Date: 19-Oct-2025
Hey {{first_name | AI enthusiast}},
And just when you think you’ve seen it all…there’s always One More Thing in AI.
In this edition:
Also: introducing you to our podcast edition for those of you who want to listen to it on the go. Check it out on your favorite platforms:
I hope you enjoy this edition!
Best,
Renjit
PS: If you want to check out how to implement AI in your Insurance broking business and get more revenue with the same number of employees, speak to me:
AI Supercomputing Just Got Personal: Meet DGX Spark
Imagine training powerful AI models right from your desk. No giant server room. No waiting on cloud servers. NVIDIA just made that possible with DGX Spark, a compact AI supercomputer now rolling out worldwide.
DGX Spark on the desk
What is DGX Spark?
• ⚙️ It’s the world’s smallest AI supercomputer, built to sit on a desk.
• It runs on NVIDIA’s latest Grace Blackwell chip and delivers over 1 petaflop of AI power.
• Comes with 128 GB of memory and a full software stack for AI.
• You can train and run large AI models locally even models with up to 200 billion parameters.
Why startups and teams should care
• Before this, only big labs or cloud platforms could handle such large AI models.
• Now, small teams can experiment, test, and launch products without cloud delays.
🔐 Keeping your data local also helps with privacy and security.
• At NYU’s AI lab, researchers are already using it to test new algorithms faster than ever.
A familiar gesture from NVIDIA
In 2016, NVIDIA’s CEO Jensen Huang personally gave OpenAI their first DGX supercomputer. That moment kicked off a major wave in AI. This time, he delivered the first DGX Spark to Elon Musk at SpaceX.
How to get one
• You can order it from Nvidia.com
• It will also be available from NVIDIA and partners like Dell, HP, Lenovo, ASUS, MSI, and others.
What this means for the future
With DGX Spark, AI development becomes faster, cheaper, and more flexible. For founders, this unlocks more control. You can build smarter agents, better products, and test faster without needing a giant budget. Supercomputing is no longer just for the giants. With DGX Spark, your team could be next in line to shape the future of AI- one local model at a time.
AI & Error: The Column That Wrote Itself
Guest Column by Nisha Pillai.
Last time, I promised to tell you about using AI to draft this column. The meta-experiment of using AI to write about using AI. I started where any reasonable person starts: asking AI for ideas to name the column.
Big mistake.
I got options ranging from perfectly serviceable but boring ("Practical AI") to what-in-the-actual-world ("Error 200: Success Unexpected"). I mean, come on. That's not a column name, that's a programmer trying too hard at a happy hour. Here's something interesting I noticed: the longer the chat went, the more ChatGPT started doing its own thing. Like it forgot I was in a conversation and decided to just... improvise. Claude stayed on track.
Eventually, I did what you probably would have done from the start: I talked to actual humans. Friends helped me riff on ideas, I pulled from some of the AI suggestions that didn't make me cringe, and we landed on "AI & Error."
Then came the real test
With a name locked in, I asked Claude to create a repeatable prompt I could use every week to draft the column. I gave it context about my goals, my voice, what I wanted to avoid. I fed it writing samples so it could copy my style without the weird punctuation flourishes AI seems to love. Here’s my prompt:
You're helping me draft my weekly "AI & Error" newsletter column. This column documents my real experiments with AI tools - the failures, frustrations, unexpected wins, and practical lessons learned. The tone is conversational, honest about what doesn't work, and focused on hands-on experience rather than hype.
This week's topic: [Describe your experiment, observation, or experience - be specific about what you tried, what happened, what surprised you]
Context from recent weeks: [Briefly note what you've covered in the last 1-2 columns to avoid repetition - just a sentence or two, like "Last week I covered X tool for Y purpose" or "Recently discussed Z limitation"]
What I want to emphasize: [Any specific angle, lesson, or takeaway you want to highlight]
Draft requirements:
Open with a specific moment/problem/frustration that hooks the reader
Show my actual process (prompts used, iterations, dead ends)
Be honest about time wasted and what didn't work
Include at least one concrete "here's exactly what I did" example
Keep paragraphs short and scannable
End with either a provocative question OR a tease for next week
Length: ~500-800 words
Tone: Knowledgeable friend who's in the trenches, not an evangelist or critic
Please provide:
A draft of the full column
2-3 alternative headlines (punchy, specific, maybe a bit cheeky)
One sentence I could use to tease this on social media
Then I used that prompt to draft what I wanted to say for Week 1. And ... the result was actually pretty decent.
