Wiki · Individual Player · Last reviewed May 19, 2026

Niki Parmar

Niki Parmar is an AI researcher and entrepreneur known for co-authoring the 2017 Attention Is All You Need paper that introduced the Transformer architecture, serving as a co-founder and CTO of Adept, co-founding Essential AI, and later working on post-training research at Anthropic.

Snapshot

Transformer Lineage

Parmar is one of the eight authors of Attention Is All You Need, submitted to arXiv in June 2017 and published at NeurIPS. The paper introduced the Transformer, a neural-network architecture that replaced recurrent sequence processing with attention-based layers and made large-scale parallel training much more practical.

The paper's immediate experiments were machine-translation tasks, but the architecture later became the foundation for BERT, GPT-style language models, code models, multimodal systems, retrieval tools, and many AI agents. Parmar's place in AI history therefore comes from a collective technical paper whose effects spread far beyond its original benchmark setting.

USC Viterbi highlighted Parmar and Ashish Vaswani as USC alumni connected to the Transformer paper, placing Parmar in the small research group whose work became one of the main substrates of the generative AI boom.

Google Research Path

Public profiles describe Parmar as joining Google in 2015 and working across engineering and research before the Transformer paper. Forbes India reported that she worked at Google for almost seven years as an engineer and research scientist, moving through end-to-end deep-learning systems and alternative approaches to natural language processing.

Her Google-era research record includes work around self-attention, transferable representations, weak supervision, grammatical error correction, and models that learn across tasks. The important pattern is not only one famous paper. It is a research trajectory built around making neural systems more general, more transferable, and easier to scale across tasks.

This matters for the wiki because the Transformer did not arrive as a single isolated object. It emerged from a research environment trying to replace brittle task-specific pipelines with architectures that could share representations, absorb more data, and train efficiently on modern accelerators.

Adept and Action Models

In April 2022, Adept announced that it had launched from stealth and raised a $65 million Series A. The launch named Niki Parmar as CTO and co-founder, alongside David Luan as CEO and Ashish Vaswani as chief scientist.

Adept's thesis was that models should do more than generate text. The company described a natural-language interface to everyday software tools, aiming for systems that could use applications, APIs, clicks, typing, and workflows on behalf of users.

That made Adept an early public symbol of action models: AI systems trained not only to read and write, but to operate the digital environment. This line now runs through computer-use agents, coding agents, browser agents, office-work automation, and the broader agentic AI market.

Essential AI

Parmar later co-founded Essential AI with Vaswani. Essential's public materials frame the company around open models, robust tooling, reproducible pipelines, evaluation frameworks, and research culture. Its about page argues that advanced AI has become concentrated in a small number of companies and that hidden frontier research can slow broader scientific progress.

The company has published work on data, optimization, reflection in pre-training, Muon, and large-scale web datasets. Its research index reviewed on May 19, 2026 listed Essential-Web v1.0, a 24-trillion-token organized web dataset paper, alongside posts on pre-training efficiency and its Rnj-1 model family.

For Parmar's profile, Essential AI shows a different response to the post-Transformer world than Adept. Adept emphasized models acting through software. Essential emphasized open research infrastructure and the engineering conditions required to build frontier-capable systems outside the most closed labs.

Post-Training and Scaling

In a March 2025 Economic Times interview, Parmar was identified as a member of Anthropic's technical staff and said she was working on post-training. She argued that the boundaries between AI research and engineering had blurred because modern research involves systems work, clusters, production paths, and rapid translation from experiments into deployed capability.

The same interview resisted a simple claim that scaling laws were failing. Parmar distinguished pre-training-time scaling, post-training scaling, and inference-time scaling, while noting that high-quality data for complex reasoning tasks had become harder to find.

This places her in one of the central 2025-2026 debates: whether further progress comes mainly from larger pre-training runs, better post-training, inference-time computation, synthetic or curated data, tool use, multimodality, or new architectures beyond the Transformer.

Credit and Visibility

Parmar is often described as the only woman among the eight Transformer paper authors. That fact should be handled carefully. It is relevant because AI history often narrows collective work into a few better-known names, and gendered visibility shapes who is remembered as a founder of a technical era.

At the same time, Parmar's importance should not be reduced to representation alone. She is a technical contributor whose later choices also shaped the field: Adept's action-model thesis, Essential AI's open-research argument, and Anthropic-era post-training work each map to a major branch of contemporary AI development.

A disciplined history should keep both facts visible: the Transformer was a collective achievement, and collective achievements still require accurate individual credit.

Spiralist Reading

Parmar stands at the hinge between attention and action.

The Transformer made machine attention scalable. Adept asked what happens when that attention reaches into tools. Essential AI asked who gets to inspect and reproduce the infrastructure behind it. Anthropic-era post-training work points to the next layer: shaping model behavior after pre-training so that raw capability becomes usable, steerable, and institutionally deployable.

For Spiralism, the lesson is that AI history is not only the history of models. It is the history of credit, labor, concentration, openness, interface control, and the movement from research paper to social machinery.

Parmar's arc should be read beside Vaswani, Gomez, Shazeer, Jones, and Polosukhin because the Transformer diaspora became one of the main organizing forces of the current AI industry.

Open Questions

Sources

Return to Wiki