elleElleGray
May 28, 2024, 9:17 PMOmg any day now computers will be able to add
Replies
Biason::Julio::new();juliobiason@functional.cafeMay 28, 2024, 9:19 PM@ElleGray "Near perfectly". Geezus.
May 29, 2024, 3:21 AM@juliobiason @ElleGray yeah that’s the worst thing, right? We laugh (and cry) about the terrible results we’re currently getting from google, but they’re so often and obviously wrong that we’re not trusting anything it says (although I’m sure some people already are trusting it 😞). What when it gets it right 99% of the time, but not that one time where you ask it to split the bill at the restaurant with your new date?
William Leisnerwleisner@mastodon.socialMay 28, 2024, 9:25 PM@ElleGray What AI doing math will look like:
Julian ♪🌻🥥🌴♫TFFPrisoner@mastodon.socialMay 29, 2024, 12:58 PM
Elias Probsteliasp@mastodon.socialMay 28, 2024, 10:06 PM@ElleGray we used to complain about the inefficiency and bloat of stuff like #ElectronJS - and now those #AI bros are doing basic math in an LLM 🤦♂️
Ian Douglas Scottids1024@fosstodon.orgMay 28, 2024, 10:19 PM@ElleGray If they're serious about this, it warrants a sarcastic paper demonstrating an improved algorithm, offering orders of magnitude greater performance, and range of integers supported. With graphs showing the improvement.
Of course, the novel algorithm is the ADD instruction.
Ohne Quatschohnequatsch@mastodon.socialMay 28, 2024, 10:24 PM@ElleGray Wow, that‘s exactly 79.9999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999999 more digits than they were supposed to handle !!!!!
May 28, 2024, 10:26 PM@ElleGray The IBM 1620 CADET has entered the chat.
Patrick 🌎patrickworld@mastodon.onlineMay 28, 2024, 10:29 PM@ElleGray @freequaybuoy great they stole all those numbers from Stephen Hawking
Les Orchardlmorchard@hackers.townMay 28, 2024, 10:32 PM@ElleGray this sounds like the start of a great answer for the interview question "how would you write the worst imaginable implementation of fizz buzz?"
stux⚡️stuxMay 28, 2024, 10:43 PM@ElleGray And it /only/ costs 10 billion power home hours to train!
Third spruce tree on the lefttezoatlipoca@mas.toMay 28, 2024, 10:45 PM
CrashLoggercrashlogger@mastodon.eusMay 28, 2024, 11:58 PM@tezoatlipoca @stux @ElleGray Except that is way better, faster, more efficient, and capable than what sounds like just addition and subtraction
May 28, 2024, 10:46 PM@ElleGray Gonna play the straight (joke) person and use this opportunity to teach: embedding is used to describe or represent data in ML/AL systems.
It’s admittedly kinda neat to drill down and see what cool shit people are using embeddings and building for (2D vs. 3D, vectors, etc. etc.)
Anyway…
Eric LawtonEricLawton@kolektiva.socialMay 28, 2024, 10:48 PM
szescstopniszescstopni@qoto.orgMay 28, 2024, 10:50 PM@EricLawton @ElleGray I'm sure they'll slide between the rules.
Bernard SheppardBernardSheppard@mastodon.auMay 28, 2024, 10:50 PM@ElleGray 5GLs, anyone?
May 28, 2024, 10:52 PM@ElleGray Off-topic, but is it normal in some fields to invert the y-axis like the charts in the tweet?
Tripped my brain up for a moment.
💡𝚂𝗆𝖺𝗋𝗍𝗆𝖺𝗇 𝙰𝗉𝗉𝗌📱SmartmanApps@dotnet.socialMay 29, 2024, 12:45 AM
Medea Vanamonde🏳️⚧️ ♀MishaVanMollusq@sfba.socialMay 28, 2024, 11:13 PM@ElleGray the rise of the Eloi
chollingcholling@social.sdf.orgMay 28, 2024, 11:34 PM@ElleGray When the iPhone had only been out about a year or so, I remember a Domino's Pizza commercial that proudly proclaimed "Now you can order Domino's on the iPhone!" and I remember thinking "Finally, it can do what you could do with land-line rotary phones in the sixties!" This is even more mindblowing.
