Sunday — January 18, 2026

AI insiders launch Poison Fountain to corrupt training data, Video-to-Grid slashes video analysis expenses by 600x and RTX 5090s enable private LLM inference at 200x lower cost than APIs.

Interested in AI engineering? Let's talk

News

Crypto grifters are recruiting open-source AI developers

AI engineers Geoff Huntley and Steve Yegge are promoting $RALPH and $GAS, crypto coins linked to their respective projects: the Ralph Wiggum loop for Claude Code automation and the Gas Town LLM agent platform. Despite the branding, these tokens have no technical utility within the software and are created via the Bags platform to funnel trading fees to the developers. This mechanism functions as a predatory pump-and-dump scheme that exploits the reputation of open-source AI contributors to bootstrap speculative memecoins.

I built a tool to assist AI agents to know when a PR is good to go

Good To Go is a CLI tool providing deterministic PR readiness detection for AI coding agents to prevent infinite polling and premature task completion. It aggregates CI status, classifies review comments into actionable categories, and tracks thread resolution to provide structured JSON output for agent consumption. By offering definitive states like READY or ACTION_REQUIRED, it enables LLM-driven workflows to programmatically decide when to merge or continue iterations.

Reality Is Breaking the "AI Revolution"

Salesforce CEO Marc Benioff’s aggressive strategy to replace nearly 50% of support staff with agentic AI has resulted in a significant decline in service quality and operational stability. The AI failed to handle nuanced customer escalations and long-tail problems, forcing remaining employees to spend more time correcting AI-generated errors than performing their original duties. Consequently, Salesforce is pivoting from a replacement-first model to a worker augmentation strategy while attempting to re-hire for the critical institutional expertise lost during the layoffs.

Why AI Doesn't Think: We Need to Stop Calling It "Cognition"

The document argues against the anthropomorphization of AI by challenging the use of the term "cognition" to describe model outputs. It suggests that LLMs lack genuine thought processes, advocating for a more rigorous distinction between statistical inference and human-like reasoning.

AI industry insiders launch site to poison the data that feeds them

AI industry insiders have launched "Poison Fountain," a project designed to undermine LLM development through coordinated data poisoning. The initiative encourages web operators to serve crawlers corrupted training data, such as code with subtle logic errors, to degrade model performance and cognitive integrity. Citing research that even small samples of poisoned data can be effective, the group aims to use these "information weapons" to counter the perceived existential threats of machine intelligence and accelerate model collapse.

Research

VaultGemma: A Differentially Private LLM

VaultGemma 1B is a 1B-parameter LLM in the Gemma family trained with differential privacy using the Gemma 2 pretraining data mixture. This open-release model represents a significant advancement in privacy-preserving LLM development.

Private LLM Inference on Consumer Blackwell GPUs

SMEs are seeking alternatives to cloud LLM APIs due to privacy and cost. A systematic evaluation of NVIDIA Blackwell consumer GPUs (RTX 5060 Ti, 5070 Ti, 5090) for production LLM inference found that the RTX 5090 delivers high throughput and low latency, while budget GPUs offer superior throughput-per-dollar for API workloads. NVFP4 quantization provides 1.6x throughput over BF16 with 41% energy reduction and minimal quality loss. Self-hosted inference is 40-200x cheaper than cloud APIs, with hardware ROI in under four months. Consumer GPUs can reliably replace cloud inference for most SME workloads, though high-end GPUs are crucial for latency-critical long-context RAG.

Does AI-Assisted Coding Deliver? A Study of Cursor's Impact on Software Projects

A difference-in-differences analysis of GitHub projects shows that adopting the LLM agent Cursor yields a significant but transient increase in development velocity. This initial gain is offset by a persistent rise in code complexity and static analysis warnings, which GMM estimation identifies as the primary cause of long-term velocity slowdown.

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

Mode collapse in post-training alignment is driven by typicality bias in preference data, where human annotators systematically favor familiar outputs. Verbalized Sampling (VS) is a training-free inference strategy that mitigates this by prompting LLMs to generate multiple responses alongside a verbalized probability distribution. VS improves diversity by up to 2.1x across creative and open-ended tasks while maintaining safety and factual integrity, with performance gains scaling alongside model capability.

Restoring Locality:Heisenberg Picture as Separable Description of Quantum Theory

Everettian quantum theory resolves the apparent conflict between General Relativity’s locality and Quantum Theory’s perceived non-locality, a proof first established by Deutsch & Hayden (2000). The paper reviews how local realism is maintained through mechanisms like local branching and updated analytical frameworks, reconciling these two fundamental physical theories.

Code

Video-to-Grid – Analyze videos with one Vision API call

VAM Seek × AI optimizes video analysis by compressing 48 frames into a single 8x6 grid image, reducing API costs by ~600x compared to traditional frame-by-frame processing. Using Claude Sonnet, the system maps visual content to specific timestamps and grid coordinates for verifiable retrieval. Key features include prompt caching for efficient follow-up queries and adaptive grid density, with future plans for recursive zooming and Whisper-based transcript integration.

Monitor Claude/Codex usage on Linux via browser cookies (no API keys)

Waybar AI Usage monitors Claude Code and OpenAI Codex CLI usage limits directly in the Waybar status bar. It utilizes browser_cookie3 and curl_cffi to extract session cookies from browsers, enabling real-time tracking of 5-hour and 7-day usage windows without requiring API keys. The tool outputs JSON for Waybar custom modules, featuring color-coded status indicators and countdown timers for quota resets.

30min video analysis for $0.003 via frame-tiling and Vision API

My way – 18-agent autonomous workflow for ClaudeCode – issues to deploy

Awesome Slash Commands is a cross-platform plugin providing professional-grade workflow automation for AI coding assistants like Claude Code, Codex CLI, and OpenCode. It offers powerful slash commands for autonomous development, including a /next-task orchestrator that manages task-to-production workflows using 18 specialist agents (optimized with Opus/Sonnet models) with minimal human intervention. Other features include multi-agent code review, AI slop cleanup, documentation synchronization, and /reality-check for detecting plan drift through parallel analysis of issues, documentation, and code to generate prioritized reconstruction plans.

Whisper Money – a zero-knowledge personal finance app (E2E encrypted)

Whisper Money is a privacy-first personal finance application designed for end-to-end encrypted tracking, categorization, and analysis of spending. It features automated transaction rules and is built on a modern tech stack including Laravel 12, React 19 with TypeScript, MySQL, and Redis, offering straightforward local setup and Docker-based deployment options.