Nirmal Krishnan
@0xnirmal
Followers
3K
Following
5K
Media
220
Statuses
1K
ceo @chakra_ai // fmr: hedge fund, @Microsoft // data, markets
New York, NY
Joined August 2020
the incredible part is I did meet vitalik once roaming the streets of LES. no security, bucket hat on - moisturized, happy - in his own lane, flourishing
5
0
34
when we started building this dataset 2 months ago, half of these logos didn't exist. cambrian explosion . better sft datasets + better environments => better CUA models. better CUA models => more white-collar automation. more white-collar automation => ?
Last week, we published a blog surveying the computer use agent landscape. Today, we're excited to share our interactive, live market map of the major players in the space.
2
1
24
I expect the big labs to continue marching down (1) because (3) is a narrative violation on the "one model to rule them all". where I think the OSS can win is in these modalities where latencies / model configuration really matter to the quality of the end consumer experience.
0
0
1
the ideal optima from my purview is (3) a small model that can run locally on the user's computer:. a) lower inference times from smaller model configuration.b) no network latencies.c) high degree of personalization .d) no captcha blockers.e) no privacy concerns.
1
0
2
in concept (2) is promising but challenges with latency walls for network operations (i.e. capturing screenshot, sending to server, waiting for response) are way too clunky for synchronous operation. add'l, most users are uncomfortable streaming their entire computer state to a.
1
0
1
while in practice (1) is "faster", it defeats the whole premise of the AI operating *your* machine.
1
0
1
the problem with (1) is that it lacks all context of the user's machine. with operator / agent, users get hit with login walls constantly. ergonomics get worse from "clanker blockers" - more websites are captcha walling discovery.
1
0
1
there are 2 ways to run computer use agents in production for a consumer and they both suck. 1) remote sandboxed environment collocated with inference servers (operator / claude style).2) on the user's PC (UI-TARS style) with inference over network.
1
0
7
RT @willccbb: @snowclipsed who said anything about 2026. in one year, gpt-6 and opus-5 will be popular and pretty good and widely adopted a….
0
1
0
amazing opportunity to be a part of a cracked team w/ @francedot @dhruv2038.
We’re hiring - in San Francisco and Spain (Remote or Madrid hybrid). Cua is building the infrastructure for general-purpose AI agents to safely and scalably use real computers and applications. 1/6
0
0
6