yeah so, assuming that where civs meet, they just stop expanding and respect their neighbors, you get the following distribution over civ size. because the x axis is log-scaled, this implies a fairly high level of inequality in the distribution of civ size.
i think i was subconsciously combining "my own personal vietnam" with bsky.app/profile/norv...
this is my sigmoid
I've previously found interesting results from using -log(Euclidean distance) attention. On shallow networks, it seems to train more easily and learn interpretable keys. arxiv.org/abs/2310.18805
github.com/calvinmccart...
exactly. (having said all this, wearing my substack special purpose vehicle shareholder hat, i am rather pleased...)
so I'm purely a browser user, but yes it does (screenshot). the problem is that the feed surfaces "substack notes" and replies thereto, but not "substack articles" and replies thereto.
it's also hard for me to get the full list of people i follow -- like what if i decide i want to subscribe instead?
another bad thing is that if i follow someone, i don't find out about their comments on various articles. an even worse thing is that if someone follows me, they don't find out about my comments on various articles.
they've reached parity on the dual problem (unfortunately with degenerate solutions)
Of all sad words of LLM,
The saddest are these: "Failed to fetch arxiv.org" again
more seriously, this one:
does this count?
yes, allegedly the Munich Conference was good because it bought time for GB to prepare for war, while clarifying that the next takeover on Eastern Europe would bring GB to intervene
interesting article: big for Neville Chamberlain if true
speaking of which, it would be cool for TMLR to be a place which tries out various improved reviewing ideas!
a similar tweak would be to require submissions be posted on arxiv (to be verified in some automated fashion), while still keeping the venue itself double-blinded. authors shouldn't be able to force reviewers to read something, if they don't want anyone else to see that they wrote it.
"SDG"
tangential, but "set the mood for people to have communal religious experiences at church" would make almost certainly make Bach flip out. if we're going to insist that his motivations matter, then we should at least be honest about what those were
well, but the lesson from that is that you need to time the market correctly in order to make money; and normal people shouldn't try to time the market.
a metaphor for x (formerly twitter) @norvid-studies.bsky.social
both of these have wikipedia articles featuring what TR thought about it -- coincidence?
2026: celebrating 250 years of having taxes that are "very low by european standards"
what's the incentive here? there are no creator payouts...
i suspect that GDM has a solution that people in academia are either unable to invent or unable to afford (guessing the latter)
it's hard to learn what CDRs do due to their messy structures and the lack of evolutionary info about them in standard ab sequence datasets
we are so back (to dalle2 trve art)
obnoxiously tagging @kevinkaichuang.bsky.social @delalamo.xyz @austinjtripp.bsky.social about our new PLM paper -- hope you find it interesting and/or useful!
When supervised models were included via multi-objective guidance, we achieved a 100% synthesizability-and-binding success rate in vitro. But a cautionary note: guidance improved target objectives at the cost of reduced humanness. All the details are here: arxiv.org/abs/2603.10302
Another surprise: ESM-2, trained on generic proteins, was highly competitive with antibody-specialized models for antibody optimization. Meanwhile, AbLang2 — trained on human antibodies — sometimes produced less human sequences than ESM-2. Training data ≠ output bias in the ways you'd expect. 6/7
Perhaps the most provocative finding from the in vitro experiments: choice of sampling algorithm matters at least as much as choice of model. Beam search consistently outperformed Gibbs across every model where both were tested. 5/7