AI AgentsSpeculative Agent Execution: Running Ahead of the Decision
Agents spend most of their wall-clock time waiting. Speculative execution starts likely next steps early and discards the wrong guesses. Here is when it pays.
ReadNotes on models, agents, and what inference costs
Practical writing for developers building with large language models — how they work, how to pick one, and how to keep the bill predictable.
AI Agents
FundamentalsSpeculative Decoding in Practice: When It Pays Off
Speculative decoding is standard in serving stacks now, but the speedup is workload-dependent. What decides whether it helps you, and how to tell.
Read
FundamentalsStop Sequences Explained: Ending Generation on Purpose
How stop sequences work, why they interact badly with tokenisation and streaming, and the finish_reason check most integrations forget to make.
Read
GuidesStreaming an OpenAI-Compatible API in TypeScript
Consuming SSE from a chat completions endpoint in TypeScript: reading the body stream, buffering partial frames, assembling tool-call deltas and cancelling cleanly.
Read
GuidesStreaming in Python: Reading SSE From an LLM API
Consume an OpenAI-compatible token stream in Python — the raw wire format, parsing data lines, partial chunks, tool-call deltas, timeouts and what buffering breaks.
Read
AI AgentsSummarisation in Agent Loops: Compressing Without Forgetting
Long sessions have to shed context. What to summarise, what must survive untouched, when to trigger it, and how to tell a good summary from a lossy one.
Read
ModelsSWE-bench Explained: What the Percentage Actually Measures
SWE-bench scores get quoted constantly and understood rarely. What the benchmark does, why its variants are not comparable, and what it cannot tell you.
Read
ModelsSWE-bench Leaders Compared: Reading the Scores Correctly
Verified and Pro are different benchmarks and the scores are not comparable. Who leads each variant in 2026, and what the numbers do and do not tell you.
ReadFundamentalsSynthetic Training Data: What It Does and Where It Fails
How generated data is used to train modern models, what distillation and self-instruct pipelines look like, and what the model-collapse evidence actually supports.
Read
ModelsTerminal-Bench Explained: Scoring Agents That Run Commands
Terminal-Bench measures whether a model can be left alone in a shell and still finish the job. What it runs, what it misses, and how to read a score.
Read
ModelsTerminal-Bench Leaders: The Benchmark That Predicts Agents
Terminal-Bench puts a model in a real shell and scores what happens. Who leads version 2.1, why the ordering differs from SWE-bench, and how to read it.
Read
FundamentalsTest-Time Compute: Buying Accuracy With Tokens Instead of Size
Spending more computation at inference can substitute for a larger model. How the trade works, where it pays off, and what it does to your latency budget.
ReadShowing 565–576 of 731 articles