tominous

Here's one (CIFSwitch) from a couple of days ago: https://heyitsas.im/posts/cifswitch/

Performance and efficiency are important, but we need you to invent the monitoring tools and visualisations that will underpin alignment!

Yes there are metabolic changes in the mother herself during pregnancy but that's not why it's hard to research. The main fear is that drugs will cross the placenta and affect the growing fetus, or similarly be transmitted through breast milk to an infant. Very young humans are uniquely vulnerable to disruption in their growth that can cause life-long problems.

Maybe not due to attention, but it is certainly possible for chat content to get leaked into other conversations due to bugs in the stack, and in fact it has happened before.

https://openai.com/index/march-20-chatgpt-outage/

"We took ChatGPT offline earlier this week due to a bug in an open-source library which allowed some users to see titles from another active user’s chat history. It’s also possible that the first message of a newly-created conversation was visible in someone else’s chat history if both users were active around the same time."

You are probably right about this particular LaTeX issue though.

In my case I had hundreds of invoices in a not-very-consistent PDF format which I had contemporaneously tracked in spreadsheets. After data extraction (pdftotext + OpenAI API), I cross-checked against the spreadsheets, and for any discrepancies I reviewed the original PDFs and old bank statements.

The main issue I had was it was surprisingly hard to get the model to consistently strip commas from dollar values, which broke the csv output I asked for. I gave up on prompt engineering it to perfection, and just looped around it with a regex check.

Otherwise, accuracy was extremely good and it surfaced a few errors in my spreadsheets over the years.

If you're trying to say that immunostimulant is a term of art distinct from immune checkpoint inhibitor, then fair point.

But what it seems like you're saying is checkpoint inhibitors don't "enrage/activate" the immune system, and they only target tumours, and both points are misleading if not outright wrong.

Immunostimulant is a reasonable way to put it: it takes the brakes off one part of the immune system. (Similarly, we call caffeine a central nervous system stimulant and it works by blocking adenosine, which is one of the brakes on brain activity.)

And it's not always specific to tumours. My wife's thyroid got wiped out. The endocrinologist said it was like bombing a paint factory: first a massive spike in thyroid hormones, then a crash as no more were produced.

More info on the potential side effects of immune checkpoint inhibitors: https://www.cancer.gov/about-cancer/treatment/types/immunoth...

I have very quickly picked up the habit of pasting snippets of my code into GPT-4 and simply asking "Why is this not working?" Almost every time, it succinctly explains the apparent purpose of the code, and how it is subtly wrong.

It's so good that I often do this preemptively to avoid a compile/deploy/test cycle.

I was dying on the GPT-4 API waitlist too. I built a proof-of-concept with GPT-3.5, got some ada embeddings, played around with some common patterns for a couple of weeks, spent less than $20. I then applied to the waitlist again with a few short sentences about what I'd done, how GPT-4 would make it better, and how it would enable something new and valuable for a particular market. Approved that day.

It's not exactly a shortcut, and maybe it was just luck, but I suspect the key is just to start building with what you have and show a trajectory. The best part is that coding with ChatGPT-4 as a "colleague" has made the whole thing super fun.

Same for me. I see it as the combination of 5 things:

1. It lowers the activation energy of getting started; for me motivation tends to come after action, and it gets me to that point.

2. It taps into the intense drive of "someone is wrong on the internet!" as per https://xkcd.com/386/

3. It's a new and different way of working so it doesn't trigger the same mental scar tissue.

4. It's vaguely social and high-status (delegating tasks, correcting and reviewing), as opposed to feeling solitary, so it tickles different reward centres.

5. More gets done for the same effort, so the cost/benefit equation is better and the feedback loop tighter.

I just spent a month doing an E-Business Suite platform migration and it was very similar: follow the step-by-step instructions to apply patches and run commands. Each patch has a README file with dependent patches or commands which need to be completed first. It works mind-numbingly great until you run into the first of many circular dependencies.

