Snack On AI

Your Daily AI Snack On - AI Research, Tools, Tutorials & Insights

Join 10,000+ AI Engineers & Enthusiasts, Subscribe & Grow Together

SnackOnAI Blog

OpenSRE Builds the AI SRE Agent First and the Answer Key Later

Oct 25, 2026

•

9 min read

OpenSRE Builds the AI SRE Agent First and the Answer Key Later

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 hard part turns out to be knowing whether the agent's root cause is correct, and that part is still a roadmap item.

Mohinish S

SnackOnAI Blog

Zonos Shows Expressive Voice Cloning Is Priced in Memory Bandwidth

Oct 24, 2026

•

10 min read

Zonos Shows Expressive Voice Cloning Is Priced in Memory Bandwidth

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.

Mohinish S

SnackOnAI Blog

Kokoro Proves the Best Small TTS Model Is Not a Language Model

Oct 23, 2026

•

9 min read

Kokoro Proves the Best Small TTS Model Is Not a Language Model

One of the most downloaded open text to speech models never predicts a token. It predicts how long each sound lasts, then paints the whole sentence at once.

Mohinish S

SnackOnAI Blog

DeepSeek's DeepInfer Shows the Engine Was Never the Hard Part

Oct 21, 2026

•

10 min read

DeepSeek's DeepInfer Shows the Engine Was Never the Hard Part

The most valuable thing in DeepSeek's inference stack is not a kernel. It is a storage cluster that remembers your prompt for three days. DeepInfer ships the kernels and scheduler. The KV store, load balancers, and disaggregation stay internal, and they are where the economics live.

Mohinish S

SnackOnAI Blog

Google's Harness Squared Buys Agent Accuracy With Extra Runs, Not New Weights

Oct 20, 2026

•

9 min read

Google's Harness Squared Buys Agent Accuracy With Extra Runs, Not New Weights

Harness2 lifts frozen models by up to 11.8 points. It costs four agent runs per task, and the editor is Claude Opus. The fastest way to improve an agent is not a better model. It is a better wrapper, rewritten while the task is still running.

Mohinish S

SnackOnAI Blog

MiniMax's Flash Model Is Cheap Because of the Cache, Not the Price Tag

Oct 19, 2026

•

9 min read

MiniMax's Flash Model Is Cheap Because of the Cache, Not the Price Tag

Price explains half of the claimed 70% savings. The rest has to come from tokens, and an independent benchmark says this model is verbose. A 50% price cut cannot produce a 70% cost cut. Something else has to do the work, and on agent workloads that something is the prompt cache.

Mohinish S

...

Snack On AI

OpenSRE Builds the AI SRE Agent First and the Answer Key Later

Zonos Shows Expressive Voice Cloning Is Priced in Memory Bandwidth

Kokoro Proves the Best Small TTS Model Is Not a Language Model

DeepSeek's DeepInfer Shows the Engine Was Never the Hard Part

Google's Harness Squared Buys Agent Accuracy With Extra Runs, Not New Weights

MiniMax's Flash Model Is Cheap Because of the Cache, Not the Price Tag

Snack on the latest in AI delivered to your inbox.

Quick Links

Subscription

Socials