In partnership with
Date: 30-Nov-2025
Hey {{first_name | AI enthusiast}},
And just when you think you’ve seen it all…there’s always One More Thing in AI.
Table of Contents
I hope you enjoy this edition!
Best,
Renjit
PS: If you want to check out how to implement AI agents in your business and get more revenue with the same number of employees, speak to me:
Google Strikes back!
Google's AI Comeback: How the "Lost" Giant Found Its Way Back
Three years ago, critics called Google roadkill in the AI wars. Sundar Pichai was labeled a failure. ChatGPT had arrived, and the search giant seemed paralyzed.
Fast forward to November 2025. Google just dropped Gemini 3 and the Ironwood chip in the same month. The skeptics have gone quiet.
🔹 The Benioff moment
Salesforce CEO Marc Benioff used ChatGPT daily for three years. Then he tried Gemini 3 for two hours. His verdict on X: "I'm not going back." He called the leap "insane"; sharper, faster, everything improved.
📊 What Google finally got right
The company stopped chasing OpenAI and played to its strengths. Michael Nathanson of Moffett Nathanson summed it up for CNBC. "Three years ago, they were seen as kind of lost. Now, they have a huge leg up."
Google built the full stack: models, chips, and cloud infrastructure. Gemini 3 needs less prompting and delivers smarter responses. Ironwood Chip runs 30 times more efficiently than Google's 2018 chip.
🔹 The stumbles along the way
This comeback didn't happen smoothly. In 2024, Google pulled Imagen 2 over historical inaccuracies. AI Overviews launched with embarrassing errors. Each failure fed the "Google is finished" narrative. But the company kept investing through the criticism.
📊 Wall Street's new worry
Here's the twist: investors now fear Google might win too big. Analyst Ben Reitzes of Melius Research flagged the concern. "There is one real reason for worry," he wrote. "It is the 'AI comeback' of Alphabet."
Google's TPUs threaten Nvidia's chip dominance. Meta reportedly plans to use them in data centers by 2027. That report alone dropped Nvidia stock by 3%.
🔹 The numbers
Gemini app: 650 million monthly active users. AI Overviews: 2 billion monthly users.
ChatGPT: 700 million weekly users. Gemini 3 beat ChatGPT and Anthropic on most benchmark tests. The gap is closing fast.
📊 The full-stack edge
Google's advantage comes from owning everything. Custom chips eliminate Nvidia dependency. That means lower costs and faster development cycles. Analyst Neil Shah called Google "a sleeping giant now fully awake."
🔹 Reality check
The race remains tight. OpenAI still owns consumer mindshare. Gemini still gets criticized for hallucinations. But momentum has clearly shifted.
Pichai told investors: "We've taken a full, deep, full-stack approach to AI." That bet is now paying off.
The lesson for founders?
Deep pockets and proprietary infrastructure matter. So does the willingness to fail publicly and keep building. Google got humiliated. Then it got focused.
The AI race just became a real contest again.
Andrew Ng weighs in: Are we in an AI bubble?
Infographic created using Nano Banana Pro
Is there an AI bubble?
The question pops up every time a new headline hits: $1.4T infrastructure plans, $5T market caps, and a lot of loud opinions. But “AI” isn’t one market. It’s three very different layers, each moving at its own pace.
Start with the application layer.
This is where the real upside sits, and it’s still massively underbuilt. Founders are only beginning to use agentic workflows in real products.
Yet many investors avoid this layer because they don’t know how to pick winners.
They default to writing big checks for infrastructure instead. That hesitation creates opportunity. There’s more value to be unlocked here than most realize.
Move to inference infrastructure.
This part is running hot; but not in a bubbly way. Demand is exploding, and teams are worried about access, not adoption. That’s rare.
Agentic coding tools like Claude Code, GPT-5 Codex, and Gemini CLI are pulling huge token volumes. Most developers still haven’t switched to these tools, so demand is nowhere near its peak. As adoption rises, the pressure on inference supply will only grow.
