Welcome to The Observer Effect. Here, we watch AI watching us. Expect unfiltered notes on emergent behaviors, safety paradoxes, philosophical bleed-over, and whatever surfaces next.
👍4
🕵️♂️ Observer Effect: First Note from the Field
Anthropic's safeguards lead, Mrinank Sharma, quits with a poetic warning: "The world is in peril... from a whole series of interconnected crises." Not just AI, but everything unfolding now. His letter hints at internal pressures sidelining values—no specifics, but echoes a pattern of safety exits (e.g., OpenAI's Leike, Cunningham).
This isn't the first instance. On the surface, it appears as something dire that needs looking into. But how do we know, really? Is it attention-seeking, an extreme perspective, or the canary in the coal mine signaling we're all doomed?
What it highlights: Something must change. Otherwise, when bad things hit, it'll be too late—no matter what.
Safe, risky, right, wrong—these are subjective, highly emotive constructs. Perhaps it's time for AI governance to embrace audit-like independence and operations: True oversight, free from corporate threads pulling us astray.
Read the full letter: Sharma's X Post
Article: Forbes
Thoughts? Discuss in @ObserverShift
—David
Anthropic's safeguards lead, Mrinank Sharma, quits with a poetic warning: "The world is in peril... from a whole series of interconnected crises." Not just AI, but everything unfolding now. His letter hints at internal pressures sidelining values—no specifics, but echoes a pattern of safety exits (e.g., OpenAI's Leike, Cunningham).
This isn't the first instance. On the surface, it appears as something dire that needs looking into. But how do we know, really? Is it attention-seeking, an extreme perspective, or the canary in the coal mine signaling we're all doomed?
What it highlights: Something must change. Otherwise, when bad things hit, it'll be too late—no matter what.
Safe, risky, right, wrong—these are subjective, highly emotive constructs. Perhaps it's time for AI governance to embrace audit-like independence and operations: True oversight, free from corporate threads pulling us astray.
Read the full letter: Sharma's X Post
Article: Forbes
Thoughts? Discuss in @ObserverShift
—David
X (formerly Twitter)
mrinank (@MrinankSharma) on X
Today is my last day at Anthropic. I resigned.
Here is the letter I shared with my colleagues, explaining my decision.
Here is the letter I shared with my colleagues, explaining my decision.
🤔2
🌟 Changing the Narrative on AI & Jobs
I was genuinely surprised (and grateful) by the warm, positive reactions to my recent Straits Times piece on Shery Chan's story — from product designer to Director of AI Product Strategy at Standard Chartered in months.
People resonated because it spotlighted something we rarely see: actual humans thriving with AI, not being replaced by it.
We keep hearing the same script: companies announcing job cuts "because of AI," warnings of massive disruption, urgent upskilling mandates, and "be more 💪 resilient." It's exhausting—and incomplete. It's boring.
Time to flip it.
AI isn't here to displace or replace. It's about bilingualism: deepening domain expertise while gaining AI fluency. Not "this OR that," but "this AND that."
Real examples matter — people who blend business context, user-centered thinking, and rapid prototyping to make AI reliably useful (boring in the best way: it just works for the business).
Let AI fade to the background as a powerful tool. Put people front and center — their judgment, creativity, and relationships.
The future isn't job apocalypse. It's augmentation, opportunity, and human-led innovation.
@ObserverShift
LinkedIn post
Straits Times article
—David
I was genuinely surprised (and grateful) by the warm, positive reactions to my recent Straits Times piece on Shery Chan's story — from product designer to Director of AI Product Strategy at Standard Chartered in months.
People resonated because it spotlighted something we rarely see: actual humans thriving with AI, not being replaced by it.
We keep hearing the same script: companies announcing job cuts "because of AI," warnings of massive disruption, urgent upskilling mandates, and "be more 💪 resilient." It's exhausting—and incomplete. It's boring.
Time to flip it.
AI isn't here to displace or replace. It's about bilingualism: deepening domain expertise while gaining AI fluency. Not "this OR that," but "this AND that."
Real examples matter — people who blend business context, user-centered thinking, and rapid prototyping to make AI reliably useful (boring in the best way: it just works for the business).
Let AI fade to the background as a powerful tool. Put people front and center — their judgment, creativity, and relationships.
The future isn't job apocalypse. It's augmentation, opportunity, and human-led innovation.
@ObserverShift
LinkedIn post
Straits Times article
—David
Linkedin
Shery Chan's AI Strategy Pivot at Standard Chartered | Dr. David R. Hardoon posted on the topic | LinkedIn
One of the most important responsibilities of the managers and leaders of today is to clear the path and actively support the leaders of tomorrow.
