Block's Dessa runs locally only on a MacBook Pro M4 Max or better, and the model carries the name of the Toronto startup Square bought in February 2020

A 59.2 s portrait clip, 720x1280, delivered as an MP4 in a QuickTime/MOV container (format_name mov,mp4,m4a,3gp,3g2,mj2), video stream HEVC Main profile at 30/1 fps carrying 1,776 frames at 301,856 bps, audio stream AAC HE-AACv2 at 44,100 Hz stereo and 64,175 bps across 1,278 frames, overall bitrate 372,844 bps, filesize 2,759,049 bytes, yt-dlp format id bytevc1_720p_372844-1; posted at epoch 1792926563, which is 2026-10-25T11:09:23Z, by @20vc_tok under the nickname 20VC, audio credited as "original sound" by 20VC, with a 136-character description that reads "Jack Dorsey, the founder of Twitter, just unveiled Dessa An AI voice model that sounds almost indistinguishable from a real human being…"; engagement at capture time (metadata.json mtime 2026-10-25 18:02:15 local) was 39,000 views, 1,254 likes, 24 comments and 161 reposts. I read 16 frames: frame_001, frame_002, frame_003, frame_005, frame_007, frame_010, frame_011, frame_012, frame_013, frame_014, frame_017, frame_021, frame_025 and frame_030 from the 2 s set in tokdoc_frames/ (frame N sits at (N-1) x 2 s), plus two single-frame extracts I pulled at 17.0 s and 25.0 s to cover the cut into and out of the screen recording. The transcript is 174 words across 10 Whisper segments. The clip opens on a split screen: a black lower-third card reading "THIS NEW AI VOICE MODEL IS INDISTINGUISHABLE FROM A REAL HUMAN", with "INDISTINGUISHABLE" set in red, sits above stock footage of Jack Dorsey in a grey t-shirt and gold chain against a dry hillside, while the narrator, a man in a pale grey quarter-zip holding a wireless lavalier capsule up to his mouth, speaks from an inset window in the lower right (frames 001, 002, 003 at 0 s, 2 s and 4 s). A blue Twitter bird icon overlays the Dorsey footage at 0 s and 2 s and swaps to a black nine-square grid, the Square logo, at 4 s. Word-by-word burned-in captions run across the center: "THE MAN WHO" (0 s), "CONTROLS" (2 s), "POWERHOUSE" (4 s), "OF ELEVENLABS" (8 s), "NATURAL" (12 s), "SNIPPET HERE" (17 s), "TO SEE" (18 s), "CRAZIEST PART" (26 s), "SO IT" (32 s), "IN REAL TIME." (40 s), "DOES TWITTER." (48 s) and "YOUR VOICE." (58 s). From 18 s to 25 s the narrator cuts to a full-width screen recording of Block's own Dessa demo player on a black background: a lowercase handwritten "dessa" wordmark at the top, two stacked waveform rows labelled "Human" (with a photo avatar) and "Dessa" (with the Dessa glyph), five lime-green pill markers along the Dessa track reading "Laugh", "Blown away", "Laugh", "All the secrets" and "Speaking louder", and a running caption line underneath that reads "Human Winter is coming for sure, but you" at 18 s, "this year?" at 22 s, "Dessa Yeah, I can see that." over "Human Well, anyway, I've had such an exciting" at 24 s, and "Human Well, anyway, I've had such an exciting" at 25 s; the embedded sample the narrator plays ("is that kind of a sore spot this year yeah i can see that") is therefore sourced on screen to Block's published demo, not to an unattributed recording. No terminal, no model card, no repository name, no URL and no file size appear in any frame I read, so nothing in the video demonstrates the local-execution claim it makes. The spoken argument is that Jack Dorsey founded Twitter, now controls Block, and has shipped a speech model called Dessa that beats ElevenLabs and OpenAI on naturalness and runs locally on a MacBook in real time. The last transcript segment spans 57.46 s to 58.86 s, inside the 59.187007 s audio stream, so there is no Whisper overrun and every quoted line is genuine; the names, however, are mangled, with "11 labs" standing in for ElevenLabs and "dessa" rendered lowercase.

The spelling is correct, and the name is recycled from a company Square acquired in February 2020

Whisper writes the product as lowercase "dessa" at 25.36 s, and the only on-screen spelling is the lowercase handwritten wordmark in Block's demo player at 18 s through 25 s. Block's own page settles it: the product page at block.xyz/dessa presents the model as "Dessa" and describes it as "the most authentic full-duplex speech model ever built", operating as "audio in, audio out". So the video has the name right.

