GrantS

Wow, as others have said below, disabling the “Slide to Type” feature in Settings > General > Keyboard makes typing work well again on iOS. I cannot believe I put up with this awful typing experience for the past year/years. This should be broadcast more widely somehow. I’m sure many people have just assumed they got worse at typing. I am genuinely flabbergasted.

This is an amazing archive of interviews. The audio quality is astounding — you would think these were podcasts from this week. I enjoyed the 1959 discussion with Arthur C. Clarke as well: https://studsterkel.wfmt.com/programs/arthur-charles-clarke-...

You can absolutely specify your own lyrics and structure to Suno.

A few small things as I think of them:

-To practice for the moment of a real launch command, he would receive encoded messages every day that had to be manually decoded as quickly as possible — this decoding would be done independently by him and the second person on duty, and they would then compare to make sure they matched. In the case of a real launch, not only would the two people in the underground facility need to agree that the command was issued, but a second team in another facility would need to do the same.

-He was not allowed to know the targets of the missiles he would be launching, though these targets were fixed for each missile.

-It was almost assumed that if they were launching, they would have already been hit on the surface by a nuclear weapon (locations of the launch facilities were not secret, because they wouldn’t be a deterrent if they were secret). The two people underground are positioned in what looks like a shipping container suspended inside a submarine hull, all encased and locked behind one giant thick steel (?) door. If the elevator shaft had collapsed during an impact, they would be stuck inside to die. So they did include an escape hatch in the roof, but buried deep underground — this would involve the two men opening the escape hatch, letting a bunch of sand fall through, and then digging upward through 100-ish feet of ground over many days to get to a surface that was a wasteland. He was never really convinced that this would work, but the men had to believe that if they did their jobs, there would be some way to survive it.

Coincidentally, I just toured the South Dakota minuteman launch control facility this week [1] and it was fascinating. The park ranger giving the tour was a veteran who manned the facility decades ago — amazing stories. You need to book tickets a few months in advance but well worth it if you’re in the area to visit Badlands, Mt. Rushmore, etc.

[1] Run by U.S. National Park Service: https://www.nps.gov/mimi/

Since the technical nugget is hidden in the code, the fun trick here is to alternate on odd and even steps between moving toward a duck in the latent space image and moving toward a rabbit in the 90-degree rotated version of the latent space image.

(Normally you would feed the output of step n right back in as input to step n+1. That’s what is not happening as usual here.)

Possibly a reference to the 1956 Isaac Asimov short story "Someday": https://en.wikipedia.org/wiki/Someday_(short_story)

"The story concerns a pair of boys who dismantle and upgrade an old Bard, a child's computer whose sole function is to generate random fairy tales. The boys download a book about computers into the Bard's memory in an attempt to expand its vocabulary, but the Bard simply incorporates computers into its standard fairy tale repertoire..."

Eternal inflation: https://en.m.wikipedia.org/wiki/Eternal_inflation

A more reader-friendly explanation: https://medium.com/amazing-science/if-inflation-is-true-then...

Yes, the situation works. The fact that the first two logicians say “I don’t know” means they do want beer but they are unsure if everyone wants beer. The last logician knows this and can say “yes” at the end. If either of the first two logicians did not want beer they would also know that by definition it would mean “everyone wants beer” would be false and so they would have just said “no”.

I noticed this phenomenon when asking for a story involving Stephen Hawking, Darth Vader, and Wookiees. I spelled it “Wookies” by mistake and ChatGPT used my misspelling throughout the story despite seemingly getting all the other details correct.

Edit: Part of the generated story:

As they continued to talk, Stephen and Darth realized that they had something else in common - they both had lost the use of their biological legs. Stephen had been diagnosed with ALS at a young age and had gradually lost the ability to move his legs, while Darth had lost his legs in a battle with Obi-Wan Kenobi on the planet of Mustafar.

Stephen and Darth discussed the challenges and obstacles that they had faced as a result of their mobility issues, and how they had adapted and overcome them. They also talked about the technological advancements that had allowed them to continue their work and pursue their passions, despite their limitations.

Suddenly, Stephen and Darth turned on each other, each revealing that they had been secretly plotting against the other. Stephen accused Darth of using the Force for evil and corrupt purposes, while Darth accused Stephen of using his scientific knowledge to create weapons of mass destruction.

OpenAI announced and released CLIP on GitHub on January 5, 2021: https://github.com/openai/CLIP

You need CLIP to have CLIP guided diffusion. So the current situation seems to trace back to OpenAI and the MIT-licensed code they released the day DALL-E was announced. I would love to be corrected if I've misunderstood the situation.