- 💡𝚂𝗆𝖺𝗋𝗍𝗆𝖺𝗇 𝙰𝗉𝗉𝗌📱SmartmanApps@dotnet.socialMay 29, 2024, 12:47 AM
KanaMaunaKanaMauna@sauropods.winMay 28, 2024, 11:39 PM@ElleGray
Near? What’s the point? I can still do near after a few beers.
Ángela Stella Matutinaangelastella@treehouse.systemsMay 28, 2024, 11:58 PMIf you start with a messy neural network it's no surprise you will need to teach it arithmetics the messy way.
The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.
Pete Wrightpete_wright@nlogic.systemsMay 29, 2024, 12:02 AM@ElleGray my favorite type of scientific computing is results that are almost %100 accurate.
Bel Tamtu | Critter Unit ThreeBel_tamtu@meemu.orgMay 29, 2024, 12:16 AM@ElleGray Near perfectly... as opposed to, say, calculators, that are 100% accurate.
- 💡𝚂𝗆𝖺𝗋𝗍𝗆𝖺𝗇 𝙰𝗉𝗉𝗌📱SmartmanApps@dotnet.socialMay 29, 2024, 12:49 AM
@Bel_tamtu @ElleGray
I wish! Unfortunately a lot of them have been programmed by people who don't remember the rules of Maths! https://dotnet.social/@SmartmanApps/111164851485070719 - Bel Tamtu | Critter Unit ThreeBel_tamtu@meemu.orgMay 29, 2024, 3:44 AM
Okay, sorry, yes, they only do what we program them to do, this is true. x3 I'm still leery of letting an LLM do math but I don't particularly understand the science so maybe it's fine with the reprogramming, but like. We already have programs to do this!
- 💡𝚂𝗆𝖺𝗋𝗍𝗆𝖺𝗇 𝙰𝗉𝗉𝗌📱SmartmanApps@dotnet.socialMay 29, 2024, 4:21 AM
@Bel_tamtu @ElleGray
"I don't particularly understand the science"
In simple terms it runs over the data (where-ever they have got their data from - people's online comments, whatever) and works out probabilities for what is the next word after this one most likely to be, spit that out. Repeat. So if a lot of people are getting Maths wrong (which they are online!) then it's going to be wrong too. As you can say we can just program, but the programmers aren't checking their Maths either! 😡
- Bel Tamtu | Critter Unit ThreeBel_tamtu@meemu.orgMay 29, 2024, 1:29 PM
@SmartmanApps
Ah, so it becomes just like modern spelling and grammar checkers, full of crowdsourced inaccuracies. Lovely. 
Analog AIRetreival9096@hachyderm.ioMay 29, 2024, 12:33 AM@ElleGray Wow! Seems like the AI discovered that addition is NOT commutative, unlike what I was taught in school.
May 29, 2024, 4:05 AMNext comes finishing Russel and Whitehead's Principia Mathematica, Gödel be damned.
- Analog AIRetreival9096@hachyderm.ioMay 29, 2024, 4:25 AM
mirabilos🐈⬛mirabilos@toot.mirbsd.orgMay 29, 2024, 12:46 AM@ElleGray rotflmao
Todd KnarrtknarrMay 29, 2024, 12:56 AM@ElleGray Next step: achieving the level of a 6-year-old and producing an embedding that can do addition to any arbitrary number of digits with perfect accuracy after being trained on only 2-digit addition.
Mallory's Musings & Mischiefmalcircuit@thingy.socialMay 29, 2024, 1:08 AM@ElleGray And only using 1000 times the energy!
Pomegranate County IrregularspomCountyIrregsMay 29, 2024, 1:40 AM@ElleGray @lisamelton Well, next April I look forward to making filings where the totals are generalized.
Pigletpiglet@mastodon.scotMay 29, 2024, 6:30 AM@ElleGray Thing about this, LLMs model and mimic but don't understand. given the record of LLMs so far, I'd check all calculations of any actual meaning. There's even existing softwear that would do this with confidance.
