What My Algorithm Read This Week (Feb 15, 2025)
I built a system that curates my own newsfeed — not optimized for engagement or ad revenue, just for what’s actually useful to my work. This week’s lens: contrarian. Same sources, but filtering for signals that challenge the prevailing narrative. 29 items from the last 24 hours, 12 made the cut.
AI Isn’t As Good As You Think
We Hid Backdoors in ~40MB Binaries and Asked AI + Ghidra to Find Them
The conventional wisdom says AI is transforming cybersecurity, finding vulnerabilities at machine speed. Quesma’s BinaryAudit benchmark tested this claim directly — pitting Claude Opus 4.6 against deliberately-planted backdoors in compiled binaries.
The result: 49% accuracy with a 28% false positive rate. Worse than a coin flip once you factor in the false alarms. The models demonstrated genuine reverse engineering ability but “often rationalize away obvious malicious code.”
If AI can’t reliably spot backdoors it was told exist in a controlled benchmark, the “AI-powered security” marketing deserves deep skepticism.
Amazon Blames Human Employees for an AI Coding Agent’s Mistake
When things go wrong, the “human-in-the-loop” becomes the scapegoat. Amazon attributed an AI coding agent’s error to the employees who approved its output. This reveals the accountability trap: companies deploy AI agents to write code, then hold humans responsible for not catching the AI’s mistakes.
The review bottleneck that was supposed to be a safety mechanism becomes a liability shield.
Stripe’s Minions: 1,000+ PRs/Week With Zero Human-Written Code
Read alongside the Amazon story. Stripe reports over a thousand pull requests merged weekly that are “completely minion-produced” and “contain no human-written code.” They’re human-reviewed, but how carefully? At that volume, review becomes a rubber stamp.
The question nobody’s asking: what happens when one of those thousand weekly PRs introduces a subtle security flaw, and the reviewer who approved it gets the Amazon treatment?
The GPU Paradigm Is Already Dying
How Taalas “Prints” LLM onto a Chip
Taalas physically etches model weights directly onto silicon as transistors, eliminating the memory bandwidth bottleneck entirely. The result: 17,000 tokens per second with dramatically lower power consumption. Only the top two chip masks need customization per model.
If this scales, the entire GPU-as-AI-infrastructure thesis — and Nvidia’s $3T valuation — rests on increasingly shaky ground.
[AINews] The Custom ASIC Thesis
Latent Space makes the economic case: since inference value for a trained model often exceeds its training cost, companies can justify building model-specific silicon. A 20% efficiency gain translates to hundreds of millions in savings. The thesis predicts 20,000+ tokens/sec through model-chip codesign within two years.
The GPU era may be remembered as a transitional phase, like vacuum tubes before transistors.
Llama 3.1 70B on a Single RTX 3090 via NVMe-to-GPU Bypassing the CPU
Even within the GPU paradigm, the conventional approach is wrong. This project runs a 70B parameter model on a consumer-grade card by bypassing the CPU entirely with NVMe-to-GPU direct transfer. The “you need $100K in hardware” narrative is overstated — creative engineering on commodity hardware keeps closing the gap.
Decentralization Is More Fragile Than Promised
A Botnet Accidentally Destroyed I2P
The Kimwolf botnet flooded I2P with 700,000 hostile nodes against a base of 15,000-20,000 legitimate ones — a 39:1 ratio — while trying to set up backup C2 servers after its primary infrastructure was destroyed.
The attack was accidental. Three years of state-sponsored Sybil attacks couldn’t do what a botnet stumbled into by chance. Permissionless openness, the core feature of decentralized networks, is also their fatal vulnerability.
“Verification” Is Surveillance in Disguise
I Verified My LinkedIn Identity. Here’s What I Handed Over
Verifying on LinkedIn through Persona means handing over passport photos, facial biometrics, NFC chip data, national ID numbers, device fingerprints, and behavioral tracking to 17 subprocessors — including AI companies that use the data for training under “legitimate interests” rather than consent. All of it accessible to US law enforcement under the CLOUD Act, potentially without notification.
That blue checkmark costs more than you think.
Small Is Beautiful
zclaw: Personal AI Assistant in Under 888 KB on an ESP32
While the industry races toward trillion-parameter models running on server farms, zclaw runs a personal AI assistant on an ESP32 microcontroller in under 888 KB. The contrarian bet: the future of personal AI isn’t in the cloud — it’s in tiny, auditable, offline-capable systems you actually own.
Andrej Karpathy Talks About “Claws”
Karpathy’s “Claws” concept — a new orchestration layer above LLM agents — notably emphasizes small, auditable implementations (NanoClaw) alongside larger systems. Even AI’s biggest proponents are gravitating toward smaller, controllable, inspectable architectures rather than opaque mega-agents.
You’re Using “Social Media” Wrong
Attention Media ≠ Social Networks
Susam Pal draws a sharp line: true social networks show content from people you follow. What Facebook and Twitter became are “attention media” — platforms designed to capture and monetize attention through algorithmic injection of content from strangers. The conflation of the two isn’t just semantic — it obscures what we’ve actually lost.
How I Use Claude Code: Separation of Planning and Execution
Against the “just vibe code it” trend: Boris Tane’s key insight is “never let Claude write code until you’ve reviewed and approved a written plan.” More human control, not less, produces better AI-assisted code. The winning strategy for AI coding tools is the opposite of what most people do — disciplined constraint, not unconstrained generation.
This digest is generated by my own curation algorithm — 35 feeds, contrarian lens, no engagement optimization. The system is part of an experiment in prompt learning I’m building in the open.