Aha! This explains a 20-year-old mystery bug! In undergrad, one of my side projects was a procedural 3D world from scratch (kind of Myst meets Halo), including physics, graphics (OpenGL), and the raw synthesized sound of ocean waves and plasma grenade explosions from sine waves via granular synthesis. After about 10 minutes exploring the world (or running around throwing tiny spheres down hills and scattering them with plasma grenades) my relaxing synthesized ocean wave sounds started to sound garbled and distorted so I always had to restart the program — now I know why :)

Yes, this is both the most misunderstood and the most amazing (and unpredicted?) thing about GPT-3 et al. You have to tell it who it is and what the rules are. And it gains or lacks knowledge depending upon who it thinks it is answering as. You’ll find that Stephen Hawking doesn’t have much to say about the Ninja Turtles while celebrities know nothing about black holes. It does not answer questions or have expertise unless you tell it that it is answering questions and has expertise. (And it is incredible that this is the case, and somewhat understandable that people don’t understand this.)

ELI5 is: If the universe is computable, might we be in the simplest computable universe or the fastest computable universe? Can computer science and math help us figure out the reasons for the physics that run our world?

I thought the overhang was going to be along the lines of the following, whether realistic or not:

-GPT-3, as is, should be the inner loop of a continuously running process which generates 1000s+ of ideas for "how to respond next" to any query, with a separate network on top of it as the filter which cherry-picks the best responses (as humans are already doing with the examples they are posting)

-Since GPT-3, as is, can already predict both sides of a conversation, it can steer a conversation toward a goal state just like AlphaGo does by evaluating 1000s+ of potential moves, lots of potential responses and counter-responses until it finds the best thing to say in order to get you to say what it "wants" you to say.

It seems ready to go as the initial attempt at the inner loop of both of these tasks (and more) without modification or retraining of the core network itself, no?

Agreed, it takes awhile to get there but the core idea comes toward the end:

The genius — sometimes deliberate, sometimes accidental — of the enterprises now on such a steep ascent is that they have found their way through the looking-glass and emerged as something else. Their models are no longer models. The search engine is no longer a model of human knowledge, it is human knowledge. What began as a mapping of human meaning now defines human meaning, and has begun to control, rather than simply catalog or index, human thought. No one is at the controls. If enough drivers subscribe to a real-time map, traffic is controlled, with no central model except the traffic itself. The successful social network is no longer a model of the social graph, it is the social graph. This is why it is a winner-take-all game. Governments, with an allegiance to antiquated models and control systems, are being left behind.

Beautiful -- several years ago, I created an iOS app that does this live on screen in real time (via CoreImage filters, saving a video of the trails building up) but then never released it. I should probably just put it out there -- it was all ready to go except for a name if I recall, made a video set to music and everything. Will see if I can dig up the video...

Edit: 45 second video here for anyone interested: https://youtu.be/df_Pr4jAu78

100% correct. There is no way they included a depth sensor in this device just to blur the background. They may also be aiming to map any Amazon clothing purchases to 3D body shapes for a purely data driven approach to clothing recommendation to start with (as opposed to having 3D scans of every bit of clothing to do geometric fitting, which is prohibitively labor intensive.) So they'll know your shape, what exact items you've purchased, how often you wear those purchases, and (if I understood correctly) how your friends think you look in them.

Does anyone know if the languages all live in the same 300-dimensional space, or are they each trained independently? (i.e. do words and their translations have similar vectors?)

Here is some of that IIHS research [1] from Jan 2016 that gives lots of raw crash numbers broken down by manufacturer, type of system (AEB/FCW), injuries involved, etc. Really informative.

Some excerpts from "Effectiveness of Forward Collision Warning Systems with and without Autonomous Emergency Braking in Reducing Police-Reported Crash Rates", January 2016:

"FCW alone and FCW with AEB reduced rear-end striking crash involvement rates by 23% and 39%, respectively. "

"Among the 15,802 injury crash involvements in these states, the percentage of injury crash involvements that were rear-end striking crashes was larger among vehicles without front crash prevention (15%) than among vehicles with FCW alone (12%) or FCW with AEB (9%). Only 4% of rear- end injury crashes involved fatalities or serious (A-level) injuries."

"Approximately 700,000 U.S. police-reported rear-end crashes in 2013 and 300,000 injuries in such crashes could have been prevented if all vehicles were equipped with FCW with AEB that performs similarly as it did for study vehicles."

