[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]

@ljupc0 "Lucky for us then that minimal NN forward pass in C is XXX lines or so Didn't @karpathy have some repo llm.c or some such So 20yrs is not a problem pfft. can run MS-DOS XXXX programs afaik. Nah - NN software implements a function is a platonic idea. So software + weights are immortal all right. To come to life though they need hardware any hardware. Like a soul seeking a body haha 😱 More amusing to me was the connection of low(er) energy analogue human (and animal and any living thing really) intelligence versus high(er) energy digital intelligence. Mortal-in-time analogue intelligence"
@ljupc0 on X 2025-07-24 19:49:02 UTC 5312 followers, XX engagements

"Heh - no idea what MMT you are referring to as there are probably as many as there are people's heads. 😆 This to me seems eminently true and uncontroversial - "gov monopoly on currency is achieved via monopoly on violence and taxation asking the citizens to return back to the state the currency the state issued; and everyone is forbidden from creating any other currency""
@ljupc0 on X 2025-07-23 16:01:50 UTC 5305 followers, XX engagements

"They themselves don't know for sure. Still - not futile to ask the Q as it may prod them to think. Not the worst of things. 😆 This is what I think is going on with cryptos goldbugs libertarians rejecting "the MMT" completely and vociferously. Whereas I thought they will have easy time with the MMT premises. For - how can I be a libertarian and think "gov monopoly on currency is achieved via monopoly on violence and taxation asking the citizens to return back to the state the currency the state issued; and everyone is forbidden from creating any other currency" is not the right story How can"
@ljupc0 on X 2025-07-23 12:01:08 UTC 5306 followers, XXX engagements

"It's a head-scratcher to me how the algofeed puts on my followers screens my forwards of my own posts in a way that it doesn't put the original posts Puzzling. Related but not on algofeed. When I spend time to reply (like now) I never know whether my (this) reply was pushed to OP (your) screen or not. So IDK I'm in the dark if the OP (unless s/he explicitly scans the replies) has even gotten a non trivial chance to seeing it. The only way I can think of requires a cooperation protocol: for the OP to Like the reply and for I to turn on notifications so I'm notified of the OP Like. Than I know"
@ljupc0 on X 2025-07-25 04:15:41 UTC 5312 followers, XX engagements

"@casper_hansen_ .and we got the big Qwen3 new refresh yesterday. (can't edit replies - still)"
@ljupc0 on X 2025-07-22 18:42:08 UTC 5303 followers, XX engagements

"Hey thanks for that will watch next. But even before that - aren't discrete symbols a world of pain Like - we can't have ordering in the space of symbols we can't have derivatives like with continuous values no symbol has a particular relation to any other symbol. I did language models Ngrams (20yrs ago - welp.) and just counting and then smoothing and falling back N-grams to combo of (N-1)-grams to combo of (N-2)-grams. - all that was a world of pain. And bodge on a top of a bodge. I disliked it without much math to guide us just seemed unfounded and a bit desperate tbh. Search and count -"
@ljupc0 on X 2025-07-22 12:32:29 UTC 5299 followers, XX engagements

"Hey - thanks for that. Didn't occur to me - AI for HF-s After 20yrs in quant trading I got really bored with it in the past X years. Tinkering with llama.cpp and running open source open weights models in the last year rekindled the 1st love - ML :-) Did 10yrs of ML until 2004 it was then HMMs based ASR Ngrams LMs some NNs but only X hidden layer. Now I'm on X weeks notice - decided to try go back to ML/AI next. ML for HF-s may be a niche I should look into"
@ljupc0 on X 2025-07-20 18:29:09 UTC 5297 followers, XX engagements

"@casper_hansen_ 1M context Then forecast Qwen3-Coder. We need 1M for coding the most. And coding agents are all the rage atm. (admit maths olympiad excitement over the weekend took the spot light and that's chain or reasoning CoR; but that's in research still not productionised)"
@ljupc0 on X 2025-07-22 18:40:48 UTC 5303 followers, XXX engagements

"If not coding then guessing reasoning on top of existing model. Recent massive non-reasoning is K2. Maybe Kimi release variant of K2 but with reasoning on top. But feels too early+ don't know the people if their group is doing it. GLM guy the ones with GLM-4 GLM-Z1 - we have not heard from them since April So maybe them"
@ljupc0 on X 2025-07-24 06:56:21 UTC 5310 followers, XX engagements