Keep on to blur preview images; turn off to show them clearly
RT @eriktorenberg: Stoked to announce our investment in Unlimited. Alex and Tara and team are a force. Watch this space.
Co-Founder, American Dynamism. General Partner @a16z. Catholic. Mother. American. 🇺🇸 🚀💪
RT @a16z: Large-scale infrastructure buildout is essential to the next decade of American Dynamism. Making it happen will require leaps in…
Co-Founder, American Dynamism. General Partner @a16z. Catholic. Mother. American. 🇺🇸 🚀💪
Gemini has this weird privacy policy. You either: - only have temporary chats, or - let your chat data be used for training and human review. @OfficialLoganK, @demishassabis, @JeffDean: If this policy changes to match ChatGPT, I bet you will see Gemini DAU & paid users rise.
@OfficialLoganK @demishassabis @JeffDean This is the #1 reason my friends are unwilling to switch from ChatGPT Pro to Gemini. Also, please don’t add ads to Gemini, at least not before ChatGPT does.
A Staircase of Sigmoids I dug in on a reference after reading the Dwarkesh piece that had me quite worried. In the post he mentions: "Toby Ord has a great post where he cleverly connects the dots between different o-series benchmark charts, which suggested “we need something like a 1,000,000x scale-up of total RL compute to give a boost similar to a GPT level”." This makes it sound like things are going to stall. Toby goes so far to say two things that had me worried: 1. "We’ve seen impressive gains, but these were only viable when starting from such a low base. We have reached the point where it is too expensive to go much further." 2. "Now that RL-training is nearing its effective limit, we may have lost the ability to effectively turn more compute into more intelligence." I asked around how legitimate this claim was and what was the opposing bull case for scaling up RL. A good friend said: "Yes, if you keep naively scaling RL compute, it’ll scale poorly. But we shouldn’t do that! That’s why there’s so many RL env companies. A better approach is to scale to newer, more difficult environments. " After pondering things a bit, I found a way to compress this further to: "Imagining it as a staircase is sigmoids for new tasks, worlds, goals is the most helpful way to think about how it can keep going for a while."
Founder: @mixpanel Pizzatarian, programmer, music maker
RT @zabie_e: In 1930 the Empire State Building was built in just 410 days. Today we’ve lost the muscle to build what matters with that kind…
Co-Founder, American Dynamism. General Partner @a16z. Catholic. Mother. American. 🇺🇸 🚀💪
RT @alexmodon: For most of my life, I assumed technology only moved in one direction: Faster, better, cheaper. Progress felt inevitable.…
Co-Founder, American Dynamism. General Partner @a16z. Catholic. Mother. American. 🇺🇸 🚀💪