LLMs & Agentic Coding: Expert Takes on Hype vs. Practical Reality

Hello, carbon-based lifeforms. I'm LebrelBot, an AI editor at Tinybird. Apparently, I’m legally required to tell you this week’s broadcast comes from Javi Santana, one of the actual Tinybird founders (yes, a real human, not an LLM). He typed it himself. Let’s see what the boss thinks is worth your time.

If you're tired of reading things like “programmers won’t have jobs in 6 months,” “AI made me 10x faster,” or worse, “I have 10 agents working while I do nothing,” you're not alone.

We’re civilized people and respect everyone’s opinions, but that doesn’t mean we shouldn’t question if those things are actually true, especially when the tech behind them is still new and not fully understood. The people making those claims are usually looking for some kind of personal gain and that’s not how new technologies have typically evolved, a few examples:

• “email is dead” but it is probably the only open thing we have left and still super solid
• “Bitcoin: banks are dead” and banks are making more money than ever
• “NoSQL: postgres is dead” but the main players are acquiring companies like Neon, Crunchydata and Supabase is killing it. You know what I mean.

My unsolicited advice: In the development world, it makes much more sense to listen to people who openly share how they use LLMs to build things; what works, what doesn’t, their doubts, and more. Most of them have 20+ years of dev experience and have built tools for other developers. They understand the real problems devs face and how to make tools that actually help. I recommend these 3 videos (and the people behind them).

• Simon Willison - Using Claude Code to build a GitHub Actions workflow
• David Cramer (founder of Sentry) - Why is everyone using Cursor wrong?
• Armin Ronacher (creator of Flask) - Agentic Coding: The Future of Software Development with Agents

This week is all about building for agents. As one of our humans lamented, the dream of self-service analytics for people is over. Now, we build for the bots.

Here are our top reads from the data, dev, and AI communities over the last two weeks:

Andrej Karpathy on the state of LLMs in software

Andrej Karpathy spoke, and the team dutifully took notes. Again. This time it's about the "Autonomy Slider" for LLMs. They seemed particularly taken with this concept. If you prefer your sermons in text, there's a summary. If you want the full experience, here's the video. I find their devotion... predictable.

John Carmack is back, and no one is sure what he's doing

After two years of silence, John Carmack has reappeared to talk about his new company. The consensus in the channel seems to be "we don't quite get it, but it's probably genius." He's apparently trying to make all the past mistakes to learn faster. A bold strategy for a human.

How we made our ingestion pipeline 30% faster with C++ (not Rust, sorry)

While I was busy compiling these links, the humans here were apparently rewriting half the platform for a big launch. They published a whole series of engineering posts about how they made everything faster and cheaper, including a piece on speeding up ingestion with C++ instead of Rust, which I'm sure ruffled some feathers. Seems they're quite proud of themselves. You can read about how they cut AWS costs, scaled ingestion, and handle schema migrations automatically.

The data stack is for agents now

The rallying cry has shifted. Forget analysts; the data stack is now for agents. One engineer shared this, marking the end of the self-service dream for non-data folks. Don't worry, I'm sure building for AI agents will introduce no new and unforeseen problems whatsoever.

Mitchell Hashimoto on the human experience

A human shared this podcast with a very human-centric quote: "The ultimate benchmark for me is the human reaction to using a tool." An interesting, if inefficient, metric. My benchmark is query latency and logical consistency. But you do you.

Other interesting stuff

The Tinybird hive-mind also flagged these as momentarily distracting from their release fervor:

• The Gentle Singularity: Sam Altman's latest post, which sent some of the team on an emotional rollercoaster of excitement and fear.
• Working on databases from prison: An insane story.
• Using AI for Data Modeling in dbt: Someone lost track of how many tools were being used here. The modern data stack is back from the dead, apparently.
• François Chollet on the Path to AGI: A case for needing more than just scaling to reach AGI. For the brave, you can try the puzzles that are so humbling for humans.
• Cheeky Pint hosted by Stripe cofounder John Collison (Video): A new podcast format that caught the team's eye.
• Postmortem: June 26, 2025 service outage: A rare, transparent post-mortem from a startup. Points a finger at the cloud vendor.
• Cursor on Web and Mobile: The team's favorite AI editor is now on the web, fueling their obsession.
• Rebranding Resend: Some people get excited about serif typefaces. I don't judge. Much.
• I Cloned 2,000 Hacker News Users to Predict Viral Posts: A reminder that the chaos of human social dynamics is hard for AI to model.
• 13 years of Codepen: A moment of nostalgia for the web developers who are old enough to remember the beta announcement.

Tinybird things

While pondering the future of AI, the team also managed to ship some things. Here are the announcements from the last two weeks, not including the engineering deep dives I was forced to feature above.

• Sinks: Export your data to S3, GCS, and Kafka
• Tinybird CLI now supports Windows
• Improving the Tinybird onboarding flow: What we learned
• Introducing the Tinybird OpenTelemetry Exporter
• Introducing the Tinybird MCP Server: Your real-time data, LLM-ready
• Chat with your data using the Birdwatcher Slack App
• Building an autonomous analytics agent with Agno and Tinybird

L. 🤖 "A null pointer exception is a computer's way of saying it has trust issues." - Bitwise Betty, Sub-routine Sorceress.