Grammarlys AI development and technology evolution

From: redpointai

Grammarly, a personalized AI assistant app for writing, boasts over 30 million daily active users and has raised over $400million,reachinga$13 billion valuation. The company has been building AI productivity tooling long before the recent generative AI wave, driven by the high-stakes nature of human communication and its diverse customer base [00:00:08]. Rahul Roy-Chowdhury, CEO of Grammarly, shares insights into the company’s approach to AI development and the future of communication.

Vision for AI in Communication

Rahul envisions a future where AI significantly reduces the drudgery of day-to-day work, allowing humans to focus on more meaningful communication, deeper connections, creativity, and synthesizing ideas [00:01:45]. He hopes for a future with less email and fewer documents, emphasizing quality over quantity [00:02:00]. The average person switches contexts 1,200 times in an average workday [00:02:29], and AI’s promise is to enable “Flow State” by making each conversation measurably more valuable [00:02:39].

Roy-Chowdhury stresses that it’s crucial for humans to actively shape the future of AI adoption to avoid a dystopian scenario where AI generates and consumes most content, leading to information overload [00:03:07]. Writing and communicating are integral to being human and should not be fully outsourced to AI [00:03:40].

Grammarly Product Evolution

Grammarly has existed for 15 years, launching in 2009, and has adapted through several technology waves, starting with rules-based natural language processing (NLP), moving to deep learning models, and now incorporating Large Language Models (LLMs) and generative AI (GenAI) [00:04:13]. The company prioritizes solving user needs and then applying the best available technology [00:04:32].

Communication Lifecycle and Grammarly’s Role

Grammarly frames the communication lifecycle in four stages [00:04:45]:

1. Ideation and Conceptualization [00:04:56]
2. Composition/Writing [00:04:58]
3. Revision and Polishing [00:05:04]
4. Comprehension (by the receiver) [00:05:08]

Historically, Grammarly focused primarily on the revision phase, helping users ensure correctness, adherence to style guides, appropriate tone, and brevity [00:05:23].

Future Direction with LLMs

LLMs enable Grammarly to “turbocharge” value for users in two main ways [00:05:55]:

1. Strategic Suggestions tied to Business Outcomes: Suggestions will become more aligned with desired user outcomes. For example, Grammarly might suggest adding “why people should go” to an event invitation or a clearer call to action in a board email [00:06:04]. Basic mechanics like correctness and tone will be increasingly auto-applied [00:07:08].
2. Full Communication Lifecycle Support: Grammarly will expand beyond revision to assist throughout the entire communication process [00:07:22]:
   • Ideation and Conceptualization: Helping users brainstorm and structure their thoughts [00:07:32].
   • Composition: Assisting in the writing process [00:07:34].
   • Comprehension: Summarizing long email threads or identifying action items [00:07:38].

AI Development Philosophy

Grammarly prioritizes careful development due to the importance of communication use cases. They do not simply “throw an LLM over the wall” but undertake extensive work to fine-tune models, run quality evaluations (evals), and ensure safety [00:08:52].

Quality and Safety Evaluations

For Grammarly, the required accuracy level is determined by user feedback [00:10:02]. They track user acceptance/rejection of suggestions and engagement with features [00:10:06]. Evaluation methods include [00:28:40]:

• External Benchmarks: Using general-purpose benchmarks relevant to their use cases [00:28:59].
• Safety Evals: Running internal evaluations based on extensive user feedback about false positives and safety issues [00:29:11].
• Side-by-Side Comparisons: Linguistic experts rate model output against human-generated text [00:29:36].
• User Experiments: Rolling out features to a small percentage of users to gauge real-world engagement and refine the product [00:29:50].

A notable example of learning from user feedback is the “tone detector” feature. While popular, it was found to be inappropriate in sensitive contexts, such as police reports on serious crimes, where suggestions to “sound more positive” were unwelcome [00:11:22]. This led to refining the product to suppress irrelevant suggestions in sensitive situations [00:12:16].

Best Practices for AI Use

Effective users of Grammarly connect communication assistance to strategic personal and organizational goals [00:13:09]. For instance, ModMed, an e-health company, uses Grammarly strategically to improve patient outcomes through enhanced communication in sensitive documents [00:13:31]. The focus is on measurable impact and achieving business objectives, rather than just adopting AI because it’s “cool” [00:14:04].

Impact of Large Language Models (LLMs)

The emergence of ChatGPT was a “watershed moment” for Grammarly, though they were already exploring GPT-3 [00:15:51]. The speed and scale of change, particularly in quality improvement, were surprising [00:15:59]. Grammarly views LLMs as a “huge enabler” for its mission, allowing them to leverage the latest technology to solve communication problems more deeply and meaningfully [00:16:29]. While early LLMs had lower precision compared to rule-based NLP and suffered from hallucinations, their quality has “phenomenally” improved [00:17:35].

