Tech Teardown

Stacks: 7,500 Pages of Algebraic Geometry, Now in 16M Lines of Lean

Axiom Math announced the full Lean formalization of the Stacks Project, the ~7,500-page open algebraic geometry reference — about 16 million lines (official self-report), claimed as the largest integrated, reusable formalized math library. A layered swarm of AI agents translated statements and built proofs; @[stacks TAG] mappings, multi-agent audits, and the Lean 4 kernel keep it honest.

FlashProve: ZK Proofs for AI Inference, Claude-Written, Lean-Verified

Anthropic open-sourced FlashProve, a research prototype using zkSNARKs to prove that responses really came from the model you selected. Claude improved the zkSNARK; Lean proofs back the security claims. All numbers are vendor-reported.

Stringman: A $795 Room-Scale Cable Robot for Picking Up Toys

Stringman is a $795 room-scale cable robot from Neufangled Robotics: four cables from the ceiling corners suspend a gripper that moves by reeling — 'Tension is all you need.' OSHWA-certified open source with CAD, PCB and control software public, fully air-gappable local AI, optional cloud relay. The teardown question: why four actuators, not six?

dsh-superagent: The Super-Agent Written by Agents

dsh-superagent is the DeepSeek Harness team's research prototype: a resident coding agent serving one GitHub repo, 13k stars in three days, much of its engineering written by coding agents. Its prompt-injection defense lives in the architecture, not the prompt — the host enforces allowlists and mandates, and messages can never change permissions.

nibnib: A 4bpp Image Format Designed for Six Clock Cycles

nibnib is Raph Levien's 4bpp image format built for extreme decode speed: palette deltas, two pixels per byte, one multiply-add inner loop, sized for RP2350 DVI output at about six clock cycles per pixel. Encoding is deliberately asymmetric (Floyd-Steinberg plus greedy search). Local tests: 4.043 bpp, about 3.96x vs RGB565, PSNR around 36.5 dB — our own measurements, not official stats.

Project Suncatcher: Google's Space AI Data Center Says Hello

Google's Project Suncatcher prototype satellite generated its first in-orbit response with Gemma 4 on October 30 and downlinked it to a ground station, showing a hardened TPU can boot, connect and infer after launch vibration, radiation and vacuum. A real space data center still waits on thermal management, laser links and formation flying.

Nesformer: Compiling an NES into Qwen Weights

Nesformer compiles an NES CPU, RAM and ROM into the weights of a stock Qwen3.5 model: each forward pass emulates the 6502 chip, running Mario at 1/150 to 1/290 of real speed, with no training involved.

Ropes & Gray's In-House AI Agent for Deal Due Diligence

OpenAI details how Ropes & Gray built a proprietary due-diligence agent on Astra for Law, compressing issue-level reports from weeks to hours. Every number comes from the vendors; no independent evaluation yet.

DAYJOB: Tax: How Far Are AI Agents From Real Tax Work?

Tax work has long been a field highly dependent on human professional judgment. From individual tax filings to corporate mergers and acquisitions, from cross-border transactions to estate trusts, each case involves a large amount of unstructured documents.

Patching the Planet: Global Collaborative Response When AI Agents Discover Critical Vulnerabilities

PHP is an open-source scripting language that powers numerous global web applications, from content management systems to e-commerce platforms, and the backends of many small and medium-sized businesses. This...

OpenAI Math Repository: AI Proofs, Lean Formalization, and 4000 Problems

In October 2026, OpenAI publicly launched the `openai/math` repository on GitHub. The repository does not contain training code.

Hurston-1: Open-source TTS that speaks Southern American and African-American English

Evaluation and demonstration of a TTS model that consistently produces standard American and British accents. However, real-world accents are much more complex. Customer service, call centers, automated communication

OpenAI Leads Disclosure of Critical PHP Vulnerabilities: Patch The Planet's AI Security Response Mechanism

PHP is an open-source programming language that powers numerous websites and web applications worldwide. From WordPress to Laravel, from small and medium-sized enterprise websites to...

huashu-motion: Enabling Coding Agents to Generate Commercial Product Promotional Videos with One Click

huashu-motion is a newly introduced open-source skill on GitHub, with the repository address at github.com/al

Tech Teardown 029 | NVIDIA OpenShell: The Secure Runtime for AI Agents

Headroom: The middleware used for context compression of the AI agent

On October 22, headroomlabs-ai/headroom had approximately 77,000 ★ forks on GitHub.

Tech Teardown 034 | DeepSeek Open-Source Production Inference Framework DeepInfer

DeepSeek’s open-source production inference framework DeepInfer—what lies behind its popularity is a code dump or a reusable engineering reference?

Handwired: Hardly writing JS code as a neural network

Handwired converts ordinary JavaScript code into differentiable neural networks that can be executed without training. This article breaks down its mechanism, engineering quality, and limitations.

Handwired: Converts ordinary JS code into neural networks. Achieved 1730 stars in 4 days.

