This is an opportunity to learn even more cool facts in the exquisite form of the corrections of my errors. Hooray!
Also I just made a patreon! I want to hire an editor so I can spend more time writing and recording and less time smashing my head into the computer.
Patreon: https://www.patreon.com/c/HGModernism
Excellent podcast: https://brainsciencepodcast.com/bsp/2023/210-neurotransmitters
Citations:
Coanda Effect Harvard Natural Sciences Lecture Demonstrations “Coanda Effect”
Armenian P, Whitman JD, Badea A, et al. Notes from the Field: Unintentional Fentanyl Overdoses Among Persons Who Thought They Were Snorting Cocaine — Fresno, California, January 7, 2019. MMWR Morb Mortal Wkly Rep 2019;68:687–688. DOI: http://dx.doi.org/10.15585/mmwr.mm6831a2. https://www.cdc.gov/mmwr/volumes/68/wr/mm6831a2.htm
Moss MJ, Warrick BJ, Nelson LS, McKay CA, Dubé PA, Gosselin S, Palmer RB, Stolbach AI. ACMT and AACT Position Statement: Preventing Occupational Fentanyl and Fentanyl Analog Exposure to Emergency Responders. J Med Toxicol. 2017 Dec;13(4):347-351. doi: 10.1007/s13181-017-0628-2. Epub 2017 Aug 25. PMID: 28842825; PMCID: PMC5711758. https://pmc.ncbi.nlm.nih.gov/articles/PMC5711758/
Del Pozo B, Rich JD, Carroll JJ. Reports of accidental fentanyl overdose among police in the field: Toward correcting a harmful culture-bound syndrome. Int J Drug Policy. 2022 Feb;100:103520. doi: 10.1016/j.drugpo.2021.103520. Epub 2021 Nov 14. PMID: 34785420; PMCID: PMC8810663.
Kurczewski, Nick “Science of New Car Smell” Car and Driver https://www.caranddriver.com/features/a36970626/science-new-car-smell/
Daley, Sherri “Secrets of that New Car Smell” Car and Driver https://www.caranddriver.com/features/a15133792/new-car-smell/
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems, pages 3111–3119 https://arxiv.org/abs/1310.4546
Farrant, Theo & AP “Denmark’s National Museum explores the life of the formidable Viking sorceress Völva” EuroNews https://www.euronews.com/culture/2024/06/26/denmarks-national-museum-explores-the-life-of-the-formidable-viking-sorceress-volva
Leroy, Fernand (2001). Histoire de naître: de l’enfantement primitif à l’accouchement médicalisé. De Boeck Supérieur. pp. 100–101.
Zarevich, Emily “The Myth of the Papal Toilet Chair” JSTOR Daily https://daily.jstor.org/the-myth-of-the-papal-toilet-chair/
Chapters:
0:00 Intro
0:38 Asymmetrical Animals: Cyclops and Fenestrons
4:20 Smell Facts: Puke Trees and VOCs
9:24 Unreliable Memories: Tuning Fork Watches and Hyperthymesia
13:45 Arguing is Hard: Lithium Mining and Electric Cars
15:18 Interlude: Thank You
16:07 Why AI? Rampant Self Indulgence
22:33 Immersive Theater: No More Sleep No More
23:21 Boy Smell: Bias and Bonferonni
25:51 Narwhals: Unicorns and Protogynous Pseudo-Hermaphroditism
28:25 Pig Butchering: Explore Exploit
31:00 Moose: Tycho Brahe and Proboscis
35:56 Gastromancy: Haruspex and Hecate
38:04 Conclave: Papal Testicles and Pope Joan
41:36 Lactose: Colonels and Cheddar
43:45 Gut Biomes: MINIMUM Serotonin
47:27 Long video is long: This year I need to be less wrong
source
Errata on the Errata… so far… 😭 I'm just going to pin a comment instead of another video on this video, I cannot start down such a path — there would be no ending. Only a Piranesi prison of my own design.
–My google translate attempt at determining if words had smell/smell was done by translating the phrase "She smelled the iPad and it smelled like an iPhone." No I'm not an apple fanboi, I just wanted words that would be unlikely to be translated so I could easily pick out the verbs. The errata is that this method was NOT effective based on many many comments.
–Aircraft are not asymmetrical. They handle harmonics using good structural design and vibration dampening.
