To use RLHF you need a dataset that includes instructions with good & bad answers - do many of those exist? I know there are a few datasets of just plain instructions-with-responses, but I'm not aware of any that have both good and bad (or ranked) responses. Is that trivial, or an important missing element here?
I agree - LangChain is totally awesome and unlocks amazing new capabilities of LLMs (easily!), but the code itself is pretty awkward. No logging framework, for instance, and the strong typing enforcement with pydantic seems un-pythonic and brittle to me (what if I want to use a different kind of Mappable instead of a dict for something?).
Still, great work overall and congrats to the team!
Somewhat OT, but I feel like this is such an underappreciated aspect of recent LLMs: Not just their ability to generate text - but their apparent effectiveness in making use of arbitrary tools to interact with their environment to achieve some goal. It seems like we're just at the beginning of this with the MRKL and ReAct papers, there will be a ton more awesomeness coming in this area I'm sure.
I'm still a little unclear on the benefits that fusion offers compared to things like wind and solar. I understand that we need to develop better storage technologies for the energy produced by wind and solar, but that seems so much easier than the challenges currently facing fusion. Wind and solar just seem so far ahead of fusion already - they're pretty cheap and very widely deployed on a global scale. In comparison fusion seems very expensive and unproven and even when we get everything to work it might not be much better than a solar farm with a big battery pack. But maybe I'm missing something important about the economics?
Most of the discussion here seems to assume bioinformatics / genomics jobs are academic, but I work for a clinical testing lab where production-quality code is a must. We're probably a 10/12 on the Joel test.
Just to clarify, exons are the portions of genes that code for protein sequence (they are expressed). Introns span the distance between exons and may contain regulatory or splicing information. Areas between genes are referred to as intergenic, and also may contain regulatory sequences that affect how genes are expressed.
Part of the issue with centromeric or telomeric sequence is that not only is it hard to sequence (being super repetitive), little is known about what a sequence variant in such an area might mean. It's kind of a chicken-and-egg issue: it doesn't get much attention because no one knows how to interpret variants there, and since no one knows how to interpret variants there it doesn't get much attention
It continues to annoy me that fusion companies make claims like "enabling a future with unlimited clean electricity.". Current, broadly used technologies such as wind and solar already deliver unlimited clean energy. What fusion promises is clean energy that is delivered more consistently - at night, or with no wind. But there are multiple paths to consistency, and improving energy storage technologies and improving the grid seem easier to do than inventing an entirely new and extremely expensive new class of energy production.
I'm honestly not sure why there seems to be so much interest in fusion these days. Wind and solar seem to offer a limitless, carbon-free energy supply with relatively cheap, well understood technology that is already price competitive with coal and gas. By contrast fusion seems super expensive and technologically very complex - even fission plants take 10+ years to bring on line. Does fusion offer some advantage over wind + solar that I'm missing?
In response to your final question: To my ear using "men" as a substitute for "all people" doesn't flow better, I found it rather jarring and antiquated and, yeah, a little insensitive. I'm also not convinced it's a purely stylistic decision, as you claim... There's a rather long history of using men as a substitute for everyone in, say, medicine, politics, finance, and many (most?) disciplines, with outcomes most of us agree were not so great. It can also muddy the interpretation of your arguments, as (asterisks aside) it's not perfectly clear when you say "men" if you're referring to everyone or just actually just males. It's an easy thing to change, and your rhetoric will be clearer, more persuasive, and less likely to be immediately dismissed by a fraction of your audience.
So curious to see what happens to salaries, and where people live, if WFH is common in many sectors. Will salaries go up or down on average if employers must compete nationally to find good employees? Will people flee high cost of living areas? Anyone know of good data regarding this?
I'm wondering how hard it would be to reverse-engineer the FLoC algorithm that assigns ids based on browsing history.. could one just have a bunch of headless browsers randomly visit sites, and compute the FLoC ids periodically to see what types of sites end up producing which ids?
