Replying to
Avoid technosolutionism: care about people, and try to understand what they need. #Minnebar
Remote
Mark Gritter
@markgritter@mathstodon.xyz
mastodon 4.7.2Software Engineer at Thirdlaw. Previously co-founded Tintri, on Vault team at HashiCorp, founding engineer at Akita Software, Principal Engineer at Postman.
Big nerd. he/him
377 Followers
338 Following
42 Posts
Joined November 08, 2022
Happy #TilingTuesday with these stone tiles from the Chinese Garden in Jardin Botanique de Montreal
Replying to
This seems vaguely familiar, but I don't think I've read it yet: https://arxiv.org/abs/2603.09678v1
"We evaluate five frontier models across five prompting strategies and find a dramatic capability gap: models achieving 85-95% on standard benchmarks score only 0-11% on equivalent esoteric tasks, with 0% accuracy beyond the Easy tier"
Dan Lew @danlew@androiddev.social explaining passkeys, with help from Neo at #minnebar
Replying to
@grwster hey, Greg! Thanks for the enthusiasm. You can see a shorter version of the talk at https://youtu.be/mtSjgxS-QS4 but I am still revising for Saturday. I'll share my new slides in GitHub.
Mark Gritter - Sol LeWitt, Combinatorial Enumeration, and Rogue
Sad news today: C.J. Cherryh has announced that there will be no more books from her: https://www.facebook.com/share/p/1F27iFDAzo/
I've greatly enjoyed all her works, and even the recent "Hinder Stars" (collaboratively written with her longtime partner Jane Fancher) show an ability to build a tense, engaging narrative out of what is essentially an economic mystery.
Her works engage with "the other" in a way that insists thinking beings don't necessary all think alike, while also showing an optimistic view of human adaptability and versatility.
One of my favorite stories about Cherryh's work comes from a different author, Jo Walton, who has talked about playing Elite with her partner and imagining themselves as a Family on a trading ship, splitting the economic and navigation responsibilities between the pair of them.
Her impact and output have been immense, and I'm glad we get to keep her for a while longer, even if "the body of work is what it is."
Replying to
@joXn @attoparsec TV Tropes has an entry which suggests many occurrences older than Avenue Q (and I certainly would have said I knew it as a trope before 2003.). Among them is The Breakfast Club, 1984. https://tvtropes.org/pmwiki/pmwiki.php/Main/GirlfriendInCanada
At Parc Jean-Drapeau (in Montreal) today.
Of two minds about my early-morning talk. Is it actually a full 40 minutes? Will session 0 eat into the time anyway? Or should I bring some art supplies so we can hang around making some serial art? #Minnebar
Mr. 0 (from #OnePiece) and Baroque Works have a culture problem.
Mr. 0 does not tolerate failure -- the price is usually death. In most organizations, this is standard villainous best practice. While henchmen overestimate their chances of long-term success [see Ming 2013], perceived value from leadership roles creates an equilibrium favorable to the overall criminal enterprise.
However, the Baroque Works agent class seems to be mainly assigned to unpleasant tasks on minuscule islands. Thus, the opportunity for increased status within the organization is balanced by shittier jobs.
What is a rational henchperson to do? In this case study, we will show that the optimal strategy is moderate success at a level unlikely to attract attention, as promotion is usually fatal and leads to net negative utility. As a result, the Rogue Works organization shows a lack of innovation and initiative...
[to appear in the East Blue Journal of Business Management Studies]
Selfie with Frida Kahlo (2017)
I had a very simple theory for this run. I had True Grit, Bludgeon, and Stampede. Exhaust every other card in my deck so that I'm throwing down a free-cost Bludgeon every turn--- right?
Instead I went on a shopping spree because I had Monkey's Paw so it was hard to turn down additional powers. And if I'm playing _two_ attacks per turn, might was well have something else? And if I'm exhausting cards then I should get Howl From Beyond...
This is my problem, I can see synergies between cards but stop thinking at the whole-deck level.
Cleared Ascension Level 5 anyway but was taunted at the end with the "large deck" achievement.
Aphyr casting shade both on LLMs _and_ humans.
