Archive
Selected essays on GPU computing, mobile deep learning, MapReduce, search engines, and software engineering recovered from memkite.com, medium.com, amundtveit.com, atbrox.com, amundblog.blogspot.com, and LinkedIn.
2025
2017
Thoughts on AI Replacing Coders by 2040
Five observations on automating software engineering — from stagnant IDEs and GitHub test corpora to learned index structures, program synthesis, and formal verification.
A Closer Look at Startup Equity Crowdfunding
An investor and founder perspective on startup equity crowdfunding — comparing platforms like Crowdcube, Seedrs, and Wefunder, return distributions, and due diligence.
Keras Deep Learning with Apple's CoreMLTools on iOS 11
Hands-on walkthrough converting a Keras neural network to CoreML and running on-device inference on iOS 11.
2016
Deep Learning with Generative Adversarial Networks — ICLR 2017 Discoveries
Survey of generative modeling and GAN papers submitted to ICLR 2017 during the early explosion of adversarial networks.
Deep Learning for Natural Language Processing — ICLR 2017 Discoveries
Curated digest of standout NLP and sequence-modeling submissions from the ICLR 2017 open review pool.
Why Deep Learning Matters
How Convolutional Neural Networks improved ImageNet top-5 error from 17.0% in 2012 to 3.08% in 2016 — surpassing human performance — and why real-world deployment lagged.
2015
Announcing DeepLearningKit — Open Source Deep Learning Framework for Apple's iOS, OS X and tvOS
Launching DeepLearningKit, an open-source deep learning framework written in Swift and Metal for running pretrained Convolutional Neural Networks on iPhone, iPad, Mac, and Apple TV.
Swift and Metal GPU Programming on tvOS for the New Apple TV
Running custom Metal compute shaders from Swift on Apple's tvOS simulator and hardware on launch day.
Example of Swift and Metal GPU Programming on OS X 10.11 El Capitan
Porting Swift and Metal GPGPU vector computation to OS X 10.11 El Capitan right after Apple brought Metal from iOS to the Mac.
Deeplearning.University — Bibliographies from LISA Labs (Yoshua Bengio's Lab)
Adding curated deep learning bibliographies from Yoshua Bengio's LISA Lab at the University of Montreal to Deeplearning.University.
Example of Sharing Memory Between GPU and CPU with Swift and Metal for iOS 8
Demonstrating zero-copy shared CPU/GPU memory buffers with Swift and Apple Metal on iOS 8 devices — an early precursor to unified memory on Apple Silicon.
2014
GPGPU Performance of Swift/Metal vs Accelerate on iPhone 6 & 5S, iPad Air and iPad Mini
Benchmarking Swift + Metal GPU compute against Apple's Accelerate framework and pure Swift for Sigmoid evaluation on iPhone 5S, iPhone 6, iPad Air, and iPad Mini.
Data-Parallel Programming with Metal and Swift for iPhone/iPad GPU
An 8-step walkthrough for writing general-purpose GPU compute kernels in Metal Shading Language and invoking them from Swift on iOS 8.
Recent Experiences from Being on Hacker News' First Page with Startup Blog Posts
Lessons and traffic observations from landing multiple Memkite technical blog posts on the Hacker News front page.
Mobile Eats the Cloud
Why 64-bit mobile CPUs, mobile GPUs, and fast local flash storage shift computation back from the cloud to on-device supercomputers.
Technical Feasibility of Building Hitchhiker's Guide to the Galaxy (Offline Web Search)
Back-of-the-envelope storage, memory, and RRAM feasibility analysis of fitting offline web search (Wikipedia, books, and web indexes) onto a pocket device.
2011
MapReduce & Hadoop Algorithms in Academic Papers (4th Update)
Comprehensive curated bibliography of academic papers on MapReduce and Hadoop algorithms across search, graphs, and machine learning.
Using F# and C# (.NET/Mono) with Amazon's Elastic MapReduce (Hadoop)
Step-by-step example running compiled F# and C# mappers and reducers under Mono on Amazon Elastic MapReduce.
2010
Word Count with MapReduce on a GPU — A Python Example
Implementing parallel MapReduce word count on an NVIDIA Tesla C1060 GPU in Python with PyCUDA, reaching ~12 GB/s throughput on ~0.5 GB of Shakespeare text.
Evaluation of Search Predictions Made in May 2000
A 10-year retrospective grading three May 2000 predictions about search specialization, ISP/network players, and distributed multimedia indexing.
Towards Cloud Supercomputing
Examining early 2010 cloud-scale compute workloads across Zynga, Animoto, Myspace, NYT TimesMachine, Reddit, Justin.tv, and biotech on Amazon EC2.
2009
Preliminary Experiences Crawling the Web with 80legs
Early engineering notes and benchmarks from running distributed web crawls using the 80legs platform.
How to Use C++ Compiled Python (Shed Skin) for Amazon's Elastic MapReduce
Compiling Python mappers and reducers to native C++ via Shed Skin for faster Hadoop Streaming jobs on Amazon EMR.
Snakes on a Cloud — Mobile Agents with Python
Dealing with data gravity by moving compact Python and Cinpy-compiled C code to massive datasets instead of hauling data over the network.
2008
Cinpy — Writing and JIT-Compiling C Inside Python
Using Cinpy and TinyCC to embed and compile C functions directly inside Python with ctypes for dramatic inner-loop speedups.
Tools for Accelerating Python
A 2008 field guide to speeding up Python programs — from parallelizing with Hadoop, MPI, and Stackless to JIT/compilers (Psyco, Cython, Shedskin) and native C/Fortran/Assembly extensions.
Example of Microsimulation with Stackless Python
Simulating 100,000 concurrent robot agents playing musical chairs on a 10,000×10,000 grid using Stackless Python microthreads (tasklets).
Test-Driven Parsing with Python, Dparser and Doctest
Building GLR grammars incrementally in Python using Dparser paired with inline doctest specifications.
Future Number of Programming Languages — Singularity or Infinity?
Why monolithic do-everything languages collapse under complexity, and why small, Python-like declarative domain-specific languages (DSLs) are the better path.
Increase Automation of Test-Driven Development?
Early 2008 thoughts on automating both sides of the TDD loop with program synthesis, machine learning, and automated corner-case test generation.