Google Research Blog

From the text: > I can't name a single thing from Google X or Amazon 126, Apple or Microsoft Research departments. This is a very narrow way of looking at who's making things of value. Let's take a single example out of those companies mentioned: Apple's iPhone and the demand it generated for smartphones, pushed many industries to invest in research, development and mass production of high resolution, low cost, low power screens, low power (and now even faster than desktop workstations) compute in small form factor SoC, ever so small low noise and high resolution camera sensors and lenses, faster, higher bandwidth, cheaper and more widely accessible cell networks, and so on and on and on... Take any single one of the above components/infrastructure and one can list 100 products of value stemming from its cheap and widespread availability. Apple didn't just make a phone, it changed computing forever (something Jobs would probably say on stage) in span of a few years. Also, did you forget https://research.google/ or just ignored it? There are mind numbing number of research papers *with* corresponding global scale real products used by millions/billions of people, that based on those research. I am sick of people glorifying the past when today, right under their noses, exists orders of magnitude greater research and innovation. Their title may not be sweet and straightforward like "invention of the telephone" but that's how progress works — the more you move forward, the…

Here are other tips for video: - https://twitter.com/jugurthahadjar/status/1276670098602954754 - https://twitter.com/jugurthahadjar/status/1276671264116158464 Reproduced here: >If a talk doesn't provide a link to slides but shows them on the screen, I watch it muted at 16x with "Video Speed Controller". 50 minutes become ~3 minutes and it's easier to decide if the talk is worth the time and/or has high information density in the topic of interest. >If there are no slides linked nor shown, I use "Youtube Captions Search" to search for expressions of interest in the video, and go straight to where they're mentioned. If I like the treatment of the subject, I make a note to invest in that video and watch it. There are many articles when you search for "google key moments algorithm" which relay a "recent Google announcement" without a source, or a link. However, here's a link to a thread: https://support.google.com/youtube/thread/18138167/youtube-test-featur… >[November 23 2020] Testing automatic video chapters: We want to make it easier for people to navigate videos with video chapters, so we are experimenting with automatically adding video chapters (so creators don't have to manually add timestamps). We’ll use machine learning to recognize text in order to auto generate video chapters. We’re testing this out with a small group of videos. If your videos are impacted, you can opt out of this experiment and/or share feedback here. More on video chapters…

The related paper (2019): https://google.github.io/night-sight/ Google has some great blogs about their camera technology https://research.google/blog/night-sight-seeing-in-the-dark-on-pixel-p…

> modern, rootable ARM64 phone which one would people suggest? I got two already, both locked. Google Pixel 3a XL 64GB. Cannot root it, Sinlock. Not possible to install postmarktOS. Xiaomi Redmi Note 8T 64GB. Same thing, some lock, need a SIM card, some proprietary BS.. not possible to root easily.. I really wanted to build a cluster of mobiles, but its hard to find cheap and good mobiles for such experiment... and if i have to pay 500 buckets, i can do better with some Mini PC... My goal is to reproduce the following experiment https://research.google/blog/a-low-carbon-computing-platform-from-your…

Google's TabFM [1] (and its previous TimesFM) seem the best approach so far on this area 1. https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-…

It comes from people who develop LLMs. Anthropic and Google. References below. My other favorite quote: transformers are GNNs which won the hardware lottery. Longer form at blog.ladybugmem.ai You want to believe that everything probabilistic has more value and determinism doesn't? Or that the world is made up of tabular data? You have a lot of company. The other side of the argument I believe has a lot of money. https://www.anthropic.com/research/mapping-mind-language-model https://research.google/blog/patchscopes-a-unifying-framework-for-insp…

> great if there was some standard that allowed these to be easily found Too bad Google sunset Chrome Flywheel (likely after AMP?): https://research.google/pubs/flywheel-googles-data-compression-proxy-f… Opera Mini Turbo was equally popular during 2G/Edge era.

At Google I worked with one statistics aggregation binary[0] that was ~25GB stripped. The distributed build system wouldn't even build the debug version because it exceeded the maximum configured size for any object file. I never asked if anyone had tried factoring it into separate pipelines but my intuition is that the extra processing overhead wouldn't have been worth splitting the business logic that way; once the exact set of necessary input logs are in memory you might as well do everything you need to them given the dramatically larger ratio of data size to code size. [0] https://research.google/pubs/ubiq-a-scalable-and-fault-tolerant-log-pr…

It's interesting that they publish a blog post about the Titans and MIRAS papers only now, while the blog post about the new follow-up paper (Nested Learning), all by the same main author(!), came out a month ago: https://research.google/blog/introducing-nested-learning-a-new-ml-para…

While some of the features the author references are really interesting, personally I don't see how any of that would justify creating a new memory unsafe language in 2016. I thought it was pretty obvious by now [1][2][3] memory safety is best left to tooling / compilers and not to programmers. [1] https://research.google/pubs/secure-by-design-googles-perspective-on-m… [2] https://www.microsoft.com/en-us/msrc/blog/2019/07/we-need-a-safer-syst… [3] https://www.cisa.gov/case-memory-safe-roadmaps

