AI Intelligence Briefing — Saturday, March 28, 2026

Top Stories

My minute-by-minute response to the LiteLLM malware attack

Source: Simon Willison (Tier 1) | Category: tools | Relevance: 9/10

Simon Willison documents a supply-chain malware attack on LiteLLM, a popular proxy used to route requests across multiple LLM providers.

Why this matters: If you use any open-source AI libraries (and most developers do), a malware attack on one of them could silently compromise your entire system. This is a real-world cautionary tale about the risks hiding in your software dependencies.

So What: LiteLLM is widely used in agentic workflows and MCP setups to unify LLM API calls. If you’re using it — or anything that depends on it — you need to check your versions immediately and review Simon’s incident timeline for remediation steps. This also reinforces the case for pinning dependencies and auditing your AI toolchain’s supply chain regularly.

[AINews] Everything is CLI

Source: Latent Space (Tier 1) | Category: patterns | Relevance: 8/10

Latent Space (swyx) reports on the accelerating trend of CLI-first interfaces becoming the dominant paradigm for AI agents.

Why this matters: The way we interact with AI tools is shifting — instead of fancy web dashboards, more and more AI agents are being built to run from the command line, which is faster and easier to automate. This matters because it changes how you’d design and connect AI-powered workflows.

So What: As a Claude Code user, you’re already in this paradigm. This trend validates investing deeper in CLI-based agent orchestration rather than building custom UIs. Consider whether your current Astro/Vercel workflows could benefit from exposing more functionality as CLI commands that agents can invoke directly.

We Rewrote JSONata with AI in a Day, Saved $500K/Year

Source: Simon Willison (Tier 1) | Category: patterns | Relevance: 8/10

A team used AI-assisted coding to rewrite a critical data-transformation library in one day, eliminating a $500K/year dependency.

Why this matters: This shows that AI coding tools aren’t just for writing new features — they can help you replace expensive third-party services by rewriting core components surprisingly fast. It’s a concrete example of AI saving real money at a business level.

So What: This is a playbook for identifying expensive runtime dependencies in your own stack and using Claude Code or similar tools to port or rewrite them. Look at your Vercel bills and third-party API costs — any data transformation or processing layer is now a candidate for an AI-assisted rewrite sprint.

Vibe coding SwiftUI apps is a lot of fun

Source: Simon Willison (Tier 1) | Category: patterns | Relevance: 7/10

Simon Willison shares his experience using AI-assisted ‘vibe coding’ to rapidly build SwiftUI applications.

Why this matters: Even if you don’t build iOS apps, this demonstrates how conversational AI coding is maturing — you describe what you want, and the AI handles the implementation details of frameworks you may not deeply know. It’s expanding what a single developer can ship.

So What: The vibe coding pattern — describing intent and letting AI handle framework-specific syntax — transfers directly to your Astro/Vercel workflow. If Simon can productively build in an unfamiliar framework (SwiftUI), you can apply the same approach to adjacent technologies you haven’t yet invested in learning deeply.

Gemini 3.1 Flash Live: Making audio AI more natural and reliable

Source: Google DeepMind Blog (Tier 1) | Category: models | Relevance: 7/10

Google releases Gemini 3.1 Flash Live with improved real-time audio capabilities across its product ecosystem.

Why this matters: Real-time voice AI is getting meaningfully better and cheaper. If you’re building any kind of conversational or voice-enabled feature for clients, this model is designed to be fast and affordable enough for production use.

So What: Gemini 3.1 Flash Live signals that real-time multimodal (especially audio) is becoming commodity-level. If you build business workflows, voice-driven interfaces and real-time audio processing are now viable features to offer clients without massive infrastructure costs. Evaluate whether any of your current text-based workflows could benefit from a voice layer.

[AINews] H100 prices are melting UP (Latent Space (Tier 1)) — GPU prices are rising again, reversing the expected downward trend as demand for AI compute continues to outstrip supply. When the hardware that powers AI gets more expensive, it eventually trickles down to higher API prices or longer wait times for the services you use. It’s a signal about the economics of the whole AI ecosystem. →
Show HN: Robust LLM extractor for websites in TypeScript (Hacker News AI (Tier 3)) — A TypeScript library that uses LLMs to extract structured data from messy website HTML, replacing brittle CSS selectors with intelligent parsing. If you’ve ever tried scraping websites for data to feed into an AI workflow, you know how fragile traditional scrapers are. This tool uses LLMs to understand page content and pull out structured JSON, so you don’t have to rewrite parsers every time a site changes its layout. →
Quantization from the ground up (Simon Willison (Tier 1)) — A shared educational resource explaining model quantization techniques from first principles. Quantization is how big AI models get shrunk down to run on smaller, cheaper hardware. Understanding it helps you make smarter decisions about which models to use and when local/smaller models might be good enough for your needs. →
The Kitchen Loop: User-Spec-Driven Development for a Self-Evolving Codebase (arXiv cs.AI (Tier 3)) — A research paper exploring a development loop where user specifications drive an AI system that continuously evolves its own codebase. This touches on the idea of AI agents that can maintain and improve code on their own based on what users ask for — which is the direction tools like Claude Code are heading. It’s still academic, but the concept of spec-driven self-evolving code is worth tracking. →
datasette-showboat 0.1a2 (Simon Willison (Tier 1)) — Simon Willison releases an early alpha of datasette-showboat, a new Datasette plugin. Datasette is a handy tool for exploring and publishing data. This is an early-stage plugin release — worth knowing about if you use Datasette, but not broadly impactful yet. →
Natural-Language Agent Harnesses (arXiv cs.AI (Tier 3)) — Research paper exploring frameworks for constraining and guiding AI agents using natural language specifications. As AI agents get more autonomous, figuring out how to keep them on track without rigid code is a real challenge. This paper explores using plain language rules as guardrails, which could eventually shape how agent frameworks work. →
STADLER reshapes knowledge work at a 230-year-old company (OpenAI Blog (Tier 1)) — OpenAI showcases how STADLER deployed ChatGPT across 650 employees to transform knowledge work at a traditional manufacturing company. It’s a case study of AI adoption in an old-school industry. Interesting if you’re pitching AI workflow solutions to non-tech companies, but it’s marketing content with limited technical depth. →

📚 5 new items added to your learning queue →

Signal Scan

Items scanned: 34
Sources checked: 6
High relevance (7+): 5
Generated: 2026-03-28T11:35:34.323Z