Powerful and Limited, Rishabh Singh Builds Toward AGI in New York

Rishabh Singh, a New York founder who runs the Helix Mind research lab and is co-founder and CTO of the social app JoinMeIn, argues that anyone building toward artificial general intelligence should design for what today's machines cannot do.

His New York story starts on a Manhattan sidewalk. In an essay on building here as an immigrant founder, he remembers standing outside a co-working space in one of his first weeks, holding "a coffee I couldn't really afford yet," with half his mind on a visa timeline and the other half on a product roadmap. He came, he writes, because the people, capital and problems he wanted to work on were concentrated in this city. The city, he adds, "asked something extra" of him in return.

What Does He Build?

Singh reached AI by way of the trading floor. His site says he spent four years engineering C++ trading systems for global investment banks, including "long nights on cache alignment and thread affinity," the low-level tuning that decides how quickly a processor reaches its data. Markets do not forgive overconfidence, he writes, and neither does production AI.

The work now runs on two tracks. Helix Mind, his research lab, works on language model architectures and training systems. His companies are where that research meets ordinary users. They include JoinMeIn, a meet-up-tonight app published on the App Store by Joinmenow Inc., and Nova Reach, a marketing platform that lists its address at 1600 Broadway, a few blocks up from a free show at 1500 Broadway that revisits Times Square's adult past. He writes that patents he filed sit underneath the shared AI layer his products run on.

What Does He Mean by AGI?

His answer is a librarian. In a September explainer on LLMs and AGI, he describes a large language model as a librarian who has read nearly everything and answers fluently by recombining it. AGI, in his telling, is the same librarian walking out of the building to start a business, pick up a skill no one taught them and reason through a problem no book covers, without being retrained. It does not exist yet, he writes, and is not close.

The distinction is practical for him, not philosophical. A team that quietly believes it is building on something near AGI, he argues, trusts outputs it should not and skips guardrails it needs. A team that knows it is building on a language model adds human review, escalation paths and graceful fallbacks. He admits he has caught himself believing the hype.

"The language we use shapes the products we build."

How Do You Build Toward It Safely?

Slowly, by his account, and with a lot of testing. In a second September essay he describes AGI research as closer to astronaut training than a breakthrough montage. In his telling, a typical day began with the results of experiments run overnight. He spent whole days designing a single test to separate real generalization from a shortcut that only looks like it, and read wrong answers one at a time. The honest version of the work, he writes, is a team "trying very hard not to fool themselves."

He carries that into product. In an October post, he recalls a demo that had run smoothly three times before he tried to break it on purpose. It failed within twenty minutes. Two weeks of deliberate attacks turned up four failure modes, and the feature shipped a month late. Before any AI feature ships now, he writes, someone's job is to make it fail.

His shortest version of the idea is a rule from another September essay. Build as if today's systems are "genuinely powerful and genuinely limited at the same time," and make them fail safely however capable the models become.

**Suite 646's read.** Over eight days in September, Singh posted three essays about AGI. All three end up in the same place, with guardrails, evaluations and human checkpoints built for what the system cannot yet do.

When Does the Work Happen?

Mostly after dark. In his published schedule, the hours before 6:30 a.m. go to reading and code review, and twice a week he protects a block for primary sources, papers rather than summaries. The evening is the deep shift.

"Evenings are reserved, as consistently as I can manage, for whichever single company has the hardest unresolved problem that week."

He does not pretend the arrangement is restful. "Do I sleep enough? Some weeks, barely," he writes. Across town the night runs on its own clock, from the Blue Note's two-shows-a-night October to the Lower East Side and Williamsburg blocks that JoinMeIn's New York page names as its turf. Singh's version of a late set is quieter. It is one problem, one evening, and a test built to prove him wrong.

"This city is full of people building on these models at two in the morning. The ones who last are the ones who find out what the model can't do before a customer does," Singh says.

*Rishabh Singh is a client of Media Morpheus, the PR firm owned by the publisher of Suite 646.*