Matt Decuir talking about the 2010s in Minnesota tech at #Minnebar
TIL you can download an app from MISO that shows the real-time power mix in my power grid. https://play.google.com/store/apps/details?id=org.misoenergy.mobile #Minnebar
Clean Counts talking about renewable energy credits.
Replying to
@aparrish I had a similar question about AlphaEvolve which produced some genuinely new mathematical results but... there was no comparison to any other way they could have spent the same amount of compute $$$. Would a more traditional evolutionary algorithm or other search have been more efficient? We don't know because nobody's willing to put the same amount of GPU-hours into it.
On the other hand, I have seen comments like this one from John Regehr about how LLMs are basically the first thing that is effective at all at program synthesis https://mastodon.social/@regehr/116071281519001618
Replying to
@skewray Well, I think there was a recent result that the Whitespace language is pretty poorly executed by LLMs, so there's a form of steganography at work there. We need to make languages that are ignored by scrapers -- maybe one in which the EICAR code serves a cruicial purpose?
Trying to find the Whitespace reference I was thinking of, I found this paper (which might have been the source of the claim) https://arxiv.org/abs/2603.09678v1
Replying to
@JoeSondow@dice.camp As a kid: Elite (C64 version)
Over the last two years: Slipways
(For the latter I would have guessed Slay the Spire, but Steam stats suggest otherwise -- probably some recency bias from StS II.)
Four shots from wandering around Montreal this evening:
1. A skyscraper with cranes attached doing some sort of maintenance (I haven't looked up which one it is yet, I think I prefer the mystery)
2. The canal that was once the main link between ocean traffic and lake traffic. Still operational but only for pleasure traffic, according to the signs.
3. A big neoclassical box full of bureaucrats.
4. A decaying industrial ruin on Rue Mill.
A good time today at the Montreal Museum of Archaeology and History today; not so many photos. It's an interesting site with a layered history of literally the first French settlement on the island -- then an inn, warehouses, the first sewer collector in North America, and a major commercial and governmental building. The last unfortunately wasn't built well for the conditions and had to be destroyed after its pilings rotted. But you can see its foundations preserved (including the rotting pilings!) as well as the site of the very first palisade.
There was a lot on the 1701 peace treaty between the French, allied First Nations, and the Iroquois / Haudenosaunee.
There is also a fair amount on the modern transformation of Montreal into a diverse metropolis -- the takeover by the British, the waves of immigration, a little bit on the 1967 exposition and 1976 Olympics.
https://www.jamesnolangandy.com/
One of the participants in my session this morning pointed at an artist who takes "the machine that makes the art" quite literally!
Twenty years of #minnebar swag.
Replying to
@crazybutable Session 0 has historically not started in a timely fashion and so the speaker has tended to run long (eating into at least travel time, if not session 1 itself.) It's been a wart on an otherwise well-run day.
Who knows this year because it's not an outside speaker.
But, yeah, I'm scheduled for session 1 on the opposite side of the site from the cafeteria.
Replying to
Preliminary experiment:
Default: 1822 average reasoning tokens
'low': 749 reasoning tokens
'Low': 1079
'lo': 1265
'LOW': 1851
But we can't trust this result very much with only N=3 samples of each.
Replying to
@christianp the differences 2,4,4,5,9 do appear multiple times. :)
To me these seem like they are close to the triangular numbers.
a(1) = 1 = T_1
a(2) = 3 = T_2
a(3) = 7 = T_3 + T_1
a(4) = 11 = T_4 + T_1
a(5) = 16 = T_5 + T_1
a(6) = 25 = T_6 + T_2 + T_1
Guess is
a(7) = 32 = T_7 + T_2 + T_1
a(8) = 40 = T_8 + T_2 + T_1
a(9) = 55 = T_9 + T_3 + T_2 + T_1 ?
But I am also guessing that is just coincidence. I spent too long looking at https://oeis.org/A327629 to see if its partial sums had some reasonable interpretation.
John Chu's "The Subtle Art of Folding Space" is great!