What the video never mentions is that this is the second thing Dorsey's company has called Dessa. On February 7, 2020, Square published a press release announcing it was acquiring Dessa, a Toronto machine learning company. BetaKit's report the same day records that Dessa was founded in 2016 under the name DeepLearni.ng by Stephen Piron and Eric Kin-Ho Lee as co-CEOs, Vince Wong as COO and Ragavan Thurairatnam as head of machine learning. Square's release stated that "The Dessa team will continue working out of their office in Toronto, with no immediate changes to their existing products, customers, or business operations."

Block's Dessa page closes the loop. It describes the model as the work of a small AI research lab at Block founded by former members of Dessa, the Toronto deep learning startup, and lists Ragavan Thurairatnam among the researchers alongside Louay Hazami, Hashiam Kadhim, Qi Liu, Rayhane Mama, Aydin Polat and Marc Tyndel. Unite.ai, reporting on the launch, identifies Thurairatnam as the lab's lead and the person who posted the announcement on X. The 2020 acquisition and the 2026 model are therefore the same people and the same name, six years apart. The video's framing of this as a new Dorsey venture skips the acquisition that supplied both.

Block unveiled Dessa; Dorsey promoted it, and the lab reports to him

The description asserts that "Jack Dorsey, the founder of Twitter, just unveiled Dessa", and at 0.00 s the narrator opens with "jack dorsey the man who founded twitter and now controls block". The primary source credits the company and the lab, not Dorsey as the announcer. Block's Dessa page attributes the work to the research lab and notes the team reports to Jack Dorsey. Unite.ai reports the announcement was posted by Ragavan Thurairatnam, the lab's lead, on October 23, 2026. CoinDesk, writing on October 24, 2026, reports that Dorsey personally promoted it on X, quoting him as saying "Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence."

The announcement date is the one figure I could not pin to a first-party source. Block's Dessa page carries no launch date; the only date on it is an October 21, 2026 effective date on the privacy notice. TechCrunch, published October 24, 2026, puts the announcement at October 23, 2026, and Unite.ai agrees. CoinDesk, also published October 24, 2026, places it on October 24. The video went up on October 25, 2026 at 11:09:23Z, one or two days after. I treat October 23 as the better supported date because two independent reports give it, and I mark the exact day Unverified against first-party evidence.

CoinDesk also calls Dorsey "Block's co-founder and CEO". Block's 2026 proxy statement does not use that title. It states that Dorsey "has served as our principal executive officer and as a member of our board of directors since July 2009, having previously served as our Chief Executive Officer and President from July 2009 until his title changed to Block Head as of April 2022."

"You can run this locally on your MacBook" is a first-party claim, narrowed to one hardware tier

At 29.4 s the narrator says "the craziest part of this whole story is the fact that you can run this locally on your macbook", and the burned-in caption at 26 s reads "CRAZIEST PART". Block's page supports the shape of the claim and tightens it: "And it runs locally on a MacBook Pro M4 Max or better." That is not any MacBook. It excludes every MacBook Air, every Intel MacBook Pro, and the base, Pro and Max-adjacent tiers below M4 Max. CoinDesk renders the same requirement as "compact enough to run locally on a MacBook Pro equipped with an M4 Max chip or better", and TechCrunch describes it as consumer-grade hardware "such as a MacBook M4 Max or equivalent". I treat the requirement as verified and the video's generalisation to "your macbook" as a partial correction.

Two things I could not verify against any first-party artifact. Block discloses no parameter count and no model size on the Dessa page, so "compact" has no number behind it. Block also publishes no weights, no code and no license on that page, and access is gated: the page offers "Join the waitlist" and says "We're giving a small group of AI safety researchers early access to find the risks we haven't anticipated." Nobody outside that group can run it locally today, whatever MacBook they own.

"Runs in real time" and "processes audio directly" both hold

At 36.1 s the narrator says "so it processes audio directly it listens to you speaking picks up on laughs and tonal shifts", and the caption at 40 s reads "IN REAL TIME." Both halves check out against Block's page. The page describes Dessa as "audio in, audio out" and "trained from scratch on nothing but human conversation" with "no text, no pretrained models", which is the direct-audio architecture the narrator describes. On latency the page states: "Inference runs faster than real time, so there's no awkward wait." Faster than real time is a stronger claim than real time, so the video understates this one rather than overstating it.

The laugh and tone detection is also visible rather than asserted. The demo player in frames 010 through 013 and the 25.0 s extract carries five lime-green annotation pills on the Dessa waveform reading "Laugh", "Blown away", "Laugh", "All the secrets" and "Speaking louder". Block's page lists the demo set those come from, including titles such as "Artist Dessa", "Bathroom disaster", "Hobbies talk", "Noisy GPT live", "Safety lab", "Singing nonsense Spanish", "Wise Dessa" and "Zelda discussions". One training-hardware figure is worth recording because the video omits it: Block says "we figured out how to train Dessa on only one NVIDIA GB200 NVL72."

