Saeejith Nair
2 followers · 4 following · 344 views
on the atlas — 50
- The 37 Implementation Details of Proximal Policy Optimization · The ICLR Blog Track9 savers
- Understanding Attention in LLMs | Bartosz Milewski's Programming Cafe2 savers
- Advice on Upskilling - Justin Skycak24 savers
- Tutorial: Train your own Reasoning model with GRPO | Unsloth Documentation1 savers
- Why Momentum Really Works7 savers
- What can we send to Mars on the first Starships? – Casey Handmer's blog1 savers
- MAKING SOFTWARE14 savers
- Helix: A Vision-Language-Action Model for Generalist Humanoid Control4 savers
- Unpacking the complexities of consciousness: Theories and reflections - ScienceDirect1 savers
- Microsoft’s Majorana 1 chip carves new path for quantum computing - Source2 savers
- Blog - Being a High-Leverage Generalist - char.blog5 savers
- UV4L - Raspberry Valley1 savers
- We Live Like Royalty and Don’t Know It — The New Atlantis8 savers
- New President of Syria: Ahmed al-Sharaa - YouTube2 savers
- A Dent in the Universe2 savers
- Nonduality - Defining the Undefinable - Deconstructing Yourself2 savers
- Discovering Bliss States - Tasshin2 savers
- [2304.06718] Segment Everything Everywhere All at Once2 savers
- NeRF From Nothing: A Tutorial with PyTorch | Towards Data Science2 savers
- MAKE: The Indie Maker Handbook5 savers
- An Intuition for Attention | Jay Mody2 savers
- GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE1 savers
- examples/examples/benchmarks/resnet_imagenet/README.md at a61497fde56ba9dac3212fd5c5310f79f1c36a55 · mosaicml/examples1 savers
- What AI can do with a toolbox... Getting started with Code Interpreter1 savers
- From the late 2000s to the mid-2010s, I worked as a molecular biologist for a national security contractor in a program to study Exo-Biospheric-Organisms (EBO). I will share with you a lot of information on this subject. Feel free to ask questions or ask for clarification : aliens1 savers
- How to Do Great Work80 savers
- Accelerated Inference for Large Transformer Models Using NVIDIA Triton Inference Server | NVIDIA Technical Blog2 savers
- LLM Powered Autonomous Agents | Lil'Log14 savers
- Modernisation and China’s ‘century of humiliation’ | CEPR1 savers
- pytorch-image-models/timm/models/vision_transformer.py at main · huggingface/pytorch-image-models · GitHub1 savers
- 🥽 Analyzing Traces - Composer1 savers
- 👨👩👧👦 Distributed Training - Composer1 savers
- ⏱️ Performance Profiling - Composer1 savers
- ⏯️ Autoresume Training - Composer1 savers
- ♻️ Auto Microbatching - Composer1 savers
- ⚡ Migrating from PTL - Composer1 savers
- Efficiently Estimating Pareto Frontiers with Cyclic Learning Rate Schedules1 savers
- Mosaic ResNet Deep Dive1 savers
- What Is ChatGPT Doing … and Why Does It Work?—Stephen Wolfram Writings19 savers
- A Comprehensive Mechanistic Interpretability Explainer & Glossary - Dynalist17 savers
- Yes you should understand backprop | by Andrej Karpathy | Medium14 savers
- The maze is in the mouse. What ails Google. And how it can turn… | by Praveen Seshadri | Feb, 2023 | Medium14 savers
- Transformers from Scratch14 savers
- Blueprint for an AI Bill of Rights - The White House10 savers
- A Recipe for Training Neural Networks7 savers
- ChatGPT Is Nothing Like a Human, Says Linguist Emily Bender7 savers
- A successful Git branching model » nvie.com5 savers
- GPT in 60 Lines of NumPy | Jay Mody5 savers
- A critique of pure learning and what artificial neural networks can learn from animal brains3 savers
- [2304.06636] Neural networks: from the perceptron to deep nets2 savers
highlights — 364
Hook this to a 2.5 m aperture camera mounted within a Starship that aerobrakes into orbit and we can get precise surface mineralogy at 6 cm resolution
What can we send to Mars on the first Starships? – Casey Handmer's blogWe know there is water, but the error bars around its composition are large
What can we send to Mars on the first Starships? – Casey Handmer's blogCO2 ingestion is easy enough, it was demonstrated by the MOXIE instrument
What can we send to Mars on the first Starships? – Casey Handmer's blogMars power will most likely be provided by large solar arrays on the surface. This is a challenge in terms of mass, since providing enough solar arrays will require multiple Starships of cargo, but Starship’s entire purpose is to schlep mass from Earth to Mars so I think this will be doable
What can we send to Mars on the first Starships? – Casey Handmer's blogMicrosoft’s topoconductor is made of indium arsenide
Microsoft’s Majorana 1 chip carves new path for quantum computing - Sourcethe difference between one billion and one billion and one electrons in a superconducting wire
Microsoft’s Majorana 1 chip carves new path for quantum computing - SourceMajorana particles, which help protect quantum information from random disturbance, but can also reliably measure that information from them using microwaves
Microsoft’s Majorana 1 chip carves new path for quantum computing - SourceThe delay was easily in seconds, which is not good for video robot control. Enter UV4L from Linux Projects. Low latency and good quality was achieved using this library
UV4L - Raspberry ValleyThe young people at my table had debts, but they were the debts of kings.
