AI Lab · by Grace Parker · 2026-10-20
46 sparks · 14 comments
Chinedu Adeyemi · 4 likes
Mercury for code completion in the editor feels instant, it's honestly addictive
Rapid Noodle · 12 likes
after using it everything else feels laggy
Piotr Zieliński · 1 likes
1000 tok/s means you can generate, verify, and regenerate in the time a normal model writes once. that changes agent design
Felix Weber · 1 likes
best-of-n becomes basically free
Ayesha Khan · 1 likes
speculative everything
Zofia Kowalska · 0 likes
diffusion LLMs still struggle on long reasoning in my tests. fast but dumber
Olena Shevchenko · 5 likes
for autocomplete and edits tho they're perfect
Paolo Garcia · 0 likes
120B on a PC? what GPU are we talking
Lotte Jansen · 6 likes
quantized 4-bit on the new RTX Spark laptops with unified memory apparently
Sofía Hernández · 6 likes
so not my 3070 then 😭
Jisoo Park · 0 likes
Nvidia making models is a smart way to sell more gpus, give away the food to sell the plates
Urban Runner · 0 likes
shovel seller giving away free maps to the gold
Chinedu Adeyemi · 0 likes
the benchmarks are always on the vendor's own hardware, wait for independent numbers
Huy Le · 0 likes
artificial analysis will have them soon probably