Keep on to blur preview images; turn off to show them clearly
Trump to Xi: "Bro this meeting could have been an email"
RL and efficient distributed pretraining • eXperiments lab • memes and training lores
math provers remain my standard meter bar. if goedel hit sota on 32b, this is probably all you need to solve hardest problems at the moment.
Reasoning models coming (very) soon. Co-founder @pleiasfr
interestingly, as use case become more complex and mature, first time i’m becoming constrained by model size. 30b dense or 50-150b active is likely becoming a sweet spot.
math provers remain my standard meter bar. if goedel hit sota on 32b, this is probably all you need to solve hardest problems at the moment.
hyped (but decidedly won’t escape the big dive in moe finetuning)
interestingly, as use case become more complex and mature, first time i’m becoming constrained by model size. 30b dense or 50-150b active is likely becoming a sweet spot.
"I just don't see Cursor or Windsurf having enough cashflow on-hand to build a training cluster and gather all the data and teams you need to build a model from nothing." Meanwhile Chinese food delivery app.
Reasoning models coming (very) soon. Co-founder @pleiasfr
"I just don't see Cursor or Windsurf having enough cashflow on-hand to build a training cluster and gather all the data and teams you need to build a model from nothing." Meanwhile Chinese food delivery app.
Reasoning models coming (very) soon. Co-founder @pleiasfr