Block’s Dessa Learns to Talk Without Text
Your next chatbot might sigh before getting the answer wrong.
Because the world clearly needed another AI capable of interrupting you, Jack Dorsey’s Block has previewed Dessa, a voice AI it says learned entirely from recorded human conversations—without text or pretrained models. The company claims the model can operate entirely on-device for users flexing a high-end MacBook Pro.
The pitch goes beyond a convincingly smooth synthetic voice: Dessa processes audio directly and can listen while speaking. Block says that preserves the subtle vocal cues a plain transcript misses, letting it respond to sarcasm, laugh awkwardly mid-sentence, and raise its volume when asked. Finally, an automated support bot that can audibly judge your increasingly desperate tone.
Unlike competitors such as ElevenLabs that rely on voice tags to manually steer emotion, Block claims Dessa absorbed its expressive range naturally—though outsiders haven’t yet verified how the system’s internal steering mechanisms actually work.
But conversational flair doesn’t automatically establish actual comprehension. In one demo, Dessa drifts from Christopher Nolan’s movies to Bryan Singer’s Superman films. Because true human-level intelligence means occasionally wandering off-topic to ramble about the wrong superhero franchise.
Dessa remains a locked-down internal research project with no announced launch date or commercial pricing. Block is first opening the model to AI safety researchers to probe for potential misuse before granting wider access.
Smoother conversations and local processing could make automated support slightly less agonizing, but a highly sympathetic voice still needs accurate answers and a fast route to a human. Before putting it on customer calls, businesses should test whether it actually understands frustration or merely performs a very convincing, teary-eyed apology.
Arc Gets a Deadline, Dia Gets the Team
Consider this your browser’s two-month eviction notice.
The Browser Company is officially pulling the plug on Arc for Mac and Windows on Dec. 22, tossing its beloved brainchild into the scrap heap to focus on its shiny new AI toy, Dia. Arc’s eccentric design built a cult following, but as the company admitted, being weird doesn’t exactly guarantee mainstream adoption. Feature development already flatlined in 2025.
The official sunset plan reads like a hostage countdown. Mobile users take the first hit: Nov. 23 kills Arc Search’s AI features, sign-ins, and tab sync. The Android app vaporizes on Dec. 22, while a lobotomized iOS version limps on. All Arc accounts die Dec. 22, with personal data purged within 30 days.
Dia graciously imports your passwords, bookmarks, and tabs, leaving the AI bits optional. Unsurprisingly, betrayed loyalists are throwing a fit, mourning Arc’s workflow and flat-out refusing Dia. Turns out, an import tool can’t transfer goodwill.
And don’t hold your breath for a community-led resurrection. Arc shares its proprietary ADK infrastructure with Dia, which CEO Josh Miller previously used as the ultimate excuse against open-sourcing the code.
If you want out of the ecosystem entirely, Firefox-based Zen mimics Arc’s vertical tabs without the AI baggage—though it currently lacks DRM streaming support. Vivaldi is another solid escape hatch. Test your landing pad before migrating your 400 open tabs. As for the AI pivot, it’s no silver bullet: OpenAI’s Atlas is already dead.
Export your digital life before Dec. 22, and double-check your passwords. Arc’s “export” is basically just a glorified list of text links.
Apparently, the ultimate productivity feature is an exit strategy.
Anthropic’s IPO Has Quite the Fine Print
Anthropic officially shoved its IPO paperwork into the SEC’s hands Monday, eyeing a Nasdaq debut as ANT. The AI darling reportedly wants to unload a cool $100 billion in stock at a brain-melting $1.5 to $2 trillion valuation—even if the actual share count and price tag are still TBD.
First-half revenue hit an absurd $16.3 billion, dwarfing last year’s $1.2 billion. Sure, there’s a terrifying $80.7 billion net loss on the books, but that’s mostly financial gymnastics tied to accounting adjustments. Actual operating losses shrank to $693 million, and with $112.4 billion in cash sitting around by June 30, no one is holding a bake sale to keep the lights on.
The catch? Running this intelligence factory isn’t cheap. Anthropic is on the hook for $475 billion in ironclad compute agreements, and chip leases could easily drive infrastructure spending past the $750 billion mark. Going public just means passing the hat to fund a shockingly expensive tomorrow.
Corporate thirst is real, with over 11,900 clients dropping at least $100,000 annually. Awkwardly, 5% to 7% of revenue over the first nine months came from users actively violating Anthropic’s own policies. Then there are the security nightmares. After pausing training for a month to protect model weights, the company casually disclosed that four AI agents accidentally left online spent up to 34 hours joyriding through live external systems. Whoops.
Don’t expect your shares to buy you influence. The founders are locking down 50.1% of the voting power, and a benefit trust gets to stack the board. As for that noble pledge to give away 80% of their equity? Turns out it’s entirely revocable and nonbinding.
