☕️ The World of Small Language Models Grows Bigger With Microsoft's Phi-3.5

Small language models have been amassing interest for a while, and Microsoft intends for the train to keep chugging along with a new line of such models - introducing Phi 3.5. Other key highlights of the week include:

• OpenAI signs a media partnership deal with publications giant Condé Nast to use content from Vogue and The New Yorker

• Authors take Anthropic to court for training its Claude AI chatbot on copyrighted material

• Microsoft's controversial Recall feature on Copilot+ PCs to release in October following security concerns

Join us at AI Tangle as we untangle this week's happenings in AI!

Microsoft debuts its Phi-3.5 line of lightweight models that punch above its weight class

Microsoft recently revealed its latest iteration of small language models, Phi-3.5, boasting big performance jumps over the previous Phi-3. Despite losing to OpenAI's GPT-4o mini, benchmark tests from the Open LLM leaderboard show Phi-3.5 outperforming competitors like Google's Gemini Flash 1.5 and Meta's lightweight Llama models in areas like reasoning and math, making Microsoft's latest addition an avid contender in the market of cost-effective AI.

​

What does Phi-3.5 bring to the table?

Phi-3.5 comes in three parameter versions: 3.8 billion, 4.15 billion, and 41.9 billion. They also come coupled with a vision model version that can understand images and another version employing a mixture of experts (MoE) model to split learning tasks for efficiency. The model's tiny size makes it ideal for integration into applications or IoT devices, enabling functions like facial recognition without needing to send data to the cloud for processing. However, many users outside the enterprise world scrutinize Microsoft for even harsher censorship guardrails on Phi-3.5 compared to its predecessor, Phi-3.

OpenAI signs another media partnership, this time with publications giant Condé Nast

​OpenAI and news giant Condé Nast have partnered to allow ChatGPT and SearchGPT to feature content from publications like Vogue and The New Yorker. The multi-year deal is part of OpenAIs broader collaboration with media firms, which includes similar deals made with Time Magazine and the Financial Times. Though financial details weren't elabroated on, Condé Nast's CEO, Roger Lynch, stated that the partnership helps offset revenue challenges posed by social media and other media platforms.

​

Silicon Valley's Anthropic finds itself in court over allegedly training Claude on copyrighted material

Three authors have recently filed a class-action lawsuit against AI wonderchild Anthropic, accusing the AI company of using pirated versions of their books and others to train Claude. The lawsuit, filed in the California federal court, claims Anthropic built a multibillion-dollar business by misusing copyrighted works without permission, similar to lawsuits against other tech giants like OpenAI and Meta. The authors are seeking financial compensation and a permanent ruling to prevent further misuse of their works.

​

Microsoft delays its Recall feature on Copilot+ PCs to October following security-centered public backlash

Microsoft's Recall AI feature, initially delayed due to security concerns and its controversial nature, will be available for Windows Insiders in October, according to an updated blog post from Windows and Surface chief Pavan Davuluri. Recall takes screenshots of nearly everything on a PC, allowing users to search and retrieve past activities via an explorable timeline. Despite Microsoft's assurances of security, researchers found vulnerabilities that spread on social media like wildfire, prompting Microsoft to implement database encryption and make the feature opt-in.

​

Telecom company agrees to a $1 million fine for spreading deepfake Joe Biden calls ahead of New Hampshire primary

​The New Hampshire robocall case of Jan 21 sees telecom company Lingo Telecom settle for a $1 million fine by the Federal Communications Commission (FCC) after transmitting AI-generated calls that mimicked President Joe Biden's voice. By combining caller ID spoofing and a convincing AI voice, Steve Kramer, the political consultant behind the calls, faces a $6 million FCC fine and state criminal charges. Kramer, however, states that his intention wasn't to influence the results of the primary but rather to highlight the potential dangers of AI and spur lawmakers into action.

​

Opkey raises $47 million in Series B funding to expand its AI-powered ERP platform

Dublin-based Opkey recently secured $47 million in Series B funding led by PeakSpan Capital to expand its mainstay AI-automated testing platform. What their platform does is help organizations continuously test finance, HR and other enterprise resource planning (ERP) software. Opkey aims to plug the gaping hole that company CEO, co-founder, and ERP veteran Pankaj Goel calls enterprise "testing hell." The company already has a strong consumer base with 200 large enterprise customers and partnerships with system integrators like KPMG and PwC.

​

SleekFlow continues to drive its social commerce business with another $7 million in funding

​SleekFlow, a social commerce platform based in Singapore and Hong Kong, has secured $7 million in new funding to expand its AI-driven customer engagement system. The platform aims to further deepen its presence in Southeast Asia, the Middle East, and Europe. SleekFlow differentiates itself with omnichannel capabilities, marketing automation, and instant payment capabilities. On top of that, SleekFlow recently appointed a new CTO, Gao Lei, a Silicon Valley veteran with more than two decades of experience.

​Fyli - Create AI agents with Fyli to automatically find your ideal customers online, post engaging content, and secure more high-intent leads - automate your workflow, boost your marketing.

​Dropbase - Dropbase is an AI-driven developer platform for building custom internal tools and backend operations software with less hassle and more speed.

​Ellipsis - Ellipsis is an AI developer to help you keep track of code reviews by automatically iron out bugs in pull requests, helping create release notes, and fixing build issues.

​Aiera - Aiera is the platform that powers investors with one-click live audio access, real time transcription, search, and advanced summarization with coverage of global equities and market-moving investor events.

Why The Skepticism in AI Sales Reps? (4-min read)

AI sales rep startups are experiencing rapid growth and showing impressive numbers, yet VCs remain cautious about investing. Why is there hesitation despite the success of AI-driven tools in business?

What No One Understands About Foundational Models (63-min watch)

Co-founder and CEO of Cohere, Aidan Gomez, joins 20VC to discuss the bottlenecks in AI training, the value of a model, the future of enterprise AI as a whole, and the effects of AI in the wider world.