This seems important since being assigned a group ID that includes a bunch of people might not be so bad, but (as the article suggests) if its well known which web sites are included in the group, that's a more disturbing story
This is a common misconception - "averaging out" errors only works if the errors are pretty rare at any given site. This is true for some types of errors & sequencing technologies, but not universally true. Some types of DNA sequences (most notably homopolymers and other simple repeats) are very difficult to sequence correctly, and X% of the reads there will be incorrect. If X>20% of so, then it may look like real germline variation no matter how many reads are sequenced
I guess I don't understand what the fundamental disagreement is here. Most astronomers believe the probability that Oumuamua was an alien visitor is very small. Loeb's argument is that the probability may be a bit larger. But we'll (very likely) never actually know which it was, so the disagreement is about the unknown likelihood of a single observed event. It just seems obvious that we'll have to settle for "probably a funnily-shaped rock, but there's a small chance it was something more interesting" and live with that.
Trail runs are different for a few reasons. First, long endurance runs take way more time, maybe 16-18 hours for top-tier athletes to do 100 miles, while a fast marathon isn't much over 2 hours. Also, there are typically lots o really big hills, which require more leg strength to run up and down. There's a lot of additional challenges to running in the mountains (rocky / uneven terrain, snow, river crossings, rain, darkness, etc), the pace is typically much slower. Even pros don't often run much faster than 8 minute / mile pace in a long ultramarathon, while Kipchoge can run most of his miles at 4:30 pace - as long as it's a flat road and it'll all be over in a couple hours
Many folks, especially those in academia, really seem to value academic positions, but I'm not really sure where that comes from. Teaching isn't often the greatest part (though I recognize that not every professor has to teach, much), the success of much of your work is at the whimsy of a small handful of reviewers who typically aren't held accountable for the quality of their reviews, and the schedule can be very demanding pre-tenure, at least. Of course, many annoyances also exist outside academia, I'm just saying that academic jobs are no panacea, and I'm not sure why so many people value them so highly
I have a PhD but ended up working in "The Industry" and it's been awesome - good salary, interesting problems, smart co-workers, real-world impact.
As someone who loves to read but has no formal background in Literature, I'm very intrigued by the ways in which people rank authors so as to find the one most deserving of some award. The article provides a handful of criteria, including things like "striking and realized style", "consistency of excellence".. how are these measured? Can reasonable people disagree? I don't mean to say that it's all arbitrary. I have this feeling that it's not, but at the same time it's impossible to directly quantify or measure in a manner persuasive to everyone. I have no idea if Delillo is deserving or not, but the discussions sounding who is and who is not, what criteria are used, and how authors are measured against those criteria are the most fascinating part to me.
As a full-time bioinformatician I'm totally excited about Julia for bioinformatics. Currently BioJulia needs some pretty serious work before it can be used for routine bioinformatics tasks, for instance there's no good support for the CRAM format. Nonetheless if some of this basic functionality were implemented I wouldn't be surprised to see much greater adoption
Here's a good example, take a look at the recently published articles in PloS Computational Biology: https://journals.plos.org/ploscompbiol/ Just scanning through them... there aren't really any that are a simple "we made a single hypothesis, performed a significance test, and because p < 0.05 we attempted to publish it". In my experience that's just not the usual way science is actually performed (but my experience is very limited to certain branches of biology). I don't mean to say that p-values aren't used at all, just that their application seems to be limited and used mostly to bolster very specific sub-arguments buried in a larger story. I guess the story is that the unit of work that a particular journal article represents is often the union of many statistical tests/ hypotheses / models / simulations / etc that together form a possibly-compelling story about how something works. Not really sure if that's better or worse from a statistical sanity perspective...
Sometimes I wonder if the discussion about p < 0.05 has diverged a bit from practical considerations. In my field (population genetics and bioinformatics) for instance, I'm not sure any current journal would reject a paper where primary result has p=0.051 but accept an (almost) identical paper with p=0.049. Most papers seem to involve many separate analyses that together tell some story, and that story may even be an interesting negative result (where p >> 0.05).
Whether or not statistical significance is a useful concept at all is a separate question, but I suspect the discussion of whether the threshold of 0.05 is useful might be out of touch with actual practice.