4 ms·
> When the machines were done ... they could classify a cell they had never seen before as one of over 1,000 different types. One of those was the Norn cell. S
by timr 3y ago
> When the machines were done ... they could classify a cell they had never seen before as one of over 1,000 different types. One of those was the Norn cell.
So, unsupervised classification. Ten years ago we were using K-means clustering to do the same thing, but I didn't know we were learning what it means to be alive! </s>
- uoaei 3y agoNo, you don't understand! It uses words in plausibly English ways! Everyone knows that natural language is composed of meaning!
- Vecr 3y agoThe sequence prediction arguments are a lot more convincing than they were a few years ago, if you care about realized progress. Sequence prediction is intelligence, the philosophers can deal with the fuzzy edge stuff that's less important.
- timr 3y agoMarkov models were intelligent? Neat!
- Vecr 3y agoNo, they aren't efficient enough. And the "typical" toy versions pretty much don't work. LLMs demonstrated the feasibility of sequence predictors in a much stronger way.
- uoaei 3y ago> aren't efficient enough Unless you can point to a hard boundary, a line that distinguishes "enough" from "not enough", this doesn't mean anything.
- ProllyInfamous 3y agoPlease define consciousness [impossible].
- ProllyInfamous 3y agoWhat are your thoughts on this part, which had a 50% "novel correct" rate: >GeneFormer recommended reducing the activity of four genes that had never before been linked to heart disease. ?
- timr 3y agoLiterally every gene expression model ever published has come with some claim of novel biological discovery. If they didn't, they wouldn't be published in the biological literature -- nobody serious cares that a classifier was applied to a dataset. Good papers will go so far as to take a random sample (or even better: all) of such "recommendations" and validate them using experiments. Then you attempt to quantify the success rate, the sensitivity, selectivity, etc. Did they do that here? We go back to the article: > When her team put the prediction to the test in real cells called cardiomyocytes, the beating of the heart cells grew weaker. OK. How many predictions did they test? How many failed for the one that succeeded? ...and later: > Dr. Theodoris’s team followed the model’s advice, knocking down each of the four genes. In two out of the four cases, the treatment improved how the cells contracted. So we have two fairly classic examples of cherry-picked experimental validation. How many predictions did they validate? Did they just select the ones they thought made sense? All we know, from this, is that they selected some subset of results to test, and failed roughly half the time in the latter experiment. That's not a great batting average for this kind of work. None of this is new. We've been applying ML to these problems for decades. Wrapping this kind of stuff up in fuzzy claims of AGI is just another way of escaping scientific rigor via application of hype [1]. [1] To be clear, I'm not accusing the researcher of this. For all I know, the underlying paper is good. But this article is dumb.