noashavit

It’s not just cost per seat. It’s lock in, eroding skills, latencies. I worry about this a lot. There are companies that rely on Claude or Cursor in a way that is not easy to rip out, even if rates 10x.

How is this different from using local models with Ollama?

Claude academy has some courses for free and couple certifications too. Otherwise, I found the best way for me to learn is by doing. So I just use oss models through ollama as part of my workflow to keep costs in check. Local models on consumer electronics are much more capable than you might think!

I just hired an IC and heard horror stories from candidates about ghosting and code tests for non-devs. Also recent mass layoffs in tech have flooded the market with high quality talent. Feels like a “sellers market”

You’re not alone and I really hope this isn’t the future of work. It’s unsustainable to push out code you dont understand. It’s one thing to prototype, or build internal tools, but this cannot scale to commercial offerings (for much longer). With agents gone rouge in the news eroding trust. Just my 2 cents.

I think human in the loop will still be a thing for critical workloads for a while, even if it’s agent-first and human pulled as escalations are needed.

An architecture where compute grows linearly with context length seems dangerous. It can get very expensive as context grows and performance degrades

Agent-native CLIs will need to handle state, stitch together conversation history across sessions, and execute multi-step workflows. I

Perimeter defense is outdated. AI changes the attack surface from network edges to the data pipeline itself. We need to securing the prompt, the model, and the data.

Add a downvote option!

100% Agents need reliable state management, conditional logic, and structured execution paths. Prompt engineering is a surface layer solution to a deeper problem

In real communities bots gets exposed

Gemma4:e4b is a huge upgrade

I think there is a massive divide in the market between large enterprises and smaller org.

The large enterprises are their typically laggard self. Prioritizing governance over innovation.

Everyone else, at least in tech or tech adjacent, is building. Build vs buy is the conversation.

I feel for Gen Z. If it’s hard for those with years of experience, judgment and taste how can you even get into the game?

If this is repeatable and holds true across testing groups and practitioners that would be amazing! Doctors could finally spend time with patients rather than rushing to probe, document, test and diagnose. They are so pressed to maximize their time that any time back could go straight into real care. Am I being blindly optimistic here?

Why are we still discussing which model is “the best”. The model is just one small part of what you need. Think agentic harness, data governance, AI guardrails, machine access controls, etc

I feel like build vs buy is the conversation now. I’m not a developer but I’ve built agents I use daily. When most people can vibe code their way to a custom app, value will most likely hinge on support and other “services”. Just my 2 cents, feel free to tell me I’m wrong!

I built a search intelligence tool with AI and vibes. The vibes ran out in 30 minutes. Here’s what happened.

It started as a Claude artifact. Hardcoded queries for topics I track closely, auto-refreshing regularly. It worked, but lacked flexibility.

So I upgraded to Claude Code. Built a frontend and a backend. On-demand search, multiple API calls, trend charts, and AI summaries.

Then I hit a wall. 30 mins in I was out of API credits.

My prototype was burning API calls in ways I never imagined. It had no real data architecture (just vibes ).

I described features, not a system. Claude Code built exactly what I asked for. The architecture was completely on me.

The unlock: SerpAPI. One clean API that returns real-time search data with a free tier that's actually usable. Once I had that, everything fell into place.

I took the refined prompt and plugged it into a couple AI builders.

Same prompt, different outcomes:

1/ bolt.new got me moving fast, but stalled when I pushed it.

2/ Lovable got me to a working prototype in the first session. A couple more and I now have something sharable.

A week of building at night. One avoidable mistake. A tool I now use every day.

→ Live site: https://noas-search.lovable.app/

→ Git repo: https://github.com/noashavit/noas-search

Type in any keyword, topic, or person to get:

1/ AI analyst overview

2/ Search trends

3/ Most relevant search results

4/ Most recent LinkedIn posts about Topic / by Person

Just bring your own SerpAPI key (free).

I'm sure there's a lot to improve and would love the feedback! What would you have done differently?

It’s hard keeping up with leading players in a fast-paced market like the data. I built this artifact to scrape relevant sources (docs, site pages, press release, etc) to surface the most recent, trusted and relevant news I should know about (with clear priority for product updates)

Use it to keep up with these or other leading players in an industry you pay close attention to.

Type in any topic, or person to get:

→ AI analyst overview → Search trends → Most relevant search results → Most recent LinkedIn posts about Topic / by Person

Just bring your own SerpAPI key (free).

→ Live app: https://noas-search.lovable.app/ → Git rep: https://github.com/noashavit/noas-search

I built this as I experiment with a couple coding agents. I'm sure there's a lot to improve and would love the feedback!

Congrats on the acquisition! Any lessons learned you want to share? Sounds like the app went viral, what would the path look like if it did not?

20 System Table queries to understand Databricks usage across the organization

SEO, reputation building, and a dedicated place to share your news. You could always use medium/other platforms, but then they get the SEO boost and added search mojo, rather than the startup's site

A free resource for data engineers using Databricks. +20 queries to help you better understand how the org uses the data platform and what areas are prime for optimization.

This blog post goes more into why each query is useful: https://synccomputing.com/databricks-health-sql-toolkit/

The contrast between true black and white or even light gray font is just too harsh on most readers' eyes.

Use free stock video to create your own videos / leverage AI to create full video clips

Looks interesting, but I'm not sure I understand how it is different from Raycast

Main learning is to transition form 0-1 manual, rapid iteration to one to many, repeatable processes.

When you are still figuring our your product market fit and gtm motion, you can go deep on experimentation, knowing that you aren't being the most efficient. Your challenge is to solve a problem, not operationalize it. But when you start scaling inefficiencies cost dearly.

New benchmark study! We ran the TPC-DI benchmark across 3 popular Databricks compute platforms to find out what's cheapest and fastest!

Findings include: DBSQL was consistently 2x cheaper than jobs serverless Optimized classic was up to 5x cheaper than jobs serverless Optimized classic was up to 2.5x cheaper than DBSQL serverless

Identify 2-3 core questions you’re looking to get answers for and rephrase them a couple times. You’d be surprised how many times the answer changes based on how you ask the question