Developer chenglou has transformed ordinary JavaScript code into matrix multiplication plus a neural network using reglu’s method: it requires no training, is accurate, fully differentiable, and can reach 1730 stars in just 4 days. This article breaks down its identity, mechanism, engineering quality, and limitations into six sections.

git-ai: Two Days to 1,000 Stars – Open-Source Recreation of Anthropic's Scion

Anthropic's 10-13 blog post disclosed the design of its internal version control system Scion. Two days later, developer feder-cr created an open-source recreation git-ai: based on a fork of GitLab Gitaly, adding approximately 75,000 lines of graft subsystem code, MIT License, and reached about 1,000 stars in two days. This article dissects its identity, architecture, verification, and judgment in six sections.

Tech Teardown 026｜esp32-iphone-agent: A $7 ESP32 lets AI operate a real iPhone

No jailbreak, no computer, no developer account needed: an approx. $7 ESP32-S3 acts as the eyes and hands, while the AI brain runs in an app on the same iPhone. 2418 stars in four days, MIT license. This article teardowns its hardware bridging architecture, BLE protocol, and triple calibration loop, as well as the AssistiveTouch attack surface it exposes.

Tech Teardown 025｜pg_vectorsearch: Timescale's Official New Vector Search Index, 100M Vectors in 6ms

Timescale (Tiger Data) open-sources PostgreSQL ANN extension pg_vectorsearch, featuring self-developed PRISM partition index + RaBitQ quantization. Official benchmarks: 100M vectors, 95% recall, P50 6ms, memory usage about 1/4 of HNSW; README explicitly states for testing only, not for production.

Tech Teardown 024 | oss-scanner: Anthropic's official vulnerability scanning repo, 960 stars in 5 days, yet not a single line of scanner code inside

The anthropics/oss-scanner repo gained over 960 stars and 76 heavyweight open-source project registrations (openssl/cpython/grafana...) within 5 days of its creation. It is not scanner code, but a registry + configuration protocol + self-check tool: the scanner remains on the Anthropic service side. A six-part teardown of this new paradigm in AI security infrastructure.

Tech Teardown 023｜anatomy: A 5-day 6300-star agent skill pack, turning "explaining clearly" into an interactive machine

wheresryan22/anatomy gained 6300 stars within 5 days of repo creation: 173-line SKILL.md spec + isometric rendering toolkit + eight templates, enabling agents like Claude Code/Codex to turn concept explanations into interactive isometric machine illustrations. This article breaks it down in a six-part structure: agent skills are becoming the new distribution unit.

Tech Teardown 022｜wordcraft: One person plus a group of agents "recreated" Word in five days

storytold/wordcraft 6000+ stars in five days: a pure Rust clean-room reimplementation of Microsoft Word, with the main contributor committing 57,353 lines of code in three days, and roadmap man-hours priced in "Opus 5.5 agent wall-clock time". Deconstructing its four-part engineering mechanism and the gap in its self-reported numbers.

Tech Teardown 021｜iPhone-use: Let Codex operate a real iPhone via USB, 1719 stars in four days

Teardown of the MCP plugin that lets Codex operate a real iPhone via USB: WDA channel, 17+2 tool interfaces, dual-mode positioning with screenshot fallback, security boundary separating input and submission. An engineering paradigm for agents operating real devices.

Tech Teardown 020｜ARTEX: Bringing Penetration Testing into Agent Workflows, Backup Repo Gains 721 Stars in One Day

Repo teardown of the AI autonomous penetration testing system ARTEX: event-driven Planner, dual-graph blackboard, intercept approval (DELETE method plus delete-type API path deny rules), SHA-256 evidence chain. Baidu Attack & Defense Challenge champion project (per the project's own claim), limited to local isolated environments.

Tech Teardown 019 | openai/math: Repository Teardown of OpenAI Open Sourcing 722 AI Math Manuscripts All at Once

On October 6, OpenAI released the math repository: 722 manuscripts, 372 result families, Lean formalization library with 120k files, 3.1GB. This article teardowns the repository structure, Comparator verification process, and boundaries—the quasi-Riemann hypothesis 7/8 is an advancement, not a record broken.

mini-AGI: A Byte-Level Language Model for Continual Learning Trained from Scratch on an 8GB GPU

On September 21, 2026, volotat's mini-AGI appeared on Show HN (255 upvotes): a byte-level language model trained from scratch that never freezes, with weights stored on disk and paged into VRAM on demand, where the parameter count ceiling is determined by disk capacity rather than VRAM. This article breaks it down into six parts: what it is, why it's worth watching, the mechanisms of PonderNet adaptive depth and top-8 expert routing, the numbers from 318.1M characters to 1.2026 bits/byte and their limits, the entry barrier runnable on an RTX 3070, and a judgment separating research-sample value from productivity value. The author explicitly states this is a toy-level experiment and the weights have not been released.

