@emollick Isn’t this debate irrelevant? They’ve proven that MLA+MOE+RL w/o proprietary training sets outperform at trivial cost. Those ideas are everyone’s to take now, nobody needs to use the actual weights when training is this cheap. If not deepseek, everyone will use some derivative.
@benspringwater I think some of the confusion is thinking of Lewis as anything _but_ a storyteller. He very openly wants to be Tom Wolfe
https://t.co/w3UC6HiMGB
@lessin Challenge with the “Money” approach is that you will attract the worst candidates (terrible but interview well or prestige backgrounds) and you have to be excellent at spotting the difference. With mission/misfits you have an advantage of high signal to noise
@goodside I mean...Not surprising? They tested OOD using Swahili/Filipino/etc while the multilingual models used here were trained on datasets (wikipedia for mbert) with <0.1% representation of those languages.
@Broncho24 Richard Rodes “Energy” is much more of a general history than most of those above going from wood to nuclear with a lot of context around each transformation.
@headphoneDas Had the reverse experience with my kids listening to the lion king. “Wait…is this the gladiator guy”. That guy does everything! Dad lives everyone dies, dad dies everyone else lives, he can score it all
@parkerconrad@natfriedman@OpenAI +1! From the self-censored examples it seems they might have a secondary model (or separate model head) that detects offense and then produces a stock response - or - a smaller model generates a trained response. Stable diffusion does something similar (but nulls the output)
@CJHandmer I thought it was pretty fair. He was kind to many of the early projects like Owens Valley (from a water use standpoint at least) and Hoover. He just let it rip on the later projects that were just keeping engineers busy at otherwise negative net value
@hkarthik Let’s be careful here - you and I can’t both be top 25% in our org! Rather than working too hard, we should pull straws to decide who survives
@JayNDonde I agree with your point on unclear goals. We seem to aim to increase the mean education level but I think society is better served if we split resources b/w the 25th percentile (foundational skills for all) and 99th (more Nobel prize winners). Average is a waste for most topics
@albrgr The Making of the Atomic Bomb is such an incredible combination of history and science (and the history of the science). Very under rated. If you liked it Dark Sun is an awesome follow up on the science while American Prometheus is a great afterword on the politics.
@pt@rsg Ive always thought taxes should have some gift comparable to a NPR fundraising campaign. Would be cool to see billionaires flex 💪🏽 by pulling out their IRS umbrella
@kaehler1920@typesfast Tax meat to lower demand/cost for grains, suspend the tax for Ukrainian livestock so they can convert their wheat->meat and march it across the land border.