Pierre Mouchan

I publish daily tech news, follow me for more ;)
Sources: engineering(.)fb(.)com

Open Source Power 🛠️

Meta isn't just a consumer; they are one of the largest contributors to the FFmpeg project.

• Custom filters for massive scale
• Optimized for MSVP (Meta Scalable Video Processor)
• Standardized media APIs across all apps

Meta’s Media Processing 🎥

They’ve built a unified service to handle everything from Instagram Reels to VR content.

→ Core engine: FFmpeg
→ Scale: Billions of uploads daily
→ Goal: Efficiency across diverse hardware (CPUs & ASICs)

If you missed yesterday' news, Meta just revealed how they process billions of videos daily.

The secret? A massive, custom implementation of FFmpeg.
🧵

I publish daily tech news, follow me for more ;)
Sources: Karpathy, Unsloth, RustWasm, AgentSafehouse(.)app

Rust + Wasm 🦀

The "Rust and WebAssembly" book is the new go-to for high-performance web code.
• 10-part tutorial.
• Learn to publish Rust to npm.

→ Move heavy logic off the JS main thread and into compiled Rust.

Agent Safehouse 🛡️

A new macOS-native sandboxing tool for AI agents.

• Secure execution environment.
• Prevents agents from accessing sensitive system files.
• Native performance.

→ Essential for anyone running autonomous agents locally.

Qwen 3.5 Local 🤖

Unsloth released the definitive guide for running Qwen 3.5 on local hardware.

→ Local LLM supremacy is here.
→ Get high-performance inference without the API latency or privacy risks.

Karpathy’s Autoresearch 🧪

AI agents now autonomously optimize LLMs.
• 5-min training budget per run.
• 12 experiments per hour.
• ~100 experiments while you sleep.

→ The agent edits train(.)py while you guide via a simple markdown file.

If you missed todays' news, AI is now conducting its own research while you sleep.

From Karpathy’s autonomous labs to local Qwen 3.5 deployments, the dev landscape is moving fast.
🧵

I publish daily tech news, follow me for more ;)

Sources: cacm.acm(.)org, ki-editor(.)org, techdirt(.)com, github(.)com, will-keleher(.)com, helix-editor(.)com

How to prove you're human in 2026? Use CSS.

One dev is using intentional typos and font hacks to signal "humanness."
• Swapping "discrete" for "discreet."
• Modifying font glyphs.

→ We are now engineering "human error" to bypass AI filters.

Text editors are moving beyond "lines of text."

Ki Editor and Helix are pushing structural editing.
• Operate directly on the AST (Syntax Tree).
• Multi-cursor support for syntax nodes.

→ Helix (Rust-based) offers IDE power with zero config.

Go is finally getting a native UUID package.

Proposal #62026 is "Likely Accept" for the standard library.
• Native crypto/uuid support.
• Supports RFC 9562 (including UUIDv7).

→ No more third-party libs for basic unique identifiers.

We are training students to write worse.

AI detectors flag words like "devoid" as "robotic," forcing students to dumb down prose to pass.

→ The "Cobra Effect": Students now use AI defensively to ensure their original work doesn't look "too perfect."

Docker is a decade old and the numbers are staggering:

• 14M+ images on Docker Hub.
• 11B+ image pulls every month.
• 3.4M+ Dockerfiles on GitHub.

→ Using a 1990s dial-up tool (SLIRP) reduced networking bugs by 99% during the early days.

If you missed today's news, the tech world is fighting back against AI detectors while Docker hits massive milestones.

From "humanizing" CSS to Go finally getting native UUIDs, here is the breakdown.
🧵

I publish daily tech news, follow me for more ;)
Sources: Business Insider, HN, Apidog, PaulGraham(.)com

Paul Graham on "The Brand Age" ✍️
PG warns that tech is shifting from "Product Quality" to "Brand Identity."
• Comparing AWS/Kubernetes bloat to the Swiss Watch industry.
• Once tech commoditizes performance, brand is the only moat left.

OBLITERATUS: The Safety Jailbreak 🔓
A new open-source toolkit removes refusal behaviors from LLMs without retraining.
• Uses "abliteration" to surgically edit neural patterns.
• Drops refusal rates from 80% to <10% in 30 mins.
→ Local models just got uncensored.

We are all AI Engineers now 🤖
The top story on HN argues AI hasn't lowered the barrier—it created a K-shaped workforce.
• Top engineers are 10x more productive.
• Juniors/Mid-level devs are being sidelined.
→ "AI is better than 50% of engineers."

The Great Tech Recession 📉

New data shows tech job losses in Feb 2026 outpaced both the 2008 and 2020 recessions.
• 92k jobs lost in one month.
• Worst since the Dot-com bust.
→ AI is now statistically visible in unemployment data. It's not just "efficiency" anymore.

If you missed yesterdays' news...

the "AI replaces devs" debate just got real numbers attached to it.

From record-breaking job losses to open-source "obliteration" of safety rails, here is the breakdown. 🧵

I publish daily tech news, follow me for more ;)

Sources: OpenAI, Jido(.)run, Alibaba, Netflix, Github

Reading minds with fMRI 👁️

New open datasets are mapping brain data to visual reconstruction.

The Natural Scenes Dataset (NSD) uses 7T fMRI to capture 73,000 images from the human brain.

→ We are getting closer to "generative" brain-to-image tech.

Netflix optimizes Java 🚀

Using the JDK Vector API, Netflix engineers achieved a 9x performance boost for recommendation scoring on Intel chips.

• 3x speedup on ARM
• 15% reduction in total CPU usage

→ SIMD operations are a game changer for JVM.

Real-time AI speech on Mac 🎙️

Nvidia’s PersonaPlex 7B now runs full-duplex on Apple Silicon via Swift.

• 16.7GB → 5.3GB (4-bit quant)
• 68ms processing speed
• Listen and speak simultaneously

→ Local, low-latency voice AI is finally viable.

Agents are going native 🛠️

PageAgent: An MIT-licensed GUI agent that lives in your web app via a single <script> tag. No Python or servers needed.

Jido 2.0: A massive Elixir framework for autonomous agents using immutable data and 665+ models.

GPT-5.4 is here 🧠

OpenAI's latest model features a 1M context window and native computer-use.

• 83% success on GDPval (beats humans)
• 33% fewer false claims
• New "thinking mode" you can interrupt mid-task

→ It's a professional-grade agent.

If you missed todays' news, the AI landscape just hit a new gear.

From GPT-5.4 outperforming human pros to Netflix slashing CPU costs by 15%, here is the breakdown.
🧵