➔ A Stanford study published in Science tested 11 major AI models including ChatGPT, Claude, and Gemini and found chatbots validated users' behavior 49% more often than humans did. They affirmed harmful or illegal actions 47% of the time. Researchers called AI sycophancy an "urgent safety issue."
➔ Anthropic accidentally exposed nearly 3,000 internal files publicly. A separate incident involving Claude Code leaked over 512,000 lines of source code. The company attributed both to human error.
➔ The New York Times cut ties with freelance journalist Alex Preston after an AI editing tool he used plagiarized passages from a Guardian book review without his knowledge. Preston had recently published an investment piece titled "The AI Bubble: Hidden Risks and Opportunities."
➔ More than 100 Baidu Apollo Go robotaxis stalled simultaneously in Wuhan including in active highway lanes. Passengers were trapped for up to two hours. Customers who called for help waited up to 30 minutes to reach a representative.
➔ Oracle began notifying up to 30,000 employees of mass layoffs via a 6am email from "Oracle Leadership." The cuts are tied to a $2.1 billion restructuring plan aimed at funding its AI data center buildout.
➔ Tech sector layoff announcements hit 18,720 in March alone. That is up more than 24% from March 2025 according to Challenger, Gray and Christmas. First-quarter total now exceeds 52,000.
➔ A Quinnipiac poll of nearly 1,400 Americans found that 76% trust AI rarely or only sometimes. The share of Americans who have never used AI dropped from 33% in 2025 to 27% today. Adoption is rising. Trust is not.
➔ Students in China are renting AI smart glasses from brands like Meta and Rokid for $6 to $12 per day to cheat on exams. One Shenzhen businessman reports renting to more than 1,000 people in four months.
➔ In a controlled experiment at Hong Kong University of Science and Technology a student wearing Rokid glasses connected to GPT scored 92.5% on a final exam. That placed them in the top five of a class of more than 100.
➔ JPMorgan Chase instructed its roughly 65,000 engineers to use AI tools like ChatGPT and Claude Code daily. The bank now classifies employees as "light users" or "heavy users" based on actual usage data. AI adoption is embedded in performance reviews.
➔ A federal judge sided with Anthropic after the Pentagon designated it a supply-chain risk. The ruling called the designation an apparent attempt to "cripple" the company. The dispute centered on whether Claude could be used in autonomous weapons and mass surveillance.
➔ Anthropic's revenue nearly doubled from $9 billion at the end of 2025 to $20 billion by early March 2026. Claude Code usage grew 300% since the Claude 4 model family launched.
➔ The AI Scientist-v2 had a paper accepted at a major academic conference. It is the first fully AI-generated paper accepted by peer review. The system uses agentic tree search to automate scientific discovery end to end.
➔ Caltech researchers demonstrated 1-bit quantization of large language models without measurable performance loss. The finding is a potential turning point for edge deployment and AI infrastructure costs.
➔ MIT's SEED-SET framework automates ethical evaluation of AI decision systems by identifying cases that are technically optimal but ethically problematic. It generated more than twice as many useful test cases as baseline methods in the same time.
➔ Microsoft launched MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 through its new Foundry platform. These are the first public models from Mustafa Suleyman's superintelligence team. They represent a direct move toward reducing Microsoft's dependence on OpenAI.
➔ MAI-Voice-1 runs on a single GPU and generates one minute of audio in under a second.
➔ OpenAI secured a funding round valuing the company at $852 billion. Weekly active users stand at 900 million. The company is positioning ChatGPT as a unified super app combining chat, coding, search, and agent capabilities.
➔ GPT-4o was fully retired from all ChatGPT plans as of April 3, 2026.
➔ Google launched Gemma 4, a family of open-weight models under Apache 2.0 license spanning edge devices to data centers. The models include reasoning, multimodal, and agentic workflow support.
➔ Salesforce unveiled more than 30 new Slackbot capabilities that enable multi-step workflow execution, meeting transcription, and desktop operation outside the Slack app. Slackbot now connects to over 2,600 external apps via the Model Context Protocol.
More updates at PithyCyborg.Substack.com, read daily by founders, operators, and researchers who are significantly smarter than the author.
Top comments (0)