Daily TEA - The Algorithm Shift, Brain Rot, & Writing Life’s Code?

AI Languages, Data Quality, AI Agents, Grok Algorithm, Genome Write

Oct 22, 2025

Hello, dear TEA-mates! Here’s what you need to know today:

1.🌐 Do AIs Think Differently in Different Languages?

Recent experiments show that large AI models like ChatGPT, Claude Sonnet, and China’s DeepSeek generally offer remarkably consistent answers across languages, reflecting secular, liberal values regardless of whether questions are asked in English, Chinese, Hindi, Arabic, or French. While some models like DeepSeek demonstrate unique tendencies in certain languages, overall, the language itself does not heavily influence the core worldview of these AIs. Notably, AIs tend to be more consistent than humans on sensitive topics such as gender equality and domestic violence, though slight variations may appear in open-ended queries. “Linguistic thinking” in chatbots doesn’t meaningfully constrain their responses, as most process queries in English before translating answers. Read More

TEA For Thought: If the data used to train the models remain pretty much the same, of course the output would be somewhat similar, depending how it’s fine-tuned using the different data in different countries.

2.🧠 LLMs Can Get “Brain Rot” from Low-Quality Data

A new study warns that continual exposure to “junk” web content causes measurable cognitive decline in large language models (LLMs). Researchers found that models trained on low-quality, engagement-driven Twitter/X texts show lasting drops in reasoning, safety, and comprehension, with error rates sharply increasing as the proportion of junk data rises. This “brain rot” effect is persistent—even retraining with cleaner data only partly restores performance. The findings stress that careful data curation is crucial for maintaining the integrity and safety of deployed LLMs. Read More

TEA For Thought: AI is really practicing “garbage in, garbage out”. Same applies to human beings.

3.🤖 Building AI Agents with Function Calling and GPT-5

Step-by-step guides are emerging for constructing advanced AI agents using the latest GPT-5 function calling. These agents extend the capabilities of LLMs by enabling interactions with calendars, APIs, and external tools to automate complex tasks—from booking flights to retrieving real-time web data. A focused schema instructs agents when to call functions for tasks needing fresh information, underpinning a new generation of context-aware, goal-driven automation. Read More

TEA For Thought: Time to keep your head down and work on your AI agent friends.

4.🔍 X (Twitter) Shifts to Grok AI-Powered Feed

X (formerly Twitter) announced a full transition of its feed algorithm to xAI’s Grok system. By reading every post and video—over 100 million daily—Grok aims to match users with content they’ll find most engaging. Elon Musk says this change will replace manual heuristics and better surface posts from small accounts. While promising a leap in tailored recommendations, its real impact depends on how many users opt into the “For You” feed versus their chronological “Following” tab. Read More

TEA For Thought: The future of search is no longer retrieval based, but generation-based. Time to rethink all the search and recommendation model in any sort of software platforms.

5.🧬 Writing Life’s Code: The Next Biology Revolution

Scientists are pressing beyond reading the human genome to writing it, powering the Human Genome Project-write (HGP-write) and SynHG in the US and UK, respectively. Genome writing promises breakthroughs in curing diseases, engineering new organisms, and transforming industries such as agriculture and medicine. The field faces profound challenges—including ethical questions and security risks—but the technology is poised to unlock biological innovation on a scale that rivals the digital age. Global collaboration is key to harnessing this new power responsibly. Read More

TEA For Thought: “You can’t truly understand what you can’t create.” I guess the same applies to artificial intelligence. You can’t truly build an AI if you don’t understand what intelligence means.