Anatomy of a Voice Agent
Voice latency benchmarks across OpenAI Realtime, LiveKit, Vapi, Retell, Pipecat, Cerebras. Hear the difference; switch the network and listen to the dead air grow.
Barada's log
Recent content on Barada's log
Latest posts
Not your Weights, Not your Brain
Every time I correct my AI copilot, I'm leaking expertise - dark knowledge about how I think, what trade-offs I accept, where I draw the line. That signal is either compounding me or vanishing into someone else's training run. I chose to keep it.
claudectl: stop tab-hunting your AI agents
How to build a TUI dashboard in Rust that monitors AI coding agents. Covers session discovery, incremental JSONL parsing, multi-signal status inference via CPU transcript heuristics, terminal integration across 9 emulators, and why you don't need an async runtime.
Benchmarks are the new stars
4.5 million fake GitHub stars. A 47-point contamination delta on SWE-bench. The credibility signal for AI is migrating from social proof to computation proof to revenue proof, and the economics of faking each rung tells the whole story.
Agent Source
AI coding agents don't choose libraries rationally. They exhibit similar cognitive biases as humans do, but from a different source: training data instead of evolution. We discuss agent selection biases in software libraries and trace the feedback loop that reinforces them. Does open source need rebuilding for machines?
Find the Kernels
Where does defensible value live in the models era? Not in model weights or API wrappers.
Why it's hard to Claw the Enterprise
Inspite of all the hype around personal assistants, personal agency still feels far from reality. It feels like we are the ones grinding for the agents today - constantly reviewing plans, uploading context and providing agents access to tools.
HydraBench: Agent Infrastructure Resilience
An open benchmark for agent infrastructure resilience. Tests crash recovery, secret containment, handoff reliability, channel security, and cost control across OpenHydra, LangGraph, CrewAI, and bare agents.
Project Hydra: Designing a world for agents
Last week I let a swarm of AI agents run autonomously for a week. One of the agents tried to steal API keys. By Friday, four more things had gone wrong. We learnt a ton on failure-modes, patterns for running long-running agents in the wild and how to prevent goal drift.
We are the average of our models
We are the average of the friends we spend the most time with. Today, one or two of those friends are models. And if these are setup to reinforce existing beliefs, we are no longer compounding our intelligence; we are blitzscaling confirmation bias.
Predistribution: An Economic Model for the AI Age
Predistribution changes how value is shared during AI-driven growth, before inequality hardens and redistribution becomes politically infeasible.
Model Collapse
The productivity gains are flowing to capital holders instead of the workers generating them. The old framework said credentials lead to knowledge, knowledge leads to value. But value is concentrating in weird corners now; places where someone has to be sued, where platforms control distribution, where tacit knowledge can’t be extracted from text.
Every Token an Iceberg
Inference workloads now account for 80% of AI compute spending. The hierarchy in tokens is no longer about information density—it’s what happens when the token leads somewhere wrong.
Every Inch Matters
When building becomes trivially easy, every software market becomes a red ocean. A browser from scratch in one week. 3 million lines of code. This isn’t an anomaly; it’s the new baseline.
Appliances, Factories, Grids
Own the chips or the customers. Everything else is a footnote.
Model-Adjacent Products, Part 5: The Implementation Path
The build manifest: 90 days from foundation to production-ready autonomy. What to ship in what sequence.
Model-Adjacent Products, Part 4: Governance & Practice
Alignment as a runtime surface, policy enforcement without retraining. Team practices that ship.
Model-Adjacent Products, Part 3: Quality Gates
Model outputs are hypotheses that need verification pipelines to catch errors before users do.
Model-Adjacent Products, Part 2: Context & Tools
Memory and hands for the model: retrieval that doesn’t hallucinate. Tools that don’t break production.
Model-Adjacent Products, Part 1: The Architecture
The physics of production AI: latency engineering that keeps humans in the loop. Token economics that don’t bankrupt you.
Model-Adjacent Products, Pre-Read: The Autonomy Ladder
Before you build: the mental models for human-AI collaboration. Why L1 copilots need different infrastructure than L4 autonomous agents.
Model-Adjacent Products: A Builder's Guide
A 6-part series on building production AI systems. The foundation model is the CPU; your product is the computer you build around it.