That's one problem with treating the implementer as a machine to run code. The whole procedure can't be tested, so when parts are changed they can break the whole. It relies on the human in the loop to resolve the conflicts, which is not repeatable.

The other problem is the "mind-numbing" part. No-one can maintain 100% perfection all of the time. And in the context of presenting to people who don't know what it all means, I can see why mistakes would be made.

It's ok, because they often make you give your postcode as a second factor.

As a native speaker of English, I didn't know about mass nouns until I was an adult and I was trying to explain to a non-native speaker that they were using "the" too much.

ProxyJump (or the -J option as a shortcut) plus TCP port forwarding (-L/-R) is ridiculously powerful. You can chain jump hosts with commas, threading a port forward deep into (or out of) multiple server hops with a single command.

In animals some things are definitely hardwired. I had a pet rabbit and made the mistake of watching a nature documentary about snakes when she was in the room. The mere sound of a rattle snake on TV caused the rabbit to stamp a few times, spray urine, and run away to her hutch (skidding on the floor as the left).

I'm sure humans have similar instincts. For example, the tendency of young kids to avoid plants until they are shown that they are edible - possibly leading to the general preference of artificial foods over anything that might be found in a garden. [1]

I've also noticed that toddlers don't gravitate towards knives the same way they do everything else. I certainly wouldn't rely on this behaviour, but I've noticed it in my own children and others.

[1] https://www.theatlantic.com/health/archive/2013/12/why-babie...

Yes I said that. Still -- many, many hours of cryptic crosswords being solved and explained if you want to see how the best in the world does it.

Cracking the Cryptic, e.g.: https://youtu.be/ecytVCusIeY

It's mostly a variant sudoku channel but Mark Goodliffe is an insanely good cryptic crossword solver, and both Simon and Mark are good teachers.

Beyond the intro the key skills are: parsing the clues into definition and wordplay, knowing signifiers for different types of wordplay, recognising common abbreviations (there are lists a web search away), recalling short synonyms, and anagramming.

If you want to get your mind blown I highly recommend spending the $2 for Cracking the Cryptic's Patreon to watch Mark solve the Monthly Club Special using almost pure wordplay and whatever string manipulation coprocessor he has in his head (which also seems to work for sudoku too).

My mind was blown when visiting Sweden in 2002 and finding out they pronounce "www" as "vee vee vee" (not just because the letter name double-v makes much more sense than double-u, but because they were smart enough to shorten it in this case).

Since then I've pronounced it in English as "dub dub dub" and found people generally understand it. I hope it catches on.

Congrats to Scott. I'm still somewhat disappointed Brendan Gregg didn't pop up there making the AI Alignment equivalent of flame graphs!

Not denying that there are many ways in which development can go wrong, whether that's genetic or the environment. What I'm responding to is the sudden change seen in many populations as per the article.

Re: the chromatic aberration spread, see e.g. https://pubmed.ncbi.nlm.nih.gov/3716229/

They measured 1.82 diopters between 420nm and 660nm. Visible light is a bit wider, 380nm to 720nm, so 2 diopters is about right.

You easily can see this in a dark data centre. If your vision is perfectly corrected (which normally means perfect for red-green light) you won't be able to focus on a blue LED unless you are about 50cm away.

Almost all children are born hypermetropic (far-sighted). The growth of the eyeball is calibrated by the amount of defocus (blur) experienced. More defocus, more growth, which reduces hypermetropia.

My hypothesis is that less sunlight means wider pupils and more defocus (contrast to a pinhole camera), so more eye growth.

Even with perfect eyesight, there's a 2-diopter difference between focussing blue light and red light due to chromatic aberration. Only narrow pupils can reduce this blur in white light.

I'm also curious whether older TVs with big blurry pixels (or low res images upscaled and smeared onto newer screens) trigger the same mechanism. These days I certainly feel physically uncomfortable looking at media like that, like my eyes aren't focussing quite right.

