ENFR
8news

Tech • IA • Crypto

TodayMy briefingVideosTop articles 24hArchivesFavoritesMy topics

AI Breakthroughs: Stepaudio 2.5, Exa $250M Raise, Microsoft Webwright Lead May 2026

AISunday, May 24, 2026

50 articles analyzed by AI / 348 total

Key points

Audio player
0:00 / 0:00
  • StepFun released StepAudio 2.5 Realtime, a cutting-edge real-time speech large language model supporting Chinese and English with roleplay-specific RLHF, achieving top ranks across five benchmarks. This demonstrates advancements in AI speech comprehension and customizable persona capabilities critical for interactive voice applications.[Marktechpost]
  • AI infrastructure company Exa raised $250 million in a Series C funding round to develop AI search technologies, highlighting strong investor confidence in AI-enabled information retrieval systems. This influx of capital is set to accelerate enhancements in AI-driven search infrastructure.[Pulse 2.0]
  • Palabra.ai achieved rapid growth with a $1 million ARR and a 17x increase in half a year for its real-time AI voice translation platform, reflecting skyrocketing demand for multilingual AI communication tools in global markets.[Pulse 2.0]
  • The U.S. Federal Trade Commission targeted AI marketing practices with new settlements emphasizing the importance of consumer consent and vendor liability, marking increased regulatory focus on ethical AI use and marketing compliance.[Frankfurt Kurnit Klein & Selz PC]
  • Experts warn that AI advancements are accelerating the quantum computing threat to cryptocurrency security, raising urgent cybersecurity concerns over potential quantum-enabled attacks against cryptographic protocols.[CoinDesk]
  • An AI analysis of 400,000 Reddit posts uncovered previously unknown side effects of the weight loss drug Ozempic, demonstrating AI's powerful capability to mine large-scale unstructured social media data for medical insights.[ScienceDaily]
  • Leading AI figures diverge on AI progress; DeepMind’s Demis Hassabis sees humanity nearing the singularity, while Yann LeCun argues current AI lacks true intelligence. Comparative insights from Oriol Vinyals reveal today's models resemble AGI models from seven years ago, underscoring ongoing debates in AI capabilities.[The Decoder]
  • Microsoft Research introduced Webwright, a terminal-native web agent powered by GPT-5.4, which improved the Odysseys benchmark success rate from 33.5% to 60.1%, marking a significant leap in applying large language models to web-based interactive tasks.[Marktechpost]
  • Meta is shifting workforce priorities in Israel by reassigning 200 employees to AI projects and laying off 100 staff members, illustrating a strategic pivot toward AI development amid broader organizational changes.[The Jerusalem Post]
  • Google launched AI tools that enable users without programming expertise to build Android apps, democratizing software development and expanding AI's practical application in mobile ecosystems.[KTVB]
  • Meta, Broadcom, and partners invested $125 million to establish an AI chip research hub at UCLA, aiming to accelerate innovation in AI hardware and strengthen the semiconductor foundation for future AI systems.[Yahoo Finance]

Relevant articles

StepFun Releases StepAudio 2.5 Realtime: An End-to-End Voice Model with Roleplay-Specific RLHF and Paralinguistic Comprehension

8/10

StepFun, a Shanghai-based AI lab, released StepAudio 2.5 Realtime in May 2026, an end-to-end real-time speech large language model with roleplay-specific RLHF and paralinguistic comprehension. It supports both Chinese and English via a WebSocket API and ranked first across all five benchmark dimensions, demonstrating state-of-the-art performance.

Marktechpost · 5/24/2026, 10:51:05 PM

Microsoft Research Releases Webwright: A Terminal-Native Web Agent Framework That Scores 60.1% on Odysseys, Up from Base GPT-5.4’s 33.5%

8/10

Microsoft Research released Webwright, a terminal-native web agent framework powered by GPT-5.4, which achieved a 60.1% success rate on the Odysseys benchmark, significantly improving from the base GPT-5.4 score of 33.5%. This advancement showcases major progress in applying large language models to interactive web tasks.

Marktechpost · 5/24/2026, 8:56:19 AM