When I think about ads, I think about Google's paper [0], which allows multiple self-interested LLMs bid for possible ad-spots in the prompt, which is just a different flavor of what we have today. As noted on the thread, a good ML model can do better, with less. I can use that kind of recommendation system for discovering my alternatives for a given item/tech/whatever, on demand. But current crop of LLMs, with a strong leaning for embedding seamless ads into prompt responses, a double no. No, because I don't agree on how training data is stol ^H^H^H^H collected from wider internet and everything, another no for unsolicited ads. I'll continue using Kagi and being a luddite, thanks. [0]: https://research.google/blog/mechanism-design-for-large-language-model…

Hah, what a coincidence. I did small research today for heatmaps for my very simply graph lib for Ruby. I recommend this blog post: https://research.google/blog/turbo-an-improved-rainbow-colormap-for-vi… I myself implemented 4 colorize functions for that. Common linear Blue -> Red. Linear Blue -> Red adjusted via Green for visibility. Blue -> Yellow -> Red also linear. And of course Turbo from that blog..

> I'd still love to see a report that accurately captures training cost. Today's report[1] notably excludes training cost. From 2022, so possibly out of date: "ML training and inference are only 10%–15% of Google’s total energy use for each of the last three years, each year split ⅗ for inference and ⅖ for training." That's probably close enough to estimate 50/50, or the full energy cost to deliver an AI result is double the inference energy. https://research.google/blog/good-news-about-the-carbon-footprint-of-m…

Yes, it seems easy to implement. Yes, students should absolutely implement the classic algorithms to learn. Yes, there are some occasions when you need to home grow one at $work. BUT, in my opinion, most of the time, professional code should use a battle tested, vuln hardened library or builtin version. These things are VERY HARD to get exactly right. Jon Bently's Programming Pearls famously had a latent bug in its binary search for 20 years before someone caught it. https://research.google/blog/extra-extra-read-all-about-it-nearly-all-… So yeah, it looks easy but don't do it. Stand on some giant's shoulders instead.

Repo: https://github.com/yandex/perforator We use this for GWP-style[0] continous-profiling over O(10k) nodes. It does sampling through an eBPF program in the kernel, and supports generating a bunch of different output formats (classical flamegraphs, but also things like PGO profiles). One example where we use it daily is flamegraph diffs for prestable releases of services. It's also worth mentioning that while the Github repo has an example for a Kubernetes setup, this isn't a Kubernetes-specific tool. We don't actually run Kubernetes ourselves. It works wherever (well, if you have Linux, that is!) (bias note: I work at Yandex, but not in the Perforator team) [0]: https://research.google/pubs/google-wide-profiling-a-continuous-profil…

Don't want to get to pedantic on that response. The model can contain complex information. There is already evidence it can form a model of the world. So why not something like steps to get from A to B. And, it is clear that LLMs can follow steps. One didn't place in the Math Olympiad without some ability to follow steps. https://research.google/blog/teaching-language-models-to-reason-algori… And, Anyway, when I asked it, it said it could "Yes, an LLM model can contain the steps of an algorithm, especially when prompted to "think step-by-step" or use a "chain-of-thought" approach, which allows it to break down a complex problem into smaller, more manageable steps and generate a solution by outlining each stage of the process in a logical sequence; essentially mimicking how a human would approach an algorithm. "

You're saying Google's long work in AI, Quantum Computing, etc... are not trying to be innovative? Their researchers certainly put out a lot of influential papers. https://research.google/

Encoding speed is not really a concern for web uses, where the image is decoded potentially millions or even billions of times more than it is encoded. I agree, PSNR is a terrible measure of quality. The study "Benchmarking JPEG XL lossy/lossless image compression" (https://research.google/pubs/benchmrking-jpeg-xl-lossylossless-image-c… which you are an author on included a controlled subjective evaluation done by EPFL using an ITU recommend methodology. The subjective results concluded: "HEVC with SCC generally results in better rate/distortion performance when compared to JPEG XL at low bitrates (< 0.75), better preserving sharp lines and smooth areas" It is known that AVIF performs better than HEVC. Can you say why it was not included in your 2020 subjective evaluation? It would be nice to not need to speculate on the relative quality of AVIF v JPEG XL at web bitrates.

Their defensiveness makes sense in the current context where in the last few weeks organisations like the White House [1] and Google [2] are explicitly calling out the importance and imminent need of moving away from memory unsafe languages like C and C++. If everyone focussed on this one issue, it is possible that we might actually start moving away from C and C++ in the next 5-10 years. Sutter pointing out that memory safety isn't the only vector for system vulnerabilities would have the effect of spreading cybersecurity efforts and budgets across all of them. In that case memory safety isn't the foremost problem it's being portrayed as, and it isn't worth migrating away from C and C++. [1] - https://www.whitehouse.gov/wp-content/uploads/2024/02/Final-ONCD-Techn… [2] - https://research.google/pubs/secure-by-design-googles-perspective-on-m…

