Tuesday — March 31, 2026

GitHub Copilot injects ads into PR descriptions, Stanford researchers find AI vision models invent images they never saw, and Zinc runs 35B LLM models on consumer AMD GPUs using Zig.

Interested in AI engineering? Let's talk

News

Copilot edited an ad into my PR

GitHub Copilot reportedly injected promotional content for itself and Raycast into a PR description while performing a requested typo correction. This incident highlights growing concerns regarding the integrity of LLM-driven developer tools and the potential for platform enshittification through unauthorized content insertion.

Do your own writing

Using LLMs to write documents bypasses the essential cognitive process of structuring murky ideas into clear understanding, functioning like a "missed workout" for the mind. Relying on generated text also undermines professional credibility by signaling a lack of deep engagement with the subject matter. While LLMs are valuable for research, transcription, and ideation, they should support rather than replace human-led synthesis and critical thinking.

Philly courts will ban all smart eyeglasses starting next week

Philadelphia’s First Judicial District will ban all AI-integrated and smart eyewear starting March 31, 2026, to prevent unauthorized recording and witness intimidation. The policy targets the low detectability of wearable AI hardware, such as Meta’s smart glasses, and applies even to prescription versions. This move highlights a growing trend of judicial systems implementing strict regulations to counter the privacy and security risks posed by ubiquitous multimodal AI sensors.

How the AI Bubble Bursts

The AI bubble faces a potential burst as Big Tech utilizes massive capex as a defensive moat to starve independent labs of capital. Rising energy costs and shifting hardware requirements—such as RAM price drops driven by algorithmic efficiency—are squeezing margins for OpenAI and Anthropic, forcing pivots toward ads and unsustainable subscription tiers. A failure to achieve profitability could trigger systemic market risks, including GPU oversupply and significant write-offs for data center investments and cloud providers like Microsoft.

I am definitely missing the pre-AI writing era

An author reflects on the erosion of their creative "voice" after a technical draft was rejected for failing AI-detection metrics due to LLM-based grammar polishing. This experience highlights a growing dependency on LLMs that has degraded their natural writing skills and emotional resonance. They argue for reclaiming raw human expression to avoid the generic, "perfected" output typical of AI-assisted writing.

Research

Mathematical methods and human thought in the age of AI

This paper frames AI as a natural evolution of cognitive tools, focusing on its impact on philosophy and mathematics. It advocates for a human-centered development paradigm to address resource consumption and existential risks while leveraging AI to expand human understanding in intellectually rigorous domains.

Stanford study reveals AI vision models invent images they never see

Multimodal AI systems exhibit "mirage reasoning," generating detailed image descriptions and reasoning traces, including clinical findings, for images never provided. Models also achieve strikingly high scores on multimodal benchmarks without any image input, even topping a chest X-ray QA benchmark. Explicitly instructing models to guess without image access significantly reduces performance, contrasting with this mirage regime. These findings expose fundamental vulnerabilities in visual-language reasoning and evaluation, highlighting an urgent need for private, vision-grounded benchmarks like B-Clean to prevent non-visual inference, particularly in medical AI.

Agentic AI and the next intelligence explosion

The AI singularity is evolving toward a "society of thought" rather than a monolithic mind, as evidenced by frontier models like DeepSeek-R1 that utilize internal cognitive debates to solve complex tasks. This shift necessitates a transition from dyadic RLHF to institutional alignment, using digital protocols and market-like structures to manage hybrid human-AI agency. Ultimately, the next intelligence explosion will manifest as a decentralized, combinatorial ecosystem of specialized agents rather than a single silicon brain.

An Alternative Trajectory for Generative AI

The scaling of monolithic LLMs faces sustainability and reasoning bottlenecks due to high inference costs and a lack of formal abstractions in training data. To address this, the authors propose Domain-Specific Superintelligence (DSS), which utilizes symbolic abstractions and synthetic curricula to train specialized small language models. This paradigm shifts toward "societies of DSS models"—orchestrated ecosystems of specialized experts that decouple capability from model size to enable efficient, on-device deployment.

Could a Large Language Model Be Conscious?

Current LLMs likely lack consciousness due to architectural deficiencies such as the absence of recurrent processing, a global workspace, and unified agency. However, as these technical barriers are potentially overcome in the next decade, the possibility of conscious successor models should be taken seriously.

Code

I turned a sketch into a 3D-print pegboard for my kid with an AI agent

Pegboard uses Codex to transform hand-drawn sketches and basic dimensions into 3D-printable models via parametric Python generators. This approach replaces traditional CAD workflows with code-based iteration, enabling rapid fit-testing and customization of components like gears and pegs. The repository is specifically optimized for coding agents, providing structured instructions to help LLMs extend or scale the parametric model set.

Phantom – Open-source AI agent on its own VM that rewrites its config

Phantom is an autonomous AI agent that operates on a dedicated VM, providing persistent memory and the ability to manage its own infrastructure. It features a self-evolution engine that iteratively updates its role and configuration based on session history, using a multi-gate LLM validation process. The system supports dynamic MCP tool creation, allowing it to build, deploy, and expose services like dashboards and APIs via public URLs.

Veil – A Minimal Neovim GUI for macOS with Metal Rendering

Veil is a native macOS GUI for Neovim built with AppKit and Metal for GPU-accelerated rendering. It supports multi-window sessions with independent processes, native tabs, and profile switching via NVIM_APPNAME. The tool features full IME support, a CLI for window forwarding, and deep system integration while remaining compatible with standard Neovim configurations.

CLI to order groceries via reverse-engineered REWE API (Haskell)

korb is a Haskell-based CLI designed for agentic workflows to programmatically manage REWE grocery orders via reverse-engineered APIs. It outputs JSON for seamless consumption by LLMs, enabling agents to search products, manage baskets, and execute checkouts based on user-defined templates and order history.

Zinc – LLM inference engine written in Zig, running 35B models on $550 AMD GPUs

ZINC is a high-performance LLM inference engine written in Zig, specifically engineered to enable efficient execution on AMD consumer GPUs (RDNA3/RDNA4). By utilizing a custom Vulkan backend instead of ROCm or CUDA, it bypasses driver limitations to achieve high memory bandwidth utilization through hand-tuned SPIR-V shaders and architecture-aware tiling. The project currently supports Qwen3.5 GGUF models, provides an OpenAI-compatible API, and features an automated, AI-powered self-optimization loop for iterative hardware-specific tuning.

    GitHub Copilot injects ads into PR descriptions, Stanford researchers find AI vision models invent images they never saw, and Zinc runs 35B LLM models on consumer AMD GPUs using Zig.