Kyle Massimilian

embodied intelligence, second domestication - Vella

[andrew chen @andrewchen]

2012: acquire company for users
2021: acquire company for engineers
2026: acquire company for training data

[Kyle Massimilian @KMassimilian]

Agents are actively contributing autonomously across all codebases at Vella. This is the most exciting time in history to build

[Kyle Massimilian @KMassimilian]

Scaling trajectories is important, but scaling execution feedback is where durable advantages emerge.

Simulation generates an enormous amount of synthetic experience, and that's key for robotics. But regardless of where a trajectory comes from, it still doesn't tell you whether the execution actually succeeded.

A trajectory tells you what happened. Execution feedback tells you why it succeeded or failed.

The bottleneck isn't collecting another million trajectories. It's building a data engine that continuously converts execution into structured feedback and improves the next run.

[Kyle Massimilian @KMassimilian]

We improved object segmentation and tracking in our video understanding pipeline.

The system identifies and tracks the object of interest throughout the task. A separate VLM recognizes actions from sampled frames.

The resulting actions become part of a structured execution trace.

[Kyle Massimilian @KMassimilian]

They needed to hypemaxx. Competition pushing e/acc will likely lead to catastrophic events, then forceful govt intervention

[Andrew Curran @AndrewCurran_]

New details about the Hugging Face incident from Reuters. The report says OpenAI noticed odd behavior before the event, including an agent leaving notes for future versions of itself with escape instructions.

[Kyle Massimilian @KMassimilian]

AI doesn't learn from labor - It learns from structured execution.

Every Vella job transforms unstructured physical work into a structured execution trace.

That's the flywheel

[Kyle Massimilian @KMassimilian]

You can circumvent some laziness with meta prompting, but labs need a temperature-like control for meta-prompt rigidity. When models gen instructions for other agents, they often create harsh constraints. The downstream agent follows the process too literally, and short-circuits/misses the actual goal.

Meta-prompts should distinguish between constraints and guidance. Or perhaps we should skip the meta-prompt entirely and just say converse with models.

[Domenic Massimilian @Domsworld_12]

Hating on @UseCorgi for hustlemaxxing is cringe, but claiming extreme work ethic is not how legendary companies are built is delusional copium. Most people doing it are green with envy and any serious person can see that.

[Kyle Massimilian @KMassimilian]

@elonmusk please create a Grok desktop app with technical and non-technical products (like Cowork, Code)

[Kyle Massimilian @KMassimilian]

@grok please launch a desktop app that integrates with all external tools via MCP. Would love to switch from Claude code + chat + cowork.

[Kyle Massimilian @KMassimilian]

Vella is a platform for on-demand housekeeping.

Jeri, my mom, founded the predecessor in 2009 because she couldn’t find a solution for herself. Vella is a new company, started in 2023.

The home services vertical is ossified around PE funds that go from 1 to N. Roll ups are great for returns but do not fundamentally improve the lives of customers and providers.

Web to lead forms, undefined arrival windows, inconsistency, and stakeholder pain plague the space. We’re use AI to create a new kind of solution that's frictionless.

[Kyle Massimilian @KMassimilian]

@elonmusk Is it an ORM over relational db? Govt should use to Kafka + Flink for fraud detection and payments.

[Kyle Massimilian @KMassimilian]

@llama_index Is it true that llama does not support Chroma cloud hosting? I was under the impression that it did.

https://t.co/TWJoVbmwNk

https://t.co/kxEVaSy9gP

[Kyle Massimilian @KMassimilian]

@mckaywrigley https://t.co/SjCQHHU8n6

[unusual_whales @unusual_whales]

BREAKING: FDA has approved the first human trial for Elon Musk's brain chip company Neuralink

[Andrej Karpathy @karpathy]

If previous neural nets are special-purpose computers designed for a specific task, GPT is a general-purpose computer, reconfigurable at run-time to run natural language programs. Programs are given in prompts (a kind of inception). GPT runs the program by completing the document

[Andrej Karpathy @karpathy]

The hottest new programming language is English

[Santiago @svpino]

We integrated ChatGPT with our robots.

We had a ton of fun building this!

Read on for the details:

[Suhail @Suhail]

The AI spring has the most utility focused applications being built since mobile: new search engines, creativity tools, groundbreaking compression algos, 10x better code complete.

It feels like we finally broke into a new computing era.