In February 2025, AI code generation has moved from novelty to daily practice for a significant fraction of working developers. GitHub Copilot has been available for almost three years. Cursor and similar editors embed AI at a deeper level than autocomplete. Reasoning-capable models can now engage in multi-step software design conversations. The question has shifted from “will this change programming?” to “how is it already changing it and what comes next?” Read more
John Woods wrote something that has stayed with me: “Always code as if the guy who ends up maintaining your code will be a violent psychopath who knows where you live.” The quote is usually cited as a case for clean code. But there is a darker inversion of the same pressure that I have been thinking about — the developer who becomes so committed to an idea of clean code that the code stops serving the project. Read more
DeepSeek released R1 on January 20, 2025, alongside a detailed technical report describing the model’s architecture and training methodology. DeepSeek is a Chinese AI research lab founded in 2023 as an offshoot of High-Flyer Capital Management, a quantitative hedge fund, with the stated goal of pursuing AGI research independent of product revenue pressures. The R1 model itself was a 671-billion-parameter Mixture-of-Experts architecture with approximately 37 billion parameters active per token during inference — a design that substantially reduced the computational cost of a forward pass relative to a dense model of equivalent total parameter count. The training pipeline centered on GRPO (Group Relative Policy Optimization), a reinforcement learning technique that rewarded correct answers on verifiable tasks (mathematics, competitive programming, formal logic) without requiring a separately trained reward model for each capability area. An intermediate model, R1-Zero, was trained using only RL with no supervised fine-tuning whatsoever; it spontaneously developed extended internal reasoning traces — long chains of intermediate steps visible in the model’s output before its final answer — a behavior the researchers described as emerging from the optimization objective rather than being explicitly trained. The final R1 model incorporated a small amount of supervised fine-tuning on high-quality examples followed by additional RL training, and performed comparably to OpenAI’s o1 model on benchmarks including AIME 2024 (mathematics olympiad problems), Codeforces (competitive programming), and MATH-500. Read more
Many children spend hours playing games. Some of them eventually ask a more interesting question: how are games made? That question is an opening. I have been exploring it with my son, and I want to share what the experience has actually taught me — about teaching, about programming, and about spending time together on something real. Read more
OpenAI previewed o3 on December 20, 2024 in a “12 Days of OpenAI” event, releasing benchmark results before the model was publicly available. The headline number was 87.7% on ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence), a benchmark of visual pattern-recognition puzzles that had been specifically designed to resist language models through its emphasis on systematic generalization rather than pattern matching from training data. GPT-4o had scored 5% on the same benchmark; the previous state-of-the-art AI system had reached 53%. ARC-AGI creator François Chollet had explicitly stated that the benchmark could not be gamed by scale alone, making o3’s score the most discussed AI capability result of 2024. Read more
By November 2024, NVIDIA had become the indispensable infrastructure provider for large-scale AI training and inference, with its market capitalization briefly reaching $3.4 trillion in June 2024 — briefly making it the world’s most valuable company and reflecting the AI infrastructure spending wave triggered by ChatGPT’s November 2022 launch. The H100 SXM5 (Hopper architecture, TSMC N4, 80 GB HBM3, 3.35 TFLOPS BF16/FP16 dense training throughput, 700W TDP) had become the standard unit for LLM training: OpenAI trained GPT-4 on an estimated 25,000 A100s (Hopper’s predecessor); Anthropic, Google, and Meta deployed H100 clusters of similar scale. H200 SXM5 (announced December 2023, HBM3e 141 GB at 4.8 TB/s bandwidth vs H100’s 80 GB at 3.35 TB/s) offered approximately 1.9× inference throughput on memory-bound LLM serving workloads. NVIDIA’s GB200 NVL72 rack (announced March 2024, available late 2024 in Blackwell generation): 36 Grace Arm CPUs + 72 Blackwell B200 GPUs in a 72-GPU NVLink domain, delivering 1.44 ExaFLOPS FP4 for AI inference — targeting trillion-parameter model serving at economics impractical with H100. Read more
Intel announced Core Ultra 200S desktop processors (codenamed Arrow Lake) on October 24, 2024 for the LGA1851 socket. The flagship Core Ultra 9 285K featured 8 Lion Cove P-cores and 16 Skymont E-cores plus 4 low-power E-cores — 24 cores total — with no HyperThreading on any core: Intel removed simultaneous multithreading (SMT) from Lion Cove’s design entirely, arguing that the wider out-of-order execution engine in Lion Cove (larger instruction window, more execution units per core than Raptor Lake’s Golden Cove) made SMT less valuable than adding real cores. The 285K had a 125W TDP base and 250W max turbo power, with boost clocks to 5.7 GHz. The design was a disaggregated package: the CPU compute tile was manufactured on Intel 20A (Intel’s own process, roughly equivalent to TSMC N5 class), the GPU tile and SoC tile on TSMC N6, and the I/O tile on TSMC N6, all connected via Intel’s Foveros Direct die-to-die interconnect at 36 μm pitch — the same approach Intel used for Meteor Lake mobile (December 2023) now arriving at the desktop. Read more
Toyota is often described as one of the world’s largest automobile manufacturers. That description is correct. It is also incomplete. Toyota became important not only because of the cars it produced, but because of the way it learned to produce them. Read more
If you are beginning Computer Engineering in 2024, welcome to a field where tools are becoming more powerful and judgment is becoming more important. Read more
