I've been working on RVW, my adaptation of the standard transformer model that is capable of online continual learning without catastrophic forgetting. I finally published the first pre-print of my early experiments: https://doi.org/10.5281/zenodo.20064617
Now I'm working on expanding the work into more parameters and improving performance. I just finished an extremely harsh test of a Nemotron-flavored RVW that consisted of stretches of a random assortment of domains interspersed with long runs of single domains. Across all of it the model didn't forget (and actually improved on some of the more challenging domains). PPL on SmolTalk is still in the ~18 range, which I'd like to get lower, but this is all with only 4B params.
Currently, I'm training a Llama 3.2-flavored RVW with only about 2B params to see how that turns out. Depending on results of that, I may take it to Gemma 4 next.
IIRC, there was always a way to filter out certain messages (or that may be an alt.org customization, but it's been a part of my config file for a while now).
For real! Valkyrie is the perfect "just bash things while only half paying attention" class. Great for when I'm playing to unwind (as opposed to playing as a challenge to myself).
I think Douglas Adams had one of the best quotes regarding observing infinity:
"Infinity itself looks flat and uninteresting. Looking up into the night sky is looking into infinity – distance is incomprehensible and therefore meaningless."
It's been a while since I worked at Apple, but back in the day the entire OS X Server team made extensive use of kerberized NFS shares for moving around large files...
...the last version of Server shipped in 2021 (and the last real version shipped almost a decade before that).
My first job after finishing my undergrad degree was performing quality analysis on corn starch. As a condition of employment, I had to sign a paper saying anything I invented related to corn was property of my employer.
It's been more than a few years since I worked at Apple, but they were always unique in the tech space in that their retail division dwarfed headcount. If I recall correctly all of OS X Lion was produced by around 3,000 engineers (and probably less, since I think that count included iLife and iWork).
I've been working on an ML model capable of robust continuous learning, resistant to catastrophic forgetting without relying on replay, an external memory system, or unbounded parameter growth. Last week I confirmed the first non-toy, 580M parameter version soundly beat LoRA, EWC, and full fine tuning. This week I'm scaling up to 4.4B parameters...
Based on what you've already mentioned, there's a good chance you're familiar, but on the off chance you're not: "Funkungfusion" (or, really, anything off the Ninja Tune label) might be right up your alley.
Eh, I'm not so sure it'll be that big a deal. The whole supply chain is so twisted and tangled all the way up and down. Shuffling out one piece doesn't seem like it will, on its own, be so major. Samsung made the chips for the iPhone, then made their own phone, then Apple designed their own chips made by TSMC, now Apple is exploring the possibility of having Samsung make those chips again.
Also, it takes a willful ignorance of history for ARM to claim this is the first time they've manufactured hardware. I mean, maaaaybe, teeeeechnically that's true, but ARM was the Acorn RISC Machine, and Acorn was in the hardware business...at least as much as Apple was for the first iPhone.
Fun fact about that cannon: it took so long for the cannon to cool off between shots that the Byzantines were able to patch each hole it caused before the next shot.
I exited academia for industry 15 years ago, and since then I haven't had nearly as much time to read review papers as I would like. For that reason, my view may be a bit outdated, but one thing I remember finding incredibly useful about review papers is that they provided a venue for speculation.
In the typical "experimental report" sort of paper, the focus is typically narrowed to a knifes edge around the hypothesis, the methods, the results, and analysis. Yes, there is the "Introduction" and a "Discussion", but increasingly I saw "Introductions" become a venue to do citation bartering (I'll cite your paper in the intro to my next paper if you cite that paper in the intro to your next paper) and "Discussion" turn into a place to float your next grant proposal before formal scoring.
Review papers, on the other hand, were more open to speculation. I remember reading a number that were framed as "here's what has been reported, here's what that likely means...and here's where I think the field could push forward in meaningful ways". Since the veracity of a review is generally judged on how well it covers and summarizes what's already been reported, and since no one is getting their next grant from a review, there's more space for the author to bring in their own thoughts and opinions.
I agree that LLMs have largely removed the need for review papers as a reference for the current state of a field...but I'll miss the forward-looking speculation.
Science is staring down the barrel of a looming crisis that looks like an echo chamber of epic proportions, and the only way out is to figure out how to motivate reporting negative results and sharing speculative outsider thinking.
When I was a young kid, my mother was a “stay at home mom”, which meant that she babysat the kids of 5 or 6 of the other families in our neighborhood where both parents worked. For me, it was a wonderful experience growing up having a ready-made group of close friends and my mother close at hand. It did mean that my mother effectively sacrificed her career (though she eventually went to work for my father as his office manager and was instrumental to his success), but I’m certain she was not charging $20k/yr/kid (or whatever the equivalent in 1980s dollars would be).
What Americans seem to only just now be waking up to is that lack of work/life balance, lack of family leave accommodations, and loss of community has a very real, very tangible dollar amount cost. I’m very, very tired of the knee-jerk response to every “socialist” proposal being, “yeah, that’s great, but how are you going to pay for it?”
How are you going to pay for not having family leave? How are you going to pay for not having universal healthcare? How are you going to pay for not having tuition-free college for all? These choices have a cost, and Americans are paying that cost every day!
Reporting blatant criminal violations is not the same thing as moderating otherwise-protected speech that could be construed as misleading, offensive, or objectionable in some other way.
The problem with this is that section 230 was specifically created to promote editorializing. Before section 230, online platforms were loath to engage in any moderation because they feared that a hint of moderation would jump them over into the realm of "publisher" where they could be held liable for the veracity of the content they published and, given the choice between no moderation at all or full editorial responsibility, many of the early internet platforms would have chosen no moderation (as full editorial responsibility would have been cost prohibitive).
In other words, that filter that keeps Nazis, child predators, doxing, etc. off your favorite platform only exists because of section 230.
Now, one could argue that the biggest platforms (Meta, Youtube, etc.) can, at this point, afford the cost of full editorial responsibility, but repealing section 230 under this logic only serves to put up a barrier to entry to any smaller competitor that might dislodge these platforms from their high, and lucrative, perch. I used to believe that the better fix would be to amend section 230 to shield filtering/removal, but not selective promotion, but TikTok has shown (rather cleverly) that selective filtering/removal can be just as effective as selective promotion of content.
Now I'm working on expanding the work into more parameters and improving performance. I just finished an extremely harsh test of a Nemotron-flavored RVW that consisted of stretches of a random assortment of domains interspersed with long runs of single domains. Across all of it the model didn't forget (and actually improved on some of the more challenging domains). PPL on SmolTalk is still in the ~18 range, which I'd like to get lower, but this is all with only 4B params.
Currently, I'm training a Llama 3.2-flavored RVW with only about 2B params to see how that turns out. Depending on results of that, I may take it to Gemma 4 next.