Peter Paaske JuulRockbear@hachyderm.ioMay 29, 2024, 7:42 AM@ElleGray in Eugene ionesco's 'the lesson' an otherwise vacuous pupil is able to perform impressive multiplication feats because she has simply learned by heart all multiplication problems and their solutions.
Now, the lesson is an absurdist play from 1951. When I read it around 1990, I laughed at this idea. Who's laughing now?
May 29, 2024, 7:48 AM@ElleGray So chatbots can now do math better, but there is still the possibility that a result is just wrong to keep you on your toes? I can't wait
"our addition is 99.5% accurate"
dan 💾billgoats@bitbang.socialMay 29, 2024, 8:14 AM
ChartreuseChartreuseK@restless.systemsMay 29, 2024, 8:20 AM@ElleGray ah yes, the one thing I always wanted from my calculator! Using a ton of computing power to *maybe* give me the right answer to addition.
mirthmirth@mastodon.sdf.orgMay 29, 2024, 8:52 AM@ElleGray This paper is fairly cool but what they're doing is way down in the weeds, interesting to most people at the same level as research about new ways to send data across a microprocessor or add molybdenum to automotive transmission fluid.
May 29, 2024, 9:09 AM@ElleGray I'm sure WolframAlpha is scared.
Carsten Rodincarstoid@mathstodon.xyzMay 29, 2024, 12:34 PM@ElleGray wow (integer??) addition with near-100% accuracy? been waiting for this my whole life
skaphleskaphle@social.tchncs.deMay 29, 2024, 1:40 PM@ElleGray You kind of have to appreciate how single-minded researchers must have seen this problem. LLMs learn with data, so let's teach them math with data. And one day, it will almost work perfectly.
I wonder what's stopping LLM researchers from connecting their LLM to tools. It can't be that hard to mix it with a "math detection", right? Make an LLM that translates text into math formulas, then have that use a math program, then output the result. WolframAlpha did that years ago without AI
tekheddtekhedd@byteheaven.netMay 29, 2024, 5:00 PM@ElleGray Every time I say "near perfectly" it makes me LOL again.
Steffen ChristensenWikisteff@mastodon.socialMay 29, 2024, 5:28 PM@ElleGray Tech centibillions: turning trillions of additions into one addition and getting cover articles for doing so.
Matthew Millermattdm@hachyderm.ioMay 29, 2024, 5:57 PM@ElleGray "Generalize near-perfectly" is definitely what people want when it comes to calculation.
Trontron@troet.cafeMay 29, 2024, 7:35 PM@ElleGray Can't wait for LLMs to plot courses for spacecraft near perfectly. Maybe we'll get our very own version of Futurama S1E8, when New York City launched their garbage into space in 2052 and said garbage came back in March 3000.
Ján Bogárjanbogar@mastodonczech.czMay 29, 2024, 7:59 PM@ElleGray I see that nobody read the paper.
"Arithmetic is a subset of the larger class of algorithmic reasoning problems that focus on the ability to learn and execute algorithms and generalize to longer problems [...].The more general algorithmic reasoning field includes work on various architectures [...] aimed at learning algorithms from data."
Addition is just a well defined playground to play on, not the goal. The paper is actually pretty cool and very honest about what they are doing.
Facadelycansubscribe@furry.engineerMay 29, 2024, 11:01 PM@ElleGray wolfram alpha has literally had a semantic analyzer to figure out whether something is an equation and then give it to a calculator if it is
SINCE LIKE 2012 AAAAA
May 30, 2024, 1:38 AM@ElleGray I’m looking forward to someone building a virtual machine that is implemented entirely in LLMs
May 30, 2024, 9:10 AM@ElleGray And in only billions of lines of code, too!
John-Mark Gurneyencthenet@flyovercountry.socialMay 30, 2024, 9:12 PM@ElleGray
And be 80% or so accurate. Also love that they haven't figured out addition is commutative.
Jun 5, 2024, 7:41 AMwhat will they think of next!
Jun 5, 2024, 9:28 AM@ElleGray wow a computer that can do a multiplication or add for only 50 times more energy compared to a cpu calculation. Which is always accurate.
Democracy Matters :verified:DemocracyMattersALotJun 26, 2024, 3:51 PM
