@BillyJacobson
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
Living in the year 3000 of building with AI and bringing you along | Technomindfulness
Brooklyn
Joined April 2009
- Tweets154
- Following419
- Followers500
- Likes342
this feels so 2013 data mining coded but also feels like the optimization that ai is promising all of us
I ran a simulation built on 700 million real MTA trips. It said to close the 3rd Avenue L stop.
stopthestop.com
🔥
Lots of discussion out there about our next model(!), so I wanted to give an early look as soon as possible. Introducing Gemini 4 Argon!
It shows frontier performance in complex workflows, cyber defense and software engineering. Teams are using it extensively at Google, from coding to quantum computing, great feedback.
Here’s a look at the benchmarks:
kinda just been loving my no reservation neighborhood spots without sending an AI agent to war for a mediocre $30 burger with no sides
Instinct CEO Noah Shinn says agents will let restaurants give tables to birthdays and anniversaries over whoever clicks first
"And traditionally, because of the way that the internet has worked, at least for reservations, it's been first come first serve. So anyone that comes in, no matter how important or how not important it might be, whoever gets in first gets the reservation."
"But what if you had the ability for the agent to be able to communicate on both sides, communicate, hey, this is important. This is actually the person's spouse's birthday, the 30th birthday, a very special event is coming up."
"And the restaurant too can say, let's actually prioritize birthdays or major decade birthdays or major anniversaries or these certain people, I don't know, whatever it might be, right?"
"So now with the ability, with all of the logistical burden of having to describe exactly what your case is and what's happening and why it's important. Now we have the ability to perfectly match what does the restaurant want? What does the user want?"
"And that will result in more special events for the users being able to actually have a spot in the restaurant."
coded a jevved up pomodoro app that forces me to stay on task 🍅
type the task, then jev gets asked "is this related to {task}?" for every window and tab, with a 4s warning before closing anything that isn't. really taking advantage of jev's low latency decision making
Awesome experimentation here. Sometimes the question can be answered with a simpler model and sometimes not. This is a great way to combine and so cool to see it on a 10M row dataset
Tested @typesafeai's Jev against BigQuery’s AI.IF and AI.CLASSIFY from 1 to 100k rows (and pushed BQ to 10,000,000). Findings:
- 1-50k rows: Jev via Cloud Run is fastest
- 50k-10M rows: BigQuery optimized flattens model cost
- Cascade (Jev + Gemini): top accuracy at 1/3 the cost
Great options for devs at any scale!
Blog: medium.com/@jeffonelson/jev-…
Billy Jacobson retweeted
this is the best explainer on jev I've seen
Jev from @typesafeai is basically the "Guess Who" of AI.
Instead of forcing a full LLM into a specific JSON format and waiting seconds, the outputs are strictly primitives (multiple choice, score, or yes/no) in ~100ms.
Demos by @heystefan_, @mattdesl, and @saragordic.
Jev from @typesafeai is basically the "Guess Who" of AI.
Instead of forcing a full LLM into a specific JSON format and waiting seconds, the outputs are strictly primitives (multiple choice, score, or yes/no) in ~100ms.
Demos by @heystefan_, @mattdesl, and @saragordic.
Jev from @typesafeai is basically the "Guess Who" of AI.
Instead of forcing a full LLM into a specific JSON format and waiting seconds, the outputs are strictly primitives (multiple choice, score, or yes/no) in ~100ms.
Demos by @heystefan_, @mattdesl, and @saragordic.
there's always another problem or another feature, we don't need to let ai drain us like this
we need to be applying the same work-life balance principles for ai development. perhaps a new metric needs to be around amount of context used per day and setting limits on that
yo no sé vosotros pero yo termino el día de trabajo exhausto, muchísimo más que antes de la IA
antes me tiraba todo el día debugeando una cosa, siguiendo el rastro paso a paso hasta encontrar el fallo, lo fixeabas... o estabas una semana entera focus solo con una feature nueva.
ahora estoy con herdr con 4 tabs y 4 terminales en cada uno, tirando agentes a fixear y otros creando. en paralelo. a la vez que estás de reuniones y respondiendo slack, con el pc al 99% de cpu porque ya no caben más worktrees
no escribo ni una línea de código pero acabo con la cabeza totalmente rota
The Google Developer Knowledge API is free of charge and only uses project quota:
- Search doc snippets: 100 req/min
- Fetch full Markdown pages: 100 req/min
- AI-generated answers from docs: 50 req/day
Happy prompting!
developers.google.com/knowle…
Claude Code + Dev Knowledge MCP.
My code, my models, not gonna be out of date.
I ask for Gemini and it gives me the version from two weeks ago, not the version from months ago when the LLM was trained.
Set this up: docs.cloud.google.com/docs/g…
An effective classroom is a system that you can master, and it's easy to play catch-up. The same principle applies for agentic software development.
Here are some lessons I took away from @andrewzigler's keynote at @TheLeadDev's LDX3 on why Ms. Frizzle would be a fantastic SRE.
What makes Antigravity, Claude Code, and Cursor feel so smooth is the harness engineering around the models.
Full Agent Harnesses episode on @GoogleCloudTech:
youtu.be/F8EZJAm9iO8