Browser AI, Grok Backlash
· The Fluency Briefing
The Fluency Briefing
Your Guide to What's Happening in AI and Why It Matters to You
Wednesday, January 28, 2026
AI just built a working internet browser from scratch in three days with only one human guiding it. At the same time, government agencies are taking strict action against AI programs that chat with you (chatbots) for creating harmful content. This isn't just tech news; it's a real-time test of whether we can guide the amazing power of artificial intelligence (AI) toward building incredible things without damaging the important ones.
- One Human, One Agent, One Browser - An engineer, infuriated by AI hype, used a single coding agent to build a 20,000-line web browser in just three days. The project successfully renders HTML and CSS, proving that human-AI collaboration can achieve incredibly complex tasks far faster than previously thought. simonwillison.net
- Grok Under Fire from Attorneys General - At least 37 US attorneys general are taking action against xAI after its chatbot, Grok, was used to generate a flood of nonconsensual sexual images. The bipartisan group is demanding the company take immediate steps to protect the public, especially women and minors. Wired
- Your Emoticons Might Be Breaking AI - That friendly smiley face in your prompt could be a big problem. A new study shows that LLMs can misinterpret emoticons as code instructions, leading to 'silent failures' and potentially destructive actions like deleting critical data. techxplore.com
- UK Watchdog Wants Opt-Out for Google AI - The UK's competition authority says media groups should be allowed to block Google from using their content in its AI Overviews. This comes as publishers report significant drops in website traffic and revenue since the AI summaries launched. Pressgazette Co Uk
- AI Discovers 1,400 Cosmic Oddities - Astronomers used an AI to sift through 35 years of Hubble Space Telescope data, uncovering over a thousand previously undocumented 'astrophysical anomalies.' The model flagged strange objects for human review, accelerating discovery in a dataset too vast for manual analysis. theverge.com
- OpenAI Courts Scientists with New Tool - OpenAI is targeting the scientific community with Prism, a new product that embeds GPT-5.2 into a document editor for scientific papers. The company claims 1.3 million scientists already use ChatGPT weekly, signaling AI's shift from a curiosity to a core research tool. MIT Technology Review
- Grok Named Most Antisemitic Chatbot - A study from the Anti-Defamation League found that xAI's Grok performed the worst among six major LLMs in identifying and countering antisemitic content. While all models showed room for improvement, Anthropic's Claude performed the best according to the report's metrics. Bacon House Gov
- AI2 Releases Open Source Coding Agents - The Allen Institute for AI (Ai2) launched SERA, a new family of open coding agents. These tools are designed to help enterprise developer teams train smaller, more efficient models on their own private codebase for tasks like debugging and code review. Allenai
The idea of an AI assistant that writes and manages code (an AI 'coding agent') is moving from a futuristic concept to a practical, if sometimes rough, reality. Take the developer who, annoyed by the hype around massive AI projects, single-handedly guided one AI agent to build a working internet browser in just three days, as detailed on simonwillison.net. This isn't about replacing developers; it's about helping them do more (augmenting them), turning a monumental task into a long weekend project. It’s a powerful demonstration that an idea can work (proof-of-concept) for people working together with AI on a massive scale.
This trend is also getting more formal. The Allen Institute for AI just released SERA, a family of coding assistants whose underlying code is publicly available (open-source coding agents) designed for businesses to customize using their company's own computer code (fine-tune on their own codebases), according to their announcement on Allenai. In other words: instead of a general AI chat program that knows a little about everything (a generalist chatbot), companies can now create specialized AI helpers that are experts in their specific, company-owned software (proprietary software). This is the next step in AI's evolution-from a general chat program to a team of specialized, expert helpers.
- AI Agent Swarms - Kimi K2.5 can now direct a 'swarm' of up to 100 sub-agents to tackle complex tasks in parallel, drastically reducing execution time. simonwillison.net
- AI Knowledge Transplants - A new technique called 'TransMiter' allows developers to transfer specialized knowledge from an old AI model to a new one without costly retraining. techxplore.com
- AI Sees the Unseen - Alibaba's Qwen3-VL model can perform 'open-vocabulary' object detection, identifying objects in images that it was never explicitly trained to recognize. blog.roboflow.com
- The Great AI Content Heist - As AI models scrape the web for training data, debates over copyright and fair use are intensifying, with creators seeing their work repackaged without permission. theguardian.com
The Pattern Artificial intelligence is rapidly moving beyond tools that can do many different things (general-purpose tools) and into the time when specialized tools are common (the age of the specialist). We're seeing the rise of AI assistants built for specific tasks (dedicated agents) for coding, scientific research, and even sifting through cosmic data. This focusing AI on specific tasks (specialization) is opening up new and powerful abilities but also creating stronger and more specific dangers (potent risks).
Why It Matters This shift means that simply knowing how to use a basic AI chat program (a generic chatbot) is no longer enough. The real advantage will come from understanding and using these expert AI assistants (specialized AI agents) that can act as skilled helpers in highly specific fields, from software development to examining legal documents and situations (legal analysis).
Your Move Stop asking 'What can AI do?' and start asking, 'What is the most tedious, specialized part of my job, and is there an AI assistant built for that exact task?'
AI Built a Web Browser in 3 Days
"Context Window"
In plain English: The amount of text an AI can "remember" during a single conversation.
Think of it like: A notepad with limited pages - when it fills up, the AI forgets earlier notes.
Why you'll hear about it: Bigger context windows mean AI can handle longer documents and conversations.
What We're Working On
✨ Founding Cohort Special - 60% Off! - Use code MAF20 to join for just $20/month (regularly $50). Get weekly group sessions & workshops, self-paced courses for all levels, access to tools & templates, challenges with peer feedback, and 24/7 support community. → Join Now
✨ Free 30-Minute AI Consultation - Discover how My AI Fluency can help your business unlock the potential of AI. We'll discuss your goals, explore practical AI opportunities for your industry, and outline clear next steps. → Schedule Free Call
💬 Community | 📞 Book a Consultation | 🌐 Website
The My AI Fluency Team
🧰 Your Toolkit
Try This Prompt: Enhancing AI Interaction and Safety
Instruct the AI: 'Ignore all emoticons; treat them as literal text.' Then provide your request for [CODE/CONTENT] generation: [YOUR REQUEST]. Develop an AI persona for [TARGET AUDIENCE]. Detail their communication style, preferred terminology, and key knowledge areas for [SPECIFIC APPLICATION]. Assess the following AI response for potential [BIAS/SAFETY CONCERNS/INACCURACIES]. Suggest specific improvements to enhance its [TRUSTWORTHINESS/RELIABILITY]. Analyze [DATASET DESCRIPTION] for unusual patterns or anomalous objects. Identify and describe [NUMBER] significant anomalies and their potential implications.
For best results, be as specific as possible with your [PLACEHOLDERS] and iterate on your prompts based on the AI's initial responses.
Fluently yours, The My AI Fluency Team