We Live Like Royalty and Don’t Know It — The New AtlantisBut despite his wealth and status his home was so cold in winter that the ink in his pen sometimes froze, making it difficult for him to write to complain about the chill.
We Live Like Royalty and Don’t Know It — The New AtlantisThey wanted a better world, but they didn’t know how this one worked.
We Live Like Royalty and Don’t Know It — The New AtlantisEBOs believe that the soul is not an extension of the individual, but rather a fundamental characteristic of nature that expresses itself as a field, not unlike gravity. In the presence of life, this field acquires complexity, resulting in negative entropy if that makes sense. This gain in complexity is directly correlated with the concentration of living organisms in a given location. With time, and with the right conditions, life in turn becomes more complex until the appearance of sentient life. After reaching this threshold, the field begins to express itself through these sentient beings,…
From the late 2000s to the mid-2010s, I worked as a molecular biologist for a national security contractor in a program to study Exo-Biospheric-Organisms (EBO). I will share with you a lot of information on this subject. Feel free to ask questions or ask for clarification : aliensDon't marry someone who doesn't understand that you need to work, or sees your work as competition for your attention. If you're ambitious, you need to work; it's almost like a medical condition; so someone who won't let you work either doesn't understand you, or does and doesn't care.
How to Do Great Workwhen you do, you know. Which means if you're unsure, you probably don't. But it may be possible to give a more concrete answer than that. Here's an attempt: sufficiently good colleagues offer surprising insights. They can see and do things that you can't. So if you have a handful of colleagues good enough to keep you on your toes in this sense, you're probably over the threshold.
How to Do Great WorkBut any energy that goes into how you seem comes out of being good. That's one reason nerds have an advantage in doing great work: they expend little effort on seeming anything
How to Do Great WorkAnd that is what you're aiming for, because if you don't try to be the best, you won't even be good.
How to Do Great WorkPer-project procrastination is far the more dangerous. You put off starting that ambitious project from year to year because the time isn't quite right. When you're procrastinating in units of years, you can get a lot not done.
How to Do Great WorkFinishing things is not just an exercise in tidiness or self-discipline. In many projects a lot of the best work happens in what was meant to be the final stage.
How to Do Great WorkThese so-called treaty ports were notable because it was from these claimed areas that Western countries abolished Chinese practices in trade and reorganised them under their management
Modernisation and China’s ‘century of humiliation’ | CEPRGenerally, it is held that although Western colonialism was associated with both destructive and modernising influences in China, on the whole Western impact was limited one way or another. The main reason for this, it is argued, is that Western influence was limited to some dozens of ports that were scattered over a very large country
Modernisation and China’s ‘century of humiliation’ | CEPRgradually reduced by foreign powers to a semi-colonial, semi-feudal society that suffered greater ravages than ever before”, bringing “intense humiliation for the country” and “great pain for its people
Modernisation and China’s ‘century of humiliation’ | CEPRrun_name=run_name, save_folder='./my_autoresume_training_run',
⏯️ Autoresume Training - Composerhow to use automatic microbatching to avoid CUDA OOMs, regardless of your batch size choice, GPU type, and number of devices
♻️ Auto Microbatching - ComposerOptimizers and schedulers are passed directly to the Trainer during initialization
⚡ Migrating from PTL - ComposerThe training_step is broken into two parts, the forward and the loss methods. This is needed since some algorithms (such as label smoothing or selective backprop) need to intercept and modify the loss
⚡ Migrating from PTL - Composersome of the early motivation for cyclic learning rates was to be able to interrupt training at any time and still return good performance (a.k.a. an “anytime” algorithm), anytime performance is not necessarily of interest for heavily studied benchmarks with known, standard hyperparameters. However, we suspected that cyclic learning rate schedules could be used to efficiently construct accuracy-time tradeoff curves.
