flâneur

Chloe Chia

69 followers · 41 following · 6620 views

on the atlas — 93

highlights — 501

  • robotic demonstrations with actions, and "high-level" robot examples, consisting of observations labeled with the appropriate semantic behavior (e.g., an observation of an unmade bed with the label "pick up the pillow").
    A VLA with Open-World Generalization
  • This includes general multimodal tasks, such as image captioning, visual question answering, or object detection
    A VLA with Open-World Generalization
  • training our VLA model on a variety of different data sources, we can teach it not only how to physically perform diverse skills, but also how to understand the semantic context of each skill (e.g., if the task is to clean the kitchen, what are appropriate objects to pick up and put away, and where to put them),
    A VLA with Open-World Generalization
  • π0.5 that exhibits meaningful generalization to entirely new environments
    A VLA with Open-World Generalization
  • t a higher level, the robot must understand the semantics of each task—where to put clothes and shoes (ideally in the laundry hamper or closet, not on the bed), and what kind of tool is appropriate for wiping down a spill.
    A VLA with Open-World Generalization
  • Many problems are the results of trade-offs and that there is no clear Pareto improvement. Make this clear when you send the friction log. Your feedback is about the product, not them. This is common advice when giving feedback ("talk about the work, not the person") but it is worth repeating.
    sbensu: How to: friction logs
  • OpenGL is by itself a large state machine: a collection of variables that define how OpenGL should currently operate
    LearnOpenGL - OpenGL
  • Many have spent much time together working in the "war room" (a place in the Grays Basement)
    Math 55 - Wikipedia
  • We decided to only rely on the VM-based system for spotting really large regressions (e.g. a rendering algorithm accidentally turning from O(n) into O(n^2)
    Keeping Figma Fast | Figma Blog
  • : A cloud-based system would handle mass testing, covering our bases for the majority of situations, and a hardware system would be highly targeted, tackling situations that required more precision
    Keeping Figma Fast | Figma Blog
  • ounts CPU instruction
    Keeping Figma Fast | Figma Blog
  • d doing our own research
    Keeping Figma Fast | Figma Blog
  • Even if we ran all tests on virtual machines, we’d still need seemingly endless engineering resources to tame the performance variance.
    Keeping Figma Fast | Figma Blog
  • need to finish in under 10 minutes—anything beyond that would get in the way of the fast-paced development that’s so central to our engineering culture
    Keeping Figma Fast | Figma Blog
  • . This would allow us to approach performance proactively, instead of reactively
    Keeping Figma Fast | Figma Blog
  • First, we wanted a system that could test every proposed code change in our main monorepo so we could spot performance regressions across all product features early in the development cycle
    Keeping Figma Fast | Figma Blog
  • A granular performance test checks detail at scale. In one of our newer tests, we simulated rapid panning around a file with 100 multiplayer editors moving layers and typing new text simultaneously
    Keeping Figma Fast | Figma Blog
  • few large design files we used for testing could no longer represent an ever-growing number of product features, not to mention their edge cases.
    Keeping Figma Fast | Figma Blog
  • Running tests on a single laptop is common practice for smaller companies. We always try to keep our processes lean and avoid over-engineering, so this approach worked for us until recently.
    Keeping Figma Fast | Figma Blog
  • . The laptop looped the same couple of test scenarios over and over, and reported any changes and timings every hour or so to a shared dashboard
    Keeping Figma Fast | Figma Blog
  • The first system runs in GPU-enabled virtual machines, in a headless Chromium process on every code change in every pull reques
    Keeping Figma Fast | Figma Blog
  • Virtual machines (VMs) can be tricky for testing performance due to the noise levels of virtualized hardware, loud neighbors on the same underlying host, and inconsistencies in measurement, which all pose problems for accurately timing complex applications
    Keeping Figma Fast | Figma Blog
  • Government would provide the infrastructure, but beyond that the residents would be free to build what they needed on the footprint of the city that once was, neighbourhood by neighbourhood.”
    Tokyo’s incredible path to redevelopment - BBC Future
  • They suggest that “when the war ended, Tokyo’s municipal government, bankrupt and in crisis mode, was in no condition to launch a citywide reconstruction effort. So, without ever stating it explicitly, it nevertheless made one thing clear: the citizens would rebuild the city
    Tokyo’s incredible path to redevelopment - BBC Future
  • 100-metre wide boulevards with extensive promenades and linear parks would subdivide the city in clusters of 300,000 residents each.
    Tokyo’s incredible path to redevelopment - BBC Future
  • For example, 10% of the urban areas were to be allocated to parks,
    Tokyo’s incredible path to redevelopment - BBC Future
  • we don’t want stemming; we don’t want stop words to be stripped from queries; and, we want to search with regular expressions.
