mcguire

8 ms·

61.
Benchmarking the Energy Costs of Large Language Model Inference (2023)(arxiv.org)
2 points by mcguire 2y ago|0 comments
62.
by mcguire 2y ago
Hmm. Let's say you're looking at that 5000 TEU (Twenty-foot Equivalent Units) ship with (the most common, IIUC) 40' containers. The state space for optimization has something like 2500! or 10^7400 states. Given 24 hours to op
63.
by mcguire 2y ago
Don't store some classes of hazardous materials near specific other classes of hazardous materials (keep your oxidizers away from the fuels), don't carry too much hazardous material, keep good records of what's in each contai
64.
by mcguire 2y ago
It's very much wackier than that. In addition to mere physical constraints, there are the laws and regulations of individual countries and a goodly stack of international treaties and standards.
65.
by mcguire 2y ago
" My prompt after ChatGPT gave me an answer of 5: 'Yes but you aren't including Alice in your logic. Shouldn't the correct answer be 6 (since Alice is also a sister)?' " Useful, if you know what the answer is.
66.
by mcguire 2y ago
I just tried that with ChatGPT 3.5 (4o stopped responding after I asked the initial question and it produced the wrong answer). Here's the Prolog it generated: % Define the number of brothers and sisters brothers(4). sist
67.
by mcguire 2y ago
Reference 40 is scaring me. A page and a half of authors?
68.
by mcguire 2y ago
And don't forget that the Houthis recently struck the USS Eisenhower with ballistic missiles. https://x.com/iAmTheWarax/status/1797732901846692117 (Yeah, @iAmTheWarax's tweet is a shitpost. It's app
69.
by mcguire 2y ago
...incomingpain writes, on one of the more heavily moderated of the current internet forums.
70.
by mcguire 2y ago
I actually don't know; I just saw that they were available, I didn't try them. I assume it would be horrible, though. :-)
71.
by mcguire 2y ago
I recently decided to try to actually learn to use Scheme, really, in at least a practical sense. I had previously never gotten too far into it. I ran into the System Crafter's Spring Lisp Game Jam livestream ( https://www.yo
72.
by mcguire 2y ago
Oh, no, it's a well-known approach to handling bug reports. Years ago, I worked with several of IBM's AIX kernel team; they talked about a well-defined, 3-level triage process where the third level was the actual developers. Unfor
73.
by mcguire 2y ago
Both of these ideas also cut down on the number of bug reports you have to handle.
74.
by mcguire 2y ago
Do you have a citation for the existence of an LLM "world model"? My understanding of retrieval-augmented generation is that it is an attempt to add a world model (based on a domain-specific knowledge database) to the LLM; the res
75.
by mcguire 2y ago
(An aside: Writing is a representation of speech, not the other way around.)
76.
A good argument against using Creative Commons licensing(bsky.app)
3 points by mcguire 2y ago|1 comments
77.
by mcguire 2y ago
The actual paper: https://onlinelibrary.wiley.com/doi/full/10.1002/arp.1940 Seems like I recall reading about Egyptian tombs with a rock-cut outer chamber and a shaft leading to a lower burial chamber. Hypoth
78.
OpenAI threatening former employees with loss of equity(twitter.com)
8 points by mcguire 2y ago|1 comments
79.
by mcguire 2y ago
Those were all designed after monadic IO was introduced in Haskell. The ability to mark IO operations in types (and the do notation) was a game-changer.
80.
by mcguire 2y ago
" ...a new (often anticipatable) requirement seems to appear... " And often not, particularly when you are building a new system. And if you are replacing an existing system, the requirements that you discover are usually, "i
81.
by mcguire 2y ago
Eventually, yes. Unless they sell to another company and the technical debt goes into the trash with the company's product. On the other hand, "the market can remain irrational longer than you can remain solvent."
82.
by mcguire 2y ago
From the article: " But now NHTSA's raw data set is available, and, if anything, it appears to contradict Musk's claims. The majority of the vehicles in the Tesla data set suffered from missing data or other problems that mad
83.
by mcguire 2y ago
That's exactly why I went through the exercise of translating Kernighan and Plauger's Software Tools in Pascal into Haskell. https://web.archive.org/web/20171020034308/https://crsr.net/..
84.
by mcguire 2y ago
As long as you remember to turn it off after you're finished refactoring. :-) (Yeah, I'm the guy with the co-worker who kept turning off -Wall in C++ because "nobody has time for that".)
85.
by mcguire 2y ago
In a lazy language, you have a significant difficulty in ensuring the order of operations, specifically printing a prompt and reading the answer. Before someone (Mogi?) realized that a category theoretical monad with specific rules made IO
86.
by mcguire 2y ago
Side effects like IO are very difficult to handle in a pure language. I believe Simon Peyton Jones' argument was that, if it's strict, there is too much of a temptation to say, "screw it", and abandon purity. https:&#x2
87.
by mcguire 2y ago
"The National Highway Traffic Safety Administration has egg on its face after a small research and consulting firm called Quality Control Systems produced a devastating critique of a 2017 agency report finding that Tesla's Autopil
88.
In 2017, the feds said Tesla Autopilot cut crashes 40%–that was bogus (2019)(arstechnica.com)
57 points by mcguire 2y ago|24 comments
89.
by mcguire 2y ago
In case the author sees this, I enjoyed the nod to the Acme editor.
90.
by mcguire 2y ago
Really? It's literally been decades, but the last I learned about neural networks, a non-linear activation function was important for Turing completeness.

More ›