A daily AI snack: research, tools and the engineering economics behind them, in one sharp read.
Here are the most recent issues, each with the hook, the argument, and the one number worth remembering.
“Everyone assumes the hard part of an AI SRE is wiring an LLM to Grafana, Datadog and Kubernetes. OpenSRE has done that wiring for 60+ tools.”
The integrations are the easy part. What OpenSRE lacks is a set of scored incident scenarios to prove its diagnoses are right, and research suggests root-cause agents fail on architecture, not prompts: frontier models get only 3.9 to 12.5% of real incidents perfectly right.
The surprising bit: 14 of its 20 shipped workflows target CI and GitHub work. Only one targets incident investigation.
“Zonos lets you dial in happiness, fear, pitch variation and speaking rate for a cloned voice. Every second of that speech also requires streaming the full model through memory 86 times, twice.”
Zonos v0.1 is a 1.62B parameter codec language model that clones a voice from a short sample and renders expressive 44 kHz speech. Its real ceiling is memory bandwidth, not compute, which keeps it on high-end GPUs.
The surprising bit: Real-time playback needs about 279 GB/s of weight reads, roughly 65x Pocket TTS. A laptop CPU manages 0.03x real time.
“One of the most downloaded open text to speech models never predicts a token.”
Kokoro reaches broadcast-quality speech with 82M parameters by predicting explicit phoneme durations and rendering the whole utterance in parallel. Its decoder holds 65% of the weights and 89 to 98% of the runtime, so better small TTS is a vocoder problem, not an NLP one.
The surprising bit: The 12-layer PL-BERT text encoder uses only 1 to 5% of CPU runtime.
“The most valuable thing in DeepSeek's inference stack is not a kernel.”
DeepSeek open-sourced its production inference engine under Apache 2.0 but kept the distributed layer: a 100+ petabyte KV cache and the load balancing that sets serving cost. The kernels make one GPU fast; the storage and disaggregation make the fleet cheap.
The surprising bit: At 890 bytes per token, a full 131K-token context is 117 MB and crosses the network in about 2.3 ms.
“Harness² lifts frozen models by up to 11.8 points. It costs four agent runs per task, and the editor is Claude Opus.”
Harness² improves agents by rewriting the instructions and tools around a model, never the model itself. It judges edits by comparing behaviour rather than checking answers, so it works on real professional tasks with no ground truth.
The surprising bit: With a tuned harness, the weaker GLM-5 beat Gemini 3.5 Flash running on its default instructions.
MiniMax M3.1 Flash gets its savings from prompt caching, not the rate card. You only collect them if your agent replays full assistant turns to keep the cache warm, so the discount is an engineering outcome, not a given.
The surprising bit: On a 30-turn coding run, caching moves the bill 4.5x ($0.31 to $0.07). Thinking effort moves it about 1.3x.
One issue a day, straight to your inbox. No hype cycle recaps: just the mechanism, the trade-off, and the number that changes how you build. Free, and you can unsubscribe any time.