"Taking on the likes of ElevenLabs and OpenAI" is the narrator's framing, with no benchmark behind it

At 8.9 s the transcript reads "he's now taking on the likes of 11 labs and open ai", and the burned-in caption at 8 s reads "OF ELEVENLABS", which is the correct spelling and the reason I can attribute the mangling to Whisper rather than to the video. At 55.9 s the narrator repeats the framing: "now he's taking on some of the biggest companies in the world to replicate your voice."

Block publishes no comparison. I fetched block.xyz/dessa twice and found no benchmark, no evaluation table, no MOS score and no named competitor claim. Unite.ai states plainly that no benchmarks comparing Dessa to competing systems are provided. The only third-party model Block names on its page is in a demo disclaimer, which appears twice, after "Dessa in conversation with GPT-Live-1" and after "Noisy GPT live": "In this demo, Dessa is talking with GPT-Live-1, a third-party AI model. Both voices are AI-generated. The developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa." That is a disclaimer distancing Block from the other model, not a competitive claim against it.

TechCrunch supplies the nearest thing to a comparison, and it is a mechanism contrast rather than a score: it reports that ElevenLabs v4 and Fish Audio Drama 3 rely on voice tags to drive emotional expression while Block claims emotion is built directly into Dessa, and that OpenAI's GPT-Live supports full-duplex conversation as well. The video's "arguably the most natural ai speech model to date" at 12.5 s rests on the narrator's ear and Block's own marketing line, and I mark it Unverified.

Dorsey co-founded Twitter with three others and holds 42.2 percent of Block's voting power, not control

The description calls Dorsey "the founder of Twitter" and the narrator at 0.00 s says "the man who founded twitter". Twitter had four co-founders. Jack Dorsey, Noah Glass, Biz Stone and Evan Williams started it in March 2006, and Dorsey sent the first tweet on March 21, 2006. Block's own proxy statement records his Twitter tenure precisely: "From May 2007 to October 2008, Mr. Dorsey served as President and Chief Executive Officer of Twitter, Inc. … Mr. Dorsey returned to serve as Chief Executive Officer of Twitter from July 2015 until November 2021. He served on the board of directors of Twitter from May 2007 to May 2022." He has had no role at Twitter or X since May 2022, which the video's present-tense framing obscures.

"Controls block" at 2.7 s, with the caption "CONTROLS" on screen at 2 s, overstates the arithmetic. Block's 2026 proxy statement reports beneficial ownership as of March 31, 2026 against 535,370,474 Class A shares and 59,991,242 Class B shares outstanding, with Class A carrying one vote and Class B carrying ten: "each share of Class B common stock is entitled to 10 votes on each proposal." Dorsey holds 1,000,000 Class A shares, under one percent, and 47,844,566 Class B shares, which is 79.8 percent of the Class B class and 42.2 percent of total voting power. He is comfortably the largest single voting holder, ahead of The Vanguard Group at 6.0 percent, T. Rowe Price Associates at 4.2 percent and BlackRock at 3.4 percent, and he is both Chairperson, since October 2010, and principal executive officer. He does not hold a voting majority. The proxy does not describe Block as a controlled company.

The chronology the video asserts is in the right order, with the dates left out

At 47.6 s the narrator compresses the story: "first the man does twitter next the man reinvents financial services with square and block and now he's taking on some of the biggest companies in the world to replicate your voice." The captions at 48 s ("DOES TWITTER.") and 58 s ("YOUR VOICE.") bracket it. The sequence survives checking. Twitter was founded in March 2006. Block's proxy puts Dorsey as principal executive officer and a director of the Square entity since July 2009. Square, Inc. became Block, Inc., and the proxy records his title changing from Chief Executive Officer and President to Block Head in April 2022. Dessa was announced in October 2026. Twitter, then Square and Block, then Dessa is the real order.

Two gaps are worth naming. The video presents the arc as continuous Dorsey ambition, and the actual Dessa capability arrived at Block by acquisition in February 2020, between the Square and the model. And I could not reach a first-party source for the exact Square-to-Block rename date: Block's own press release page at block.xyz/inside/press-release-block failed to parse on fetch with a header overflow, and the Businesswire copy returned HTTP 403. I therefore cite only the April 2022 title change, which the proxy states directly, and leave the rename date out.