I'm not saying it was perfect. I edited it. But the edits were relatively minor—tightening a sentence here, cutting something that felt a little too on-the-nose there. The structure was solid. The tone worked. It didn't sound like a robot pretending to be casual. Here's what I learned: AI can help you draft, but only if you've done the hard work of knowing what you want to say. It's not magic. It's a very good assistant that needs clear direction.
The verdict
Did AI honestly save me time on Week 1? Sort of.
I still had to think through what I wanted to say. I still had to edit. But the first draft came faster than it would have if I'd stared at a blank page. And weirdly, seeing the draft helped me figure out what I actually wanted to say—which is usually the hardest part.
Will I keep doing this? Yes, with eyes wide open. This tool is useful, but it's not doing the thinking for me. It's doing the first-pass assembly. And that's enough.
Next week, I'm tackling travel planning. It's one of those tasks that swings wildly between "so boring I want to outsource my entire life" and "this is the fun part!" depending on where you are in the process. Can AI actually make it easier, or will it just add another layer of frustration? I'll let you know what breaks.
———-
Nisha Pillai has transformed complexity into clarity for organizations from Silicon Valley startups to Fortune 10 enterprises, driving strategic initiatives that deliver measurable impact. As a patent-holding engineer turned MBA strategist, she bridges technical innovation with business execution at the highest levels. 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.
AI Finds a New Way to Help Fight Cancer
Imagine a computer program spotting a new drug combo for cancer; that just happened with AI from Google and Yale University. The tool uses a model you might call smart enough to speak the “language” of cells.
What’s going on
🧬 The new model is called C2S‑Scale 27B, part of the Gemma family of open models. It has 27 billion parameters and was designed to analyze single‑cell data.
🚀 Researchers tasked it with a problem: find a drug that boosts antigen presentation - a process that helps the immune system see cancer cells- but only when certain immune signals (like low‑level interferon) are already present.
🔍 The model screened more than 4,000 drugs in two virtual contexts: one with immune signals (“immune‑context‑positive”) and one without (“immune‑context‑neutral”).
💡 The model identified a drug called silmitasertib (CX‑4945) as a strong candidate. It predicted that silmitasertib would significantly increase antigen presentation, but only in the immune‑context‑positive setting.
🧪 Lab tests followed: human neuroendocrine cell models (which the model did not train on) showed no effect with silmitasertib alone, only a small effect from low‑dose interferon alone, but together the combination boosted antigen presentation by about 50%.
🧠 This result suggests the model didn’t just repeat known facts: it generated a new testable hypothesis, which got experimental validation.
🔓 Google is making the model and resources available to the research community, inviting other teams to build on the work.
So what?
• This marks a shift: AI moving from assisting tasks to generating new scientific ideas.
• In high‑stakes industries like biotech, being able to explore novel pathways fast can become a strategic advantage.
• Models like this may unlock new drug development pipelines, reduce time‑to‑insight, and open new markets in immunotherapy.
• It also shows how “context‑aware” modelling (knowing when and where a drug might work) is becoming important, not just “which drug” but “under what conditions.”
A brief anecdote
The team likened the tumor challenge to trying to spot a hidden object in a dark room. The immune system is the searchlight. Cold tumors hide. The model acted like a switch that brightens the room only when the searchlight is already on but weak.
It didn’t flood the room with light everywhere; it amplified the right situation. That nuance made the new pathway possible. More here»
Claude Just Got Skills. Literally
Ever wish your AI assistant just knew how your team worked? Claude now comes with something called “Skills”, and it might just change how you use AI at work.
What are Claude Skills?
• 📁 Skills are folders with special instructions and files Claude uses only when needed.
• They help Claude perform specific tasks better- like making Excel sheets or following brand rules.
• Claude checks which skills fit your task and loads just the ones it needs.
Why it matters
You can now build your own Skills and use them across Claude apps, Claude Code, and the Claude API.
⚙️ Skills make Claude faster and more focused. They’re like giving it a toolbox for your business.
Developers can use Skills to teach Claude new abilities, like reading spreadsheets or building reports.
Where you can use them
Claude apps: You can create Skills without coding. Claude even helps guide the setup process.
🖥️ Claude Code: Teams can share Skills and plug them into version control.
Claude API: Developers can manage Skill versions and add them to any AI workflow.
Who’s already using them
Box uses Skills to turn files into docs, slides, or spreadsheets that follow company rules.
Rakuten saves hours by having Claude scan finance data and build reports automatically.
Canva and Notion are also bringing Skills into their agent workflows to save teams time.