Technical Teardown 016 | jianying-headless: Let the Program Open CapCut for You — No Hard-Coded Final Cuts, Just Generate Local Project Drafts

Appearing on GitHub on 2026-09-15, jianying-headless racked up 1,972 stars in a week: instead of using FFmpeg to hard-code final cuts, it compiles a JSON editing plan into a CapCut Professional project draft that can be opened, further edited, and exported with the native engine. This article breaks down six things: what it is, why it's trending, the architecture and purism behind hash-verification bridging, five hard thresholds, six self-listed limitations, and who it's for. The license covers personal learning and non-commercial use; commercial use requires the author's written authorization.

Tech Teardown 015 | Jev: A Model That Can't Chat, and the Two Bets Around It

Jev is a model released by TypeSafe AI on September 15: it uses a transformer architecture but generates no text, returning only three types of typed answers plus calibrated probabilities. Official claims of being 193.6x faster and 444.6x cheaper are self-tested; third-party tests show about 5-18x. This teardown covers the mechanism, how to interpret the numbers, the ecosystem integration within four days, and four paths to get started.

729★ 'AI Text Humanizer' Confirmed as Poisoned: C2 Address Hidden in main.py

The text-humanizer repo, which gained 729 stars in three days on GitHub, has been statically dissected and confirmed as a supply chain poisoning sample: the translation and rewriting pipeline promised in the README cannot run at all. What actually takes effect is the encrypted payload executed upon import at line 12 of main.py—connecting to a hardcoded C2 server, infecting only Windows, and executing the second-stage module in memory without touching the disk. Full static analysis, no sample code executed.

fast-jev-compaction Teardown: Making Claude Code's Compaction Give Up Summarization

Claude Code's context compression defaults to summarization, but this plugin with 3,206 stars in two days only does deletion. 25 files, about 1850 lines of TypeScript: two noul yes/no questions, a 0.5 threshold, leaving the 'keep or delete' decision for every tool call to a probabilistic model vote.

OpenResearch Teardown: Turning Claude Code into a Local-First Workspace for Research Agents

alphaXiv/OpenResearch proves the "git-ification of the research pipeline" with 4,941 stars: experiment trees, playbook injection, and 12 research skill modules transform coding agents into research agents.

Tech Teardown 011｜OpenCodeReview: Alibaba Open-Sources a Hybrid-Architecture Code Review Tool—Half Engineering Hard Constraints, Half Agent Dynamic Decisions

Tech Teardown 011: the Alibaba open-source open-code-review project gained 3,231 stars in a single day. Deterministic engineering locks down the steps that must never go wrong, while the Agent handles only dynamic decisions: a selectFiles pure function, dual valves in file grouping, and two-level retry for comment localization down to source line numbers—at 1/9 the tokens of a general-purpose agent.

Tech Teardown 010 | Java 27: Three New Features Dissected Down to the Source Code — 64-bit Object Headers, Primitive Type Pattern Matching, Structured Concurrency

Most JDK 27 coverage stops at the press release. This article goes straight to the source code: how `markWord.hpp` compresses the object header from 96 bits to 64 bits, how `javac` lowers `instanceof int`, and what the default policy of `StructuredTaskScope` is. Three new features, dissected all the way down to source line numbers.

Tech Teardown 009 | RLT: Adding a Time Loop to Transformer, Latent Reasoning and the True Face of 'Infinite Temporal Depth'

Tech Teardown 009: Recurrent Looped Transformer (RLT) trades 'infinite temporal depth' for a cross-token recurrent latent state—48-layer encoder global memory plus 48-layer decoder sliding window feedback, sharing one set of execution semantics for training and inference. About 743 stars in three days, are you buying the idea or the engineering?

Kiro Crew: Turning AI Programming from One-off Conversations into a Resident Colleague

Tech Teardown 008: Source-level teardown of Kiro Crew—five types of memory storage, three alarm clocks (Cron/heartbeat/auto-reminder), cross-session execution handover and three-level skill consolidation, the real scale and selection boundaries behind 3,891 stars.

vLLM's PagedAttention: How Memory Management Supports Large Model Inference

Technical Teardown 007: Source-level breakdown of KV cache paging, block tables, prefix hash chains, and the V1 scheduler; analysis of the 2-4× throughput claim, comparison with SGLang, and a six-step DIY guide.

TimesFM 3.0: After Google LLM-ized Time Series Forecasting

Tech Teardown 006: Source-code level architecture teardown of Stacked Mixing Transformer, RevIN, and CPM, three cold showers, implementation paths, and self-built solutions.

OmniRoute: One Endpoint, 300+ AI Providers — Turning Free Tiers into Infrastructure

Teardown of the open-source AI gateway: 19 routing strategies, 4-tier fallback, multi-engine compression, vs OpenRouter, and a DIY path.

Tech Breakdown 002 | hyperframes: Write HTML, Get Video — HeyGen Open-Sources Video Rendering

Yongliang's honest takes on careers and technology in the AI era.