@kevin_loo

deploying on premise ai and full software replacements using apple silicon MTS @ Sendblue | Ex-CTO @ Easy Street Offers

 SF, NYC, SZ, HK, LA
loo.ski
Joined August 2011
  • Tweets579
  • Following647
  • Followers269
Replying to @thsottiaux
you can do this locally, btw github.com/looskis/tarski

GitHub - looskis/tarski: Reading diffusion language models for decisions: same OpenJev model, one...

Reading diffusion language models for decisions: same OpenJev model, one pass instead of 3.65, +5 points, calibrated, no training. Plus cheap task branches on a frozen encoder. - looskis/tarski

github.com
2
5
137
23,596
this is where we're at nowadays jetkvm connected to an instance, controlled from a vnc over tailscale on my laptop from 300 miles away
30
i get a lot of my work done at night bc i swear usages go down slower at night
1
1
31
uhhhh dot?
3
46
Yes @supabase it runs doom
31
i won the google gemini track at the @supabase select hackathon it solves the problem of needing to download gigabytes of climate data just to use a few megabytes of it for multi-year analysis it's free until my credits runs out verdant-ai.com/

Climate Data API & MCP for AI Agents | Verdant AI

Query climate and agricultural data through an API or MCP. Get normalized NOAA and SILO data in JSON or CSV, with source provenance and versioned datasets.

verdant-ai.com
1
3
120
fun fact: if you're connected to yc wifi, and you upload more than 10GB of data, you get disconnected i ran into this during the hackathon trying to send climate data to my agents on mac mini to confirm findings
37
Kevin W. Loo retweeted
Supabase
@supabaseOct 2
Supabase local dev just got much better for agents • Run locally without Docker • Multiple Supabase instances per repo/worktree • Declarative Schemas 2.0 with pg-delta generating migrations • Auth, API & storage config now lives in code
86
77
31
1,203
100,853
I told Steve Wozniak that “I’m going to open source enterprise software and run it on Apple silicon. I hope one day you’ll hear about me in the news” He said “I hope I will” Definitely local maxima of my life Thanks #supabase !
6
189
Supabase is giving $1m in equity to community members
1
1
69
how many codex credits do yall have left from the 62k grant?
1
10
157
barclay says ai reached 50% of its developers what's the other half doing?
1
3
89
so far my exp with dot - indistinguishable from muse - no different from codex with computer use - value added are probably: it's sandboxed, it's free with existing plan, it's cute
1
63
I'm Kevin 👋 deploying on premise ai and full software replacements using apple silicon, MTS @ Sendblue | Ex-CTO @ Easy Street…, based in SF & NYC 🇺🇸 Looking to meet marketers, founders, early users and investors! Just put myself on the map 👇 makermap.lol/kevin_loo?v=9fg… xitter.cf/verbove/status/2104175…

Kevin W. Loo · building Make Your Firm AI-Native, With AI Built On Site · San Francisco, United...

#649 of 5,205 makers on makermap.lol. 265 points: 5 comments, 0 likes, 0 reposts.

makermap.lol
Martijn Verbove@verboveSep 27
drop your startup link 👇 i'll add it to the map
5
3
257
..
Tibo
@thsottiauxSep 29
Announcing dots. Dots work 24/7 for you, learn from your feedback, have their own computer, browser and can be connected to over 4k apps in our ecosystem. They’re powered by Astra our best model yet. Included in your Pro plan, without drawing down on any of your usage. You can even call a dot while it’s working. We’re getting you started with your primary dot today and soon you’ll be able to create entire teams of them.
1
4
131
openjev caught up to jev few days after launch i beat openjev by 5% in accuracy, made it nearly twice as fast, all without fine tuning the model github.com/looskis/tarski
3
1
8
5,847
I made the repo a harness so you can do the same, and if you're patient you don't even need to rent a gpu
26
Three things nobody had checked: • The random token OpenJev drops into each answer slot is the wrong input. The vocabulary-mean embedding beats it, and beats sixteen averaged reads, in one pass. • Question order is the biggest error source on the canvas. Reorder five questions and one answer in four changes. OpenJev's default order is the worst one. • The order effect comes from the encoder reading the state after the question list. Put the state first and it mostly disappears, with no labels.
1
81
Basis says GPT-6 Astra finished a 50-tab tax workbook in half the time of GPT-5.6 Sol question is, what's left for basis once tax companies release an mcp?
2
71
xai team bots lets you share context with team but keeps individual chats private. what about the memory snippets that they produce?
2
53
The useful pattern in Proaction’s Codex story isn’t “AI replaces engineering.” It’s a tighter sales-to-build loop: customer context becomes a tailored demo, then a clearer engineering brief. i've sent out three proposals this past week. one is seriously reviewing openai.com/index/proaction/

Proaction boosts sales 60% and saves 75+ hours with Codex

With Codex, GPT-Live-1, and GPT-6 Astra, Proaction builds, operates, and sells modern fleet management faster.

openai.com
1
3
57
Load more