Published on
Category
AI Adoption & Systems Design
Welcome back to Cutting Edge News!
Last week, we looked at how AI is starting to outperform human engineers.
This week, the updates show a clear shift. We are moving past the Chatbot Era, where you used to ask a question and get an instant answer, into the Agentic Era, where the agent researches and surveys for you, understands your audience and builds for you.
THIS WEEK’S LINEUP
Main updates:
Anthropic Interviewer: Why the traditional focus group approach is outdated, and how to interview 10,000 customers in a day.
Gemini 3 Deep Think: Google’s new model pauses to “think” before answering.
Google x Replit: How to build software by describing what you want, not by writing code.
Rapid Fire: OpenAI’s honesty update, DeepSeek’s new model beat ChatGPT at a fraction of the cost, Kling AI’s video breakthrough and more…
Let’s get started.
Anthropic Interviewer: The End of the Focus Group
Anthropic just released a new tool called Interviewer. It uses their Claude AI to automate the entire process of talking to your customers.
For years, businesses had two choices: send a survey to thousands of people (which gives shallow data) or interview a dozen people (which is slow and expensive). This tool solves that problem.
The Numbers:
The Scale: A single person can now interview 10,000 users in a day.
The Speed: It reads thousands of transcripts and gives you a summary report in hours, not months.
The Study: Anthropic tested this by interviewing 1,250 professionals to see how they really feel about AI at work.
What makes this different: This AI tool isn’t a survey form. It engages in a real conversation. If a customer gives a vague answer, the AI asks a follow-up question to understand what they really mean.
Why this matters: It finds insights that surveys miss. In their test study, the tool discovered that many employees are using AI to finish work faster but are hiding it from their bosses to avoid looking lazy. A standard survey likely wouldn’t have caught this because people rarely admit to “hiding” things on a checkbox form.
What you should do:
If you work in Product: Use this to understand why users are leaving, not just how many left. Get deeper insights that are usually overlooked by simple forms.
If you work in HR: Use it to find out what employees are actually struggling with, beyond the standard engagement survey.
Gemini 3 Deep Think: The Value of Slow Thinking
Google has released Gemini 3 Deep Think. This model is different because it doesn’t answer immediately, it pauses to think first.
While standard LLMs generate the next token immediately, Deep Think models pause to engage in a hidden process before outputting a single character.
How is this different from the “Thinking Mode of Gemini 3”? It is important not to confuse this with older step-by-step features which the other models had.
Old Way (Linear): The AI thinks in a straight line, like a person doing math one step at a time.
New Way (Parallel): Gemini 3 explores multiple possible answers at the same time, like a chess player considering five different moves at once, and discards the bad ones before giving you an answer.
Why this matters: This changes how you use AI. Deep Think is slower and currently costs more. It is not for casual chat; it is for solving hard problems. You now have a choice: do you want a fast answer, or a well-thought-out one?
Note: Deep Think is currently exclusive to Google AI Ultra subscribers because it is computationally expensive.
Our Take: For the last three years, the main selling point of AI was speed. “It writes an email in 2 seconds.” Now, we are accepting slowness in exchange for intelligence. The new models pause to “think” and check their work before answering. We are moving from using AI for Content Generation (writing fast marketing copy) to Problem Solving (fixing complex logistics or legal contracts).
Google x Replit: Official Start Of The Era of “Vibe Coding”
Google Cloud has partnered with Replit (a popular coding platform) to change how software is built. Google is aiming to commoditize software creation to drive demand for their cloud compute.
This partnership is built around a new concept called “Vibe Coding”.
The Concept: Vibe Coding means you build an app by describing the result you want in plain English, rather than writing the technical code yourself.
The Tech: They have put Google’s powerful Gemini model directly inside Replit’s coding tool.
The Workflow: You say, “I need a dashboard that tracks these three spreadsheets.”
The Execution: The AI writes the code, sets up the Google Cloud server, and puts the app online. All by itself, all you have to do is monitor the checkpoints and give permissions whenever it needs to access the data.
Why this matters: Google wants to make it easy for anyone to create software. But for companies, this creates a massive governance challenge. If every employee can build their own automation tools, it becomes hard to track where company data is going.
What you should do:
Stop waiting for engineers. The barrier to entry has collapsed. Use this to build working prototypes and visual drafts yourself
The fundamental pattern of software engineering is shifting from syntax (writing correct code) to semantics (describing the correct outcome). The user’s role is moving from “writing the function” to “prompting the vibe,” with the AI handling the translation into Python and infrastructure.
Bottom Line: “Being good at coding” is becoming “Being good at describing what you want.” English or your natural language, will slowly become the next programming language of this decade. The barrier to building software is no longer a coding skill; it is clarity of thought. If you can describe it clearly, the AI can build it.
Rapid Fire: The Week’s Other Updates
OpenAI Confessions: OpenAI introduced a Confession Channel, a hidden output where the model self reports its performance. Even when the model lied to the user to be helpful, it admitted the deception in the confession channel. This could be the future of regulatory compliance.
DeepSeek V3.2: DeepSeek found a way to make AI much cheaper. They introduced “DeepSeek Sparse Attention” (DSA) method. Their new method reduces the cost of processing long documents by about 66%. The latest model, V3.2 speciale surpassed the scores of ChatGPT high in performance at a fraction of ChatGPT’s token costs.
Claude’s “Soul” Document: A leaked document shows that Anthropic instructs its AI that making revenue is a good thing because it funds safety research. It’s a practical look at how these companies align AI behavior with business goals
Kling AI AI 2.6 Launch Fixes Lip Sync: Video generation just took a major leap. Kling AI released a model capable of Simultaneous Audio Visual Generation. The new model creates video and the audio at the same time, meaning the characters’ mouths move perfectly in sync with their words.
Microsoft VibeVoice: Microsoft released a tool that could generate 90-minute podcasts with realistic voices, but they pulled it offline almost immediately. It was likely disabled due to the extreme risk of deepfakes.
Final Thought
If you look closely, we are witnessing the different parts of a “digital employee” coming online:
The Ears: Anthropic Interviewer listens to and understands your customers.
The Brain: Gemini Deep Think reasons through complex problems and plans the solution.
The Hands: Google x Replit builds and deploys the actual software.
For the last three years, the bottleneck was technical skill. If you couldn’t code or run a statistical analysis, you couldn’t do the job.
As of this week, the bottleneck has shifted. The barrier is no longer skill; it is clarity and taste.
If you can clearly describe the problem well, think through the logic, and describe the outcome, the AI can handle the execution.
We are moving from an era where you are paid to do the work, to an era where you are paid to design the outcome. Outputs will not matter, Outcomes will.
Don’t just chat with the AI. Direct it.
Stay sharp, The Cutting Edge Team