As a counterpoint, Sydney showed aligning these models on the first go, and even discovering unsafe behavior is non-trivial.

The flip side of this is that it helps to speak in a way that’s almost repetitive to make sure someone who zoned out can catch back up.

For example, frequently once the subject/object of a sentence is known, people refer to the subject/object with pronouns instead of repeating the subject/object. If someone misses the initial definition, they are immediately lost.

(This also applies somewhat to writing — above, I’m repeating “subject/object” instead of “it”. People also zone out during reading, and this repetition saves them from having to scroll back up and figure out what a pronoun is referencing.)

Did you feel a subjective increase in your intelligence? E.g. feeling like you’re thinking faster, more clearly, having a better memory?

Cool project!

I’m curious—what does the long tail of websites look like for you? For me, it’s the small number of sites that i repeatedly go to (twitter, youtube, hackernews, etc…) that take up the vast majority of my wasted time.

(Btw, I also built my own website blocker: https://chrome.google.com/webstore/detail/webblock/jeahkphmdfbddenabgndnooheiciocka)

I think the main beneficiaries of being able to sideload apps will be incumbents, not startups. Big companies like Spotify, Netflix, and Tinder will offer users discounts if they sideload because it will spare them the 30% Apple tax.

I am confused how to square your claim of requesting extra time for incontrovertible proof, with Ben’s claim that he had a 3 hour call with you and sent the summary to Emerson, who then replied “good summary!”

Was Emerson’s full reply something like, “Good summary! We have incontrovertible proof disproving the claims made against us, please allow us one week to provide it?”

Bloomberg reported 2 weeks ago that Twitter resumed paying Google Cloud: https://www.bloomberg.com/news/articles/2023-06-21/twitter-resumes-paying-google-cloud-patching-up-relationship

You might want to clarify that, because in the post you explicitly say things like “if your goal is to predict the logits layer, then you should probably learn about Shakespearean dramas, Early Modern English, and the politics of the Late Roman Republic.”

This is probably obvious, but maybe still worth mentioning:

It’s important to take into account the ROI per unit time. In the amount of time it would take for me to grok transformers (let’s say 100 hours), I could read ~1 million tokens, which is ~0.0002% of the training set of GPT3.

The curves aren’t clear to me, but i would bet grokking transformers would be more effective than a 0.0002% increase in training set knowledge.

This might change if you only want to predict GPT’s output in certain scenarios.