@puttasync

building @sixtyfourai | prev @helixdb, neurocognitive aging research, data science @ ucsd

 sf
putt.studio
Joined June 2021
  • Tweets479
  • Following226
  • Followers192
108Media
i built a restaurant sim where 3 AI rats run the kitchen from morning to close 🐀
6
2
20
2,112
putt retweeted
hud
@hud_evals9h
We partnered with @sixtyfourai to release RECON, a benchmark for people research. Agents get a name and a few identifying details, then have to find hard-to-find facts about that person: 514 facts across 140 real people. Across 22 systems, precision was close (56–86%), but recall wasn't (8–84%). Most misses were unanswered fields, not misidentified information. Sixtyfour XHigh ranked #1 at 70.6% weighted accuracy, and High ranked #2 at 54.3%.
1
5
1
39
8,569
We ran Sixtyfour on Exa’s people and company search benchmarks. 82.8% of our people search results matched the criteria. Exa’s published score: 63.3%. That is roughly 20 more relevant people for every 100 results. We came out ahead on all six measures in the chart.
Sixtyfour versus Exa’s published scores on Exa’s people and company search benchmarks. People search, 1,400 queries: first result matches 90.4% versus 72.0%; a match in the top ten 96.6% versus 94.5%; precision 82.8% versus 63.3%. Company search, 605 queries: first result matches 86.4% versus 61.8%; a match in the top ten 94.4% versus 94.2%; precision 79.6% versus 65.9%.

ALT Sixtyfour versus Exa’s published scores on Exa’s people and company search benchmarks. People search, 1,400 queries: first result matches 90.4% versus 72.0%; a match in the top ten 96.6% versus 94.5%; precision 82.8% versus 63.3%. Company search, 605 queries: first result matches 86.4% versus 61.8%; a match in the top ten 94.4% versus 94.2%; precision 79.6% versus 65.9%.

11
16
2
58
21,280
Introducing Atlas, the world's first AI investigator.
30
12
8
425
34,443
living with @georgecurtiss for 6 days has made me want to touch rust and learn about databases again
2
7
657
Introducing the world's first way to screen for fraud on platforms you shop from every day. Every marketplace runs on sellers it has never met. Sellers sign up under stolen identities, post fake listings, and slip back under new accounts after a ban. Now our agents catch them automatically, work that used to take manual investigation. Seller fraud used to slip through. Now it turns red with our agents.
9
11
1
56
73,320
graduating feels so bittersweet
7
26
1,045
sometimes you can simply make a chat interface. but no. you need to go beyond that.
3
206
introducing folk beta the personal ai that lives in your texts and quietly runs your life. built for everyone. just text a number.
60
24
12
345
96,008
how does one find housing in sf - do i need to build a bot or something
3
2
273
been away for a bit with a good career update to share soon. now back to building projects again!
3
6
440
first thing i want to build is a nice visualization of professors across the states and finding links between their research areas or papers. worked on this last year but never got around to finishing it
2
150
putt retweeted
arlan rakh
@arlanrApr 6

Turning the entire web into a filesystem.

Article · @arlanr
37
52
19
571
193,594
naturemaxxing
7
232
gonna figure out how to build a mini version of this
1
6
247
inspired by @joon_s_pk smallville
185
rat.simugent.tech/

rat kitchen

An AI-driven restaurant simulation where LLM-controlled rats run a kitchen with personalities, needs, and autonomous decisions.

rat.simugent.tech
188
i've been using using cc/codex the way boris describes it in his blog, and it's so much better than just letting the agents do it's own thing. i think at a certain point of yolo-ing, you just don't have a mental image of how things flow/work anymore when the codebase gets large. but this could just be a skill issue boristane.com/blog/how-i-use…

How I Use Claude Code | Boris Tane

The research-plan-implement workflow I use to build software with Claude Code, and why I never let it write code until I've approved a written plan.

boristane.com
1
6
276
been finishing up my finals and have been working on a few things like > integrating pi-mono/opencode into browser > building codebase retrieval agents in cloudfare sandboxes/daytona > then testing what i build against benchmarks
1
2
257
Load more