccinations is enormous.
That paragraph is a logical fallacy. To compute a cost/benefit analysis for something you have to include all the costs, not just one. There are many papers like that one covering many different side effects.
Note that there was an attempted discussion of that paper on HN but it was immediately flagged to death. It always happens for every such paper. People still aren't ready to do a full and honest accounting. Probably they never will be.
You wouldn’t copyright an RNA sequence like this. While it is novel, I’m not sure you’d be able to get a copyright on it, as it’s a set sequence.
Instead, you’d patent it. And I’m not aware that that question has yet been asked. Especially when computationally driven drug design has a rich history, I’d expect for AI generated work to be patentable.
Not sure why this would matter, generally medication is incredibly expensive because of the research that needs to go into it not scaling well to the amount of people that might need it for rare disease. If drug discovery costs go down to almost nothing, and trialling is greatly assisted with AI driven protein folding, it would almost become trivial to cure most diseases.
Where are you seeing this? Everything I saw is that they are eligible for copyright, and the copyright is owned by the person who ran the tool and selected the output.
The news was dumb media bait where people tried to claim the AI itself owned the copyright which makes no sense since AI is not a legal entity able to own things.
Judge declared it about a month ago. Positon was the human edited parts were but machine generated parts were not. Sorry don’t have the reference at hand.
Of course it is. It's artificial and makes intelligent decisions. So is a pile of if sentences if they encode logical reasoning that applies some knowledge.
If one had a database of all answers to every question in the universe, making AGI would be as simple as fetch call.
This is an optimization "AI", like a computer chess program, right? It's not determining in advance which kinds of structures are going to be most stable and generating them, it's working toward pre-determined structural optima?
AI is usually a marketing term. They used machine learning to learn the topology of a prediction space related to desirable properties of mRNA vaccines.
I could be wrong, but that makes it sound algorithmically similar to AlphaFold 1 where they used ML to learn properties of the folded protein and then optimized the final structure to fit those properties
Readit News
Before people draw links to recent large language model breakthroughs: Although they do use techniques from computational linguistics, there are no neural networks involved. This is more like old-school AI.
They have essentially a giant optimization problem, and they (approximately) model it as a lattice parsing problem, with a stochastic context-free grammar. They can solve that to optimality in O(n^3), which is too slow for some applications. So they propose a O(n) heuristic (hence no optimality guarantees, but the model was approximate to begin with anyways, and the heuristic is a lot faster), which is the reason for the name of their code: "LinearDesign".
[1] https://www.nature.com/articles/s41586-023-06127-z
1. https://doi.org/10.21437/interspeech.2016-1583
2. https://doi.org/10.3115/997939.997950
3. https://dl.acm.org/doi/10.3115/991146.991188
4. https://doi.org/10.21236/ada105028
----
(full disclosure this might not be correct, I tried this with an LLM approach we're beta testing at my job called scite Assistant that answers with real references - no hallucinations, just curious how the response is against someone that knows the field a bit more..!)
We re-wrote their C implementation[0] of LinearFold in Go and added comments to explain how the algorithm worked: https://github.com/allyourbasepair/rbscalculator/blob/main/l...
[0] - https://github.com/LinearFold/LinearFold
Based on the official COVID-19 death reports, it is estimated that vaccinations have prevented 14.4 million excess COVID-19 deaths worldwide between December 2020 and December 202139. Thus, vaccination is the most effective method for preventing the spread of SARS-CoV-2.
The number of reported ophthalmic complications has remained low, and vaccine-related retinal vascular occlusion is very rare, although the number of COVID-19 vaccinations is enormous.
Note that there was an attempted discussion of that paper on HN but it was immediately flagged to death. It always happens for every such paper. People still aren't ready to do a full and honest accounting. Probably they never will be.
Instead, you’d patent it. And I’m not aware that that question has yet been asked. Especially when computationally driven drug design has a rich history, I’d expect for AI generated work to be patentable.
The news was dumb media bait where people tried to claim the AI itself owned the copyright which makes no sense since AI is not a legal entity able to own things.
If one had a database of all answers to every question in the universe, making AGI would be as simple as fetch call.
Dead Comment
Dead Comment