Happy Monday! ☀️
Welcome to the 176 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
Separating signal from noise in coding evaluations matters
Meerkat: an experiment in global consensus building
AI code reviewer trust at DoorDash
Understanding is the new bottleneck in software
Some new agentic patterns emerging now
🗞️ Tech and AI Trends
New method keeps kids safe from illegal AI content
Grok 4.5 launches with major improvements
Apple sues OpenAI over trade secret theft
👨🏻💻 Coding Tip
React Server Components with Suspense streamline data-fetching, reducing time-to-interactive significantly
Time-to-digest: 5 minutes
DoorDash built a unified memory platform that extracts semantic understanding from behavioral signals and serves it to every personalization surface—from traditional ML models to LLM-driven experiences. Instead of treating user data as just a feature engineering problem, they treat it as a semantic extraction problem.
The challenge: Traditional engagement-based models can't capture the why behind purchases. An embedding knows users buy kale and almond milk, but it can't express "plant-forward health-conscious" in a way that generalizes across categories or LLM systems can reason about.
Implementation highlights:
Three-layer memory architecture: Long-term memory (durable preferences), in-session context (real-time intent), and explicit context (user-stated preferences) operate on different timescales
LLM-synthesized memory blocks: Use LLMs to distill behavioral patterns into semantic narratives—"Strong affinity for organic produce; price-conscious on packaged goods"—that both ML and language models consume
Versioned manifests for decoupling: Separate memory generation from encoding via manifests. Deploy new component versions to 10% of users, revert if quality drops. Reconstruction of any consumer's memory at any historical date becomes trivial
Asymmetric dense embeddings: Map consumer memory and item descriptions into shared semantic space using block-specific retrieval instructions. Memory blocks act as high-level query expansion pulling items matching semantics users haven't purchased
Consumer context graph: Build heterogeneous graph with consumers, brands, taxonomies, and semantic concepts as nodes. Keywords extracted from memories form semantic bridges enabling multi-hop reasoning—"prefers X → implies Y → connects to untried categories"
Results and learnings:
Semantic matching primitive unlocked: Memory blocks work for ML systems too, not just LLMs—many personalization problems are fundamentally about aligning consumer intent with catalog semantics
Multiple encodings beat single representation: Dense embeddings capture semantic similarity; graphs capture relational structure. Together they capture more signal than either alone
Versioning prevents cascading chaos: Full lineage (model ID, prompt hash, response hash, timestamp) means unexpected downstream behavior becomes traceable. You can reconstruct historical states and A/B test across generations
DoorDash's move from "users click X, Y, Z" to "health-conscious, organic-preferring, premium-seeking" shows that scale isn't incompatible with semantic understanding. Their layered architecture proves you can decouple how you generate memory from how you encode it—each improving independently on different timelines.
ARTICLE (engineers go brrr)
The software engineering war
ARTICLE (shirt code chaos)
Decoding the obfuscated bash script on a Uniqlo t-shirt
ESSENTIAL (yeet yourself into jobs)
Advice to new grads: take any job you can get in the industry
ESSENTIAL (brains are the new problem)
Understanding is the new bottleneck
ARTICLE (robots go think-think)
Some new agentic patterns
ARTICLE (dumb but wise)
Be the Idiot
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: Chinese AI companies like DeepSeek and Z.ai are rapidly gaining market share among U.S. businesses, with their models now accounting for up to 46% of tokens used on developer platforms, driven by significantly lower costs (60-90% cheaper) while maintaining near-frontier performance as American labs raise prices on their most advanced systems.
Brief: Thomas Dohmke, ex-GitHub CEO, debuts Entire.io, a decentralized Git hosting network built to handle AI agent traffic that's overwhelming GitHub's infrastructure, offering agent auditing, code mirroring, and superior performance (2.1M pushes/hour vs. Cursor Origin's 81K).
Brief: Apple is in talks with PrismML about running larger AI models on iPhones, with the startup successfully shrinking Alibaba's 27-billion-parameter Qwen model to fit on iPhone 17 Pro—bigger than Apple's current 20-billion-parameter on-device model—which could enable more Apple Intelligence features to run locally and reduce reliance on cloud servers.
Brief: SpaceXAI launches Grok 4.5, an Opus-class AI model with twice the token efficiency of competitors and significantly lower costs ($2/$6 per million tokens vs. Opus's $5/$25), positioning itself as a faster, cheaper alternative just as OpenAI prepares to release GPT 5.6 this week.
Brief: OpenAI rolls out GPT-5.6 with three tiers—Sol (flagship), Terra (affordable), and Luna (fastest)—featuring multi-agent coordination, improved artifact generation for work outputs, and tiered pricing starting at $1 input per million tokens, now available globally across ChatGPT, ChatGPT Work, Codex, and the API.
Brief: OpenAI launches GPT-Live, a new voice mode model that delegates complex tasks like web search and deeper reasoning to GPT-5.5 while maintaining conversation flow, finally replacing the outdated GPT-4o era model that limited ChatGPT's usefulness as a brainstorming partner.
This week’s tip:
React Server Components with Suspense boundaries enable server-side rendering of expensive data-fetching components while streaming HTML and deferring client hydration. Use Suspense to wrap server components that fetch data and fallback UI shows while the Promise resolves server-side, reducing time-to-interactive.
Wen?
Large data-driven feeds (e-commerce, news): fetching 100+ items server-side and streaming HTML chunks avoids waterfall requests and improves CLS and LCP compared to client-side fetch-then-render.
Personalized pages with authenticated data: server components can safely access database/session context directly without exposing secrets through API routes or client bundles.
Mixed static+dynamic layouts: Suspense fallbacks allow fast static-shell rendering (header, nav) while deferred RSC chunks load personalized content below-the-fold asynchronously.
It's okay to be discouraged. It's not okay to quit.
Ryan Holiday
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).