Block unveils Dessa, a conversation-only AI model trained from scratch on human dialogue

Block unveils Dessa, a conversation-only AI model trained from scratch on human dialogue

On October 23, Jack Dorsey's Block introduced Dessa, an AI model trained from scratch on nothing but human conversation — zero web-scraped text — and released several sample chats in which it comes across as an awkward, faintly uncanny new acquaintance. It is the first time the payments company has placed a self-built foundation model at the centre of its public story, a bid to own the conversational layer closest to its merchant and consumer business. The takeaway: Dessa's significance lies

Block announced Dessa on October 23 (US time), positioning it as a dialogue model built “from scratch on nothing but human conversation.” Ragavan Thurairatnam, who leads Dessa, posted one sample exchange: an intentionally mundane stretch of small talk in which the model asks follow-up questions and answers in a faintly stiff register, closer to a new acquaintance than an assistant. Other transcripts released alongside show the same profile — syntax and local logic hold up, but pacing, humour and reading the room lag well behind commercial assistants. Block did not disclose parameter count, training tokens, context window or inference cost, nor whether the model will be open-sourced or wired into Cash App and Square seller tools. No benchmark scores were published. Three facts are confirmable: the corpus contains no web-scraped text; the model was trained in-house rather than fine-tuned

Over the past two years the competitive question in foundation models has shifted from “how big” to “where did the data come from.” Copyright suits filed by publishers against model developers have pushed up the legal cost of scraped web text, while high-quality public text approaches exhaustion. Training from scratch on a single data class has therefore moved from purism to strategy: licensed, auditable, traceable corpora. Block's position explains the focus on conversation. Square seller acquiring, Cash App transfers and Afterpay instalments are all high-frequency, short-utterance, context-heavy interactions, where the quality of a support or merchant assistant converts directly into cost

Liu Gong's take: Dessa will not ship as a general-purpose product before the first half of 2027. It reads as a tone component built for Block's own rails, and the metric that matters is first-contact resolution in merchant support, not leaderboard position.
From a technical standpoint, dialogue-only training cuts both ways. Conversational corpora are dense with turn-taking cues — pauses, back-channels, topic shifts, repair moves — so a model learns rhythm before it learns facts. That is precisely where the published samples land: grammatically sound, emotionally slightly off, like a stranger making polite small talk. The same property caps the model. Ask how long a refund takes and it has no grounded answer without retrieval or a tool call. The absence of disclosed parameters, token budget and context window fits a model whose job is to sound right in a narrow setting rather than to know things.
On the business side, the arithmetic is simple. Square's merchant base generates a continuous stream of real customer conversations with clean consent chains, transcripts no general model vendor can buy. A small in-house model trained on that corpus can push inference costs well below third-party API pricing at volume and, more importantly for a regulated payments company, keeps merchant dialogue inside Block's perimeter — a constraint that usually outweighs a few points of quality in procurement.
For the industry, the framing is the interesting part. In industrial inspection, a pattern Liu Gong encounters repeatedly is that a general vision model scores well on standard parts and loses double digits the moment reflective metal enters the frame; the winner is usually a small model trained on defect images from a single line. Dialogue models look set to repeat that path. Within twelve months, expect vendors selling models trained only on support calls, only on clinical dialogue, only on court transcripts — pitched on auditability rather than intelligence. The first line of a model card will stop being the parameter count and start being the data-licensing statement.

  • Share of web-scraped text in Dessa's training corpus: 0%, against 60–80% for typical open chat models, where Common Crawl-style web data dominates the pretraining mix.
  • Disclosed specs: Block published no parameter count, token budget or context window, while contemporary open chat models usually state a 7B–70B range.
  • Benchmark scores: zero. Block showed sample conversations instead; mainstream model cards typically list a dozen or more benchmark results.
  • Distribution surface: Square's seller ecosystem runs into the millions and Cash App has held more than 50 million monthly transacting actives, giving Dessa both a deployment path and a corpus source.

For the industry, the near-term (three-month) effect is imitation: smaller teams will try single-source corpora for compact models and put data auditability into their pitches. Over twelve months, if Block wires Dessa into Square's merchant back office and publishes support-efficiency numbers, corpus provenance shifts from a compliance talking point to a procurement criterion — buyers will start asking whether a model's training-data licensing chain is clean.
For ordinary users, little changes soon. Dessa has no consumer entry point, so Cash App users will not suddenly gain an assistant. If it reaches merchant support, replies will sound flatter and less scripted, but factual accuracy will

Questions worth exploring:

  • Dialogue-only training buys turn-taking fluency at the cost of factual density — where the trade-off line gets drawn
  • Block turns merchant conversations into proprietary model assets, balancing compliance risk against inference cost
  • Whether corpus provenance becomes a product category, spawning models trained only on support calls or clinical dialogue

Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.

This item comes from The Verge AI (published 2026-10-23 19:32:39). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.


Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private

Leave a Reply

Your email address will not be published. Required fields are marked *

中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn