Listen to today’s podcast: https://www.youtube.com/channel/UC-nqwUyvLDEvs7bV985k-gQ
AI Daily Podcast — December 25, 2025
Today’s podcast episode was created from the following stories: a mix of research breakthroughs, hardware advances, policy guardrails, practical tools, buyer beware advice, and a timely deal if you’re exploring AI more deeply.
AI supercharges scientific output while quality slips
A Cornell-led study finds large language models boosted preprint output by 33–50%, with the biggest gains for non‑native English speakers. But polish no longer signals rigor: AI-assisted papers with complex writing were less likely to be accepted by journals, complicating peer review and funding metrics. The authors urge clearer norms on how AI is used and better evaluation tools to separate slick prose from substantive science.
This New 3D Chip Could Break AI’s Biggest Bottleneck
Researchers working with SkyWater built a monolithic 3D chip that stacks compute and memory vertically, delivering up to an order‑of‑magnitude gains over comparable 2D designs in tests and simulations. By tightly integrating layers with dense vertical interconnects, the prototype attacks the “memory wall” and promises 100–1,000x improvements in energy‑delay over time. It’s also a milestone for U.S. manufacturing, proving advanced 3D architectures can be fabricated domestically.
OpenAI CEO Sam Altman claims ‘AGI’ might have already “whooshed by” — with surprisingly little societal impact compared to the hype that surrounds it
Sam Altman argues we may have crossed an AGI threshold without a dramatic societal shift, suggesting the conversation should pivot to “superintelligence” capable of outperforming top human leaders. The comments underscore the fuzzy definitions around AGI and echo broader industry debates about capability vs. real‑world impact and safety. Tech leaders from Google and Microsoft continue to push forward while warning of preparedness and risk.
A Nobel Prize-winning physicist explains how to use AI without letting it do your thinking for you
Saul Perlmutter cautions that AI’s confident tone can create the illusion of understanding, dulling critical judgment. His advice: treat AI as a tool that supports reasoning, not a substitute for it—embracing skepticism, error‑checking, and comfort with uncertainty. UC Berkeley’s coursework reflects this approach, training students to operationalize critical thinking alongside AI.
NotebookLM wants to take one of its best features to the next level
Google is testing a new “Lecture” mode for NotebookLM’s Audio Overviews, generating comprehensive spoken walkthroughs with selectable length and language. The feature would sit alongside Deep Dive, Brief, Critique, and Debate modes to help users digest sources in the format they prefer. No launch date yet, but it signals continued investment in multimodal learning workflows.
PSA: Please do not buy this dubious ‘AI translator’
Gizmodo flags serious red flags with InnAIO’s T10 translator, including evidence it may lack a microphone and perform little beyond triggering an app on your phone. The company’s messaging is murky about on‑device processing, and the device appears to rely entirely on your phone’s mic and cloud services. Bottom line: save your money and use a trusted translation app instead.
iX-Workshop: RAG-Systeme effizient evaluieren und optimieren
A two‑day hands‑on workshop teaches teams how to evaluate and tune Retrieval‑Augmented Generation systems with frameworks like RAGAS. Attendees practice techniques such as semantic chunking, hybrid search, query expansion, and re‑ranking to boost accuracy and robustness. It’s geared toward developers and AI engineers ready to move RAG from promising prototype to reliable production.
Apple Intelligence will be tested with 2,000 questions it must not answer in China
China requires AI systems to refuse at least 95% of 2,000 censorship‑triggering prompts, a test Apple’s locally‑partnered model must pass before launch. The mandate highlights tensions between capability and control, as models are encouraged to learn from global data while blocking restricted topics. It’s another example of the tradeoffs multinational tech companies face to operate in China.
Google One is offering 50% off its 2TB storage and AI Pro plans right now
Google is discounting annual Google One plans by 50%, including the 2TB storage tier and the AI Pro bundle with higher Gemini limits and extra tools. The deal is officially for new users, but you can create a fresh account and use family sharing to bring benefits to your main profile. If you’ve been waiting to expand storage or test Google’s AI features, this limited‑time offer is a low‑risk entry point.
