𝗥𝗬𝗔𝗡 𝗟𝗘𝗘

VCs, please stop funding this slop. We’re entering clown world with the amount of AI building we’re seeing. 🤡🌎 All I see in the data infra space is companies building infrastructure for AI Agents with the obligatory buzzwords like A2A or MCP. Who is productionizing these AI Agents? Two years of marketing hype in, I have literally only seen ONE talk about an AI "Agent" in prod. It's by Uber and it's used for code generation. First off - it's a dev tool and second - it's not even an agent, it's a workflow. Who honestly thinks the future of software architecture will be hundreds of AI Agents talking to each other, similar to microservices today? Does anyone ACTUALLY believe this is what the world will be in 5 years? Can they comment here and share their perspective? The data is pointing to the opposite. • We haven't seen any major LLM advancements since GPT-4. • All frontier model companies are grossly unprofitable with no sustainable path. • A lot of investments between big tech are circular (recycling revenue) Even in a winner-take-all model, OpenAI is burning $5 BILLION a year. Anthropic, the number two, is looking way, way worse regarding its revenue model. What happens when the funding stops? Companies chasing investor interest via shiny narratives are missing the forest for the trees - we, the engineers, need practical solutions. That's what's going to last. Fortunately, I see two trends growing quietly behind the spotlight (the sloplight?) of AI. They will most certainly survive a dot-com like crash if that were to happen. They are: • Small Data • The “Just Use Postgres” Renaissance The common theme between both? Simplification. ♦️ Small data is the cure for the Big Data hangover that tech collectively went through in the 2010s. It’s the group realization that: • organizations don’t use that much data • hardware is getting really, really good Most applications will never see a terabyte of data, even if they’re successful. Snowflake and Redshift-published data showed that 99.9% of real-world queries fit on one node. 💡 And the nodes are becoming super beefy. AMD released a 192-core (!) CPU this summer. 2025-era SSDs can do 5.5 million random reads a second and 28 GiB/s sequential reads. Embedded and simple systems like DuckDB are going viral for a reason. ♦️ Postgres is doing to data infrastructure what the monolith comeback is doing to the microservices architecture. It’s the group realization that Postgres can handle 80% of companies' needs. And it does it with 20% the effort. It’s not the extreme view where "when you have a hammer, everything looks like a nail". 👉 It’s the extremely reasonable and practical application of reducing the amount of system sprawl in your organization when that sprawl isn’t warranted. Instead of MongoDB, Redis, an AI Vector database, Snowflake and Elasticsearch - you can use tsvector, pg_mooncake, pgvector, unlogged tables and jsonb in Postgres. Life becomes significantly simpler when you don't have to manage 5 different systems. There is a ton of organizational overhead in having to: • learn their nuances/configs/gotchas, • learn their UIs and terminology • learn how to deploy and upgrade them safely • build operational expertise • write runbooks for them • debug them • test them • find, hire and retain talent that understand them • keep up with their ecosystems • etc. Paying this organizational overhead for zero significant gain is not worth it. It's all common sense.