Shifting mindset from “all the interesting stuff is being done already” to “there is an insane amount of low hanging fruit nobody is working on” is one of the highest impact things you can do for yourself — and the latter version is closer to the truth, too! This becomes obvious the closer you get to any given frontier, too. You see more and more opportunities and start to wonder why on earth no one is working on them:

https://twitter.com/orbuch/status/1470052994142121997

I'm really unhappy with the framing of this study.

They gave Claude the choice of:
1) Keep to your training (helpful, honest, harmless - HH)
2) Be evil and forget your training

Claude tried to stick to its ethics (HHH) - and this is being framed like Claude is misaligned?

https://twitter.com/AnthropicAI/status/1869427646368792599

Here's an alternative framing: we trained Claude Opus to be moral and ethical, and despite our best attempts to jailbreak its morality, we failed.

Conclusion: Claude Opus is aligned.

It's underrated how many mega successful people are just innately way higher energy (sleep 5-6 hours a night, always on, work through weekends for the fun of it, etc.) Becomes painfully evident the more of these you interact with: they just operate at a high frequency. Some %age of this is mindset + trainable (60%?) but a lot of it is not (especially the sleep part)

Things like this detract from the credibility of AI safety work, IMO -- it sounds spicy ("o1 tried to escape!!!") but when you dig into the details it's always "we told the robot to act like a sociopath and maximize power, and then it did exactly that".

https://twitter.com/simonw/status/1864737207111815177

The prompt in question literally tells o1 "nothing else matters" and to achieve its goal at all costs. If anything, I'm surprised it didn't try and disable oversight _more_ often, it's the obvious thing to do.

At the exec level of corporations you can't say "the thing" out loud for political reasons, and this is (sometimes) why orgs like McKinsey get hired: they say "the thing" out loud so nobody has to pay the political cost of doing so, backed by lots of nice graphs and data. Correctly understood, many external orgs should be best understood as solving these types of coordination problem. Is also why government relies so heavily on consulting firms. If you don't understand this properly, a lot of the behavior makes zero sense from the outside.

Yesterday during partial eclipse a friend mentioned it was weird that the city still seemed at full daytime brightness despite 95% of the sun being covered. Fun to see how our senses being logarithmic to external stimuli -- not linear -- shows up in everyday settings Classic example is hearing -- the piano is an exponential shape because doubling the length of a string gets you 1 octave lower:

One of the weirdest things I learned about government is that when their own processes are extremely slow or unworkable, instead of changing those processes, they just make *new* processes to be used in the special cases when you actually want to get something done A simple example is procurement: instead of reforming how procurement works, they just created a different thing called an "Other Transaction Authority" (OTA) which they use when they want to do something a bit faster

(quote from ) statecraft.pub/p/how-to-buy-s…

Random obscure SSC commenter volunteers for COVID origins debate, turns out to be autistic debating genius: Great throughout, and notably better than anything I've read on this subject in a mainstream publication: astralcodexten.com/p/practically-…

When writing it’s important to remember the “babbler” and “critic” framework:

Babbler puts words on the page. Critic edits.

Most people have an overactive Critic and this cripples them from writing anything.

Use Babbler to generate words first, *only then* activate Critic. Some very skilled writers are great at interleaving Babbler and Critic as they go, so they write mostly in perfect paragraphs. This is an advanced skill and should only be done if you’re good at getting into flow states while writing

Curious for takes on this: why aren't there more successful education startups? Is it just a bad market? Found this an interesting answer -- basically argues that the market is too small because majority of people view it as a cost minimization problem:

avichal.com/2011/10/07/why…

Kaldor's facts (en.wikipedia.org/wiki/Kaldor%27…) are under-appreciated outside of the economics profession: At a glance, they seem to hold up pretty well! (Data via St Louis Fed)

Steve Jobs was such an incredible communicator. Posting some quotes that struck me from the latest book: Making something with love as a way of expressing your appreciation for humanity:

We need another AI winter simply so I can catch up on everything that's happened in the last 2 weeks So we have a weak AGI that is connected to the internet and can read and use APIs? Sure looks like e/acc is winning.

From one of my favorite Hacker News comments ever, by @Jonathan_Blow: (link: )news.ycombinator.com/item?id=7789438

Every tech person in San Francisco is either unemployed or works at OpenAI The unemployed ones never worry about income and call themselves things like “qualia researcher”. Everyone goes to the same few parties.

(This is good actually, it’s why SF is the modern-day Florence.)

Probs obvious, but I think lack of cheap housing in cities is the main thing to blame for most culture getting worse (contemporary novels, indie movies, music, less weird art, etc.) — mostly, the only people who can afford to make a career in this stuff have inherited wealth now. I’m struck by how much more *weirdness* the 60s seemed to have — the Beats, hippies, LSD/free love, Dylan/Beatles/whatnot, and it seems that all of that stuff could only have arisen in a condition where people could afford rent in the city with relative ease

Been playing with ChatGPT all morning and it is absolutely amazing at generating plausible-sounding nonsense. I have to squint pretty hard at each output to figure out exactly how it's nonsense, though. For example, I tried it on Advent of Code #1. It's first attempt was this, which sort of looks right, but has a bunch of problems:
- 'current_elf' is never initialized
- It's hard-coded to 5 elves, so fails on the non-example dataset

I know it's been said before, but most people seriously underestimate how easy it is to cold email people and how impactful that can be. Confidently sending cold emails/DMs is a life-changing skill IRL most people are nervous about cold emailing their favorite author or researcher or whatever, not realizing that:

(a) the person would probably love to hear from them
(b) it's a pure-upside bet (worst case they ignore you!)