Dominik Kloepfer

True simultaneously: - Tariffs are stupid self-owns, like laying siege to your own country. - A major power needs to be able to run its most essential industries without a supply chain that relies on enemy powers. - Tariffs are an ineffective way to accomplish even that.

Curtis Yarvin is mistaken when he says that Apple can produce iPhones because it's a monarchy. There are millions of firms ("monarchies") in the world that can't produce anything nearly as impressive as iPhones, from the laundromat down the street to Boeing. Apple is the result of a decades long evolutionary process facilitated by the market which uplifts the very best culture, talent, processes, and ideas in the entire world. And the moment Apple slips, it'll get replaced (the average lifespan of a Fortune 500 is 15 years). Governments just don't work this way. Xi Jinping isn't competing again a million counterfactual Chinese leaders who didn't do Zero-COVID, avoided deflation, didn't kill the tech industry, and were awake to AGI. He can fuck up as much as he wants. If a monarch happens to be competent, like Lee Kuan Yew, it's merely by chance, not due to some intrinsic selection mechanism of monarchy that we can replicate. You are just as likely to get brutal dictators like Mao and Stalin by chance - this is not a reasonable gamble to take with the lives of hundreds of millions of citizens. Apple is indeed a wonderfully competent organization - if we want more of the world to be run competently, we should delegate more functions to the market, which is constantly and ruthlessly sizing down incompetence. To be clear, ton of incompetent businesses exist, but they loose access to capital, talent, and power rapidly, which is reallocated to those who can deliver. They don't drag down the fortunes of entire countries and kill millions of people, which has happened again and again in authoritarian systems.

Joao Henriques (joao.science) and I are hiring a fully funded PhD student (UK/international) for the FAIR-Oxford program. The student will spend 50% of their time @UniofOxford and 50% @AIatMeta (FAIR), while completing a DPhil (Oxford PhD). Deadline: 2nd of Dec AOE!!

@CineraVerinia It kind of sounds like von Neumann was just having fun and doing a bit to impress a woman / make her laugh / show that he has enough money to afford losing some. Asking her to buy him a drink (because he had gambled the money on him away) would fit that vibe

@robinhanson Women who are single

@robinhanson Single women?

A good heuristic about whether something can in principle be solved with deep learning: can it be done by an expert human 100% intuitively, without having to consciously think about it? If yes, then pure pattern recognition is an appropriate approach to the problem, and given enough data you should be able to produce a working DL solution. Driving -- yes, but perhaps only until the point you hit a rare edge case (which would jolt a human driver back into awareness mode). Language generation -- yes, but perhaps only at the syntax level. When you speak in a language you're fluent in, you don't think about the words to use or grammar rules, you only think about the idea you intend to convey (you might even be thinking about it in non-verbal form). Words just flow. Chess -- yes, totally. While mid-level players have to calculate a lot, a top-level player can play extremely well after only spending half a second on each move, just instantly "seeing" the right move on the board.

The curse of dl: "Maybe the idea was good, you just messed up the initialization / scheduling / nb of epochs / data set size / model dim / non-linearity"

@paulg @mattyglesias That Trump has an equal chance against Biden as against a different candidate seems surprising; I guess the markets might not be efficient enough or potential chaos might harm the non-Biden nominee

@paulg @mattyglesias Other markets had Biden being the nominee at ~25% earlier today, so having approx. no change in P(Trump wins) means that P(Trump wins | Biden) ~ P(Trump wins | not Biden) though — assuming efficient markets

Democratic vibes are excellent this evening, but on Polymarket the odds of Trump winning have only declined very slightly.

@Vinc3nt_Leroy Ah I see, so in that case MASt3R / DISt3R have learned to e.g. recognise different parts of the same bed and infer the relative camera positions using that information

@DominikKloepfer When no reliable matches can be established, we fall back to the objective function of DUSt3R. It is based on 3D regression outputs and naturally handles the impossible case.

MASt3R can still manage the impossible case, I think it will always amaze me! This was my room in Seattle, note that I had to concentrate to avoid image overlap ;)

@eigenrobot useful context to that chart

oh wonder what y/y or q/q changes in reprting were like and how that affected aggregate crime rates that showed up in Discourse vibe is crime is down from peak but im sort of skeptical of the extent now otoh one can imagine an information cascade here 🤔

