Hacker entrepreneur; Artificial Intelligence researcher
Philadelphia | HCMC Vietnam
Joined March 2009
- Tweets2.6K
- Following178
- Followers507
- Likes344
758 consultants with GPT-4: faster and better inside AI's capability, 19 points less accurate just outside it. Delegation works until you can't see the edge. Seeing it takes hands-on expertise. qcoda.com/blog/agent-boss
Bainbridge, 1983: operators who only monitor automation lose the skill to take over when it fails. "Stop producing, start orchestrating" removes the practice that makes a good orchestrator. qcoda.com/blog/agent-boss
Supervising AI agents is barely management. Agents need a precise spec and an expert who can spot bad output. That is the skilled operator's job description, not the charismatic leader's. qcoda.com/blog/agent-boss
Microsoft says every employee is now an "agent boss." Its own survey: only 21% of employees expect to manage agents within 5 years. Leaders: 36%. The appetite runs top-down. qcoda.com/blog/agent-boss
Gallup: ~1 in 10 people have high talent to manage. Firms promote top performers who then manage worse. Now AI vendors want everyone to be a manager of agents. The Peter Principle at workforce scale. qcoda.com/blog/agent-boss
A lot has shipped in LLMesh since v0.2 - the self-hosted mesh that routes OpenAI/Anthropic-compatible inference across your own nodes.
Quick tour of what's new through v0.21.0 🧵
🚀 The v0.2 bundle: real per-token streaming unified across Ollama, vLLM, and MLX, adaptive SSE batching (~80% fewer hub syscalls under fast clusters), Anthropic Messages streaming, hub state durability, and weighted routing.
🖼️ Also in v0.2: image generation v1 (BETA) - OpenAI-compatible /v1/images/generations via mflux on Apple Silicon. Read the sysreq advisory before enabling (64 GB UMA min).
All open source: github.com/qcoda-ai/llmesh
#AIEngineeringGovernance
Four industry studies. Same pattern.
95% of enterprise GenAI pilots fail (MIT NANDA).
74% can't show tangible AI value (BCG).
Recognition-vs-mitigation gap in every risk category (Stanford AI Index).
77% of CIOs say AI is outpacing governance (IBM).
Not a training problem.
IBM added the compounding version: 66% of CIOs are accountable for AI systems they don't fully control. 11% believe they're ready for the deployment scale coming by 2027.
The artifacts get produced. The decisions do not.
That's the same pattern the essay diagnoses.
Full essay with references at the end:
qcoda.com/blog/governance-th…
Every enterprise has an AI Ethics Committee.
Ask the person running it how many projects they've stopped.
Watch the pause.
The pause is the answer.
New essay on governance theater and the one question that separates it from the real thing:
qcoda.com/blog/governance-th…
Strip the compliance baggage off "governance" and what's left is:
The gap between what a company says it does and what it actually does, measured on the days nobody is looking.
Built to sell or built to hand down, the discipline is identical.
hackernoon.com/lifestyle-vs-…
Three-episode series on local AI agents:
Ep 1: Silent tool-calls — wrong layer
Ep 2: Engine × model bake-off — wrong variable
Ep 3: Task complexity + routing — wrong question
People debug the model. The problem is always one layer down.
Full series → qcoda.com/blog/right-model-r…
Neither model is the bad guy.
Transactional exits fund retirements and seed the next founders. Lifestyle companies can stagnate if they confuse loyalty with comfort.
The damage is running one while telling everyone you're running the other.
hackernoon.com/lifestyle-vs-…
Promoted in 3 years → 70% still there
Stayed in same role → 45%
External hires → 61% more likely to be fired
Internal mobility = 41% longer tenure
Replacement cost = 0.5–2x salary
The "right person, right now" model isn't cheap when "right now" ends.
hackernoon.com/lifestyle-vs-…
