In my experience you can “stop reading the code” if you adhere to prompting the AI aiming for changes that would be reasonable in a PR to review, if you’re using rather opinionated framework(s) to base your work on and explicitly ask for tests (models tend to add them on their own if there are some).
I personally still read the output and start by checking if existing tests were modified (kind of a red flag when this unexpectedly happens imo).
When I read posts about AI generating garbage nowadays it’s either because of a small prompt/big ask combination or a lack of an underlying framework.
Very naive approach but wouldn't training a model on a few specific buckets like "someone is naked", "image is explicit" and "there is a child" in the picture, would do the job without having to train explicitely on CSAM? If the model returns both a high probability of "there is a child" and a high probability of any other bucket than this image is classified as CSAM. It'd be high recall,low precision but one would be on a rather safe side.
If more and more of the web's content is AI generated, AI companies are bound to train on each other's data.
Or, what if I generate content with Claude/ChatGPT/Gemini, warp it in HTML using an open model, put this on my website conveniently dedicated to "Best practices in prompt and AI answers" for example, then train my own model that only scraps my website?
The only ban that should be put in place is on models' general and specific capabilities. If they have open or closed weigths is irrelevant. With the current state of capabilities there is already plenty we can do (in terms of creation and destruction unfortunately) before we as a society need more powerful capabilities.
Once the limit is set, any non- governmental/military entity providing/hosting/using a more capabable model can be prosecuted.
This was harder than I thought it would be. It's pretty fun to play!
May I suggest displaying the final result and/or ongoing progress as 18 circles/shapes that fill up depending on how far you made it?
This guide is also made from me (or some of the me from a couple years back).
I haven't read the whole thing yet and it's probably clearly stated at some point (though one can deduce it with the beginning already) but the surprise for me was that this field is highly statistical. Before starting I had the (very) naive view that it was possible to read the genome as one reads a file and look at what's going on. But the sequencing technics (and accompanying algorithms) only allow to statistically read the genome. So variants/mutations found are only found with a given statistical certainty. If the sample wasn't well prepared for example it could be that this certainty is ultimately not high enough to do a proper analysis/diagnostic.
It's a fascinating field (try to watch a video on sequencing by expansion, to feel how sci-fi this field actually is) that is very hard to approach with only high-school biology level and this guide is really well done to sort of bridge this first gap.
I'm working on a project in malaria genetics this summer, and I was shocked to find out that the entire analysis toolkit is entirely based on math and statistics (and some non-trivial stuff too, e.g. hidden Markov models to predict CNV). Genotype likelihoods throw an extra wrench into the process, since even basic stuff like predicting allele frequencies requires a maximum likelihood estimator instead of simple counting. This whole area was quite eye-opening, and I'm still amazed that reading billions of base-pairs in DNA sequencing reliably works.
To expand this a bit, most sequencing methods are exact, and have a low error rate (except nanopores).
But they produce short reads, and because DNA is full of repetitive fragments, it's not always clear where the read came from.
We also have two copies of genes, which also further complicates matters.
The first startup where I worked, developed synthetic long reads on top of Illumina's hardware. We could stitch together 50kbp reads, which really helped with de-novo sequencing.
And the foundations of those statistical approaches are built on heuristics and shortcuts.
For example, sequencing instruments include base quality strings in the output. Base qualities are estimates how likely the instrument got each sequenced base right. But most people don't want to store that much noise, especially when the actual data is highly compressible. So the base qualities get quantized using more or less principled methods that seem to work well empirically.
Read aligners make similar estimates of how likely they got the correct alignment for each read. Those estimates are typically based on simplistic models and a number of assumptions. There are two main components in the estimate. One is based on comparing the primary alignment the aligner chose to the secondary alignments it also found. Another is an estimate that the aligner didn't find the correct alignment, because that part of the sequenced genome is too different from the reference. The latter is obviously handwavy. And the aligner cheats in the former. Because people don't want to wait 10x or 100x longer for better results, the aligner gives up early and estimates how good secondary alignments it might have found if it had actually done the work.
And then there is variant calling. At some point, the state-of-the-art callers were statistical. But then people got better results with neural networks. Or at least the results were empirically better.
Biology is often an intensely statistics-heavy field. A remarkably large part of statistics was developed to study issues in biology, particularly dealing with evolution and ecology.
All one needs to do is look at the Claude Science thread here last week and note how many comments were surprised that it appeared to be a statistical/analysis tool.
Just a couple days ago I argued with an HN poster who quipped that biology is stamp collecting. A non-negligible number of "mathy" engineer types (not actual mathematicians, those usually understand the complexity of biology and even gladly contribute to the field) seem to think all biologists are quirky eccentrics dedicating 30 years to a single protein or a species of ants in the Kalahari desert. (Not that these don't exist or aren't worthy of respect, but they don't score high in the sophomoric 'hardness scale' of fields that these mathy types still subscribe to)
The rub is of course it is actually far harder to work on your rare species of ant than to do something more "noble" like human genetics. You may have to build the reference genome yourself working on your ant before you can begin with other work. Collecting your ant samples and processing them eventually into raw sequence reads. You may have to optimize this library preparation process yourself if you are really in uncharted territory here.
Meanwhile human geneticist doesn't even need to collect any data. Reference genomes are always being improved. People dump their data into public repositories (at least public for other credentialed researchers). A couple emails and filling of approval forms and you too can have access to 10,000 patient samples of some human disease already sequenced for you to an acceptable depth. Of course you will still need to pay for downstream compute needs in money and your time crafting the analysis to suit your reasoning, but still, half the battle is already won when you work on these well trodden paths. So much necessary groundwork has been performed by others for you already.
I think that it's just because a lot of that is grunt work, a bit like getting a PhD by studying mice, with a little bit of theory at the end of a lot of repetitive work, whereas maths is generally the complete opposite of that: no practical, all cerebral work.
I'm not a maths person; that's just what I've observed.
There always seems to be a difference between the "normal" stamp collecting biology, and the anatomy/genetics biology. The latter part is always more fascinating to me (and I'm better at it) than the former. In essence, intra-organism biology >> inter-organism biology.
I would be curious to know if the WGS was done using short read or long read. There is a lot of hope that long read is "just" what is needed to catch these type of cases nowadays (not sure about this particular one).
I'm in the process of getting my first batch of long reads, but I am skeptical that this is the "just" what's needed. There is little doubt that long read > short read, but I think that computational techniques for both need to be improved significantly.
There is already some clinical evidence to support my hypothesis. The first clinical long read trial at Kansas City Mercy showed a 10% bump in diagnostic rate, which is great but not fully solving the problem: https://news.childrensmercy.org/unlocking-answers-faster-chi...
When I read posts about AI generating garbage nowadays it’s either because of a small prompt/big ask combination or a lack of an underlying framework.
reply