Dessa AI 推特 · X (Twitter) 热门推文与学界评测精选
追踪全网对 Block Dessa AI 的实时反响——从 Jack Dorsey 极简的三词力挺、项目负责人 Ragavan Thurairatnam 关于“音频领域 GPT-3 时刻”的长文，到 GB200 单机柜训练与 M4 Max 本地推理的硬核解析。
@jack · Block Head & Cofounder, Block, Inc.
“We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. block.xyz/dessa”
Block 联合创始人 Jack Dorsey 官宣文案：“我们相信声音是与智能交互最自然的方式，因此我们打造了有史以来最真实的原生全双工语音模型 Dessa。与它交谈就像与真人对话……充满情绪、细微差别、停顿与自然韵律。”
@RagavanThurair · Dessa AI Lead at Block (Former Cofounder, Dessa Toronto)
“We made the first AI model you can have a realistic conversation with that was trained entirely on speech! Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This is a GPT-3 moment for speech.”
Dessa 项目负责人 Ragavan Thurairatnam 核心帖：“我们做出了全球首个完全仅靠语音训练、能进行真实实时对话的 AI 模型！Dessa 是可在消费级硬件运行的全双工语音模型——这是属于语音的 GPT-3 时刻。”
@RagavanThurair · Dessa AI Lead at Block
“What happens when you put Dessa on the phone with GPT-Live-1? Nobody steers it. Listen to how Dessa drops natural backchannels ('okay', 'right', 'that's relieving') while GPT-Live delivers textbook paragraphs.”
当让 Dessa 直接与 GPT-Live-1 通话会发生什么？全程无人工干预：听听 GPT-Live 在背诵教科书式段落时，Dessa 如何像真人一样在换气间隙自然附和（“okay”、“right”、“那我就放心了”）。
@RagavanThurair · Block Deep Learning Lab Engineering Note
“Trained from scratch on a single NVIDIA GB200 NVL72 rack. Runs locally faster than real-time on a MacBook Pro M4 Max or better. Zero text tokens. Hundreds of times fewer words than text LLMs.”
X 与 Hacker News 工程师社区热议的算力指标：仅用单台 NVIDIA GB200 NVL72 机架从零完成训练，词汇量比文本大模型少几百倍，且能在 MacBook Pro M4 Max 上超实时本地推理。
@RagavanThurair · Acoustic Representation & Prosody Breakdown
“Say 'I didn't say she stole my CUDA code' out loud, stressing a different word each time. Same 7 words, 7 completely different meanings. Run speech through STT -> Text LLM -> TTS and all of that acoustic truth is destroyed.”
X 上转发最广的“CUDA 代码重音测试”：大声读出“I didn't say she stole my CUDA code”，每次重读不同单词，同样一句话会产生 7 种截然不同的含义——而传统的语音转文字（STT）管线会把这些声学潜台词彻底抹杀。
@RagavanThurair · Early Access Call for AI Safety Researchers
“AI keeps gaining power and independence, and it's still nearly deaf to the emotion in our voices. Imagine being in serious pain while a voice AI has no idea and just keeps going. That's the safety problem. Calling all safety researchers!”
为什么 Block 优先向 AI 安全研究员开放早期访问？“试想当你处于剧烈痛苦中，语音 AI 却因为听不出声音里的颤抖而自顾自说下去——这就是安全隐患。”Dessa 团队正公开招募安全学者共同红队测试。
@RagavanThurair · Unsteered Audio Sample Reaction
“It laughs the moment something's funny, mid-sentence, the way people do. Other models wait for you to finish, then laugh on cue. Ask it to speak up and it gets louder. Nobody taught it that.”
X 上播放量最高的实测细节：传统语音模型会等你把话说完再按指令干笑两声，而 Dessa 会在好笑的瞬间直接笑场；你让它“大点声说”，它就会自然提高嗓门——没有任何人为它写过这条规则。
@RagavanThurair · AI-to-AI Collision & Turn Recovery Analysis
“In the Noisy GPT Live sample at 4:07, both AIs speak at the same time. GPT-Live-1 triggers a hard stop ('Sorry, go on.'), while Dessa smoothly replies: 'Oh, no, that's perfectly fine. I like what you're saying...'”
X 技术博主对《Noisy GPT Live》第 4 分 07 秒“抢话碰撞”的逐帧拆解：当两个 AI 同时开口时，GPT-Live-1 触发硬中断生硬喊停（“Sorry, go on.”），而 Dessa 却像真人一样自然接话：“哦没事没事，我很喜欢你刚才说的，你继续……”
项目负责人 Ragavan Thurairatnam 证明：仅靠在原始人类对话音频上做自回归预测，无需一行文本就能涌现出流利对话与推理能力。
作为直接向 @jack 汇报的神秘深度学习实验室，Team Dessa 的首次公开亮相迅速引爆开源与端侧 AI 社区。
开发者纷纷惊叹：能在 MacBook Pro M4 Max 上脱离云端、快于实时运行的全双工语音模型，将彻底改写语音助手的隐私与成本结构。
无论是吸气声、轻笑还是半句改口都逼真得令人起鸡皮疙瘩，这也解释了为何 Block 选择先联合安全学者评估风险再决定权重开源。