PLUS: Microsoft's new AGI team, Gamma's $2.1B valuation, and why GPT-5 fails CAPTCHAs
Good morning
Meta has just open-sourced a powerful new AI system providing speech recognition for over 1,600 languages. This release dramatically expands digital access for hundreds of previously unsupported language communities worldwide.
With a new feature allowing users to add entirely new languages with just a few audio examples, the project aims to democratize speech technology. Will this open-source foundation be enough to truly bridge the global digital divide?
In today’s Next in AI:
Meta's AI transcribes 1,600 languages
Microsoft launches new AGI team
Gamma’s $2.1B valuation
Why top AI models fail CAPTCHAs
Meta's Universal Translator
Next in AI: Meta just open-sourced Omnilingual ASR, a powerful new suite of models that provides speech recognition for over 1,600 languages, drastically expanding access for previously unsupported communities.
Decoded:
The system achieves impressive accuracy across its vast range, delivering character error rates below 10% for 78% of the supported languages, including 500 that have never been transcribed by an AI before.
It introduces a new "bring your own language" capability, allowing users to extend the model to entirely new languages with just a few audio-text examples, bypassing the need for massive training datasets.
Beyond the models, Meta is also releasing its training data and a new 7B parameter speech foundation model, empowering developers to build new tools and explore the languages with an interactive demo.
Why It Matters: This release significantly lowers the barrier for bringing speech technology to low-resource languages, which have historically been left out of the digital world. It provides a powerful, open-source foundation that will enable developers and communities to create more inclusive applications worldwide.
Microsoft's AGI Quest
Next in AI: Microsoft is accelerating its long-term ambitions by launching a new 'MAI Superintelligence Team' led by AI chief and DeepMind co-founder Mustafa Suleyman. The group's mission is to develop advanced AI explicitly designed to serve humanity.
Decoded:
The team is helmed by Mustafa Suleyman, a prominent figure in the AI space and a DeepMind co-founder, whom Microsoft hired from AI startup Inflection.
Instead of pursuing an abstract AGI, the group will focus on practical applications like creating helpful digital companions, advancing medical diagnostics, and discovering new forms of renewable energy.
This move signals Microsoft's strategic push to build its own foundational research capabilities and reduce its dependence on OpenAI, putting it in direct competition with similar superintelligence labs at Meta.
Why It Matters: The race for artificial general intelligence is entering a new, more structured phase as tech giants formally create dedicated teams. This signals a focused shift from improving existing products to tackling fundamental, real-world challenges with next-generation AI.
Gamma's AI Goldmine
Next in AI: AI presentation-builder Gamma just raised a $68 million Series B led by Andreessen Horowitz, launching it to a massive $2.1 billion valuation. The platform helps users instantly create polished presentations, websites, and documents from rough ideas.
Decoded:
Gamma has profitably hit $100 million in annual recurring revenue while serving over 70 million users worldwide.
The company’s unusual efficiency stands out in the AI space, reaching its double-unicorn status with only about 50 employees and $90 million in total funding.
Evolving beyond slide decks, Gamma is now a complete visual storytelling tool that uses over 20 different AI models to generate websites, documents, and social media content.
Why It Matters: Gamma's profitable, lean growth offers a compelling alternative to the capital-intensive models common among AI startups. This signals a growing market demand for AI tools that deliver immediate, measurable productivity gains rather than just technological promise.
AI's CAPTCHA Problem
Next in AI: A new study benchmarking top AI models found that while Claude and Gemini perform well against Google's reCAPTCHA, GPT-5 struggles significantly, often failing because its obsessive reasoning causes it to time out.
Decoded:
Claude Sonnet 4.5 led the pack with a 60% success rate, followed closely by Gemini 2.5 Pro at 56%, while GPT-5 lagged far behind at just 28%.
The study, conducted using the Browser Use framework, suggests GPT-5's failure stems from overthinking; the model generated excessively long reasoning steps, which led to repeated timeouts before it could submit an answer.
The research also revealed a universal weakness, as all models struggled with dynamic CAPTCHAs that change after a click, highlighting that current agent architectures have trouble with interactive web elements.
Why It Matters: This study highlights that for AI agents performing real-world tasks, raw intelligence is not enough. Building effective agents requires balancing deep reasoning with the speed and efficiency needed to act decisively in real-time environments.
AI Pulse
CoreWeave reported its third-quarter revenue more than doubled to $1.36 billion, with its revenue backlog surging to over $55 billion amid massive demand for AI computing infrastructure.
The European Commission plans to overhaul its landmark GDPR privacy rules to create new exceptions for AI companies, aiming to boost the region’s competitiveness.
OpenAI faces seven new lawsuits alleging that its ChatGPT chatbot encouraged suicides and other mental health crises, escalating scrutiny over the platform's safety guardrails.
Hypercubic launched its YC-backed AI platform designed to help large enterprises understand and modernize legacy mainframe systems that run on COBOL.
LinkedIn adopted the C2PA standard to label AI-generated and modified images with Content Credentials, providing users with more information about a file's origin and history.