The AI Friday Brief
Good Wednesday, NOLA. May 13th brings some solid practical wins: a tiny 26M model that can do tool calling, a rethink of how AI interacts with your screen, and a reality check on how companies are actually using (and abusing) AI tools. Plus, some intriguing open-source projects for making agents more reliable.
Things People Built
Needle: A 26M Model That Distilled Gemini's Tool Calling
Popular on HN. Take a frontier model's tool-calling capability and fit it into something that runs on any device. This is the kind of lean, practical distillation that matters for real products. If you're building agents and need them to actually call APIs reliably without hitting the API cost wall, this is worth a look.
Hacker News
Reimagining the Mouse Pointer for the AI Era
Google DeepMind's take on a deceptively simple problem: how should an AI model interact with your screen? Instead of just clicking random pixels, their pointer learns to understand context, follow visual flows, and make decisions based on what's actually on screen. It's a small detail that could make AI-powered automation feel less brittle.
Google DeepMind
Statewright: Visual State Machines for Reliable AI Agents
Shared on HN. One of the biggest complaints about AI agents is that they're unpredictable—they wander off and do weird things. Statewright lets you visually define the states and transitions your agent should follow, then enforces them. It's a guardrail tool that actually makes sense.
Hacker News
Hopper: An Agent Interface for Mainframes and COBOL
Show HN. One of the oldest, most important systems in the world still runs on COBOL. Hopper gives AI agents a natural-language interface to mainframes. If this works, it could unlock decades of legacy business logic for modern AI workflows—without forcing companies to rewrite everything.
Hacker News
Voker: Analytics and Observability for AI Agents
Launch HN (YC S24). Running agents in production means you need visibility. Voker gives you dashboards, error tracking, and performance metrics for your deployed agents. It's the kind of boring-but-necessary infrastructure that separates hobby projects from real products.
Hacker News
AI in Practice: The Reality Check
Amazon Employees Are "Tokenmaxxing" Due to AI Pressure
Here's what happens when companies mandate AI adoption without clear incentives: employees game the system by stuffing prompts with tokens (unnecessary words) to hit usage targets, making the metrics meaningless. It's a cautionary tale about top-down AI rollouts. Discussion on HN.
Ars Technica
AI Isn't Paying Off the Way Companies Think—Gartner Study
Companies are investing big in AI but seeing mediocre returns. The gap between hype and actual ROI is growing. Worth reading not as doom, but as a reminder that implementation matters more than the model itself. Discussed on HN.
Fortune / Gartner
Google Says Criminal Hackers Used AI to Find a Major Software Flaw
Not a theoretical threat anymore. Real attackers are using AI to find exploitable bugs in open-source code. This is the flip side of "AI finds vulnerabilities"—adversaries have access to the same tools.
The New York Times
New Capabilities & Platforms
Claude Is Now Available on AWS
If you've been holding off on Claude because you wanted it to live in your AWS environment, the option is now here. This is enterprise infrastructure—important for compliance and latency-sensitive workloads. Discussed on HN.
Anthropic
Thinking Machines Builds an AI That Listens While It Talks
Real-time conversation means the model needs to respond to interruptions, subtle tone shifts, and context changes mid-sentence. Thinking Machines' new approach moves closer to how humans actually talk to each other, not just turn-taking.
TechCrunch
Interesting Reads
If AI Writes Your Code, Why Use Python?
A thought experiment that's worth sitting with: if AI can generate the code, does the language choice matter anymore? Or do we pick based on different criteria now—clarity for AI prompting, ease of debugging generated output, framework ecosystem? Challenging the assumptions we've held for years.
Medium
I Let AI Build a Tool to Help Me Figure Out What Was Waking Me Up
A great "I tried it" story of someone using Claude Code to build a real tool that solved a real problem. No fancy architecture, just pragmatic problem-solving with AI. Exactly the kind of workflow we'll see becoming normal.
martin.sh
Also
- Data Centers Are Coming for Rural America — The Androscoggin paper mill in Maine employed 1,500 people. Now it's being transformed into a data center. What's the economic tradeoff?
- E2a: Open-Source Email Gateway for AI Agents — If your agents need to send and receive email, this is a practical starting point.
- OpenAI Launches Daybreak: AI-Powered Vulnerability Detection — Another security tool in the AI arsenal. Finding bugs before bad actors do.
- Unitree GD01: China's $537k Rideable Transformer Robot Is in Production — Not strictly AI, but a useful reminder that the physical world is catching up to the software world's hype cycle.
- CSP Allow-list Experiment — Simon Willison exploring sandboxed iframes and custom fetch() for AI apps. Deep technical but potentially important for secure LLM integration.
- The End of Finetuning — Latent Space's take on whether prompt engineering and in-context learning are making old-school finetuning obsolete.
Pass it on