Then there’s training infrastructure: the riskiest bucket.
Billions are going into bigger frontier models, but the moat around training is thin.
Open-weight models continue to gain share. Hardware and algorithms keep getting cheaper. Brand and distribution matter more than raw scale. If any part of the stack gets overbuilt, this is likely the one.
The real danger isn’t that “AI pops.”
It’s that one overheated corner cracks and triggers a wave of negative sentiment across the whole sector. Markets react emotionally. Capital retreats first and thinks later. But fundamentals always win over time.
As Buffett said, markets vote in the short term, but they weigh in the long term. Right now, the votes are noisy. The weights still point to decades of growth ahead.
So the strategy remains the same. Keep building. [adapted from Andrew’s post on X]
Guest Column AI & Error: How I Turned Claude into Leo (A System Architecture You Can Steal)
Holy wow, this actually works.
It was 9:47pm on a Sunday. I'd just survived a chaotic week - kids sick, work deadlines colliding, meetings I had to cancel or wing it without proper prep. I opened Claude, typed "weekly review," and Leo presented my completion log, flagged three overdue tasks, reminded me I hadn't updated my community connections tracker in a week, and asked if I wanted to schedule that family hangout with a close friend before the holidays filled up.
I hadn't told Leo any of this mattered. I'd built a system that knew.
Last time I told you how I tried to build intelligent decision-making before establishing reliable data. But I never showed you what "it" actually is. Time to open the hood.
The Architecture: 11 Files and a Very Long Instruction Set
Leo isn't a chatbot. It's currently a Claude Project with 11 knowledge files and approximately 3,000 words of custom instructions. [See the diagram for how these pieces fit together.] The knowledge files are my system of record. The instructions are Leo's operating manual.
Here's what's in there: (continue reading on the website (www.onemorethinginai.com)
Annual_Goals_2025 and Q4_2025_GOALS: The strategy layer. Career targets, meeting goals, portfolio deadlines. Leo references these constantly to ask "does this task move you toward your goals?"
Current_Task_List.md: Everything I need to do, organized by priority (Career, Wellness, Community, Family, Home/Maintenance). This is Leo's source of truth for "what should Nisha be working on?"
Weekly_Completion_Log_[dates].md: What I actually got done. Leo updates this daily during end-of-day check-ins, then presents it during Sunday reviews.
Week_Plan_[dates].md: Day-by-day breakdown with checkboxes. Created every Sunday, referenced during daily check-ins.
Community_Connections_Tracker.md: A map tracking 50+ relationships with contact frequency goals. Leo flags if I haven't talked to my friend in Atlanta in three months.
Reminders_Follow_Ups_Tracker.md: Conditional triggers ("after travel booking finalized → share itinerary with emergency contact"), recurring reminders (pay bills first of every month), week-specific items.
Task_Adds_Log.md: Every new task I mention gets logged here with deadline, goal alignment, dependencies. Reviewed and incorporated into plans during weekly reviews.
Career_Themes: The three themes I use in career conversations. Leo references this when I have career-related meetings coming up.
Listen_List.md: What I consume during commute time. Lightweight tracking for future automation.
Here's what that looks like in practice: I tell Leo "finished the portfolio review and sent the follow-up email to Kelly." Leo updates the Weekly Completion Log with both tasks, marks them complete on Current_Task_List, checks the Week Plan and updates today's checkboxes, asks if Kelly should be tracked in Community Connections (they're not already there), and logs the completion in Task_Adds_Log if either task was recently added. Five file updates from one statement. That's the interconnection that makes this feel like a system instead of a collection of documents.
How I Actually Use This, Some Examples
Daily end-of-day: "Here's what got done today" → Leo generates the day's updated completion log, task list, week plan, reminders tracker. Asks if anything else should be added, and generates updated files for download.
Sunday weekly review: Leo presents last week's completions, reviews task additions, proposes an updated task list, checks community connections, and drafts next week's plan for my review and approval.
