@karpathy · daily metrics resolved at 00:00 UTC
a 20-min walkthrough of tokenisation from scratch. no libraries, no slides, just the byte pairs. this is the part of LLMs most people skip and it explains more weird behaviour than anything else.
the hottest new programming language is english. still true two years later, and more so.
llm.c now trains gpt-2 xl on a single node in 4 hours. 2,300 lines of c. no framework.
@navaldaily metrics resolved · 14.4M impressions•
@samadaily metrics resolved · 8.3M impressions•
@patrickcfollower series updated · +1.2K•
@paulgdaily metrics resolved · 3.1M impressions•
@elonmuskengagement histogram rebuilt•
@garrytandaily metrics resolved · 1.9M impressions•
@pmarcafollower series updated · +840•
@leviedaily metrics resolved · 1.1M impressions•
@brian_armstrongengagement histogram rebuilt•
@karaswisherdaily metrics resolved · 760K impressions•@karpathydaily metrics resolved · 2.7M impressions•
@nytimesdaily metrics resolved · 11.8M impressions•
@MKBHDfollower series updated · +3.4K•
@federalreserveengagement histogram rebuilt•
@tim_cookdaily metrics resolved · 4.6M impressions•
@michaeljburryfollower series updated · +2.1K•
@levelsiodaily metrics resolved · 918K impressions•
@sundarpichaidaily metrics resolved · 3.3M impressions•
@dhhengagement histogram rebuilt•
@FTdaily metrics resolved · 2.2M impressions•
@lexfridmanfollower series updated · +1.7K•
@DarioAmodeidaily metrics resolved · 1.4M impressions•
@BillGatesdaily metrics resolved · 6.9M impressions•
@dwarkesh_spfollower series updated · +460•