Compression is prediction

Compression is prediction
19 minutes by Annie Sexton

Compression works by finding patterns in data and encoding them more efficiently, and better predictions about what comes next lead to better compression. Language models like GPT work the same way: they predict the next word based on context, which is exactly what a compression model does. A 2023 DeepMind paper confirmed this link, showing the two are mathematically equivalent. The main reason we don't use LLMs for everyday compression is cost — they are far too large and slow for practical use.

Cloud vector databases were priced for humans. AI agents just broke the model.
sponsored by Actian

Your agents query your vector database thousands of times per session. Cloud providers charge for every one. VectorAI DB runs on your own infrastructure with no per-query pricing and no egress fees. On-premises, at the edge, or air-gapped. Free to start.

Improving heuristics
13 minutes by Amit Patel

Placing "landmark" nodes on a map can make A* run faster without changing the algorithm itself. Before each search, Amit suggests running Dijkstra's from each landmark and storing the costs. Then use the triangle inequality to get a tighter lower bound on path length, replacing the standard distance heuristic. Placing more landmarks near map edges or common paths means fewer nodes are explored, though the best placement depends on your specific game and map layout.

A deep (and fuzzy) dive into search
15 minutes by Andrey Abramov

Fuzzy search helps search engines find results even when words contain typos or extra characters. Andrey explains two approaches used by SereneDB: Levenshtein automata and n-gram similarity. Levenshtein is useful for precise typo matching, while n-grams work better with longer text and partial matches. SereneDB combines these techniques to make fuzzy search practical and flexible.

Email encryption
32 minutes by J. B. Crawford

Email was invented in many places at once, and its security was never a priority. Two main encryption standards emerged: PGP, which is open and decentralized but notoriously hard to use, and S/MIME, which works well inside organizations with shared infrastructure but struggles across them. In practice, most email today is encrypted in transit between servers, which solves many real threats even without full end-to-end encryption. The core problem is that strong encryption requires key distribution, and no universal solution for that has ever taken hold.

A shell exclamation mark is not for yelling
7 minutes by Filip Roséen

Filip introduces shell event designators, a powerful but often overlooked way to reuse and modify previous commands. He explains useful shortcuts such as !!, !$, and !ssh, along with word and path modifiers. Filip also covers the POSIX fc command, safer ways to use these features, and when alternatives like Ctrl-R may be better.

And the most popular article from the last issue was:

• Growing as an engineer in a world of AI