Happy Monday! ☀️
Welcome to the 95 new hungry minds who have joined us since last Monday!
If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.
📚 Software Engineering Articles
Core dump epidemiology: how OpenAI fixed 18-year-old bug
CORS protection explained: what it actually protects
Fintech engineering handbook: comprehensive guide for builders
P99 0ms autocomplete for 240 million domains
Reddit's anti-spam internals revealed: how they fight abuse
🗞️ Tech and AI Trends
QA teams adapt to AI: high-performers lead the way
Claude Sonnet 5 arrives: Anthropic raises the bar
AI shouldn't cut headcount, it should shrink teams
👨🏻💻 Coding Tip
Master CSS
@layerto prevent Tailwind conflicts and enable predictable style overrides without!importanthacks
Time-to-digest: 5 minutes
Today's voice AI feels real-time, but it's actually a turn-based language model surrounded by helper components simulating conversation. Thinking Machines just blew that approach open by putting interactivity inside the model itself, enabling behavior that seemed impossible before.
The challenge: Turn-based models can't handle continuous input and output simultaneously. Helper components are too simple to enable tasks like "interrupt me when I say something wrong" or "count my pushups in real time"—they lack the intelligence to understand context the language model owns.
Implementation highlights:
Time-aligned micro-turns: Slice conversation into 200-millisecond chunks instead of discrete turns, treating time as the fundamental unit so input and output happen concurrently
Lightweight encoders from scratch: Skip heavy pretrained models like Whisper; train audio and video processing components end-to-end for tight latency budgets
Dual-model coordination: Pair a fast interaction model (responds in 200ms) with a slower background model handling deep reasoning, tool use, and web browsing—both share context seamlessly
Streaming inference optimization: Contribute SGLang streaming sessions to process 200ms chunks efficiently without batching delays
Custom benchmarks for new capabilities: Build TimeSpeak, CueSpeak, and RepCount-A to measure behaviors turn-based systems physically can't do (speaking while listening, proactive interjections, real-time visual tasks)
Results and learnings:
New capability class unlocked: Tasks like live translation, real-time commentary, mid-sentence corrections, and video-triggered responses work because the architecture itself changed, not because latency improved
Continuous interaction possible: The model genuinely speaks while listening, watches while speaking, and interjects mid-sentence based on visual context—all native behaviors, not hacks
Old benchmarks miss it: Existing voice AI metrics can't measure what matters here; the breakthrough is qualitative, not quantitative latency gains
Thinking Machines just proved that scaffolding around dumb models hits a ceiling. The real lesson: when external helpers become your bottleneck, move the capability inside the system. This applies way beyond AI—it's a pattern in operating systems, web browsers, and any layered architecture.
ARTICLE (youtube oopsie-daisy)
Leaking YouTube creators' private videos
ARTICLE (gene-bench-flex)
Inside Genebench-Pro
ARTICLE (ai-jobs-go-brrr)
Mapping Europe's AI Workforce Opportunity
ARTICLE (chatgpt-everywhere-now)
How ChatGPT adoption has expanded
ESSENTIAL (big-tech-censorship-shenanigans)
Pollen tried to remove my article and Google is assisting with it
ESSENTIAL (pizza-teams-are-back-baby)
A return to two-pizza culture
ARTICLE (obvious-talk-hits-different)
Saying the obvious thing
Want to reach 200,000+ engineers?
Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.
Brief: Apple released emergency security updates for iOS, iPadOS, macOS, and Safari, patching 30 WebKit vulnerabilities including four discovered using AI tools like Claude and Codex, and notably broke its usual practice by shipping standalone patches ahead of schedule to outpace attackers who can now weaponize exploits faster with AI.
Brief: Expo launches SDK 57, upgrading React Native to 0.86 with no breaking changes, and explores a new release cadence that delivers non-breaking updates as optional upgrades between major releases, making app updates as simple as running npx expo install expo@latest --fix.
Brief: Anthropic releases Claude Sonnet 5, a more autonomous AI model that matches Opus 4.8 performance on reasoning, coding, and tool use while costing significantly less, now available across all plans with introductory pricing at $2 per million input tokens through August 2026.
Brief: RevenueCat argues that AI shifts the balance of team dynamics by making code writing and review cheaper, enabling smaller, more agile product teams with clear ownership and faster decision-making—not fewer hires, but more parallelism and output per person.
Brief: As AI struggles with global code understanding while excelling at local functions, programming languages like Rust demonstrate how local reasoning about data races can enforce surprising global properties—suggesting future languages may need to help both humans and AI write more reliable software by building guarantees into language design itself.
This week’s tip:
Use CSS @layer with strategic specificity to decouple vendor utilities from component styles, preventing Tailwind overrides and enabling scoped design system tweaks. Layers establish a cascade order independent of selector weight, making overrides predictable.
Wen?
Third-party widget integration: Wrap external component styles in a lower layer so your design tokens override without !important.
Design system versioning: Move deprecated utilities to a deprecated layer and conditionally load it, enabling gradual migration.
Framework style sandboxing: Isolate component library resets in @layer reset so host app utilities don't conflict unexpectedly.
Work joyfully and peacefully, knowing that right thoughts and right efforts inevitably bring about right results.
James Allen
That’s it for today! ☀️
Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!
If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦
Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.
Icons by Icons8.
*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).