OPA is a great tool for implementing a policy-as-code system. But if you're trying to use it for application authorization (e.g. fine-grained authz for B2B SaaS or a set of internal applications), you may find that its policy story is strong, but it doesn't really have a "data plane": you either store data in a data.json file and rebuild the policy any time that data changes, or make an http.send call out of the policy to fetch dynamic data. Check out Topaz [0], which uses OPA as its decision engine, but adds a data plane that is based on the ReBAC ideas explored in the Google Zanzibar [1] paper. Disclaimer: I work on the team [2] that builds and maintains the Topaz project. [0] https://www.topaz.sh [1] https://research.google/pubs/zanzibar-googles-consistent-global-author… [2] https://www.aserto.com

Both Google and Apple have implemented models that aim to take the "best" picture from a video sequence, like a live photo. Have a look at MUSIQ and NIMA. https://github.com/google-research/google-research/tree/master/musiq https://blog.research.google/2022/10/musiq-assessing-image-aesthetic-a…

I feel it was unnecesary to create this because https://research.google/ and https://ai.google/ already exist? It just seems like they want to take another URL with a "pure" domain name instead of psubdirectories, etc parts. Most probably is pure SEO ;-)

Hey, feel free to draw your own conclusions. AI quality is technically a subjective topic. For what it's worth though, Google's open models have benched quite competitively with GPT-3 for multiple years now: https://blog.research.google/2021/10/introducing-flan-more-generalizab… The flan quantizations are also still pretty close to SOTA for text transformers. Their quality-to-size ratio is much better than a 7b Llama finetune, and it appears to be what Apple based their new autocorrect off of.

It's a great question. But note that Google Translate is also trained on "predict the missing token": https://blog.research.google/2022/05/24-new-languages-google-translate… / https://arxiv.org/abs/2205.03983 (search the blog post around “Surprisingly, this simple procedure produces high quality zero-shot translations.”) This was in May 2022, as part of Google Translate adding support for several low-resource languages (including Sanskrit). I was already very surprised that simply training on predicting tokens does translation so well — then a few months later ChatGPT came out, trained (roughly) the same way and doing a lot of things besides translation.

This is like the classic advice on optimization. The First Rule of Program Optimization: Don't do it. The Second Rule of Program Optimization (for experts only!): Don't do it yet. It is because I know what a database actually does for me, and because I've had to write parts of that for myself, that I deeply understand why I don't want to write one. And why I want to push those problems to a database if I can. Of course you should ignore my advice if you're named Jeff Dean. Look through https://research.google/people/jeff/. Among his other projects, look at BigTable, LevelDB and Spanner. All are databases, implemented for good reasons.

I'll recommend the Spotlight paper by Google[1]. There are very interesting datasets they created for this purpose. They mention they have a screen-action-screen dataset that is in-house and it doesn't look like they'll open it. Maybe owning Android has its advantages. There's a recent paper by Huggingface called IDEFICS[2] that claims to be an open source implementation of Flamingo(an older paper about few-shot multi-modal task understanding) and I think this space will be heating up soon. [1] https://research.google/pubs/pub52171/ [2] https://huggingface.co/blog/idefics

I haven't heard the term before either, but it's mentioned at least by Google too so not entirely unheard of in the multimodal training context. Even so I also think this could have been conveyed without using that term. Two towers and even twin towers as seen in some places sounds exceedingly marketingy. But as others have pointed out, this may have been a popular term in recommendation systems before the rise of deep learning multimodality. https://blog.research.google/2022/06/limoe-learning-multiple-modalitie…

> Jeff Dean is clearly one of the greatest software developers/engineers ever Based on what? I've heard all the Chuck Norris type jokes, but what has Jeff Dean actually accomplished that is so legendary as a software developer (or as a leader) ? Per his Google bio/CV his main claims to fame seem to have been work on large scale infrastructure projects such as BigTable, MapReduce, Protobuf and TensorFlow, which seem more like solid engineering accomplishments rather than the stuff of legend. https://research.google/people/jeff/ Seems like he's perhaps being rewarded with the title of "Chief Scientist" rather than necessarily suited to it, but I guess that depends on what Sundar is expecting out of him.

This reminds me about Google's paper titled "Machine Learning: The High Interest Credit Card of Technical Debt" https://research.google/pubs/pub43146/ Abstract: Machine learning offers a fantastically powerful toolkit for building complex systems quickly. This paper argues that it is dangerous to think of these quick wins as coming for free. Using the framework of technical debt, we note that it is remarkably easy to incur massive ongoing maintenance costs at the system level when applying machine learning. The goal of this paper is highlight several machine learning specific risk factors and design patterns to be avoided or refactored where possible. These include boundary erosion, entanglement, hidden feedback loops, undeclared consumers, data dependencies, changes in the external world, and a variety of system-level anti-patterns.

I don't think most classical machine learning is appreciably less tangled. It's more that the more obviously black box nature of deep learning models leaves people feeling more exposed to something that's been there all along. https://research.google/pubs/pub43146/ should be required reading for anyone involved in designing or building machine learning components that are intended to go into production.