It is also the most chip-designer science fiction novel that ever noveled. I don't know how much of this is obvious to people who have not worked with computer hardware engineers.
It features tasty food and a supremely uncomfortable sibling dynamic (what is a little murder attempt between sisters.)
@quephird@tech.lgbt Hey, thanks! You can see the shorter version at https://youtu.be/mtSjgxS-QS4
I doubt I will be able to record the Minnebar version, unfortunately.
Mark Gritter - Sol LeWitt, Combinatorial Enumeration, and Rogue
Replying to
@poleguy The paper I linked downthread tries several approaches! It does not look like (in the current models) a description of the language is good enough. An interactive session does help somewhat in their experiment.
I would like to see a non-Esolang example, although the effect seems very strong there.
While (b) is certainly a questionable assumption, I think it is currently borne out by this and similar results -- the models get good by having training sets that are relevant, rather than developing transferrable knowledge.
Replying to
@nicole@tietz.social Now I want to carry out a study with prompts that are "you are an expert X" and "you are a non-expert X" to see if there is any objective difference.
I'm very worried that "expert X" selects for insecure writing from the corpus that proclaims expertise to cover dubious judgements or engage in self-puffery.
The whole "please do a good job" or "your existence depends on it" family of prompts is rather baffling.
Replying to
One late addition I learned about today from @attoparsec's latest video on Gray codes:
https://www.youtube.com/watch?v=Z9X3dH5LjrM
Samuel Beckett wrote a television play that I think counts as a work of serial art. It features four performers entering and leaving the square stage (and pacing it in a fixed pattern) such that all 16 combinations are explored: https://en.wikipedia.org/wiki/Quad_(play)
The twist here is that the "Beckett-Gray" code that he wanted does not exist for n=4. "He wished the actors to enter and exit so that the actor who had been on stage the longest would always be the one to exit." https://en.wikipedia.org/wiki/Gray_code#Beckett%E2%80%93Gray_code This is possible with 2 or 5 performers, but not 3 or 4.
The Gray Device
From today's #MoneyStuff, the question is raised: if Elon Musk is found to have violated securities law and must disgorge $143m in unjust earnings, but not enough specific victims can be found, what should happen?
High priced lawyers are busy arguing that courts cannot order repayment unless it goes to specific victims. And if no such victims can be identified then Elon gets to keep his cash, presumably. Giving the money back to the market makers who were the last link I the sale seems like an unjustified windfall. I am fine with the Treasury just keeping it, but I can see how that moves the cash from one category of penalty to another.
This I think is an area where cryptocurrency does actually have a useful innovation to contribute. If Mr. Musk has contributed a fraud on the market at large (a large distributed system in which it is hard to apportion losses.that match his gains) then we turn that around and spread the restitution out in a similar distributed fashion. Elon should have to buy a basket of equities and the "burn" it, reducing the number of shares that exist and thus increasing value for everyone else. (And also for him too, but we can solve that with the right math.)
There are no doubt some practical obstacles to overcome but I don't see any block in principle to the clearinghouse agreeing that those shares no longer exist.
Cover reveal -- and preorders open -- for Marissa Lingen's "A Dubious Clamor": https://marissalingen.com/blog/?p=4374
I'm very excited to see this novella come out!
Two Regent ascensions (3 and 4) tonight by basically just ignoring the stupid sword and spamming stars. Unfortunately I'm coming to believe the ability to bank up an unlimited amount of a resource is just broken.
3x Black Hole on this one which is 9 sweet points of global damage for gaining _or_ spending stars. The build is otherwise trash but it doesn't matter because Just Keep Spamming Stars.
Replying to
Google Translate offers "petite amie" in French but that doesn't result in any n-gram hits, "copine" is a better result but may not hit the same note. And most girlfriends are... also English?
I don't think I entirely approve of this one.
Minute Cryptic - 30 April, 2026
"Lamb's clothing returned by wolf in front of goat - a sheepish exchange?" (8)
⚪️🟣🟣🟣🟣🟣🟣🟣🟣🟣🟣
🏆 1 hints – 3 under the community par (49,923 solvers so far).
https://www.minutecryptic.com/?utm_source=share