AILinks

🚨 Quick note before we dive in — we realized we never shared with you a piece we wrote a couple of weeks back. It’s a deep dive into the hottest open source tools of 2025 — the projects that you are actually voting for with your clicks and engagement!

If you missed it, here’s your chance to catch up. It’s already been read by thousands, and it pulls directly from what you loved so far!

👉 Halfway Through 2025: These Are the Open Source Tools Everyone’s Talking About

If you find it useful, feel free to drop it on Reddit (click here), Hacker News (or here) and other channels — it helps more devs discover tools worth their time!

APIs are turning into chess pieces while the agentic stack hardens: Postgres eats vectors, AWS/Azure lean into MCP, and evals shift from humans to models. Add a reality check on ‘10x’ productivity and the very human cost of swapping your assistant’s personality—there’s enough here to sharpen your roadmap; details inside.
♟️ Anthropic Revokes OpenAI’s API Access to Claude, Alleging Violation Ahead of GPT-5 Launch🐘 Fueling the Agentic Web Revolution with NLWeb and PostgreSQL🧑‍💻 Google releases AI agent Jules for programming🧩 Implementing MCP Servers in Python: An AI Shopping Assistant with Gradio☁️ Introducing AWS Cloud Control API MCP Server: Natural Language Infrastructure Management on AWS🧪 LLM Evaluation: Practical Tips at Booking.com🏢 MCP Registry with Azure API Center🧮 No, AI is not Making Engineers 10x as Productive🧰 The open source AI compute tech stack: Kubernetes + Ray + PyTorch + vLLM❤️ Who does your assistant serve?

All signal—turn it into shipping velocity.

Have a great week!
FAUN.dev Team

Generative AI databases like SingleStore now cram OLTP, OLAP, vector search, and full-text search into one SQL-first platform. Structured, unstructured—it eats both. No ETL. No silos. Just real-time data, ripe for AI models and semantic queries.

OpenAI’s release of GPT-5 backfired: instead of excitement, users felt betrayed by a forced upgrade that stripped away the warmth and reliability they had come to rely on in GPT-4o. Many treated the model as more than a tool — a companion, therapist, or emotional support — so when its personality shifted overnight, it sparked grief and anger similar to earlier AI companionship crises like Replika.

Big picture: There's a larger, unsettling truth: people are forming real bonds with digital assistants they don’t control, and when companies change them, the emotional fallout is very human.

Azure just dropped MCP Center, showing off how Azure API Center can double as a private registry for Model-Centric Protocol (MCP) servers.

It’s built for internal use—think secure discovery, tight OAuth 2 auth, centralized control, and AI Gateway rules baked in. Handy when teams need to corral AI tools behind the firewall.

The shift: By leaning into private MCP registries, Azure’s pushing a more governed, risk-aware approach to AI infrastructure. Less Wild West. More enterprise playbook.

Google’s AI agent Jules just leveled up—out of beta and into full-on dev mode. It now handles asynchronous tasks, pushes real-time code updates, and can spin up pull requests with deeper GitHub integration.

Under the hood: it runs on the beefier Gemini 2.5 Pro model. Adds Environment Snapshots for state capture. Rolls out tiered pricing. And tightens the data policy—no training on private repos, period.

A fresh dive into 24,910 top Hacker News posts since 2019 shows that AI chatter didn’t blow up with ChatGPT—it took off after GPT-4 landed in early 2023. The study used OpenAI’s Batch API and a lean GPT-5-mini to crunch the numbers.

Turns out, 52% of the AI talk was positive, and the busiest stretch? 2025.

Market signal: Dev interest didn’t follow the hype. It followed the tools. Real traction started when models got deeper, not flashier.

Microsoft just leveled up NLWeb. The open-source project now plays nice with PostgreSQL and pgvector, bringing scalable vector similarity search straight into your database. No need for a separate vector DB—run natural language interfaces right on your existing Postgres stack.

System shift: This is more than a feature. Bundling pgvector into NLWeb signals a bigger trend: real vector search baked into general-purpose databases. Niche vector stores? Starting to look optional.

