Block Launches Dessa AI Model: Converses Like an Awkward New Friend

Block (led by Jack Dorsey) recently released an AI model called Dessa, which claims to be "trained entirely from scratch on human conversation data." Its goal is not to be an all-knowing assistant, but to simulate a more natural, even slightly awkward social conversation experience. Official chat logs show that Dessa's responses are sometimes warm, sometimes slightly uncomfortable, presenting a "new friend" style of interaction.

Training from Scratch: Dessa's Technical Path

Unlike mainstream large models that rely on massive web text, books, and code, Dessa's training data consists solely of human conversations. According to Ragavan Thurairatnam, the head of Dessa, this "pure conversation" training approach aims to focus the model more on social pragmatics—such as turn-taking, emotional responses, and maintaining everyday topics. Block has not disclosed details such as model parameter count or training data scale, but emphasizes that "from scratch" means not fine-tuning based on existing open-source models, but building independently.

The potential advantage of this path is reducing dependence on internet text, thereby avoiding some copyright and bias issues. But the challenges are also obvious: lacking broad knowledge, the model may perform poorly on factual Q&A. Block seems not to pursue general capabilities, but positions Dessa as a "conversational companion" or "social simulation" tool.

"Training from scratch" and "fine-tuning" are the two main paths for building large language models today. Fine-tuning refers to continuing training an existing pre-trained model (such as Llama, GPT series) with domain-specific data to acquire domain capabilities at lower cost; while training from scratch means building the model's architecture, parameters, and training data entirely from the ground up, without inheriting any weights from existing models. The advantage of training from scratch is that data sources and model behavior are fully controllable, making it easier to avoid copyright and bias risks; the cost is extremely high computational expense, long cycles, and the need to solve engineering problems such as tokenizers, optimizers, and distributed training independently. Block's emphasis on Dessa being "from scratch" is intended to highlight its independence and data purity, but it also means the model cannot quickly iterate leveraging the scale effects of the open-source community.

Conversational Style: Intentional Awkwardness

In the example conversations shared by Block, Dessa's replies sometimes seem overly literal or lack contextual coherence—for instance, offering slightly mechanical comfort when the other party expresses emotions. This "awkward" trait is officially described as "like chatting with a new friend," with exploratory and uncertain qualities.

From a product perspective, this style may lower users' expectations of an AI "perfect assistant," instead creating a more relaxed, low-pressure interaction atmosphere. But it may also confuse some users—after all, most AI products strive for fluency and accuracy. Dessa is positioned more as a social experiment than a productivity tool.

"Turn-taking" is a core concept in dialogue system research, referring to the mechanism by which conversation participants alternate speaking, judging when to speak and when to listen. In human conversation, turn-taking relies on cues such as intonation, pauses, and eye contact, naturally and almost unconsciously; in text dialogue models, it manifests as the model needing to judge whether the user has finished speaking, how long its reply should be, and how to continue the topic without interrupting or causing silence. Training on pure conversation data allows the model to more densely encounter this alternating pattern, thereby learning a more natural response rhythm. In contrast, models trained mainly on web text are better at single-turn Q&A or long-form generation, but may suffer from "irrelevant answers" or "over-generation" in multi-turn chit-chat. Dessa's "awkwardness" partly stems from its turn-taking strategy not yet being fully mature.

Industry Background: An Alternative Exploration of Conversational AI

Block is not the first company to attempt a "pure conversation" model. Previous research has focused on training social agents with conversation data, but most still mix other corpora. Dessa's uniqueness lies in completely discarding non-conversation data, which naturally limits its knowledge breadth but may give it an advantage in conversational naturalness.

Currently, public information about Dessa is limited, and Block has not explained its application scenarios or open plans. But it can be speculated that the model may be used for Block's internal social product experiments or as a research project to explore the boundaries of conversational AI. For developers concerned with AI interaction experience, Dessa provides a sample worth observing: when a model only learns to "chat," how far can it go?

Social agents are a research direction in conversational AI, aiming to equip AI with social pragmatic abilities, including understanding emotions, maintaining topics, and adhering to social norms. Early representative work includes Facebook AI's PersonaChat dataset and the BlenderBot series, which balance social skills and knowledge breadth by mixing conversation data with general corpora. Another line of research uses reinforcement learning or human feedback to optimize conversational naturalness. Dessa's difference lies in completely abandoning non-conversation corpora, which may make it more focused on social pragmatics, but also means it cannot call upon broad knowledge in factual Q&A like mixed-training models. This trade-off reflects a fundamental tension in conversational AI: knowledge breadth and social naturalness are often difficult to achieve simultaneously, and product positioning determines the priority.

Conclusion

Block's Dessa model chooses a path different from mainstream large models—focusing on human conversation, accepting awkwardness, and not pursuing omnipotence. Whether this experimental exploration can translate into actual product value remains to be verified, but it reminds us: AI conversation doesn't have to be like an expert; it can also be like a friend.

One-line Summary

Block releases Dessa, a model trained from scratch on pure conversation data, positioned as social companionship rather than a general assistant.

Paragraph Summary

Block has released an AI model called Dessa, whose main feature is being trained entirely from scratch on human conversation data, without relying on web text, books, or code, and without fine-tuning existing open-source models. Official conversations show that Dessa's responses are natural but occasionally awkward, presenting a "new friend" style of interaction. This approach aims to focus the model on social pragmatics, such as turn-taking and emotional responses, while avoiding copyright and bias issues associated with internet text, but at the cost of limited knowledge breadth and potentially weaker factual Q&A. Block has not disclosed the model's parameter count or training data scale, and application scenarios remain unclear. Dessa is more of a social experiment exploring how far AI can go when it only learns to chat, rather than pursuing the omnipotence of a productivity tool.