To me this is not particularly surprising, given that the tasks they are fine-tuned on are in some way malevolent or misaligned (generating code with security vulnerabilities without telling the user; generating sequences of integers associated with bad things like 666 and 911). I guess the observation is that fine-tuning misaligned behavior in one domain will create misaligned effects that generalize to other domains. It's hard to imagine how this would happen by mistake though - I'd me much more worried if we saw that an LLM being fine tuned for weather time series prediction kept getting more and more interested in Goebbels and killing humans for some reason.
My understanding is IRB is not a governmental regulation but an institutional regulation. For example if you want to design a study and then have acme medical center take part in conducting it, you would need to pass the Acme IRB. Governmental funding may also have restrictions that require you to have some kind of IRB review.
While I have been amazed by the progress of LLMs and believe that this ultimately leads to AGI, I still think that LLM aided development is limited to things that are often done and are well documented in the training corpus. Writing code with new libraries or less documented lower level apis still requires you to actually do it. LLMs slowly separate you from your understanding of what your code is doing, and if they don’t meet a minimum threshold of proficiency on a topic, they will slowly and insidiously degrade your codebase until you realize Claude (or whoever else) has built a house of cards.
Nicely done - it's crazy that you could do this solo and the multi billion dollar device makers haven't been able to do it in many years. Would be a nice peripheral for them to sell along with their CGMs (instead of those old receiver devices that everyone stopped using in favor of smartphones). Wishful thinking, but maybe try to talk to Abbott or Dexcom to see if they will fund another open source production run of these? Output to them could be a report/collection of user feedback to help them w/ product strategy...
I cant tell what my charge level is. This is crazy though, congrats... Look at the blog to see how he did it mostly using ai. It looks like the dad set it up like a series of challenges to get the kid familiar w using ai then he was able to make a game. Nice work!
Has anyone seen an example of time series analysis via transfer learning / fine-tuning an LLM to process and predict multivariate data as xml or something? e.g. :
<speed 45>
<speed 46>
<heading 123>
<speed 47>
<speed 47>
...etc
I agree, maybe a more conservative performer than Lang Lang would've been good to hear the first run from. Love Lang Lang but he's a bit of an 'innovator' w/ the romantics.
Watch the results getting skewed in real time as night falls across the Americas and more people’s phone enter the mode with more yellow for low light conditions…
This is great and surprisingly consistent. Apparently I’m in the 98th percentile of how blue my cutoff is. I wonder if this is related to my favorite color being green (I’m perceiving more things as green because I like the color)
This is crazy I just stumbled upon this company earlier today via search when I realized that Aurora Postgres Serverless V2 no longer allows you to scale down to 0 in times of low usage. AWS makes you select 0.5 ACUs as the minimum for V2 where with V1 you could select 0... is this just a straight up money grab by AWS to make the minimum monthly cost for aurora like $50 or is there a justifiable reason behind making you run virtual servers continually when there is no usage. I may be switching soon!
So for hardware accelerated training with something like PyTorch, does anyone have a good comparison between Metal vs Cuda, both in terms of performance and capabilities?
This is great news for the healthcare system because it is going to drive down costs for people with diabetes who really need these things. Well done Dexcom!
This OTC model limits the measurements to every 15 mins instead of every 5 for the prescription version; this will limit its ability to be integrated into open source automated insulin delivery systems. Not sure why they made this choice as the hardware looks the same as the Dexcom G7. In theory they could be doing it to increase battery life or allow for cheaper batteries to be used in the consumer version.
It's not really that the pancreas itself can't handle it (though this may be true also). What happens is that having high blood glucose, even if it's not continuously high, is just bad for you by virtue of the high concentration of glucose in your blood. Having high blood glucose leads to the creation of "Advanced Glycation Endproducts" (AGEs) which bind to certain receptors (RAGEs). The AGE-RAGE interaction is linked to downstream diabetes side effects like atherosclerosis.
I think what you're saying makes sense, but you are also making an assumption that communist <=> fascist politics exist on a one dimensional spectrum, which is not really true. There are totalitarian and libertarian axes as well.. More so, the problem w/ all the examples we see is that these 'errors' are aligned particularly well with the corporate-progressive party line, while there are less errors in the other direction. You would expect to see multiracial Zulu warriors too if we were to believe the explanation you provided, but we only see multi-racial Kings of England, which amounts to a quasi-racist cleansing of historical data. This happened via either a bad system prompt, RLHF implemented by biased tutors, or tutors w/ biased instructions or some combo of all three.
If you look at who is on the board, how it's structured (they don't have equity right?), it seems like it's actually because he violated the charter. Why would Ilya Sutskever punish Sam for doing the right thing wrt AI safety?
This is a compelling way of looking at things for sure; to me it seems like you're saying math is just a language (like English) that we can use to describe the world. English is no more "right" than some other made up language, so same goes for math, and thats certainly seems true in some cases. However there are things that are intrinsic to mathematics that also seem intrinsic to existence, like prime numbers. Many things about prime numbers are 'discovered' as if they were elements of nature that need to be revealed through study. For example, the distribution of primes and the distribution of twin primes are things likely to have been discovered by an alien intelligence, assuming one exists. What would their math look like? Different for sure, but I think it's likely they still studied the primes, hypothetically.
Many fancy smart fridges are already quantified shelves. I see you are still adding stuff... let's keep it going.... Rideable Boston Dynamics Spot for commuting. An smart vape-pen line of medications. Smart Wigs. Self driving strollers.
Not seeing this in any of the comments but may have missed it: To me the obvious answer here is increased screen time across the board, precipitated by the pandemic. It's easy to see why this would lead to a perceived faster flow of time; do you form memories of your screen time? Personally I have next to no distinct memories of consuming content on the internet, especially of doom scrolling. Also interestingly, as far as I can tell we never dream about consuming content, or our devices really. If we are going into a trance like state consuming content for a larger proportion of our lives, the time when we actually are cognizant / conscious / forming memories will feel less voluminous overall, thus making perceived time feel faster. The more I think about it, the more this seems like one of the most important issues facing humanity currently... AI doomers should probably be very worried about this in the near term.