Sandeep Kumar Jha

Sandeep Kumar JhaFounder, NuroSparx3w ·

Was using Claude Code through Dispatch on my phone this morning when I caught myself doing the AI math again. Which model should I use for this job? Should I clear the context or type /compact? Should I send the browsing to a cheaper agent and save the expensive one for thinking? Should I keep outputs to /concise to save tokens? how do I ensure the limits stretch (across models ) before a reset hits. It reminded me how quickly these tricks become useless. Not long ago, I started some prompts with, “You are an expert dev with 20 years of experience.” I worried about chunk sizes. I wrote regex loops to fix broken JSON. I felt smart doing all of it. Most of those tricks are dead now. The tricks I use today, like choosing models, managing context, and saving tokens, are on the same path. So are the systems I build to preserve context and make sure each session picks up where the last one ended. I will keep using them because they help with the models we have today. But normal users should not have to think about any of this. One should not need to understand how the machine works to get good results from it. Tokenmaxxing should not be a skill for top 1% of AI users . The system should do that work for everyone. As capabilities evolve, saving tokens should become the model’s job, not another decision the user has to make.

Sandeep Kumar JhaFounder, NuroSparx2mo ·

In 1962, Kenneth Arrow, who'd go on to win a Nobel, pointed out something odd about information. "You can't value it until you've seen it, and then it's free." Satya Nadella just flipped that on its head. He wrote an article called The Reverse Information Paradox, and it put words to something I've felt for a while using these tools. With AI, the buyer is the one giving away knowledge. You pay for the model with money, then you pay again by feeding it everything that makes your business work. And if you want better answers, you have to feed it more. Forget the documents you upload for a second. The models learn from your prompts, your tool calls, and most of all your corrections. Every time someone on your team fixes a wrong answer, a little hard-won judgment walks out the door. You'd never notice it happening. These labs trained on the whole public internet under fair use, then wrote terms that ban you from distilling their models, while keeping the right to learn from how you use theirs. The learning flows in exactly one direction. Ironic!! His fix is what he calls a trust boundary. Your evals, traces, feedback, and tuned models live in one place you control, and nothing leaves without consent. Own your evals. Train inside your own tenant. Never depend on a single model. Worth noting that Microsoft owns about 27% of OpenAI, and "decouple from any single model" happens to be great business for the company selling the layer underneath. Doesn't make him wrong though. In June a frontier model got pulled offline for 19 days over an export order, and companies built on it just sat there. You pay for AI twice. Most companies only track the first bill.

Sandeep Kumar JhaFounder, NuroSparx5mo ·

Tried Claude Design. Three use cases: built a ppt, asked it to reimagine a website frontend, made a template using the Canva connector. Output looked great, as usual But .. I've done all three before using specific Claude skills. Same output. Similar quality. Those capabilities sat in the Anthropic skills repo for months. Anything you could run with code, you could already build. Claude Design feels like repackaging with new UI on top of what already existed. So why ship it? Because shipping itself is the strategy. Even sloppy first versions get attention, then improve at incredible pace. Look at the pattern. Anthropic ships micro announcements almost daily. Tiny features, small wraps around existing tech. Each one earns a news cycle. Each one feeds the algorithm. Each one tells builders "they're moving faster than everyone else, keep watching." The perception of speed compounds faster than actual speed.