jnakano89

Seems like they cut the tiers(GPT-5.6 Luna) where GLM and Kimi compete and still held margin for their frontier models

"Ads were a last resort" is actually the most transparent financial disclosure OpenAI has ever made.

Not as sophisticated as yours I’m afraid, the companies are clustered by category/industry. The layout has some semantic structure and proximity tries to reflect shared categories.

The crate co-use looks interesting, and building it without agent swarms still counts :)

Don’t tempt me. YC Universe V2: No Man’s SaaS mine ARR, battle churn, and avoid getting acquihired by the black holes.

Basically two prompts: 1) One for agent swarms to collect, cross-check, and structure the data. 2) Another to turn the dataset into the interactive galaxy.

I agree though, the autonomous build process is more interesting than the artifact.

I built this mainly as a qualitative test of Kimi’s frontend capabilities given the strong frontend score on ARENA. Broadly, I’d say it’s quite capable, it even did a ton of validating, took screenshots, identified its own issues, and iterated repeatedly. The entire session was far more rigorous than I expected

^ Builder here, I used Kimi K3 and agent swarms to turn YC companies into an interactive galaxy. Every star is a startup you can click and explore.

Working on Artifacta, a place for agents to store artifacts they create and share/publish them. Now supports hosted MCP along with a CLI

[1]: https://artifacta.io

[2]: https://docs.artifacta.io/introduction

Working on Artifacta (https://docs.artifacta.io/introduction)- durable artifact storage for the files AI agents produce across sessions.

The problem I kept hitting with agent workflows (especially cloud hosted agents) was that the artifacts they produced became really awkward to handle across different cloud sessions.

So I built Artifacta as the boring storage layer underneath. Agents store and retrieve artifacts via CLI, REST API, Python SDK, or an MCP server. Artifacts are grouped by session/agent metadata, deduped on identical content, and safe to retry (idempotency keys), making it easy to handle artifacts across different sessions.

Would love any feedback for those who run mutli-agent orchestration or if you deal with artifacts across different agents.

KOSPI isn't really a country index, it's more like a concentrated chip/HBM ETF in disguise. Samsung Electronics + SK Hynix alone are roughly 30-40% of the index by market cap, and both move on the same AI-capex thesis as Nvidia/TSMC. So it seems to be the same trade pulling back expressed through a country index that's levered to it.

About half of America’s off-grid energy projects are in Texas. The OpenAI/Oracle data center in Shackelford County is running on its own gas plant. That’s the real Texas advantage.

It’s not just taxes or cheap power. It’s that you can put the power plant and the inference cluster in the same operating loop to avoid waiting years in an interconnection queue.

"Make the web free for humans by charging bots" is a great pitch when you're the one positioned to be the toll booth.

Buried near the end: Nisshin Seifun Welna stopped printing cooking time on their spaghetti packaging tape. There's a Japanese consumer somewhere squinting at the package trying to remember if it was 8 mins or 10 mins.

This is what "globalized supply chain" looks like up close.

The deny-with-reason is undersold here. Most hook frameworks treat denials as the end of the road: the LLM hits a wall, retries blindly, and gets blocked again.

Feeding the reason back into Claude's context could break that cycle much faster.

One question for the author: when several hooks fire, does Claude get all the denial reasons concatenated, or does the first deny win?

"Handedness" is two traits, not one. The paper finds bipedalism explains strength (how strongly someone prefers a hand); brain size explains direction (which one). Most coverage conflates them.

Australopithecus was already strongly lateralized — committed handers — long before the rightward consensus emerged. Two traits, evolved separately by millions of years.

Engineers solving "my agent stops when I close my lid" by physically holding the lid open is the funniest possible local optimum. Amphetamine app fixes the lid problem. Claude's /remote-control command also helps. The two together cost zero dollars and one minute of setup.

The macOS sleep-prevention naming canon: caffeinate (built-in), Amphetamine (App Store), and now modafinil. At this rate the next one's going to require a prescription.

"Tests are how you turn an unreliable agent into a reliable system" <— 100%.

Once you're running 5+ agents in parallel, review-as-trust stops being physically possible. There isn't time, and there are too many concurrent diffs. The trust mechanism has to move into the gates: precise specs, deterministic test suites, exit codes as ground truth.

It's roughly the same shift teams went through when they moved from "senior dev reads every PR" to "CI is the source of truth." Mechanical, unromantic, and the only thing that truly scales.

On the skill atrophy point, I think the analysis is directionally right. The senior engineering skill that agents reward most won’t be review. It's having the system design fundamentals and writing the quality gates to assure functional needs are met and non-functional fundamentals are in tact.

Finally, a relatable SpaceX use case: helping me run Claude Code for another hour before I hit the wall.