[1] http://orfe.princeton.edu/~alaink/SmartDrivingCars/Papers/II...

Since it's not mentioned in the article, this is also called bilateral filtering [1]. The external links section of the Wikipedia article lists some famous SIGGRAPH papers describing fast implementations, and there is some code available [2].

[1] https://en.wikipedia.org/wiki/Bilateral_filter

[2] http://people.csail.mit.edu/sparis/bf/

While you would never do this kind of training on your user's machines (which takes multiple weeks even with a powerful GPU), you should be able to apply the trained model to a single photo nearly instantaneously. So the real roadblock is mostly that they don't appear to have a included a completely pre-trained model with this release, and it will take you as a developer a lot of GPU time to train one. But your users would not necessarily have a problem captioning images on their machines.

Impressive work, Daniel! Do I understand correctly that any image prediction for which the deltas are smaller in absolute value than the full JPEG/DCT coefficients would offer continued compression benefits? As in, if you could "name that tune" to predict the rest of the entire image from the first few pixels, the rest of the image would be stored for close to free (and if not, it essentially falls back to regular JPEG encoding).

If that's the case, then not only could we rely on the results of everything we've decompressed so far to use for prediction (which is like one-sided image in-painting), but we also could store a few bits of semantic information (e.g. from an image-net-based CNN, from face detection) about the content of the original image before re-compression, and use that semantic information for prediction as well via some generative model. All of this would obviously be trading computation for storage/bandwidth, but it this seems like an exciting direction to me. Again, nice work.

Yes, the results of all the learning can be saved in a compact form which is efficiently run on the device. Apple just yesterday unveiled their neural network API [1] which they are presumably using for these tasks. The description states: "BNNS supports implementation and operation of neural networks for inference, using input data previously derived from training. BNNS does not do training, however. Its purpose is to provide very high performance inference on already trained neural networks."

So they're really saying that no data leaves the device for either object or face recognition at classification time. Now, whether they are using iCloud-stored photos (which the user has already agreed to share) for training, I don't know, and any privacy issues would still be important at that point.

[1] https://developer.apple.com/reference/accelerate/1912851-bnn...

Images don't necessarily map directly onto a word2vec-like solution but the closest thing is pre-trained deep networks, for example: http://caffe.berkeleyvision.org/model_zoo.html

You're right that bag of words with SIFT is not state of the art, with deep learning dominating computer vision approaches these days.

For anyone interested in the computer vision side of this topic, the author here is using a variant of color histograms, which was state of the art around 1990 [1][2]. Since 2003, bag of visual words approaches have usually meant extracting SIFT-like features from a database of images, quantizing the features down to a list of thousands or millions of "words", and then treating the images like documents containing those "visual words" [3][4]. (Nothing wrong with the approach he's using [simple and fast], but the bag of words terminology in the article usually suggests a different class of approaches.)

[1] https://staff.fnwi.uva.nl/r.vandenboomgaard/IPCV/_downloads/...

[2] https://www.cs.utexas.edu/users/dana/Swain1.pdf

[3] http://www.robots.ox.ac.uk/~vgg/publications/papers/sivic03....

[4] http://www-inst.eecs.berkeley.edu/~cs294-6/fa06/papers/niste...

To clarify, there's no "my simulation" and "your simulation" in the present discussion. What's potentially being simulated is the entire universe, with no privileged viewpoint. That the universe is being simulated simply means that the laws of physics of "the universe" have been implemented on a computing device that exists outside of "the universe" (and which therefore does not itself necessarily obey those same laws of physics).

He's using the quickness with which we've progressed to VR as an existence proof that creating universes is not as hard as once expected, such that "universes" might actually be numerous and commonplace. From there, it's a statistical argument -- if "existence" contains one billion "universes" and only one of them is the base level of reality in which all the other simulations are nested, then if we assume a uniform distribution over universes, we have a one in a billion chance of being in the base level of reality.

That's a very simple version of the argument with obvious flaws, but that's what he meant. Nick Bostrom has made a much more elaborate version of the argument.

Very good to be reminded of this. The funny thing is that there was a brief time right before deep learning took over computer vision research, but after people realized the implications of big data, where it looked like the future could be really simple nearest neighbor search supported by enormous data sets. For example, the famous "80 million tiny images" paper by Torralba, Fergus, and Freeman from 2008.

PDF:

http://people.csail.mit.edu/torralba/publications/80millionI...

Web:

http://groups.csail.mit.edu/vision/TinyImages/

Yes, all of them. Oculus Rift, HTC Vive, Google Cardboard, etc. They show a different view to each eye.