The other problem is that once eyes have grown too long and are myopic (near-sighted) they experience even more defocus and grow more. See for example [1] which discusses how under-correction of myopia accelerates progression.

[1] https://reviewofmm.com/does-the-undercorrection-of-myopia-in...

OpenAI, on AI alignment instrumentation?

Third option: the test ID is generated well before the test is taken.

When I donate blood the nurse pulls out a sheet of around 20 pre-generated identical barcode stickers, and attaches one to each piece of paper and bag of blood. This uniquely identifies my donation end-to-end.

Djokovic's PCR tests were done at two separate labs. It is quite possible that each lab is bulk-allocated a unique range of IDs every day or week from the central authority. You'd see big "time jumps" from one lab to the other if you assumed the ID was always in time order.

Depending on how they handle their range (e.g. if they are handing physical stacks of barcode stickers around) the code may not even always be in time order in the same lab.

I assume they meant T^4 radiation (Stefan-Boltzmann law): https://en.m.wikipedia.org/wiki/Stefan–Boltzmann_law

Interesting analogy but it doesn't always hold. Take the greening of Ascension Island [1] where planting trees created a virtuous cycle: the trees trapped more moisture which promoted the growth of more trees.

In biology there are many complicated feedback loops and in some cases a parameter seems to get stuck outside of the normal range. In those cases, bringing the parameter back into the normal range can be enough to allow the normal feedback mechanisms to work properly again.

Some examples:

* Eczema: There can be a loop between irritation, inflammation, damage to the skin, more irritation. Treat the skin with steroids to both improve symptoms and disrupt the loop.

* Myopia: Studies have shown that undercorrecting for near-sightedness actually makes it worse [2]. Bringing it back into the normal range doesn't just help the symptoms, it also makes it progress slower.

* Depression: Maybe operating at a higher level, but there's a clear loop between changed behaviour and improved motivation.

Even something like acidity could be like this. You could speculate a link between gut bacteria and acidity, whereby changing the acidity directly could alter the balance of bacteria which locks in the change in acidity.

[1] http://faculty.washington.edu/dcatling/Catling2012_GreenMoun...

[2] http://www.journalofoptometry.org/en/under-correction-human-...

The cost isn't just in dollars though. In a PET scan they inject a radioactive dye into your bloodstream. It "exposes you to around the same amount of radiation that you would receive from the general environment over about three years." [1]

The high dollar cost may also be related to newer, more sensitive machines that don't require as much radiation exposure to the patient.

[1] https://www.betterhealth.vic.gov.au/health/conditionsandtrea...

What you say makes sense, but the people involved didn't even attempt to make me aware of the differences in the concise way you just did. And when I asked about the difference in final result, pain, and time, there was no straight answer. It was just, "This is what I do. I can't tell you about the other thing."

The standard approach to orthodontics is to extract teeth to make room and then move the remaining teeth into place. There is an alternative approach which involves enlarging the jaw, either through expanders, surgery or a combination of both.

Unfortunately this is an area where the patient or parent often has to make the decision because specialists only know about what they do. As a related example, I asked a dentist and an orthodontist what was more suitable for me, Invisalign or braces. They could each tell me about the time and likely result of their own specialty, but refused to offer any opinion on what was better for me and my situation.

Somewhat related, there was an interesting article about the wildly different advice given to someone who visited 50 different dentists [1].

Bottom line is seek out multiple opinions and don't expect someone else to guide you on the bigger picture.

[1] https://www.dentistat.com/ReaderDigestArticle.pdf

Paradoxically, more frequent fires can often mean less severe fires. That's the rationale behind controlled burns and traditional aboriginal land management practices.

And because this is Hacker News, have a play with the simulation hosted at Fourmilab: http://www.fourmilab.ch/cellab/manual/rules.html#Forest