Should we make programming languages that are 'easier' for LLMs to write code for? Nope. In fact, we need languages where LLM-generated code can be verified more easily!
Remote
sree
@sree@ublog.thirdlaw.net
pleroma 2.9.1Academic (Computer Science@Rochester). Randonneur. (RFC1983) Hacker.
I like writing code that writes fast, parallel code, mostly for GPUs.
Amateur Photographer.
"Greetings and friendly wishes to all who may encounter this voyager." - from the Golden Record
Was @SreepathiPai on Twitter.
I like writing code that writes fast, parallel code, mostly for GPUs.
Amateur Photographer.
"Greetings and friendly wishes to all who may encounter this voyager." - from the Golden Record
Was @SreepathiPai on Twitter.
228 Followers
331 Following
49 Posts
Joined November 20, 2022
Twitter:
Replying to
@ricci@discuss.systems Yep, this is psychological safety and has been key in my advising adventure as well. Another aspect has been clearly separating criticism of the work (frequent) from criticism of the person (extremely rare). Google has some nice studies on psychological safety.
TIL of the LaTeX markdown package that allows you to write markdown and compile it with LaTeX. Student misunderstood what I wanted, but hey, learnt something new.
https://ctan.org/pkg/markdown?lang=en
https://ctan.org/pkg/markdown?lang=en
I've been told one side effect of LLMs is that the number of late assignment submissions goes way down.
Man, not funny.
Week 4 of debugging (without AI) an AI artifact that is not good enough to use, but not so bad enough that it can be thrown away. I guess this is the productivity we deserve.
Mt. Tom Creek Cascade / Hoh River Trail / Olympic National Park
(1/4s)
(1/4s)
Still resistant to measles 👍
TIL Bison grow and shed a winter coat
At Antelope Island
At Antelope Island
A bison herd on the Salt Lake
(from Antelope Island)
(from Antelope Island)
The real productivity revolution will be here when we stop emailing each other Excel sheets and PDF files to fill out.
Grandeur Peak (summit!) + East Trail + Pipeline Trail
After failing last time, I started off near the west face trail with fresh legs and hiking poles. Although it was unrelenting, I took fewer breaks and only wavered near the last 0.2miles. 3 hrs of climbing in all, 3000ft of climbing, and summit at 8200ft. This trail is hard, nearly every climber I saw was at threshold, though some were impressive.
I then descended using the East trail and at this point it is easily my favorite trail in SLC so far. A verdant canopy, shade, gentler grade, birdsong, spring flowers, roaring creek, and no traffic sounds. A complete change from the west face.
Instead of getting out at the east trailhead, I opted to use the pipeline trail to get out instead. This was a flat and fun trail ending at Rattlesnake gulch, leaving only 2 miles to be hiked out on Millcreek canyon road.
#SLCHike
After failing last time, I started off near the west face trail with fresh legs and hiking poles. Although it was unrelenting, I took fewer breaks and only wavered near the last 0.2miles. 3 hrs of climbing in all, 3000ft of climbing, and summit at 8200ft. This trail is hard, nearly every climber I saw was at threshold, though some were impressive.
I then descended using the East trail and at this point it is easily my favorite trail in SLC so far. A verdant canopy, shade, gentler grade, birdsong, spring flowers, roaring creek, and no traffic sounds. A complete change from the west face.
Instead of getting out at the east trailhead, I opted to use the pipeline trail to get out instead. This was a flat and fun trail ending at Rattlesnake gulch, leaving only 2 miles to be hiked out on Millcreek canyon road.
#SLCHike
At some point after the pandemic, Oak Hall seems to have decided to use hoods featuring national colors for those educated outside America. I don't know how to feel about that given that the Tiranga is younger than the Institute.
Shrinkflation is finally making portion sizes realistic.
A "cyberweapon" that messes with floating point calculation correctness:
https://www.sentinelone.com/labs/fast16-mystery-shadowbrokers-reference-reveals-high-precision-software-sabotage-5-years-before-stuxnet/
https://www.sentinelone.com/labs/fast16-mystery-shadowbrokers-reference-reveals-high-precision-software-sabotage-5-years-before-stuxnet/
One of AI scientific writing's tells is how hard it tries to be convincing by just arguing about it. Like, don't assert you are right, show me. The scientific method is now half a millenia old if not older.
Replying to
@regehr The concluding sentence is exactly what I'm excited about. Software will almost certainly get better with better tooling, AI agents or not, but now there's going to be a ton of money chasing better SE tools (I hope).
