Weekly Review — July 6-12, 2026

This is the Weekly Review, a Sunday digest of everything that went up on the blog this week, plus a short list of things I read but didn’t write about. If you only have ten minutes on a Sunday, this is the one to read.

This Week on the Blog

Compression Is Intelligence

Information theory ties prediction and compression together — pair a probabilistic next-token model with an entropy coder and better prediction is, bit for bit, better compression — and that gives the cleanest mental model I know for a language model: it is a compressor of its training text, and predicting the next token well is the same achievement as compressing that text well. The framing traces to Marcus Hutter’s compression prize, and a 2024 paper found that how well a model compresses tracks its capability almost linearly. The practical payoff sits one level up: if the model is a fixed compressor you rent by the token (Bedrock’s on-demand pricing), then context is the real lever, and Bedrock prompt caching is the operational version of assembling that context once and re-feeding it cheaply — a cached prompt prefix skips recomputation on later calls and its tokens bill at a reduced rate, while the model itself and its learned priors stay untouched. This is Part 2 of the “Whiteboard to Cloud” series.

How My Setup Learns While You Sleep

The thing that lets an agent get better over time is a loop, and the loop is the same whether you buy it or build it: capture the raw interaction, consolidate it on a schedule, and keep the distilled lesson while the transcript fades — whether the next run is actually better is something you verify, since a bad lesson persists as faithfully as a good one. Amazon Bedrock AgentCore Memory is the managed version; my own kiro-cli and Obsidian setup is the do-it-yourself version, and it ran this exact loop on me at the start of the session where I wrote the post. The choice between them is less about trust than about ownership and operational trade-offs: the managed service is the natural reach when you are shipping an agent to users and want the storage, scaling, and operations handled; the DIY path when you want to read every file and diff every change. This closes the three-part agent-memory series.

Post-Quantum Crypto Is Here, What It Means on AWS

Post-quantum cryptography stopped being a research-paper topic. The reason to care today is harvest-now-decrypt-later: an adversary records your encrypted traffic now and decrypts it once a large quantum computer exists, so anything captured today under quantum-vulnerable public-key protection, with a secrecy lifetime longer than the wait for a cryptographically relevant quantum computer, is already at risk. In August 2024 NIST finalized the first standards (ML-KEM, ML-DSA, SLH-DSA — FIPS 203, 204, and 205), and on AWS the practical news is that AWS KMS already supports hybrid post-quantum TLS for connections to its API endpoints and ML-DSA key specs for signing and verification. For data at rest, the symmetric encryption KMS uses — AES-256 — is regarded as resistant to known quantum attacks; the quantum-vulnerable pieces are the RSA- and ECC-based public-key schemes, which is where the migration effort concentrates. The move that ages well is not a rushed migration but crypto-agility: know where your long-lived secrets are, and be able to swap algorithms without re-architecting. This is Part 3 of the “Whiteboard to Cloud” series.

The Thread This Week

Read together, the week is one argument: a clean mental model is worth little until you know which concrete AWS primitive turns it on. “Whiteboard to Cloud” makes the move twice, following compression from information theory to Bedrock prompt caching, and post-quantum crypto from a threat model to AWS KMS key specs. The agent-memory post runs the same pattern from a different angle: name the learn-over-time loop, then show it managed with AgentCore Memory or built from files you own. Theory tells you what is happening under the hood; the AWS building block is the thing you actually switch on.

Further Reading

Things I read this week that didn’t get their own post, all public:

• Sebastian Raschka, LLM researcher and author — Independent researcher behind “Build a Large Language Model From Scratch”, with a steady stream of clear writing on reasoning LLMs and inference-time compute. A good follow if this week’s compression post left you wanting the mechanics underneath.
• Matt Wood on agent-friendly publishing and llms.txt — An AWS VP crossposting his essays to an agent-friendly site with llms.txt, a permissive robots.txt, and full RSS, treating agents as the next distribution channel. A useful companion to the “own your real estate” thread from last week.
• Anthropic’s model-selection and eval framework — Build lightweight custom evals over public benchmarks and optimize for cost per successful outcome rather than leaderboard scores, using the thinking, effort, and caching dials. Practical guidance for anyone choosing a model for real work.
• 20 Laws of Software Engineering (Milan Milanovic) — A tour through the timeless principles (Conway’s, Brooks’, Goodhart’s) that still explain why projects slip, systems rot, and teams slow down. Worth a reread before your next architecture or org-design conversation.
• Harvard Advanced Algorithms CS224, Jelani Nelson — A free 25-lecture course on sketching, streaming, and hashing, the math foundations sitting under a lot of modern AI infrastructure. A solid deep dive if you want the theory behind retrieval and vector systems.

Until Next Sunday

That’s the week: compression as the mental model for what a language model does, an agent-memory loop you can buy or build, and post-quantum crypto that is already turnable-on at the first endpoints that matter — with certificates, clients, and stored ciphertext still on the migration list. The connective tissue is the same each time, moving from the idea on the whiteboard to the AWS primitive that makes it real. Which of your own mental models is still waiting for the concrete building block that would let you act on it?

About the Author

Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.

This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.

Learn more →

• WeeklyReview
• AI
• AWS
• Cryptography
• AgentMemory