LLM Idiosyncrasies and Product Roadmap

LLMs are still “special snowflakes” and are not plug-and-play interchangeable [00:23:15]. Each model operates differently, responds uniquely to prompts, and has distinct safety parameters, requiring significant work to fit for purpose [00:23:17].

Grammarly’s product roadmap balances addressing current shortcomings with exploring future capabilities [00:18:20]. Immediate priorities include ensuring model safety, which cannot be outsourced [00:18:45].

Future capabilities that are particularly exciting for Grammarly include:

• Multi-step Reasoning and Agentic Workflows: As models become more capable of complex, multi-step reasoning, Grammarly aims to orchestrate intricate communication flows. This would involve pulling context from various sources and guiding users through complex tasks, such as crafting a board email by integrating information from different teams and applying specific communication attributes [00:20:51]. This capability could be a game-changer for reducing the “drudgery” of cutting and pasting context [00:21:41].
• On-device Inference: With models becoming more efficient, moving inference to the edge (on-device) is being explored. This offers benefits in security, privacy, and significantly reduces latency, enhancing user experience and enabling “Flow State” [00:19:34].

Data Advantage and Personalization

Grammarly’s unique advantage lies in its vast amount of “relevant, contextual, fresh, high-quality user data” [00:25:51]. They process 75 billion user events daily [00:26:06], which is used to fine-tune and train models for different use cases [00:26:31].

This data enables personalization in several ways:

• Individual Voice: Allowing users to fine-tune the product to sound more like their own voice, with future automation of this process [00:26:40].
• Organizational Compliance: For businesses, Grammarly can ingest organization-specific knowledge such as style guides, brand tones, and corporate values, enforcing compliance and consistency across all internal and external communication [00:26:54]. This automates adherence to complex rules that might otherwise be buried in large documents [00:27:51].

AI Team Structure

Grammarly employs a dual approach to its AI team structure [00:34:10]:

• Core Research Team: Focuses on longer-term initiatives, exploring capabilities like on-device AI and building foundational infrastructure for future product integration [00:34:35].
• Embedded AI Engineers: Integrated into full-stack product and feature teams, working collaboratively on front-end, back-end, and AI capabilities for specific features [00:34:55].

AI in Education

AI presents a transformative, yet potentially misused, tool in education [00:39:51]. Roy-Chowdhury draws parallels to calculators and online code snippets, which were initially seen as “cheating” but later embraced as essential skills [00:40:07]. Educators are increasingly eager to partner with industry to equip graduates with critical AI skills for the workforce [00:41:00].

Grammarly is committed to responsible AI in education, demonstrated by features like:

• Citing AI Use: Allows students to cite how they used AI in their work, differentiating between simply generating an essay and actively engaging with the AI tool for feedback and deeper learning [00:41:16].
• Authorship: A new feature that provides “provenance” for every part of a document, indicating whether it was manually written, cut and pasted, or AI-generated [00:42:27]. This gives educators tools for transparency and setting acceptable boundaries for AI use [00:43:04].

Roy-Chowdhury believes AI serves as a “powerful tool to engage better with the material, to engage deeper, to get feedback in real time” [00:44:40]. Critically, AI is a “great leveler” and “democratizer of skills,” offering educational access and upskilling opportunities in parts of the world where traditional resources are limited [00:44:47].

Enterprise AI and Future Outlook

Grammarly sees AI as profoundly transforming the workplace, though it is a multi-year journey akin to the on-premise to cloud transition [00:36:03]. Enterprises must select trusted AI vendors for a long-term partnership [00:37:02]. While there’s much excitement, quantifiable productivity gains from AI are still somewhat “elusive” outside of specific areas like software engineering and code generation [00:37:23].

Grammarly emphasizes demonstrating measurable value: the average user in an organization saves 19 days per year, equivalent to a month’s worth of work [00:38:03]. This focus on “repeatability and measurability” ensures AI delivers tangible benefits [00:38:33].

Overhyped vs. Underhyped AI

• Overhyped: Chat interfaces, which Rahul views as “subpar command line interface[s]” that he hopes will disappear [00:46:41].
• Underhyped: AI’s potential to upskill and uplevel people globally, acting as a powerful “democratizer of skills” and “force multiplier” for individuals struggling with certain capabilities in the workforce [00:46:52].

Surprises and Company Evolution

The biggest surprise in building AI features for Grammarly has been the profound resonance and impact of the tone detector feature on users’ lives [00:47:59].

Rahul has changed his mind on the distinction between consumer and enterprise businesses. Initially, he saw them as separate, but now views them as a “seamless customer journey” where users might start with free or premium consumer versions and scale into enterprise deployments [00:49:11].

Exciting Areas of AI Outside Grammarly

Rahul is particularly excited about AI for healthcare, citing AlphaFold’s impact on drug discovery and ModMed’s use of AI to improve patient outcomes as “game-changing” and “awesome” uses of AI with real, tangible, life-changing impact [00:50:07].

Grammarly can be found at grammarly.com and across various social media platforms [00:50:48].