–My dear sweet Ralph is not pronounced Ralph but in fact Ralph. (Rhymes with safe)
–Gunwale rhymes with funnel
–Clarification: Most neurotransmitters themselves do not cross the blood brain barrier but there is indirect influence that scales with production or lack thereof
–Papal rhymes with staple
–Tycho Brahe rhymes with Rico pa
–I think my mispronunciation of pronunciation might actually be a dialect thing, I'm in the process of asking people I grew up with to say it and I'll get back to you. Can it still be a dialect if the source is an error because of the word "pronounce" instead of "pronunce?
Regarding AI, and specifically image generation models using diffusion: they are not producing multiple possibilities via anything inherent to the model, but rather they actually run the entire model four times with different seeds to generate the different images. There's a bit of optimization that can go on when you request a batch of two or more images, but it's mostly along the lines of putting two different letters in an envelope that have to go to the same person anyways – the actual computation is still going to scale nearly linearly with the number of images requested. This is particularly obvious if you try to run the models yourself, as increasing the number of images increases computation time quite a bit. As far as LLMs go – well, with OpenAI's API you can actually SEE all of the token probabilities (unfortunately I don't think there's any easy way to do this without signing up and paying for API access). You can see someone playing around with this if you search for the article "How outdated information hides in LLM token generation probabilities" by Anj Simmons (I'd link it but usually comments with links get filtered out on YouTube), which as an aside might be interesting to anyone looking for more knowledge on how LLMs work.
Your channel is a bright new star being born amidst a sea of dull AI slop. Wishing you good health and success this year.
I don't have any cool insights or information to add but I just wanted to say that your video on AI was the first vid that YT recommended to me and I'm glad I clicked on it, I really enjoy your videos and all the research and digging you do to make them 😊
"I haven't actually seen anything by Percy Jackson"
I love your sweater! What brand is that ?
you talked a lot I heard nothing. good job 🙂
I love this video
I love the fact that this video exists
I for one would be down for more moose content.
i remember when you could say anything in the youtube comments
good times 😶
Thank you for the content
31:50 Tameness? This is earth-shattering for me considering that in Spanish "tame" means "domesticar".
About asymmetrical planes: apart from other commenter's examples, you have NASA's AD-1 project with its oblique wings.
Your Moose facts inspired a thought for me. I considered that Moose spend a large amount of time in water, evidenced by predatory Orcas and eating whilst holding their breath. Then I recalled that whales, millions of years ago, evolved from modestly sized 4 legged land animals.
This leads me to surmise that, in millions of years, Moose will evolve into aquatic beasts. Those of enormous size and enhanced swimming abilities thanks to their hydrodynamic antlers.
Wow I really thought for ages that 'cyclopes' just meant 'giant' that the meaning of single eyed came later because the cyclopes in the story happened to be monocular but no. I'm not sure where I got that misapprehension.
16:45 Dewdney, A.K. – Armchair Universe, 1988, page 102: "Although the perceptrons discussed here operate with a fixed set of weights, the notion of programming plays a central role in the theory of perceptrons developed in the 1950s." …
page 107: "Perhaps it is not surprising that perceptrons should fail in many cases where the human visual system succeeds. I noted above that the local demons and head demon could be replaced by simple computational circuits. They could also be replaced by the formal neurons first described in the 1940s by Warren S. McCullock and Walter H. Pitts in their classic work on neural networks. These formal neurons are much simpler than human neurons; likewise the complexity of a perceptron organized as a two-layer neural network does not come close to the complexity of the first two layers of the human visual cortex".
(There is also talk of Marvin L. Minsky and Seymour Papers of MIT who published Perceptrons in 1969.)
That's the closest I could find. 🤷 The commenter would have to provide actual names and dates.
38:44 Wait until you find out you're STILL saying it wrong. 😂 It's not /raff/, it's /rayff/. (Alternately, congrats on the user-engagement-bait.) … 39:21 "Papal" = /paypal/ not /pappal/. 🤨
39:50 Pendentes bene
41:27 Shoes of the Fisherman (1968) You're welcome.
49:05 DOH!
The section on tuning forks and fourier transforms is super interesting, and there's a ton of detail that could be expanded upon. A hypothetically "pure" tone would be a single sine wave, but no system is perfect and any imperfections along the way will introduce additional overtones (or other sounds) to the waveform, even for theoretically pure sources.
Overtones are also functionally how we actually differentiate instruments. A pure tone doesn't sound like anything, but the right mix of overtones can make it sound like a guitar, piano, saxophone, or trumpet.
Unintentional overtones can be very small/weak, in well-designed systems; audio engineers often talk about "total harmonic distortion" which relates to how much extra junk you get out when you feed it a given audio signal (pure tone or otherwise).
