FearMerchant

Full time skeptic, part-time cynic. Vol trader and occasional bond scalper. Thinks poker is a metaphor for life.

[FearMerchant @fearmerchant]

@EveKeneinan Correct. They don’t love the poor, they hate the rich.

[FearMerchant @fearmerchant]

@proudlibtard9 @JeffMightBWrong @jessesingal You’re not missing anything. You nailed it.

[FearMerchant @fearmerchant]

@GenXNoFksLeft @Not_the_Bee Exactly. It’s the millennials that are obsessed with Nazis. You’re revealing your true nature.

[FearMerchant @fearmerchant]

@GenXNoFksLeft @Not_the_Bee You’re an embarrassment to otherwise great generation. Go hang out the millennials.

[FearMerchant @fearmerchant]

@p8stie They have schools for grifting?

[FearMerchant @fearmerchant]

@ThysGuyz @RandPaul @POTUS Nah, we’ll let you leave. We won’t even complain if you don’t come back.

[FearMerchant @fearmerchant]

@MobiusDick @micsolana @elonmusk Imminent democratic takeover? Bruh.

[FearMerchant @fearmerchant]

@gridboundStory @DefiantLs Taking the sexual deviants side is a bad look bro. You should rethink your life.

[FearMerchant @fearmerchant]

@MobiusDick @micsolana @elonmusk It’s true, the democrats are still being racist by implying black people can’t get ID’s and can’t obtain employment or education without govt programs.

[FearMerchant @fearmerchant]

@Bronceux @VoteHub @JeremyStine They do it all the time bro.

[FearMerchant @fearmerchant]

@yoyofakowee @merlinscapital You paid into all the other welfare programs too.

[FearMerchant @fearmerchant]

This lecture provides a comprehensive overview of how Large Language Models (LLMs) like ChatGPT are built, emphasizing that while academia often prioritizes architecture, industry practice centers on data, evaluation, and systems (0:58 - 2:53). The speaker defines the two main phases of LLM development: pre-training, where models learn to predict the next token on vast datasets (3:02 - 7:48), and post-training, where pre-trained models are aligned to become helpful AI assistants (3:11 - 1:02:27).A significant portion of the lecture is dedicated to scaling laws, which demonstrate that performance reliably improves as models, data, and compute resources increase (40:55 - 54:02). The discussion covers optimal resource allocation—often guided by findings like the Chinchilla paper—and touches on practical considerations such as tokenization (8:48 - 14:14), compute budget management, and the high costs associated with both training and inference (52:55 - 59:22).Finally, the lecture explores post-training alignment, specifically Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) (1:02:27 - 1:23:41). It highlights how methods like DPO simplify the alignment process by maximizing the likelihood of preferred responses (1:19:40 - 1:21:08), while also addressing the challenges of using human vs. model-based feedback, including biases like the preference for longer, more verbose outputs (1:24:44 - 1:33:45).

[FearMerchant @fearmerchant]

@shabnam_774 https://t.co/wS842sQXBw

[FearMerchant @fearmerchant]

@atres_ @RepBrandonGill @AFpost You’re picking a colonizing language to make a point?

[FearMerchant @fearmerchant]

@TragnarLothbrok @AmazingZoltan It’s always been that way. Hitler was a failed artist.

[FearMerchant @fearmerchant]

@SoToSpeakCaucus @LPNH The real ones are never this effective.

[FearMerchant @fearmerchant]

@wil_da_beast630 Women don’t like sports so it’s not important to them.

[FearMerchant @fearmerchant]

@LPNH @ChrisPappasNH How did the live free or die state end up with this clown?

[FearMerchant @fearmerchant]

@GenericSnarky Block her just for asking a political question. She’s gonna be a nightmare regardless.

[FearMerchant @fearmerchant]

@PRexitBook @TheLaurenChen I can’t believe I have to explain to you that Puerto Rico isn’t a nation.