News & guides
Qwen 3.6 vs MiniMax M3: Different Benchmarks, Really
One reports 77.2 on SWE-bench Verified, the other 59.0 on SWE-bench Pro. Those numbers cannot be subtracted, and the real difference is elsewhere.
· 9 min read
Qwen 3.6 vs Qwen 3.5 Coder: General or Specialised
One is a newer general model, the other a code-specialised predecessor. Whether a specialist still beats a stronger generalist at its own job.
· 8 min read
Rate Limit Tiers Explained: RPM, TPM and What Binds First
Provider rate limits are two separate meters and the token one usually binds first. How tiers are assigned, how bursts work, and how to engineer around both.
· 12 min read
Rate Limiting Agent Fleets Without Killing Throughput
One API key, many agents, one shared quota. How to build a fleet-wide limiter that respects token budgets, protects interactive work and survives bursts.
· 9 min read
Redacting Secrets From Prompts Before They Leave Your Machine
Credentials reach model context through error output, config files and git history. Pre-send scanning, why gitignore is no defence, and rotating on suspicion.
· 9 min read
Regression Testing a Model Upgrade Before It Bites
A new model version is a dependency bump with no changelog. How to gate an upgrade, what to diff, and how to roll back when behaviour shifts.
· 9 min read
Repetition Penalties: Why They Break Code Generation
Frequency and presence penalties suppress loops by punishing tokens the model already used. On code that punishes syntax, and the loop was rarely the real bug.
· 9 min read
Replaying Agent Traces: Debugging a Run You Cannot Reproduce
Agent runs rarely reproduce, so the stored trace is the only evidence. What to capture, how to replay it, and where replay stops being faithful.
· 9 min read
Residual Connections: Why Deep Models Train at All
The one-line change that made hundred-layer networks possible, why the residual stream behaves like a shared bus, and what it explains about pruning.
· 8 min read
RLHF Explained: Training on Comparisons Instead of Answers
Reinforcement learning from human feedback shapes the qualities nobody can write down. The three-stage pipeline, and the failure modes you see as a user.
· 9 min read
Rollback Strategies for Agents That Changed Things
An agent stopped halfway through eleven changes. Checkpoints, compensating actions, staging areas and the undo plan you should require before it acts.
· 9 min read
Rotary Embeddings (RoPE) Explained With Clock Hands
RoPE encodes position by rotating vectors rather than adding to them. Why that gives relative distance for free, and how it made 1M context possible.
· 9 min readReady to try it?
Every model behind one key. Start free and upgrade when it sticks.