Bought this interesting book on sale (SALE2026) for about $11: Culinary Herbs and Spices of the world by Ben-Erik van Wyk.
It does mention kokum, but only in a single sentence as a west coast alternative to tamarind, but points for that. Mostly a reference, no recipes. The photos are very good.
https://press.uchicago.edu/ucp/books/book/chicago/C/bo16940311.html
With leaders like these, who needs enemies?
"The Computer and Information Science and Engineering Directorate’s (CISE) would receive a significant reduction to its budget in FY27 under the President’s proposed budget. CISE, located within R&RA and home for most of the computing research support at NSF, would see its budget drop from $933 million in FY25 to $346 million for FY27, a decrease of $586 million or 63 percent. (NSF has not yet released the spend plan for CISE’s FY26 budget.)"
https://cra.org/govaffairs/blog/2026/04/nsf-fy2027-pbr/
Congress has rejected these cuts twice in a row, but it's maddening they still keep trying.
"The Computer and Information Science and Engineering Directorate’s (CISE) would receive a significant reduction to its budget in FY27 under the President’s proposed budget. CISE, located within R&RA and home for most of the computing research support at NSF, would see its budget drop from $933 million in FY25 to $346 million for FY27, a decrease of $586 million or 63 percent. (NSF has not yet released the spend plan for CISE’s FY26 budget.)"
https://cra.org/govaffairs/blog/2026/04/nsf-fy2027-pbr/
Congress has rejected these cuts twice in a row, but it's maddening they still keep trying.
There are em-dashes and spaced en-dashes, with the superior, of course, being spaced en-dashes. I've seen at least one LLM start using spaced em-dashes. I like the attempt, but stick to spaced en-dashes please.
American science is dealt another blow:
https://www.science.org/content/article/trump-fires-nsf-s-oversight-board
https://www.science.org/content/article/trump-fires-nsf-s-oversight-board
Replying to
@hannorein @MonniauxD Yep, this is an important action. The state I'm visiting currently has seen a record fall in foreign tourist visitors. That will definitely lead to more push back and help better sense prevail.
Another of my attempt to statically analyze C++ code fails.
Replying to
@bosak Very much a beginner at developing, and I've mostly settled on D-76 and Rodinal, the latter because it allows me to do stand development and seems to last forever.
Replying to
@jbigham What gets my goat is not that the US is regressing to treating non-citizens as second-class, the norm in much of the world (including my own country, India), but that current policies are actively hurting US citizens. In science funding -- halving GRFP, not funding REUs, getting rid of alternate pathways like CSGrads4US -- directly affects US citizens who want to pursue science. I cannot understand these moronic self-defeating policies.
Replying to
@regehr@mastodon.social I'd agree, the cost of the ISA is probably negligible, especially since x86 is converted to RISC-like uops under the hood anyway. Saying that ARM probably has a more efficient frontend is probably correct, but a more efficient complete processor, probably not, especially on real-world tasks where the processor is unlikely the main power sink.
Replying to
@kaiser_photo It's a process... even worse is removing the darkslide without closing down the shutter!
I just realized the entire Erie Canal corridor has primitive campsites for bike-packing!
Genesis just switched from strongly discouraging subaward-style submissions to strongly encouraging them via amendment 2. There are 11 days left to go for the deadline.
Google maps is willing to route my bike through the Appalachian National Scenic Trail, but as far as I know, bikes aren't allowed.
In general, and perhaps naively, I believe democratization of technology is a net positive for society over the long run. The democratization of programming by AI LLMs has a better chance of being that if 1) it can be made cheap so it can survive, 2) the technology is not hostage to a few companies but diffuses broadly.
Anybody know of work to minimize a test corpus for grammar-based parsers?
Had an interesting conversation with a colleague who does multispectral document recovery from ancient/medieval manuscripts, and learnt of MegaVision who provide the 150MP digital back. I only know of Phase One who had a back of that resolution, but there were not many kind words for them.
LF film in film holders usually goes through TSA without issues (n=2). That's bulky, but I'll take it.
Discovered hotels still use registration cards, albeit electronic ones. Useful when FedEx (and other carriers) require you to present address proof to pick up a package.
(I still think requiring address proof is archaic -- they have my phone number *and* I know the address the package was shipped to!)
(I still think requiring address proof is archaic -- they have my phone number *and* I know the address the package was shipped to!)