Loading video...
Video Failed to Load
Dwarkesh calmly shreds Zuck's argument for open-sourcing AGI. The flimsy wishful thinking behind Meta's reckless actions has been exposed. Another incredible job by Dwarkesh Patel.
115,123 views • 2 years ago •via X (Twitter)
11 Comments

Full podcast:

The moment in my clip when Zuck freezes (1:44) reminds me of how Sam Harris broke @pmarca

@dwarkesh_sp Liron, have you ever felt that you were born in a wrong era and wrong country? You'd have fit well with NKVD. Your fetish for centralization and putting kulaks in their place is pretty obscene. Re BiOWeaPOnS: Yes, we'll do a bunch of R&D, on ambient DNA sensors and UV filters.

@dwarkesh_sp I don’t have a fetish for centralization, I just think we’re fucked and the only proposal to save ourselves is bad.

@dwarkesh_sp CRISPR Cas 9 was virtually opensourced how many years ago? decade and a half? And the sequence code for ultra deadly smallpox was made public for how many decades now? The evidence for this argument is not strong & all cases outside the argument show its better to opensourcing

@dwarkesh_sp Where was the shredding

@dwarkesh_sp 1:44 Mark freezes and gives answers that no longer back up his original claim that open-source AI is better for safety.

@dwarkesh_sp How does AI significantly accelerates the production of a bioweapon? If someone has the intention and resources to produce a bioweapon they can already do so without AI, doesn't really seem like a significant factor to me.

@dwarkesh_sp Bioweapon was just an example of a situation where attack seems easier than defense, and democratizing the ultimate superpower may simply end the game when a single terrorist presses the attack button.

@dwarkesh_sp I really don’t get it. What is so dangerous about these models vs. *the internet* All this bioweapon stuff is in the training data because it’s crawled and easily searchable, and tbh searching google is easier than jailbreaking the RLHF. This seems histrionic

@dwarkesh_sp Intelligence itself is dangerous, existential past a certain threshold, and the models are getting more intelligent. It’s not about the particular algorithms doing the work, it’s about the type of work being done.