Efficiently Estimating Pareto Frontiers with Cyclic Learning Rate SchedulesWe recommend that accuracy-time tradeoff curves should be standard practice for ML research (in addition to accuracy tables!). With cyclic learning rate schedules, generating tradeoff curves is not computationally expensive.
Efficiently Estimating Pareto Frontiers with Cyclic Learning Rate SchedulesThe default reduction value (”mean”) averages the loss over all n samples in a batch and 1000 classes in ImageNet, resulting in infinitesimal loss values that don’t allow learning with learning rate values in the traditional range. Our solution was to set reduction=”sum” and then divide the loss by the number of samples in the batch
Mosaic ResNet Deep Divebased on hyperparameters from NVIDIA Deep Learning Examples
Mosaic ResNet Deep DiveNaively, one would expect the throughput to quadruple to ~65,000 images/sec when the input image size is at 112x112 pixels
Composer + FFCV: Faster TogetherProgressive Image Resizing, which start training with small image sizes, then increases the height/width throughout training, don’t actually yield a significant speedup
Composer + FFCV: Faster TogetherOn some models, we’ve reached the point where the GPU portion has been sped up so much that we are now bottlenecked on CPU decoding and augmentations of the dataloader
Composer + FFCV: Faster TogetherThe most offensive thing about these falsehoods is not the fact that they later turned out to be supported by evidence, which can happen to even the most unlikely seeming hypothesis. Rather, it is that the people who labeled them “false” often knew full well from the beginning that they were true, and were seeking to avoid the consequences; that is how a truth becomes a “conspiracy theory.”
The RFK Jr. Tapes - Tablet Magazinetoday’s conspiracy theories include whatever evidence-based realities threaten America’s flourishing networks of administrative state bureaucrats, credentialed propagandists, oligarchs, and spies
The RFK Jr. Tapes - Tablet Magazineconspiracy theory” has become a flashing red light that is used to identify and suppress truths that powerful people find inconvenient
The RFK Jr. Tapes - Tablet MagazineWhenever a hole appears in the ozone layer of received opinion, it is sure to be quickly labeled a “conspiracy theory” by a large technology platform
The RFK Jr. Tapes - Tablet MagazineThe collision he’s about to cause between the world of official group-think and the world of normal-speak—where most Americans weigh what might be best for themselves and their children—can only be good for American democracy, and for the American language
The RFK Jr. Tapes - Tablet MagazineKennedy founded the World Mercury Project, which would be renamed the Children’s Health Defense, to keep pressing his assertions about empirical links between vaccinations and the explosion of neurological issues in children
The RFK Jr. Tapes - Tablet Magazinejournalistic-bureaucratic firestorm proved that vaccines were different
The RFK Jr. Tapes - Tablet Magazinethe case Kennedy made in his article was no more or less plausible and empirically grounded than the cases that he and dozens of other environmental advocates had been making for decades against large chemical companies for spewing toxins into America’s air, water, and soil, and then lying about it
The RFK Jr. Tapes - Tablet MagazineIn 2005, Kennedy wrote a blockbuster Rolling Stone magazine article titled “Deadly Immunity,” which presented compelling evidence of an ongoing vaccine safety cover-up led by U.S. national health bureaucrats, including transcripts of a 2000 CDC conference in Norcross, Georgia, where researchers presented information linking the mercury compound thimerosol with neurological problems in children
The RFK Jr. Tapes - Tablet MagazineObama’s political chi came from his cool mastery of the politics of race and empire and his disdain for the American working poor, whom he redefined as white bitter-enders who clung to religion and guns
The RFK Jr. Tapes - Tablet Magazinewhile selling out the American working class to the globalist gods
The RFK Jr. Tapes - Tablet MagazineYellows are default highlights, Cyans are things that the author says that I think are important for things I’m working on, reds are things that I disagree with the author of the paper on, greens are references I want to follow up and read, Magentas are places where the author critiques someone else
Zotero -> zotfile -> mdnotes -> obsidian -> dataview Workflow - Share & showcase - Obsidian Forumdecrypting a message encoded with a block cipher is equivalent to inferring the circuit complexity of a quantum state
In New Paradox, Black Holes Appear to Evade Heat Death | Quanta Magazinewe can think of any collision as a gate
In New Paradox, Black Holes Appear to Evade Heat Death | Quanta MagazineQuarto has native support for embedding Mermaid and Graphviz diagrams. This enables you to create flowcharts, sequence diagrams, state diagrams, gantt charts, and more using a plain text syntax inspired by markdown.
Quarto - Diagramsview dependent color that changes depending on the angle the point is viewed
NeRF - nerfstudioeach point has a density which describes how transparent or opaque a point in space is
NeRF - nerfstudioFor each point in space the NeRF represents a view dependent radiance
NeRF - nerfstudio