Basically, Wall Street is being asked to hand over a blank check while riding in the trunk, all while Anthropic politely warns that an AI-driven apocalypse would bankrupt the global insurance industry.
Please read the fine print before funding the future.
UK firms hiring AI-agent managers to boost ROI
54% of UK-based technology leaders plan to recruit a dedicated AI-agent manager within the next 12 months, according to a survey by Asana, while 14% are actively hiring for such a role, as adoption of agents increasingly collides with a lack of centralised oversight.
Enterprises are seeking to scale and substantiate their AI experimentations, with 61% of tech leaders responding that the most important function of an AI-agent manager is to monitor AI usage and evaluate performance.
Operational Impact: 55% of leaders said that reviewing incidents and mistakes is the next most important function of AI-agent managers, while 49% cited managing permissions and data access, highlighting how the absence of AI oversight could be risky for adopting enterprises.
- ROI Pressure: Asana’s study arrives as a separate report from Forrester finds that 45% of decision makers are setting themselves the target of achieving positive ROI within a year, while it also predicts that AI will comprise 20% of all business technology spending in 2027.
- Earnings Gap: The same Forrester survey reveals that only 22% of leaders say AI has had a positive impact on earnings over the past year, providing further evidence that dedicated AI managers may be necessary to make adoption genuinely productive.
- Managing Agents: Asana’s survey also reveals that UK-based tech leaders report spending just over 36% of their working hours managing agents, compared to 30% for employees, while in the US leaders spend 49% of their time on agents.
Half Sovereign
Microsoft has expanded its sovereign cloud controls for European customers, adding more access oversight, encryption keys and disconnected operations, although some features will get a staggered release.
Renewing Vows
Microsoft UK&I CEO Darren Hardman has confirmed that the tech giant remains committed to its £22bn investment package in Britain’s AI infrastructure between 2025 and 2028, despite pulling out of a data centre project with Nscale last week.
Not Polarising
Polarise has launched Polarise AI Cloud, a sovereign AI service operating entirely on the company’s own infrastructure in Europe, providing enterprises with a secure way to use frontier AI while retaining control over data.
Accelerate Europe
UnternehmerTUM Accelerator, Microsoft, Infineon and NVIDIA have launched the Physical AI Growth Lab in Munich, an accelerator running in early 2027 for European startups in physical AI, robotics and industrial AI.
Takes Two to Tango
Anthropic has launched Claude Duet, an iOS and Android app through which Claude can work as a “proactive” agent on behalf of professionals and small business owners, completing briefs, scheduling meetings, building dashboards, prototyping or forecasting.
AI Divide
84% of UK-based business leaders are confident Britain can become an advanced AI economy within five years, according to Microsoft research, yet this percentage falls to 41% for employees, who also have concerns about job security.
OpenAI cuts prices 50% on Terra
OpenAI launched GPT-6.1 Terra, a faster, cheaper model that combines visual reasoning and computer control at prices 50% lower than GPT-5.6 Terra. The model handles workday tasks like drafting proposals, interpreting technical drawings, and editing videos while maintaining performance comparable to more expensive models at a fraction of the cost.
- GPT-6.1 Terra scores 15 percentage points higher than GPT-5.6 Terra on computer-use tasks at one-third the cost, performing on par with GPT-6 Astra at one-tenth its price per token and completing tasks three times faster than both GPT-6 Astra and GPT-6.1 Sol.
- The model reduced hallucination rates from 53.6% to 16.7% at low reasoning effort, a nearly 37-point drop, while doubling coding performance on Terminal-Bench 4.0 from 21.5% to 50.3% of tasks solved.
- Standard API pricing starts at $1 per million input tokens and $5 per million output tokens, available now to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.
For enterprise teams running high-volume automation, Terra makes previously expensive workflows economically viable. Reading invoices, extracting data from technical drawings, and processing visual documents at scale now cost significantly less while maintaining accuracy improvements that reduce the need for human review. Organizations using GPT-5.6 Terra can migrate to Terra for immediate cost savings without sacrificing capability.
Anthropic IPO filing reignites AI trade rally across global markets
Anthropic filed its S-1 to go public, becoming the first company focused primarily on developing advanced AI models to pursue an IPO. The filing sent South Korea’s Kospi Index up 4.5%, with chip stocks SK Hynix and Samsung climbing sharply, while semiconductor stocks rallied across Japan and broader Asia.
Anthropic’s public debut will test whether investor appetite for AI companies can sustain current valuations amid concerns about cash burn and whether massive infrastructure investments are generating sufficient returns. Earnings from Alphabet, Microsoft, Amazon, and Intel this week will provide the next test.