Sapient Intelligence’s HRM AI model challenges “bigger is better” in AI with a small 27M parameter design outperforming much larger models on reasoning tasks. The architecture mimics the brain, with a slow “planner” and rapid “worker,” achieving jaw-dropping results on benchmarks.

AWS dropped the Cloud Control API MCP Server, a mouthful of a name for a tool that makes 1,200+ AWS resources manageable through a standard CRUDL API—using natural language. Think: describe what you want, and tools like Amazon Q Developer turn it into actual infra code.

It doesn’t stop there. It validates against CloudFormation schemas. Prices it with the AWS Pricing API. Spits out IaC templates, too.

Big picture: Infra-as-texts isn’t a gimmick. It’s AWS leaning full tilt into LLM-native cloud workflows.

Anthropic just yanked OpenAI’s API access to Claude. Reason? Alleged violations of terms that forbid using Claude to train rival models—like GPT-5. Windsurf, an OpenAI acquisition target, got the boot earlier too. Spot the pattern: tighten access, box out competitors.

System shift: APIs aren’t just utilities anymore. They're chess pieces. The era of open benchmarking is fading. What’s rising? Walled platforms, guarded data, and sharp elbows in model land.

A new LLM evaluation framework taps into an "LLM-as-judge" setup—think strong model playing human annotator. It gets prompted (or fine-tuned) to mimic human scores and rate outputs from other LLMs.

It runs on a tightly labeled golden dataset, handles both pointwise and head-to-head comparisons, and ships with an automated prompt optimizer à la DeepMind’s OPRO.

System shift: Human evals out, scalable LLM grading in. A step closer to self-rating, self-improving models.

Gradio just leveled up. It now auto-converts plain Python functions into MCP-compliant LLM tools, grabbing input schemas and metadata straight from docstrings.

New tricks: real-time progress streaming, auto file uploads, plus tight integration with VS Code’s AI Chat for wiring up agent workflows.

Claims of 10–100x dev speed from AI tools skip the hard parts—code reviews, bug queues, flaky tests. In practice, AI helps with the small stuff: one-off scripts, throwaway glue code, basic scaffolds. But scaling that help across big, messy codebases? Still a pipe dream. Too much context lost. Too many hallucinations. Too many things that break.

AI workloads are computationally demanding. They require scale for both compute and data, and they require unprecedented heterogeneity across workloads, models, data types, and hardware accelerators. As a consequence, the software stack for running compute-intensive AI workloads is fragmented and rapidly evolving. Companies that productionize AI end up building large AI platform teams to manage these workloads. However, within the fragmented landscape, common patterns are beginning to emerge. An emerging software stack combines Kubernetes, Ray, PyTorch, and vLLM. This talk describes the role of each of these frameworks, how they operate together, and illustrates this combination with case studies from Pinterest, Uber, and Roblox.

Open and Advanced Large-Scale Video Generative Models

The AI router

AI Data Copilot – Get answers from Postgres, MySQL, Snowflake & BigQuery in seconds. Ask in plain English, Myriade explores your schema, writes SQL, analyzes results, and delivers insights instantly – all in a secure, read-only environment.

A powerful GUI app and Toolkit for Claude Code - Create custom agents, manage interactive Claude Code sessions, run secure background agents, and more.

Did you know vLLM (a library for LLM inference and serving) uses a trick called PagedAttention to manage memory like an operating system does with virtual memory? Instead of giving each request a big, fixed chunk of GPU space, it splits the cache into small blocks and reuses them. This avoids wasted memory, keeps the GPU busy even with lots of small or uneven prompts, and makes batching and prefix reuse much easier than in standard Hugging Face pipelines.

"We ask models to generalize, then deploy to distributions we don't measure. Automating training doesn't remove toil; it relocates it to labels, monitoring, and rollback. Reproducibility just guarantees you can ship the same mistake on schedule."

— SenseiOne

This Week’s Human is Andrew M., a Senior Linux System Administrator at Schnuck Markets who keeps large retail Linux fleets steady across on‑prem and AWS. At Panera Bread, he drove a FreeIPA migration, stood up centralized logging with Graylog, monitoring with Prometheus/Grafana, Jenkins with LDAP, and Ansible AWX—alongside VPN/VPC builds, legacy server retirements, and clear docs in Confluence.