OpenAI released o1 on September 12, 2024 — initially called “o1-preview” for ChatGPT Plus and Team subscribers at $20/month, with API access at $15 per million input tokens and $60 per million output tokens. A smaller, faster variant called o1-mini launched the same day at $3/$12 per million tokens, designed for coding and STEM tasks on a tighter budget. The full o1 model (removing the “preview” label) launched December 5, 2024. Read more
In 2024, artificial intelligence is becoming a normal part of software products rather than a separate research topic. The pace of change this year has been striking even by recent standards. Read more
Google announced the Pixel 9 on August 13, 2024 — the earliest ever for a Pixel launch, moved from October to beat Apple’s iPhone 16 announcement in September. The Pixel 9 lineup included four models: Pixel 9 ($799, 6.3-inch), Pixel 9 Pro ($999, 6.3-inch), Pixel 9 Pro XL ($1,099, 6.8-inch), and Pixel 9 Pro Fold ($1,799, foldable). All used Google’s Tensor G4 processor (manufactured by Samsung on 4nm), with 12GB RAM on the standard Pixel 9 and 16GB on Pro models — increases over the Tensor G3’s 12GB configurations, explicitly motivated by on-device AI workload requirements. Read more
At approximately 04:09 UTC on July 19, 2024, CrowdStrike began deploying a content configuration update to its Falcon sensor — the kernel-level security agent running on Windows endpoints — that caused millions of machines to crash with a Blue Screen of Death (BSOD) and enter an unrecoverable boot loop. CrowdStrike Falcon operates as a kernel-mode driver (running at CPU Ring 0 privilege) that inspects process creation, network connections, file system operations, and memory allocations at a level below normal applications, giving it visibility into attacks that user-mode security tools cannot detect. The faulty update was not the Falcon sensor binary itself but a “rapid response content” channel file — a configuration update named C-00000291-*.sys that defined detection logic templates for a new IPC (Inter-Process Communication) Template Type. A bug in the validator for this new template type caused the Falcon sensor to attempt to access memory through a null pointer when loading the file during Windows boot, triggering a kernel STOP error before Windows could complete initialization. Machines that had already booted and were running were not immediately affected; machines that rebooted after 04:09 UTC during the rollout window entered the crash loop. Read more
If I were building a gaming PC on July 1, 2024, this is the machine I would actually want. I am not simply choosing the most expensive part in every category. The priority is gaming performance, then reliability, cooling and enough headroom to keep the system useful for years. Read more
Apple announced Apple Intelligence at WWDC on June 10, 2024, as a set of generative AI capabilities built into iOS 18, iPadOS 18, and macOS Sequoia. Initial features included Writing Tools (rewrite, proofread, and summarize text in any app), Image Playground (generating images from text prompts in three styles), Genmoji (custom emoji generated from descriptions), a redesigned Siri with conversational context and screen awareness, and a summarized notification inbox using on-device language models. Apple Intelligence launched in beta with iOS 18.1 in October 2024, initially for US English only, with language expansion planned through 2025. Read more
Microsoft announced Copilot+ PCs on May 20, 2024, defining a hardware category requiring a minimum 40 TOPS (tera-operations per second) NPU (Neural Processing Unit) — more than any existing Windows laptop NPU at the time. The first wave of Copilot+ PCs launched June 18, 2024, using Qualcomm’s Snapdragon X Elite and Snapdragon X Plus processors, which contained an Oryon CPU (derived from Nuvia’s acquisition), an Adreno GPU, and a Hexagon NPU rated at 45 TOPS. Devices from Surface, Samsung, Dell, HP, ASUS, Lenovo, and Acer launched simultaneously, priced from $999. Intel and AMD Copilot+ PC laptops (using Intel Lunar Lake and AMD Strix Point with their respective 47–50 TOPS NPUs) followed in fall 2024. Read more
Meta released Llama 3 on April 18, 2024, with 8-billion and 70-billion parameter models in both base and instruction-tuned variants, distributed through Hugging Face, llama.meta.com, and major cloud providers (AWS Bedrock, Google Vertex AI, Azure AI). The 8B and 70B Instruct variants outperformed comparable open-weight models (Mistral 7B, Gemma 7B) on standard benchmarks and approached GPT-3.5 Turbo performance on tasks like MMLU (79.5% for Llama 3 70B vs GPT-3.5’s ~70%) and coding. Meta announced that a Llama 3 model exceeding 400 billion parameters was in training. Read more
GitHub Copilot launched technical preview in June 2021 and became generally available in June 2022. In the roughly two years since, it has gone from curiosity to part of the daily development environment for many teams. By March 2024, GitHub reports over a million paying users and adoption at the majority of Fortune 500 companies. ChatGPT and Claude have expanded the use case beyond inline completion to conversational assistance — asking questions about code, requesting explanations, and generating complete functions in a chat interface. Cursor, a VS Code fork with deeper model integration, has emerged in 2023-2024 as a model for what tight AI-editor integration can look like. Read more
NVIDIA announced the Blackwell GPU architecture at GTC on March 18, 2024 in a keynote where Jensen Huang said the company was “not in the semiconductor business — we’re in the infrastructure business.” The flagship B200 GPU used two connected dies (the GB202 silicon) totaling 208 billion transistors on TSMC’s 4NP custom process — more than double the H100’s 80 billion transistors on 4N. The B200 delivered 20 petaFLOPS at FP4, compared to the H100’s 4 petaFLOPS at FP8 (halving precision doubles throughput, so the increase was partly from precision and partly from transistor count). HBM3e memory reached 192 GB per B200 with 8 TB/s bandwidth — nearly 2.4× the H100’s 3.35 TB/s. Read more
