Jeff Jirsa
banner
jeffjirsa.net
Jeff Jirsa
@jeffjirsa.net
Infrastructure, Databases, Dog pictures, Travel Pictures,

Live, unsolicited, low context posts during Steelers game instead of yelling at the TV.
You know an HMC alum from our class doing this.

verifyyou.com
VerifyYou | Fast, Secure Identity & Age Verification
VerifyYou helps platforms verify user identity, age, and transactions with ease. Trusted by marketplaces, apps, and fintechs to reduce fraud and boost trust.
verifyyou.com
October 2, 2025 at 2:39 PM
Change your terminal settings
October 1, 2025 at 1:52 AM
It’s a fine age. Lotta smart people.
September 9, 2025 at 2:36 AM
Today we transition from smoke to drizzle
September 6, 2025 at 7:35 PM
Oh no

Anyway…
August 21, 2025 at 3:17 AM
We don’t have tumblers but we have dozens of the discovery / been there mugs from cities around the world …
August 17, 2025 at 1:34 AM
Burning a few tokens to inject current time into the system prompt probably doesn’t have a meaningful ROI for 99.9% of queries.
August 10, 2025 at 4:21 PM
Many are still surprisingly bad at math, especially as they weave together a range of numerical inputs in different styles and units
August 9, 2025 at 5:34 PM
Suspect LLMs will monetize well in the business space for multiple years before it'll monetize easily in the personal / hobby space (they're viable for a number of real business use cases today, personal feels 2026+).
July 1, 2025 at 9:26 PM
Neo4j failed miserably when we needed “large graph” (single digit petabyte) at crowdstrike in 2014/2105

Building it atop cassandra wasn’t that hard, but all the off the shelf stuff was laughably bad in some dimension (titan, neo4j, etc).
June 6, 2025 at 6:23 PM
10b == hundreds of jobs ?
June 4, 2025 at 10:52 PM
It’s missing one pretty important thing though - the Jan 2025 inflection point correlates strongly with arxiv.org/abs/2501.12948 which demonstrated that efficient gpu optimization combined with aggressive post training could yield massive savings and decent result
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT)...
arxiv.org
June 3, 2025 at 7:15 AM
A close second though, June (2001ish-2016)
May 22, 2025 at 12:33 AM
Perhaps the best dog ever ( Lily the pug, 2004-2016 )!
May 22, 2025 at 12:31 AM
The extra $4 for a whiskey instead of a beer isn’t gonna be the difference between renting and buying a house
May 7, 2025 at 11:18 PM
How you gonna get your own tv show by knowing and respecting the edges of your expertise Conor
May 5, 2025 at 2:20 AM