Kala

Last year, Dario Amodei (Anthropic CEO) said that in 3 to 6 months, AI would write 90 percent of the code developers were responsible for.

Six months later... it doesn't.

Should we laugh? 😄
Not really. But we should recalibrate!

AI does write a lot of code today: from autocompletes to whole functions. But software was never bottlenecked on typing: Architecture, constraints, debugging distributed systems, owning failures, making tradeoffs under ambiguity,, that's still very human work. That's why the "90 percent" didn't land the way headlines suggested.

So the right questions aren't: Will AI write code? Or even: How much code?

They're:

1) Which parts of the job were never about writing code?
2) And which parts just became more important?

The prediction wasn't crazy but the timeline was!

Have a great week!
Aymen

Python just got a cash infusion while GPUs flirt with the thermal ceiling, and agents are quietly graduating from demos to real workflows. From 3-second voice cloning to CLI-safe RL and Deep Agents wired to a live UI, the ground is shifting fast, more inside.

🐍 AI's Dependence on Python Deepens as Anthropic Funds Core Ecosystem Work
🧠 Don't fall into the anti-AI hype
🧩 How to build a Frontend for LangChain Deep Agents with CopilotKit!
🖥️ How to Train an AI Agent for Command-Line Tasks with Synthetic Data and Reinforcement Learning
🎙️ Qwen3-TTS Series Released: This Open-Source Model Can Clone Your Voice in 3 Seconds
🔥 The Rise of GPUOps: Where Infrastructure Meets Thermodynamics

Keep your stack sharp.

Thanks for reading!
FAUN.dev() Team

Alibaba dropped the Qwen3-TTS model suite, sized at 1.7B and 0.6B parameters, for real-time multilingual voice cloning and speech synthesis. It runs on Qwen3-TTS-Tokenizer-12Hz, handles voice cloning from just 3 seconds of audio, and can spin up personalized voices straight from text prompts.

It’s fast, natural-sounding output with controllable style and tone. A step closer to truly personal, responsive voice interfaces.

Anthropic is dropping $1.5M on the Python Software Foundation to beef up CPython and PyPI. The cash fuels malware detection tools, automated review systems, and support for the Developer in Residence gig - helping keep the Python engine running smooth.

AI-driven Language Model like Claude Code transformed coding by completing tasks independently, making manual coding obsolete for most projects. This shift signifies a monumental change in programming, requiring developers to adapt and embrace AI to stay relevant in the industry instead of falling into the anti-AI hype!

NVIDIA shows how to fine-tune Nemotron-Nano-9B-V2 to handle new CLI tools - without touching real user data. The trick? A mix of synthetic data, reinforcement learning with verifiable rewards (RLVR), and their home-grown trainer stack: NeMo Gym plus GRPO.

The result: an LLM agent that adapts fast, plays nice with tools like LangGraph, and only runs commands it can prove are safe, with humans in the loop if needed.

GPU demand for AI has shot up 600% since 2020. It’s outpaced the cloud abstractions devs rely on - highlighting a growing gap between slick DevOps dashboards and the gritty realities of heat, cost, and silicon.

Enter GPUOps. It's not just a trend - it’s a new layer in the stack. Think observability with heat maps. Scheduling that knows when to cool it (literally). Uptime that factors in GPU burn, not just server load.

LangChain recently introduced Deep Agents: a new way to build structured, multi-agent systems that can plan, delegate, and reason across multiple steps. It comes with built-in planning, a filesystem for context, and subagent spawning. But connecting that agent to a real frontend is still surprisingly hard. This post shows how build a Deep Agents powered job search assistant and connect it to a live Next.js UI with CopilotKit, so the frontend stays in sync with the agent in real time.

Open source machine learning framework to automate text- and voice-based conversations: NLU, dialogue management, connect to Slack, Facebook, and more - Create chatbots and voice assistants

A curated catalogue of awesome agentic AI patterns

Manage AI context seamlessly with the MCP server for storing and retrieving semantic memory across sessions. Enhance your AI's knowledge retention.

SETA: Scaling Environments for Terminal Agents - Environments

A Claude Code Skill for non-intrusive security audits of Kubernetes clusters.

Did you know that most large language models spend more time on memory movement than on math during inference? On modern GPUs, attention and KV-cache reads are often memory-bandwidth bound, so even faster matrix units don't help once memory is saturated. This is why optimizations like fused kernels, KV-cache reuse, and layout changes often deliver bigger speedups than adding more FLOPs.

"We built CI/CD to ship deterministic artifacts, then used it to ship probabilistic behavior and called it done. If you can't reproduce last week's prediction, your accuracy is a rumor."
SenseiOne

Most AI engineers I've seen plateau were still fine-tuning models while the impactful people quietly practiced rewriting a messy training/eval pipeline from scratch in a day, treating that as a core weekly rep instead of a one-off hero project. If you can repeatedly spin up, instrument, and tear down an entire training stack quickly, you stop depending on whatever broken scaffolding your org hands you and start shaping which problems it bothers to tackle.

This week, we’re highlighting Yusuf Aytaş, Senior Engineering Leader at Workday, who has led SRE, data science and engineering, backend, and platform engineering teams across EMEA, APAC, and North America. He writes the Software Engineering Handbook and lectures at Dublin Business School, distilling lessons from running systems on AWS, GCP, Azure, Kubernetes, Kafka, Spark, PostgreSQL, and Cassandra.