Rajesh Naroth

Distinguished Technologist at HPE.

[Rajesh Naroth @rajeshnaroth]

To ship fast, right shift Product Management and left shift QA.

[Rajesh Naroth @rajeshnaroth]

Creative writing is safe from AI. LLMs, even the frontier models now really suck at it. What happened. Wasn’t it supposed to get better?

Coding on the other hand..

[Rajesh Naroth @rajeshnaroth]

Prediction: graph engineering hype will die off and event engineering will replace it.

[Rajesh Naroth @rajeshnaroth]

Opus 4.8 for planning and Grok 4.6 for coding seems to set the right balance.

[Rajesh Naroth @rajeshnaroth]

Social media ruined my attention span. AI fixed it by making me read thousand of lines of plans and specs everyday.

[Rajesh Naroth @rajeshnaroth]

@Krishnasagrawal Graph is a CS abstraction. Karpathy talks about graphs in the attention/transformer sense. It’s misleading to present that lecture as primarily advocating the modern agentic execution/knowledge-graph paradigm.

[Rajesh Naroth @rajeshnaroth]

Don't mistake a temporary scaffold for a permanent architecture. Frameworks that work around a model's limitation will be extinct when the model absorbs that into its intelligence. Be mindful of what you're investing in.

[Rajesh Naroth @rajeshnaroth]

@reiraxbt Graph is a CS abstraction. Karpathy talks about graphs in the attention/transformer sense. It’s misleading to present that lecture as primarily advocating the modern agentic execution/knowledge-graph paradigm. Apples and oranges.

[Rajesh Naroth @rajeshnaroth]

water—mark—this—b**ch.

[Rajesh Naroth @rajeshnaroth]

During brainstorming sessions with AI for UI, try ASCII mocks. They are fast and LLMs are quite good at it even when the right border is off usually.

[Rajesh Naroth @rajeshnaroth]

Planned with Codex today. Felt like I was cheating on Opus.

[Rajesh Naroth @rajeshnaroth]

One missing element in your AI workflow may be an ontology. You can create a quasi-ontology based on Andrej Karpathy’s llm-wiki.

[Rajesh Naroth @rajeshnaroth]

Create accurate context. Execute in smaller chunks. Write better gates. Use the best coding model you can afford.

[Rajesh Naroth @rajeshnaroth]

Question for software devs: If the agent codes 10x faster but you’re only shipping 3x faster, where does the other 7x go?

[Rajesh Naroth @rajeshnaroth]

My process: each problem or feature gets split into stacked plans, then into stacked modules or checkpoints. Every plan and module has gates, and the most important is a multi-panel, multi-model review loop. It takes time, but it produces really high quality software.

[Rajesh Naroth @rajeshnaroth]

@unclebobmartin Smaller modular changes with gates has been working great for me. One of my most important gates is a multi-panel, multi-model code review, each focusing on a specific dimension like UI/style drift, code confidence/test coverage, security, feature completion etc.

[Rajesh Naroth @rajeshnaroth]

Claude Opus peaked at 4.8.

[Rajesh Naroth @rajeshnaroth]

Want to ship fast with AI? Remove the paper pushers from the pipeline.

[Rajesh Naroth @rajeshnaroth]

@trq212 @mattshumer_ It’s unfair to ask developers to rewrite their harnesses, skills and automation every time a model is updated.

[Rajesh Naroth @rajeshnaroth]

Opus 5 needs very little instructions, and actually regresses when your context is busy. Which means redoing all my skills to work with it. Too much work. Sticking with 4.8.