    The technology behind GitHub’s new code search - The GitHub Blog
  • Code is already designed to be understood by machines and we should be able to take advantage of that structure and relevance.
    The technology behind GitHub’s new code search - The GitHub Blog
  • we haven’t had a lot of luck using general text search products to power code search. The user experience is poor, indexing is slow, and it’s expensive to host
    The technology behind GitHub’s new code search - The GitHub Blog
  • The short answer is that we built our own search engine from scratch, in Rust, specifically for the domain of code search.
    The technology behind GitHub’s new code search - The GitHub Blog
  • Lisp code is made out of Lisp data objects. And not in the trivial sense that the source files contain characters, and strings are one of the data types supported by the language. Lisp code, after it's read by the parser, is made of data structures that you can traverse.
    Beating the Averages
  • Ten years ago, writing applications meant writing applications in C. But with Web-based software, especially when you have the source code of both the language and the operating system, you can use whatever language you want.
    Beating the Averages
  • It seems to me that there have been two really clean, consistent models of programming so far: the C model and the Lisp model.
    The Roots of Lisp
  • He called this language Lisp, for "List Processing," because one of his key ideas was to use a simple data structure called a list for both code and data.
    The Roots of Lisp
  • If there is more than one shortest path between a pair of nodes, each path is assigned equal weight such that the total weight of all of the paths is equal to unity
    Girvan–Newman algorithm
  • For any node � , vertex betweenness is defined as the fraction of shortest paths between pairs of nodes that run through it
    Girvan–Newman algorithm
  • progressively removing edges from the original network. The connected components of the remaining network are the communities
    Girvan–Newman algorithm
  • While the atom nodes themselves have respective feature vectors, the edges can have different edge features that encode the different types of bonds (single, double, triple).
    Math Behind Graph Neural Networks - Rishabh Anand
  • Edges can have features a i j ∈ R d ′
    Math Behind Graph Neural Networks - Rishabh Anand
  • These node features are the inputs to the GNN as we will see in the coming sections
    Math Behind Graph Neural Networks - Rishabh Anand
  • a user node has the properties
    Math Behind Graph Neural Networks - Rishabh Anand
  • Different nodes have different degrees (number of connections to other nodes) and is all over the place.
    Math Behind Graph Neural Networks - Rishabh Anand
  • An image can be considered a special graph where each pixel is a node and is connected to other pixels around it via imaginary edges
    Math Behind Graph Neural Networks - Rishabh Anand
  • That means FROs should move beyond academic proof-of-concept into standardized systems that don’t rely on graduate students to fix glitches with tweezers and tape.
    Unblock research bottlenecks with non-profit start-ups
  • Janelia separates its large-scale projects, which are professionally staffed and managed, from its academic research projects
    Unblock research bottlenecks with non-profit start-ups
  • Some organizations are set up to take on tightly specified government- or industry-sponsored applied-engineering projects. Research institutions with this purpose include the non-profit SRI International in Menlo Park, California, the Fraunhofer Society institutes in Germany and the Battelle Memorial Institute in Columbus, Ohio. The European Union has a mechanism called the European Innovation Council Accelerator that makes mid-scale investments in ‘deep tech’ start-ups. Germany’s Helmholtz Association enables large partnerships between a Helmholtz Centre and a university, bringing more-scalab…
    Unblock research bottlenecks with non-profit start-ups
  • 1925: Bell Telephone Laboratories set up in the United States for fundamental research in communications. 1933: US Rockefeller Foundation funds molecular-biology initiative. 1942: US government launches Manhattan Project to develop nuclear weapons. 1958: Advanced Research Projects Agency (ARPA) founded by US government to advance military technologies. 1970: US laboratory Xerox PARC emerges from early computer-research communities convened by ARPA. 1972: ARPA renamed Defense Advanced Research Projects Agency (DARPA). 1976: US biotechnology firm Genentech founded with venture capitalists. 1990:…
    Unblock research bottlenecks with non-profit start-ups
  • In the second half of the twentieth century, Bell Labs, Xerox PARC and other large US corporate labs famously merged aspects of fundamental research with large-scale product development and manufacturing
    Unblock research bottlenecks with non-profit start-ups
  • The Human Genome Project, for instance, cost $5.4 billion in today’s dollars and took almost 13 years.
    Unblock research bottlenecks with non-profit start-ups
  • The Human Genome Project, the Large Hadron Collider at CERN (Europe’s particle-physics laboratory near Geneva, Switzerland)
    Unblock research bottlenecks with non-profit start-ups