My guess is that real physical tuning forks are still slightly "imperfect" devices and especially when first struck, can have a variety of other unintended oscillations traveling through the fork (imagine, for example, a sound wave traveling up/down the length or side-to-side of the intended direction). The physical construction of the fork tends to damp these out quickly because it's not designed to efficiently sustain vibration in anything but the "preferred" direction, so you get something very close to a pure tone after a few seconds.
More generally, Fourier transforms are fascinating and form the underlying mathematics behind a whole ton of useful technologies. Converting between the time domain and frequency domain representations (effectively what a Fourier transform does) of a signal has wide applications in processing not only audio data, but radiofrequency waves, images, and video too. A huge number of compression and encoding schemes rely heavily on Fourier and inverse-Fourier transformations to make their processing steps tractable to implement in real devices.
i found your hair coming in and out of in front of your face between cuts amusing. Also great video, praise be to scientific integrity!
Nice sweater
I would personally rather errata videos instead of tacking errata onto other videos.
Thank you for making this video! Wish more creators used this platform to learn with their viewers!
Just a note Miss Modernism, it's impossible to pronounce everything "correctly" as different people pronounce things different ways. Technically a person with a "foreign" accent is pronouncing the "other" language wrong. And as every person has an accent, to anyone who doesn't share the accent, they're pronouncing things at least a little wrong. All that is to say, I wouldn't worry so much about it.
Most mammals are asymmetric on the inside.
Are you an AI or data scientist professionally or just very interested casually in the subject? Glad you decided to start a patreon, ill happily throw a few dollars your way.
This has quickly become one of my favorite channels on YouTube. Thanks for a great year of content, here’s to many more. It’s nice to see that other people are also seemingly addicted to learning about random topics.
The part on AI is still slightly wrong. You are way more correct than that commenter, don't get me wrong, but there are a few things I'd like to mention.
– LLMs also use encoders. It encodes language into vectors. So, you are more correct than you thought.
– Diffusion models give you 4 different results, but they are just 4 different seeds. There are not more, because that would be extremely computationally expensive, and saying that the results could be "ranked" implies that there exists a model better at understanding what you want than the model used to generate the image, which is somewhat silly, because then it would have been integrated into said model.
– Both LLMs and diffusion models use random noise, although the commenter explained it really weird. LLMs are essentially markov chain approximations sized at their context window, so without random noise, they will sound as close as possible to their training data. Turns out that's actually less useful than making it slightly different, so random noise is used on the next token probabilities. That's what "temperature" means in the context of LLMs, btw. Diffusion actually doesn't have this, the noise is just used for the original image of complete noise that it "denoises" into the generated image.
Explore-exploit problem I've heard phrased as a k-armed bandit problem: imagine you are in a casino, playing k one-armed bandits simultaneously. After initial assessment of expected winnings of each bandit you have a decision to make — is your choice of maximum expected payoff actually the maximum across the entire selection? So you can either continue using a suboptimal bandit and waste a tiny but consistent amount, or waste moves exploring other bandits in the hopes of finding an even better bandit
Regarding Overtones: When sounding a tuning fork, it vibrates. The fundamental is the whole lenght, the 1st overtone is half of the lenght etc… Because the partial vibrations are getting exponentially smaller, they also have less energy to sound and die off quicker. There is also the enery needed to propagete through air, which is why you can hear the low frequencies of your neighbors music better than the higher melodies.
About overtones: (Practically) every vibrating/sounding object produces overtones. Based on shape and material these are either following an exponential pattern (x2) we call "harmonic", or, in every other case "inharmonic".
Bonus tidbit about true sinewaves: A pure sinewave is very hard to achieve. Usually, a membrane is used to produce and record sound, these distort and produce harmonics of their own. Yes this includes your ears. Ears are not a very objective and precise microphone, since they introduce a lot of bias, errors and general weirdness.
We used to use piezo-electronic microphones, using the inverse of the clock mechanism, sound pressure producing electricity, but these have been largely replaced.
This one, right here, is still a horrible comment. You're welcome
Colonel Sandra reporting for duty
Respect for making a video like this. It is something more people .. and media, in general, should do. Make the correction of mistakes as loud as the wrong formation. Some of the stuff you went over was just minor, would have been totslly fine to leave it, like the picture thst was wrong, but it is great that you went over everything. The thank you in the middle made me think the video was over. I had it on on the side, I hope that was just me though
dude on a crusade about machine learning not learning apparently
Or your glasses x3
Imagine correcting mistakes in 2025 hahaha such silliness
Seriously, it's refreshing to see someone genuinely interested in the subjects they cover, cheers