Maybe, in-vivo genetic editing of the brain is possible. Adenoviruses that are a normal delivery mechanism for genetic therapy can pass hemo-encephalic barrier, so seems plausible to an amateur.
(Not obvious that this works in adult organisms, maybe genes activate while fetus grows or during childhood.)
Odds games against engine are played with contempt equal to matherial difference.
Sorry you didn’t know that beforehand.
Obviously fine. I posted here to get better than my single point estimate of what’s up with this thing.
The post expands on the intuition of ML field that reinforcement learning doesn’t always work and getting it to work is fiddly process.
In the final chapter, a DeepMind paper that argues that ‘one weird trick’ will work, is demolished.
The problem under consideration is very important for some possible futures of humanity.
However, author’s eudamonic wishlist is self-admittedly geared for fiction production, and don’t seem to be very enforceable.
It’s a fine overview of modern language models. Idea of scaling all the skills at the same time is highlighted, different from human developmental psychology. Since publishing 500B-PaLM models seemed to have jumps at around 25% of the tasks of BIG-bench.
Inadequacy of measuring average performance on LLM is discussed, where a proportion is good, and rest is outright failure from human PoV. Scale seems to help with rate of success.
In 7th footnote, should be 5e9, not 5e6 (doesn’t seem to impact reasoning qualitatively).
Argument against CEV seems cool, thanks for formulating it. I guess we are leaving some utility on the table with any particular approach.
Part on referring to a model to adjudicate itself seems really off. I have a hard time imagining a thing that has better performance at meta-level than on object-level. Do you have some concrete example?
Thanks for giving it a think.
Turning off is not a solved problem, e.g. https://www.lesswrong.com/posts/wxbMsGgdHEgZ65Zyi/stop-button-towards-a-causal-solution
Finite utility doesn’t help, as long as you need to use probability. So you get, 95% chance of 1 unit of utility is worse than 99%, is worse than 99.9%, etc. And then you apply the same trick to probabilities you get a quantilizer. And that doesn’t work either https://www.lesswrong.com/posts/ZjDh3BmbDrWJRckEb/quantilizer-optimizer-with-a-bounded-amount-of-output-1
Maybe people failure is caused by whatever they tweaked to avoid ‘generating realistic faces and known persons’?
Cool analysis. Sounds plausible.
So you’re out to create a new benchmark? Reading SAT is referencing text in answers with ellipsis, making it hard for me to solve in single read-through. Maybe repeating questions in the beginning and expanding ellipses would fix that for humans. Probably current format is also confusing for pretrained models like GPT.
Requiring a longer task text doesn’t seem essential. In the end, maybe, you’d like to take some curriculum learning experiment and thin out learning examples so that current memorization mechanisms wouldn’t suffice? Admittedly I don’t know much about that field.
Area of neural networks in search looks like a half of a simple long-term memory: just retrieval, but may have some useful ideas. Using an existing tool like recoll to search through the corpus doesn’t work because you can’t back-propagate through it. This lack of compositionality is always bothersome.
No particular philosophy: just add some kludge to make your life easier, then repeat until they blot out the Sun.
Non-computer tool is paper for notes & pen, filing everything useful to inbox during daily review. Everything else is based off org-mode, with Orgzly on mobile. Syncing over SFTP, not a cloud person.
Wrote an RSS reader in Python for filling inbox, along with org-capture. Wouldn’t recommend the same approach, since elfeed should do the same reasonably easy. Having a script helps since running it automatically nightly + before daily review fills up inbox enough novel stuff to motivate going through it, and avoid binging on other sites.
Other than inbox have a project list & calendar within emacs. Not maintaining a good discipline for weekly/monthly reviews, but much smoother than keeping it in your head.
I have a log file that org-mode keeps in order by date. And references file that don’t get very organized or used often. Soon will try to link contents of my massive folder of PDFs with it.
Oh, sorry. Javascript shenanigans seem to have sent me into antoher course, works fine on a clean browser.
Consider not wasting your reader’s time with having to register on grasple to be presented with 34-euro paywall.
I’d think ‘ethical’ in review board has noting to do with ethics. It’s more of PR-vary review board. Limiting science to status-quo-bordering questions doesn’t seem most efficient, but a reasonable safety precaution. However, typical view of the board might be skewed from real estimates of safety. For example, genetic modification of humans is probably minimally disruptive biological research (compared, to, say, biological weapons), though it is considered controversial.
My town …
Let 20% wards be swung by one vote, that gives each voter 1 in (5 * amount of voters) chance of affecting a vote cast on the next level, if that’s how US system works?
… elected officials change their behavior based on margins …
Which is an exercise in reinforcing prior beliefs, since margins are obviously insufficient data.
Politicians pay a lot more attention to vote-giving populations...
Are politicians equipped with a device to detect voters and their needs? If not, then it’s lobbying, not voting that matters.
...impact of your reasoning by the population who might follow it.
Population following my reasoning: me.
P.S. Thanks for hinting at other question, which might be of actual use to me.
Absolutely, shutting up and multiplying is the right thing to do.
Assume: simple majority vote, 1001 voters, 1 000 000 QALY at stake, votes binomially distributed B(p=0.4), no messing with other people’s votes, voting itself doesn’t give you QALY.
My vote swings iff 500 ⇐ B(1001, 0.4) < 501, with probability 5.16e-11, it is advised if takes less than 27 minutes.
Realistically, usefulness of voting is far less, due to:
actual populations are huge, and with them chance of swing-voting falls;
QALY are not quite utils (eg. other’s QALY counts the same way as your own);
You will rarely see such huge rewards (if 1 QALY ~ 50 000$, our scenario gave each voter free $50 M )
So, people who ‘need your vote’ in real-world scenarios are either liars or just hopeless.
I don’t care, because there’s nothing I can do about it. It also applies to all large-scale problems, like national elections.
I do understand, that that point of view creates ‘tragedy of commons’, but there’s no way I can force millions of people to do my bidding on this or that.
I also do not make interventions to my lifestyle, since I expect AGW effects to be dominated by socio-economic changes in the nearest half a century.
Things that I seem to notice about the plan:
Adjusting weights a plan for basic AIs, which can’t seek to e.g. be internally consistent, eventually landing wherever the attractors take it.
Say, you manage to give your AI enough quirks for it to go cry in a corner. Now you need to lower your AI nerfing to get more intelligence, leading to brinkmanship dynamics.
In the middle, you have a bunch of AI, trained for maximum of various aspects of incorrigibility, hoping they are incapable of cooperating; or for that any single AI will not act destructively (while trained for incorrigibility).