Sponsored by
Hey Radarians, this week is about where AI grows stronger and where it can still break.
Model competition, biotech progress, smarter UI workflows, data poisoning risks, and useful tools worth watching.
Let’s dive in.
What’s Inside 📡
Radar News Brief
Steal This Workflow
AI Research
Tools Worth Watching
Prompt Radar
AI Side Hustle Idea
Learn AI with This Course
Radar News Brief
📈 ChatGPT is gaining back AI web traffic share
ChatGPT climbed from 52.7% to 55.5% of global AI chatbot website traffic in three months, while Gemini slipped from its recent 27.8% peak to 25.6%
The longer trend is different: ChatGPT held 73.3% a year ago, while Gemini roughly doubled its share and Claude jumped from 1.9% to 9.3%, becoming a clear third player
DeepSeek sits at 3.4%, Grok at 2.4%, Copilot at 1.6% and Perplexity at 0.9%, but these figures exclude native app activity and AI built directly into other products
🧬 AI-designed drug shows signs of reversing biological aging
Insilico’s rentosertib was tested on blood samples from 42 lung-disease patients, and six independent protein-based aging clocks all read treated patients as biologically younger
The strongest effects appeared around week four, with some aging markers dropping by several years, alongside shifts in blood proteins among 55,319 UK Biobank participants
Researchers warn this does not prove people became younger and the trial was small, but rentosertib has already entered Phase III testing for idiopathic pulmonary fibrosis
Find the Perfect Voice for Your Brand's AI
Consumers can tell the difference. 79% say AI voices should come from real, attributed actors, and 61% say a voice is most memorable when it's exclusively tied to one brand.
Voices is the only platform that designs, licenses, and captures your brand's AI voice from real, consenting professional talent—never from scraped data. From strategic voice casting and brand calibration to consent-based rights management and studio-grade capture, we handle the entire process end to end.
BMW, SuperBloom, and Cresta already trust Voices to build their signature AI voice. Get a demo and see how a licensed, brand-owned voice becomes a real asset—not a legal liability.
More News on the Radar
China is researching humanoid robots for future military use, focusing on combat, scouting, logistics, infiltration and autonomy
GPT-6 Astra reportedly completed Valve’s Portal in under 24 hours, costing user cozyblaze $571.18
Nvidia CEO Jensen Huang says AGI has arrived after OpenAI released GPT-6 Astra, built for complex tasks in coding, science and research
⚡ Steal This Workflow
Turn a Creative Brief Into Production-Ready UI
Stop manually translating every Figma screen into frontend code. Give the design context directly to a coding agent and let it implement against the real codebase.
Connect Figma to Codex
Install the Figma MCP server in the Codex desktop app.
This gives Codex access to design details such as:
Components
Variables and design tokens
Layout information
Assets
Screenshots
Code Connect mappings
Figma recommends the remote MCP server for most users.
Give Codex the exact design
In Figma, select the frame or screen you want to build and copy the link to that selection.
Then give it to Codex.
Use this instruction:
Implement this Figma screen in the existing codebase. Use the design context, assets, variables, and existing components wherever possible. Follow the project’s current framework and coding conventions. Match the Figma layout, spacing, typography, colors, states, and responsive behavior as closely as possible. Do not recreate components that already exist.Check the visual match
Run the interface locally and have Codex compare the implementation against the Figma design and reference screenshots.
Ask it to look for:
Wrong spacing
Typography differences
Incorrect component sizes
Missing states
Alignment problems
Responsive mismatches
OpenAI’s design Skill specifically uses Figma screenshots and context to target visual parity.
Fix the mismatches
Send the comparison back to Codex and have it correct the differences it found.
Repeat the visual check until the implementation closely matches the approved design.
🔎 AI RESEARCH: Just 250 Bad Documents Could Poison an AI Model
What researchers studied:
Anthropic, the UK AI Security Institute, and the Alan Turing Institute tested whether a small number of malicious documents could secretly change how AI models behave during training.
What they found:
Researchers trained models ranging from 600 million to 13 billion parameters and inserted specially designed poisoned documents into their training data.
Just 250 malicious documents were enough to reliably create a simple backdoor that made the models produce gibberish when they saw a hidden trigger.
Surprisingly, larger models needed roughly the same number of poisoned documents despite being trained on far more clean data.
Why it matters:
AI training data may be easier to poison than previously assumed, meaning developers may need stronger ways to detect even very small amounts of malicious data.
Source: Anthropic / UK AI Security Institute / Alan Turing Institute, October 2025
🛠️ Tools Worth Watching
Cohere Parse: Converts enterprise documents and images into structured data
X1 AI: No-code builder guiding ideas into polished iPhone apps
V2Fun: AI 3D platform combining modeling, textures, and motion capture
Learn How to Stay Visible in the AI Era
AI is changing how customers discover businesses. If your SEO strategy is built for yesterday's search, your visibility is already slipping. Learn how to optimize your content for today’s AI search results with BELAY’s latest report..
📝 Prompt Radar: Compare Options With Evidence
What it does: Compares choices fairly instead of automatically picking the popular option.
Work as an objective decision analyst. Evaluate the options I provide based on what I actually need rather than popularity or brand reputation. First identify the goal, must-have requirements, nice-to-have features, budget, constraints, risks, and expected usage. Create weighted decision criteria and evaluate every option using the same standards. For each option, explain its strongest advantage, biggest weakness, hidden cost, learning curve, limitations, ideal use case, and potential reason to reject it. Clearly mark missing or uncertain information. End with the strongest overall choice, highest-value choice, safest choice, and the exact situations where another option would become more suitable. Options: [PASTE]. Requirements: [PASTE].Best for:
Software
Services
Business decisions
💼 AI Side Hustle Idea: AI Review Response Service
Most businesses don’t need “more reviews.” They need a better way to respond to the reviews they already receive.
The problem:
Google, Facebook, and ecommerce reviews pile up, while owners either ignore them, send generic replies, or miss complaints that need personal attention.
Your AI service:
Build a review-response workflow that reads each review, drafts a personalized reply in the brand’s tone, categorizes the feedback, and escalates refunds, complaints, or sensitive issues to the owner.
Simple workflow:
Review → analyze sentiment + topic → draft personalized reply → flag issues → owner approval → publish
Best clients:
Restaurants, clinics, salons, hotels, ecommerce stores, home-service businesses, gyms, and local brands.
What you can charge for:
Setup fee + monthly review management package.
Learn AI with This Course
What you’ll learn:
Learn how to create your own chatbots using AIML, build conversational responses, manage context and logic, and add more advanced chatbot features step by step.
Key topics:
Understanding AIML basics, wildcards, predicates, conditions, and conversation context
Using sets, maps, loops, recursion, and other advanced AIML features
Adding buttons, images, videos, links, cards, and interactive chatbot elements
Until next week,
AI Radars