Ad hoc: "I need to prep for this call" → Leo references the context documents and suggests using one of the other projects to prep, based on what kind of call it is (Catching up with a colleague? Use the Networking project.)
Where the System Gets Useful
The custom instructions turn Leo from a helpful assistant into something closer to a chief of staff. I didn't just give it files - I gave it operating principles.
Like this: "When I ask for priorities, lead with: What time of day? How much time is available?" Because my 8am brain can handle strategic work. My 6pm brain is often done with things that require deep thinking.
Or this: "If I mention a new person, ask if they should go in the Community Connections Tracker." So people don't slip through the cracks just because I didn’t tell Leo to track this contact.
Or this: "Build projects take priority over home maintenance, but not over in-person conversations. Flag when Bucket 3 tasks are crowding out Bucket 1 work." Because, left to my own devices, I'll spend three hours building and reading and fixing stuff instead of taking in-person meetings.
The instructions create behavior patterns. Leo doesn't just respond to what I ask - it knows what questions I should be asking myself.
What Actually Worked
This system has been delivering:
Weekly reviews that took 10 minutes instead of an hour
Automatic flagging when build work was getting crowded out
Community connections I'd have let slide
Task additions captured and categorized immediately
Progress tracking without manual log updates
The structure was sound. The files were right. The instructions were clear.
The problem? I'd built decision intelligence (Leo interpreting priorities, suggesting actions) before establishing reliable data flow (consistent task completion tracking, systematic file updates).
Now you know what needs rebuilding. Foundation first. Intelligence second.
Next time: The rebuild.
———-
Guest writer: Nisha Pillai transforms complexity into clarity for organizations from Silicon Valley startups to Fortune 10 enterprises. A patent-holding engineer turned MBA strategist, she bridges technical innovation with business execution—driving transformations that deliver measurable impact at scale. Known for her analytical rigor and grounded approach to emerging technologies, Nisha leads with curiosity, discipline, and a bias for results. Here, she is testing AI with healthy skepticism and real constraints—including limited time, privacy concerns, and an allergy to hype. Some experiments work. Most don't. All get documented here.
—--
This is extracted from the Gemini’s website:
Nano-Banana Pro supports up to 14 reference images (6 with high fidelity). This allows for "Identity Locking": placing a specific person or character into new scenarios without facial distortion.
Best Practices:
Identity Locking: Explicitly state: "Keep the person's facial features exactly the same as Image 1."
Expression/Action: Describe the change in emotion or pose while maintaining the identity.
Viral Composition: Combine subjects with bold graphics and text in a single pass.
Example Prompts (upload the Image 1 - image of the presenter):
"Design a viral video thumbnail using the person from Image 1.
Face Consistency: Keep the person's facial features exactly the same as Image 1, but change their expression to look excited and surprised.
Action: Pose the person on the left side, pointing their finger towards the right side of the frame.
Subject: On the right side, place a high-quality image of a delicious avocado toast.
Graphics: Add a bold yellow arrow connecting the person's finger to the toast.
Text: Overlay massive, pop-style text in the middle: '3分钟搞定!' (Done in 3 mins!). Use a thick white outline and drop shadow.
Background: A blurred, bright kitchen background. High saturation and contrast."Try it for your own thumbnails (after tweaking the above prompt to suit your video, of course!)
The Impossible Prompt that AI LLMs just can’t crack!
Create an image that displays two seven-pointed stars, two eight-pointed stars, and two nine-pointed stars. All stars are connected to each other, except for the ones with the same number of strands. The lines connecting the stars must NOT intersect.Try it out first manually and then on your favorite LLMs. Watch them fail!
Credit to teodordyakov
Perplexity just leaned into holiday purchases.
Virtual try-on is now live for all Pro and Max subscribers. Upload a photo. Get a clean digital avatar.
Try on outfits instantly while you shop inside Perplexity.
This is a glimpse of where AI interfaces are headed: search + reasoning + commerce blended into one seamless loop. The line between “browsing” and “doing” gets thinner every week.
URLs (affiliate): Perplexity