I'll also add, most food products down there are "expired" already. When I was down there, it was often a challenge to find the oldest piece of food. I think we found 5 years + beyond the Best By date.
Best by dates for shelf stable/frozen food are often not safety related, so the antarctic program just charges forward with whatever they have.
(creds: I was down to McMurdo as a researcher three times)
I suspect this has to do with space and weight constraints, and probably a touch of old-school procurement practices.
In the not-too-distant past, basically everything was flown to south pole station, so weight was at a premium. Powdered milk weights a lot less than UHT milk. Now they do a traverse to the pole with sleds and tractors, so weight is less of an issue, but volume might still be.
On top of that, procurement may be slow to change. If, in fact, weight is no longer a constraint, it might take years for procurement to change to include buying UHT milk.
My experiences were with the McMurdo Dry Valleys LTER program [1]. They have (things may have changed) a number of sites sending telemetry back to the states via the Iridium network. Nothing too fancy. It worked. Biggest challenge really was that iridium was slow and relatively high power (and somewhat flaky deep in the valley where we were). I also had some involvement with the NTL LTER program[4], but that type of work has even easier telemetry constraints (these days, just use a cell modem).
I totally agree with you on the "using a combination of different sensors and a cleverly trained algorithm to get at the parameters of interest". This is something not too far from, in a way, how many sensors work already. They are *proxies* of the actual thing being measured. From my world, the s-can DOC sensor was always a good example, using in-situ spectroscopy to estimate DOC concentration.
Crux of the challenge is "what is the parameter of interest" and "can you come up with a way to estimate it with something easily measured?
Because this is HN, I'll say there is another interesting route possible. If you can change the economics of a situation and decrease the cost of a basic sensor, then you can often increase the volume of applicable uses. I was tangentially involved with the development of the miniDOT [3], which ended up being one of the first "inexpensive" (as in less than $5k) dissolved oxygen sensor. It really changed how people used them and increased the amount of DO sensing by probably an order of magnitude.
Having been in technology, antarctic research, and remote monitoring of environmental systems down there, my take: don't read into this this too much.
This is mostly a PR piece, probably pushed by the university or non-profit researchers involved. They are trying to use some sort of partnership with NVIDIA (as loose as that partnership might be) to draw attention and show they are having "Broader Impacts" for their impact statement.
Most eco research is based on historical comparisons of months/years/decades of data So the use of real-time/streaming data down there is pretty limited. You can just as easily shove the data into storage and have a researcher pick it up next time they go down (often *much* easier as you don't have to worry about powering comms systems).
Climate/weather data may be different, if only because some of the data might go into current/real-time weather models. But even there, it's probably a stretch (I know of very little work being done with anything near real-time as far as data goes down there).
I'd argue, they aren't doing something future-proof right now because the fundamental architecture of the LLM makes it nearly impossible to guarantee the model will correctly respond event to special [system] tokens.
In your SQL example, the interpreter can deterministically distinguish between "instruct" and "data" (assuming proper escape obviously). In the LLM sense, you can only train the model to pick up on special characters. Even if [system] is a special token, the only reason the model cares about that special token is because it has been statistically trained to care, not designed to care.
You can't (??) make the LLM treat a token deterministically, at least not in my understanding of the current architectures. So there may always be an avenue for attack if you consume untrusted content into the LLM context. (At least without some aggressive model architecture changes).
This is interesting. Pondering about this, the vulnerability seems rooted in the very nature of the LLMs and how they work. They conflate instruct and data in a messy way.
My first thought here was to somehow separate instruct and data in how the models are trained. But in many ways, there is no (??) way to do that in the current model construct. If I say "Write a poem about walking through the forest", everything, including the data part of the prompt "walking through the forest" is instruct.
So you couldn't create a safe model which only takes instruct from the model owner, and can otherwise take in arbitrary information from untrusted sources.
Ultimately, this may push AI applications towards information and retrieval-focused task, and not any sort of meaningful action.
For example, I can't create a AI bot that could send a customer monetary refunds as it could be gamed in any number of ways. But I can create an AI bot to answer questions about products and store policy.
Is there an English version of the website? I just see the spanish version when I go to https://www.neuraan.com/
Also, as a former DSist in a service area of ecommerce, one challenge with automated service interactions are not just conversations (with answers) but also actions based on those conversations. Of course, something to iterate into, but it may be a question that comes up when talking to potential clients.
I'm going to finally try to start my own company. Will it be a "startup"? I don't know. But it will be technical and it will go from 0-1 (or from 0-0).
After years of lurking here, watching on the sidelines, working for larger companies, having kids, buying a house, I'm finally going to take the dive. I'm excited, nervous, lost, all at the same time. But I have enough savings and an accommodating spouse, so I don't have to work for a while.
I'm a long-time academic, turned ML-practitioner. I have no major online presence. I don't have a brand. But if anyone is interested in talking, DM me, I have lots of time and am still in the divergent phase of entrepreneurship.
Edit: Added email address to profile. Excuse the confusion, I have been a lurker too long.
It was public knowledge. Further, there is no such thing as New Zealand territory (or anyone's territory for that matter) in Antarctica. There are existing territorial claims, but they are overlapping and basically nullified by the Antarctic Treaty.
Best by dates for shelf stable/frozen food are often not safety related, so the antarctic program just charges forward with whatever they have.