Block's New Dessa AI Model Aims to Chat Like a Real Friend—But With Awkward Results

Block Launches Dessa: A Conversation-First AI Model

Jack Dorsey's Block has entered the competitive AI conversation space with Dessa, a new artificial intelligence model trained exclusively on human conversation data. Unlike many AI models that learn from diverse internet sources, Dessa takes a different approach—building its foundation entirely on dialogue patterns between real people.

The announcement, covered by The Verge, showcases actual conversations between Dessa and humans, revealing an AI that sounds surprisingly human-like in some moments, yet oddly awkward in others. Dessa's lead, Ragavan Thurairatnam, shared examples of these interactions to demonstrate the model's conversational abilities and limitations.

What Makes Dessa Different?

The core distinction of Dessa lies in its training methodology. Most conversational AI models rely on vast datasets scraped from the internet, books, articles, and code repositories. Dessa, by contrast, was built "from scratch on nothing but human conversation." This specialized approach aims to create more natural, human-like dialogue patterns.

This training philosophy reflects a broader industry shift toward more specialized AI models. Rather than creating generalist systems that attempt to excel at everything, companies are increasingly building purpose-built tools optimized for specific tasks. For conversational AI, learning purely from human dialogue could theoretically produce more authentic interactions.

The Reality of Current Conversational AI

However, the leaked conversations also highlight the current state of conversational AI—even when trained on human dialogue, these systems can still produce responses that feel uncanny or slightly off. Users reviewing Dessa's sample conversations noted moments where the AI's responses, while grammatically correct and contextually relevant, lack the natural flow of genuine human interaction.

Why This Matters for AI Users and the Industry

Dessa's launch signals several important trends in the AI landscape:

• Specialization Over Generalization: Companies are betting that focused training data produces better results than broad, generic datasets.
• Conversational AI Remains Challenging: Despite years of development, creating truly natural-sounding AI conversation partners remains difficult.
• Transparency in AI Development: Block's willingness to share awkward conversations shows a more honest approach to AI limitations than competitor marketing claims.
• Increased Competition: With OpenAI's ChatGPT, Google's Bard, and other major players dominating conversation, new entrants like Block must find differentiation angles.

Implications for AI Tool Users

For users evaluating conversational AI tools, Dessa's arrival means more options to test and compare. If Block can refine the model to produce more natural interactions, it could become a viable alternative to market leaders. The company's transparency about current limitations also serves as a reminder that even well-funded AI initiatives still face technical hurdles.

This development also reinforces an important lesson: training methodology matters. Users looking for conversational tools should consider not just the company behind the AI, but how and on what data the model was trained. Dessa's conversation-specific approach may produce different results than general-purpose models, for better or worse.

The Takeaway

Block's Dessa represents an interesting experiment in specialized AI training, but it also exposes the persistent challenges in creating truly natural conversational AI. While the model shows promise, current conversations reveal a gap between marketing promises and actual performance—a gap that remains true across the entire AI industry. For users, this means continuing to evaluate tools critically and understanding that even innovative approaches still produce imperfect results. As the conversational AI space matures, products that balance realistic expectations with genuine improvements will ultimately win user trust.