Key Takeaways

• Verified: Dessa is real, the spelling is correct, and Block's page describes it as "the most authentic full-duplex speech model ever built", operating "audio in, audio out" and "trained from scratch on nothing but human conversation" with "no text, no pretrained models".
• Verified: The real-time claim is first-party and slightly understated by the video. Block says "Inference runs faster than real time, so there's no awkward wait."
• Partial correction: The video says "you can run this locally on your macbook" (29.4 s). Block's page says "it runs locally on a MacBook Pro M4 Max or better", which excludes every MacBook Air and every Mac below that tier.
• Correction: The description and the 0.00 s line call Dorsey "the founder of Twitter". Twitter had four co-founders in March 2006: Dorsey, Noah Glass, Biz Stone and Evan Williams. Block's proxy records that Dorsey left Twitter's board in May 2022.
• Correction: The video says Dorsey "controls block" (2.7 s). Block's 2026 proxy puts him at 42.2 percent of total voting power as of March 31, 2026, from 1,000,000 Class A shares and 47,844,566 Class B shares, which is 79.8 percent of Class B. Largest holder, not a majority.
• Correction: The description says Dorsey "just unveiled Dessa". Block and its research lab announced it; Unite.ai credits the post to lab lead Ragavan Thurairatnam, and CoinDesk reports Dorsey promoted it afterward on X.
• Clarification: The model's name is reused. Square acquired a Toronto machine learning company called Dessa on February 7, 2020, and Block's page credits the current lab to former members of that startup, including Thurairatnam, who was Dessa's head of machine learning.
• Clarification: Dorsey's title at Block is Block Head, per the proxy, replacing Chief Executive Officer and President in April 2022. CoinDesk's "co-founder and CEO" does not match the filing.
• Unverified: "Taking on the likes of ElevenLabs and OpenAI" (8.9 s) and "arguably the most natural ai speech model to date" (12.5 s). Block publishes no benchmark, no evaluation table and no competitor comparison. Its only mention of another model is a disclaimer stating that GPT-Live-1's developer "is not affiliated with Block and has not reviewed or endorsed Dessa".
• Unverified: The exact announcement date. Block's page gives none; TechCrunch and Unite.ai say October 23, 2026, CoinDesk says October 24, 2026.
• Context: Dessa is a research preview with a waitlist. Block publishes no weights, no code and no license, and says it is "giving a small group of AI safety researchers early access to find the risks we haven't anticipated."
• Context: Block discloses no parameter count or model size, and says it trained Dessa "on only one NVIDIA GB200 NVL72".
• Context: The embedded audio sample the narrator plays at 22 s to 25 s is sourced on screen. Frames 012 and 013 show Block's own demo player with the caption "Dessa Yeah, I can see that.", so the snippet comes from a published Block demo.
• Context: No frame I read shows a terminal, a model card, a repository or a file size. The video demonstrates nothing about local execution.

Resources

• Dessa, Block, Inc. establishes the first-party claims I cite: "audio in, audio out", "trained from scratch on nothing but human conversation", "no text, no pretrained models", "it runs locally on a MacBook Pro M4 Max or better", "Inference runs faster than real time, so there's no awkward wait", training "on only one NVIDIA GB200 NVL72", the GPT-Live-1 non-affiliation disclaimer, the safety-researcher early access and waitlist, the demo titles, and the lab's origin in the former Dessa team reporting to Jack Dorsey.
• Block, Inc. 2026 Proxy Statement (DEF 14A) establishes Dorsey's title ("Block Head and Chairperson"), the April 2022 title change, his service as principal executive officer since July 2009 and Chairperson since October 2010, his Twitter tenure and May 2022 board departure, the ten-votes-per-share Class B structure, the 535,370,474 Class A and 59,991,242 Class B shares outstanding as of March 31, 2026, and his 1,000,000 Class A / 47,844,566 Class B holding at 79.8 percent of Class B and 42.2 percent of total voting power.
• Block's new voice model sounds eerily human, TechCrunch, 2026-10-24 establishes the October 23, 2026 announcement date, the single NVIDIA rack training setup, the MacBook M4 Max requirement, the safety-researcher recruitment, and the mechanism contrast with ElevenLabs v4, Fish Audio Drama 3 and OpenAI's GPT-Live.
• Block's new voice AI sounds eerily human, sparking awe and scam fears, CoinDesk, 2026-10-24 establishes Dorsey's own X post promoting the model, the "MacBook Pro equipped with an M4 Max chip or better" phrasing, the research-preview status, and the lab's staffing from the Dessa startup Square acquired in 2020.
• Block Unveils Dessa, a Full-Duplex Speech Model Trained on Conversation, Unite.ai establishes that Ragavan Thurairatnam posted the announcement on October 23, 2026, the GB200 NVL72 training detail, the waitlist, and that no benchmarks against competing systems were published.
• Welcoming the Dessa Team to Square, Square press release, 2020-02-07 establishes the February 7, 2020 acquisition of the Toronto company named Dessa and the commitment that the team would keep working from Toronto.
• Toronto machine learning startup Dessa acquired by Square, BetaKit, 2020-02-07 establishes Dessa's 2016 founding as DeepLearni.ng and its founders Stephen Piron, Eric Kin-Ho Lee, Vince Wong and Ragavan Thurairatnam, who was head of machine learning.