tj

part shiba part builder at @kepler_ai_hq

[tj @cerealnomics]

@RaoulGMI I wonder if the endpoint is “everything is a token” or “everything has an API-priced property right.” machines need things to be ownable, transferable and composable — tokenization may just be one way to encode that.

[tj @cerealnomics]

@Midnight_Captl Instinct took 3 hours to reply to me. Muse just makes things up. pick your poison

[tj @cerealnomics]

@turbopuffer made 100M+ docs searchable at @kepler_ai_hq. somewhere along the way, the interesting constraint stopped being scale and became boundaries.

Huge shared corpora, isolated customer namespaces, different permissions and trust boundaries — but at query time, they still need to behave like one corpus.

Now we’re pushing that architecture further: working on running turbopuffer inside our customers’ own clouds. Private data stays where it is; the retrieval infrastructure moves to it.

Retrieval wants to abstract away where data lives. Enterprise AI keeps pulling that abstraction back into the architecture.

Really cool to see some of the thinking behind what we’re building at Kepler featured in turbopuffer's latest case study

Fun to see some of the architecture that got us here featured. Thanks to turbopuffer team @Cameronarcher_ @Sirupsen

[turbopuffer @turbopuffer]

ask Kepler for a public company's revenue growth, and it traces the number to an exact line item in a verified SEC filing

to find the filing, Kepler searches on tpuf

→ 100M+ financial documents
→ 85ms p90 latency
→ 10x lower cost than alternatives

https://t.co/hc89XLqHoN

[tj @cerealnomics]

@alexandr_wang loving the bucket hat

[tj @cerealnomics]

@Rick_Zullo this is basically the question behind Kepler: what does the service look like when the client, not the intermediary, gets the AI leverage?

[tj @cerealnomics]

@nikunj @typefaceai thanks for putting this together — dangerous amount of weekend project inspiration here

[tj @cerealnomics]

@nedwize very cool — exploring something similar. what does your eval set look like, and how does Jev behave when a document belongs to a class that wasn’t in the original label set?

[tj @cerealnomics]

@MisbahSy @typesafeai very interesting! will give this a try !

[tj @cerealnomics]

@joaquinbarroso lol

[tj @cerealnomics]

1. The institution has to be allowed to be wrong. We usually judge institutions by consistency, but here consistency could be dangerous. If the facts change, it needs to reverse itself quickly without that being treated as institutional failure.

2. “Who picks them?” may matter less than “who can fire them?” If members know an unpopular conclusion can cost them their seat, you’ve built political incentives into the institution regardless of how bipartisan the appointments were.

Maybe the goal isn’t finding perfectly neutral people. It’s designing a system where changing your mind is rewarded and disagreeing with whoever appointed you is survivable.

Would you give a body like this fixed, nonrenewable terms?

[tj @cerealnomics]

@eastdakota the male gaze but it’s observability infrastructure

[tj @cerealnomics]

@AIatDoorDash have you tested how much of Vera’s performance survives if you swap models without retuning the harness?

[tj @cerealnomics]

@Jason @stregishotels was it good at least? 🧐

[tj @cerealnomics]

@ChuckRobbins @splunk “can i trust it?” might eventually become a cost question too. perfect verification is expensive. the interesting systems will know when verification is actually worth paying for

[tj @cerealnomics]

@Kitty_Mayo_ i remember those stairs 🤣 good luck!

[tj @cerealnomics]

@gokulr @Costco @DoorDash how different will it be vs costco on instacart?

[tj @cerealnomics]

@ericbahn @HustleFundVC cute!! 💜

[tj @cerealnomics]

@nicodotdev @huggingface salut!!!

[tj @cerealnomics]

@pejmannozad more signal, less time to actually know what any of it means