@loveofdoing

Philosopher, Self-taught Programmer, Self-taught Cognitive Scientist, Marine Corps Veteran, Will talk about Artificial General Intelligence with anyone

 New York, USA
loveofdoing.substack.com/
Joined September 2018
  • Tweets28K
  • Following1.5K
  • Followers5.3K
2.6KMedia
Once solutions to The Problem of Universals and The Problem of Induction are implemented within current AI technology, AGI will come within weeks.
36
5
8
157
is it possible for an llm to produce a visualization of its own weights during activation???
1
1
7
345
apparently spending on the api is death
6
256
loveofdoing retweeted
loveofdoing
@loveofdoingOct 24
Lyriccheck.com . use @pangram to check the lyrics of songs.
1
1
213
Lyriccheck.com . use @pangram to check the lyrics of songs.
1
1
213
You know the devil invented linear algebra to fight God, right?
8
4
2
39
1,590
Replying to @Memetic_Theory
MCTS over embeddings. That's so novel.
1
1
3
709
It’s kind of wild that some people think that embedding tables are conscious.
3
21
672
Whomst amongst you is brave enough to have an agent swarm decompile CUDA?
1
16
446
found an old ptxas bug: it compiled shr.s64 and shr.u64 to byte-identical SASS. a shift+add pattern got 'optimised' into a rotate even when the shift was signed, so every negative input came out off by 2. fixed in current CUDA, but it was live from sm_75 through sm_90
1
4
337
Is Metal better than CUDA?
1
8
450
cross-session inbound
180
You know government is just a centralized exchange for force.
5
21
626
loveofdoing retweeted
b
@rgbqcdOct 20
office hours: seal of approval edition nyc, dba bar, 6:45pm 2nd night back at dba! Meet your online friends irl, chat about twitter and things off twitter. Look for the rubber duck! (The bar prefers cash)
b
@rgbqcdOct 13
office hours: manhattan return edition nyc, dba bar, 6:45pm We’re back in the LES for our winter season, come hang out with your mutuals at dba! look for the rubber duck! (The bar prefers cash)
4
2
23
2,413
loveofdoing retweeted
Max Black
@Max_Black_4411Oct 19
Planning with a world model shouldn’t require thousands of rollouts. Meet HIJP: train the inverse → predict loop. No test-time search. 86.8% overall success, beating 8 baselines. Up to 16.5× faster than LeWM (CEM). Demo: hijp.demohub.live Paper: arxiv.org/abs/2610.23615
2
6
1
72
5,022
claric as the orientation, jev as the decider, the llm as the actor and observer, ooda
Muneeb shinwari
@imuneebshinwariOct 20
Replying to @michellechen
Separating extraction, decision, and action gives each stage a clearer contract. I’d keep “missing” distinct from a valid zero or false value, so an extraction gap cannot quietly become a confident downstream decision.
2
1
324
Replying to @michellechen
Separating extraction, decision, and action gives each stage a clearer contract. I’d keep “missing” distinct from a valid zero or false value, so an extraction gap cannot quietly become a confident downstream decision.
1
1
1
362
has anyone tried dictionary learning during training? like when a token or word has low prediction score just read it the definition from the dictionary
1
2
189
loveofdoing retweeted
michelle
@michellechenOct 19
a brand new model paradigm! introducing 𝚌𝚕𝚊𝚒𝚛𝚒𝚌: a system one model for extracting, computing, estimating, and predicting values from unstructured inputs. blog.cloudflare.com/clairic-…

Introducing Clairic, a new System One class of extraction models

Clairic is a new model from Cloudflare that is a System One extraction modeI, capable of extracting, computing, estimating, and forecasting data from unstructured inputs.

blog.cloudflare.com
128
206
121
2,276
328,715
Other than compute costs everyone at home can train a GPT-3 level maybe GPT-4 level model at home right now.
3
6
320
neurons aren’t superposed but continuations are superposed???
1
6
370
Load more