New blog! Notebooks are McDonalds of Code. You can come to McDonalds and order a salad, but you won't. Same with notebooks, you can write NASA-production-grade software in a notebook, but most likely you won't. Notebooks make you lazy, and encourage bad practices. **common arguments** I first watched this talk by Joel Grus and laughed hard. I later saw the extensive use of notebooks everywhere, even in production, and I don't laugh anymore. I'm scared and sad. This section will have arguments mostly from that talk, but I might add a couple more here in the future. state State is like Jesus from the Big Lebowski, you don't fuck with Jesus. But state will for sure fuck you up. You execute a cell, you change the variable value, save the colab and forget to execute the cell. Twenty minutes later you find a bug, hello there! versioning We've seen it before. tfinal_final_final_of_final.xls is back. You want to play with a notebook, you copy it, change a couple of cells. Two weeks later, you have 25 different versions, which one do you need? Good luck with that! bad habits This is exactly the McDonalds methaphor. If people are given a chance to be lazy, they will be (I will be for sure). Quickly hacking some stuff without properly testing? Sure! Linting? Pfff, living on the edge, you'll throw the colab away in an hour anyways. Writing everything in one file? Of course! Scrolling is so fun! **notebooks slow you down** Here I try making a point why using notebooks is bad for you personally. distractions I don't have a diagnosed ADHD, but I have 90% of the symptoms from the NHS website. I'm very easy to distract. And when I use colabs, I'm just one tab away from everything else. Wikipedia? Sure, let's open five more tabs. Gmail? Let's check the inbox! Youtube music? Let's change the playlist. You got it, my tmux pane with neovim or any IDE of your choice is far less destructive. execution environment Notebooks are often used as a playground to easily have access to an accelerator. In this case, the execution environment is often set up differently and with different dependencies. In this case, you make your code work in a colab, smile widely and run your experiment after on a cluster or whatever. You start crying in twenty minutes after your code crashes due to lack of dependency or versioning or anything else similar. efficiency Notebooks keep the state, you have all the data on your fingertips, you are not encouraged to optimise your code. When you rerun scripts, you want them be damn fast, you think more about the efficiency of your code. reading code I find notebooks broken for moving around the codebase. In neovim (or Pycharm, or VSCode), you can easily go to the place where the function is defined and change it, it's just one hotkey away. You can easily look at all the places where the function is used. How do you do that in a notebook? Do you go to your IDE and search for it? What do you do if you change a function? Reload the whole thing? Autoreload can help, but now you have to remember to rerun the cells you need and potentially fuck up your state if you skip a cell. **notebooks slow your team down** Enough with personal reasons, let's think about the issues affecting the whole team. breaking changes You use some function in a colab. Another developer changes the function signature and their IDE changes all other calls of this function in the code, but not in the notebooks! If the notebook is not used often (e.g. for leaderboarding), you are in for a treat. Apart from the frustration, this is also bad from the context switching and credit assignment perspective. Who is to fix this? You, who uses the notebook? Developer who changed the function signature? Both of the cases suck. awareness You don't usually check in notebooks in your version control system (if you do, I'm sorry). They usually pile up either locally or on some cloud drive. In this case, people are unaware of what's going on. When you check in your data analysis scripts or any modules, people can glance over PRs and have an idea of what's going on. Notebooks are like dark matter of development (yes, I have almost zero knowledge of physics, and still think I can use this metaphor here). fucking around -> production Some people like notebooks as they allow them to easily check some ideas and move on. However, when their ideas work out, they are having hard time moving this code to modules. Let's think about what you need to do. First, you need to move the code to modules, sometimes it's not just a single file, it can be multiple files across the codebase. Now you need to test it somehow. Personally for me, after I've moved the code, I'm already exhausted and sometimes bored. I know my code runs in a colab, why do I need to test it again? Often I end up not unit testing my code for this reason. But even if you don't test your code, you have to make sure that it runs and produces similar results to what you've had in a notebook. This also takes time and energy. sharing is caring You have hunderds of notebooks with useful utils that are accessible only to you. If that was a library within your team codebase, everyone could use that and the whole team could avoid code duplication. But as discussed earlier, you are not encouraged to move this to a module because you are lazy. **FAQ** I'll try to answer some common answers I get when I tell people I do not use notebooks. If your question is not here, let's chat on twitter. how do you do plotting? I have small utils that unify how my plots look. In case I want incrementally play with the plot, I pickle the data for it and run the plotting script for each iteration. how do you work with a remote machine? sshfs works greatly for these purposes in case you need interactivity. I am using code autoreload and write code in modules Nice! I've done this for a while as well. This is a good use-case. However, this approach does not address some of the issues, e.g. data analysis scripts should be checked by another team member. i considered you to be my friend, how could you do this to me? Hi Lucas, I'm not judging you. We can still be friends. But we can be better friends if you stop using notebooks. some of your points are valid, but why stop using notebooks completely? I don't think I'm losing much. I'm also constantly exploring other options and having fun.

I don't know when we'll have AGI, but as an outsider, the core claim that current AI is intelligent like a high schooler strikes me as false. I don't think our evals are good. I think people use scores on tasks that if humans did well on they'd be very intelligent, to say the AI is intelligent at that level. But this is like saying my pocket calculator is intelligent because a human who did maths as fast would be a maths genius. Just sayin' 🤷‍♀️

@lwangoku99066 @gabriberton If they only share part of the computational graph, del loss1 (or I believe loss1=loss1.detach()) would free the non-shared part even with retain_graph=True which might still be worth it

This simple pytorch trick will cut in half your GPU memory use / double your batch size (for real). Instead of adding losses and then computing backward, it's better to compute the backward on each loss (which frees the computational graph). Results will be exactly identical

@eric_brachmann Pretty cool! If I understand correctly, the 3D keypoints are not actually constrained to lie on any surfaces — how well does the network nonetheless learn to reconstruct pointmaps? Does e.g. depth estimation emerge “for free”?

✨ Presenting MicKey (#CVPR2024, Oral) ✨ We regress and match 3D camera coordinates rather then 2D key points, all in metric space. Gives you a scaled relative pose between two RGB images. Paper: arxiv.org/abs/2404.06337 Page: nianticlabs.github.io/mickey

Credit to @OxfordUniBC’s Lenny Jenkins post @theboatrace indirectly shaming @thameswater on @BBCSport - E.coli cases within OUBC and himself the morning of. @thameswater - you have, you must do better. Any integrity/morals? I’m not paying until your sewage dumping is stopped.

Addendum: The world is also not street view.

The world is not a turntable.

Due to appear at @3DVconf !