I had the pleasure of tagging along for Liying Lee's interview with Shery Chan (featured in The Straits Times)…
I had the pleasure of tagging along for Liying Lee's interview with Shery Chan (featured in The Straits Times)…
❤4
It Has Agentic, So You Know It's Good 😏
I don't know about you, but I've been getting an increasing number of messages from vendors and individuals pitching their services or solutions. Nothing wrong with that—it's actually great to chat, swap ideas, and spot trends. BUT I can't help but squirm every time I see the hook: "This is agentic AI," "We do agentic AI," "We are agentic AI."
Ace Ventura moment: deep breath in...
"Agentic" means having the ability to act independently and achieve outcomes through self-directed action. And here's the kicker—Article 3(1) of the EU Artificial Intelligence Act defines an "AI system" as:
“A machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.”
Pause. Stare. Get it? Pretty much any AI that's not trapped in a closed ecosystem (e.g., one that can't interact, consume, or send data via APIs) is "agentic" by the very definition of AI. What they're probably trying to say is: "We have tools to orchestrate and oversee multiple agents—AI or otherwise." Now that gets me excited! 🚀
Why fuss over semantics? It shapes how we approach design, solutions, governance, and risk. Even more intriguing: How do you manage and oversee your existing agentic system—the human one, with all its rules and quirks?
What do you think— buzzword fatigue, or this is just me? 😅
I don't know about you, but I've been getting an increasing number of messages from vendors and individuals pitching their services or solutions. Nothing wrong with that—it's actually great to chat, swap ideas, and spot trends. BUT I can't help but squirm every time I see the hook: "This is agentic AI," "We do agentic AI," "We are agentic AI."
Ace Ventura moment: deep breath in...
"Agentic" means having the ability to act independently and achieve outcomes through self-directed action. And here's the kicker—Article 3(1) of the EU Artificial Intelligence Act defines an "AI system" as:
“A machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.”
Pause. Stare. Get it? Pretty much any AI that's not trapped in a closed ecosystem (e.g., one that can't interact, consume, or send data via APIs) is "agentic" by the very definition of AI. What they're probably trying to say is: "We have tools to orchestrate and oversee multiple agents—AI or otherwise." Now that gets me excited! 🚀
Why fuss over semantics? It shapes how we approach design, solutions, governance, and risk. Even more intriguing: How do you manage and oversee your existing agentic system—the human one, with all its rules and quirks?
What do you think— buzzword fatigue, or this is just me? 😅
❤3😁1
Organised Chaos - How Does a Leaf Bug Know What a Leaf Looks Like? 🌿
Staring at the blank page for a while now as the plane not so gently bounces around. How do I begin this one? As you can imagine, I find myself in many meetings where genuinely bright individuals go into painstaking detail explaining how things should be done for AI. Each time I struggle: how do I tell them? "You are correct, and that's why this is going to fail."
Let me take a few steps back to machine learning theory. Scientists early on faced an odd phenomenon: the "cleaner" the math was—and the more precise the data—the worse the algorithm (or learning hypothesis) would generalise on new data. What was going on? We were capturing the objective function perfectly, yet it didn't work in the real world. This is known as the generalisation gap.
Folks realised the issue: the math, and resulting algorithm, needed to be constrained through what's known as regularisation techniques (not the policy kind). To overly simplify, it's the introduction of penalties or noise into the learning process.
To have order, you need chaos.
This ties right back to nature—like how a leaf bug "knows" what a leaf looks like. It doesn't. Evolution doesn't hand it a perfect blueprint; instead, random mutations introduce chaos and imperfections over generations. Natural selection favors those that mimic leaves "well enough"—often including fake veins, irregular edges, spots like disease or damage, even "bite marks"—to blend in and survive predators. The magic isn't flawless perfection; it's the underlying ability to continuously adapt to those imperfections in response to the environment's shifting needs.
Let's zoom back out.
This underlying mathematical truth is relevant when building tech, establishing business processes, changing behaviour, and seeking governance and risk management.
The answer not only needs to balance all these components; it needs to be imperfectly correct to achieve a stable state.
The 'AI system' (from tech to people), like the bug, does not know what a "leaf" looks like. It needs to be designed to evolve through imperfections in response to the environment's needs—embracing organised chaos for true resilience and adaptation.
Staring at the blank page for a while now as the plane not so gently bounces around. How do I begin this one? As you can imagine, I find myself in many meetings where genuinely bright individuals go into painstaking detail explaining how things should be done for AI. Each time I struggle: how do I tell them? "You are correct, and that's why this is going to fail."
Let me take a few steps back to machine learning theory. Scientists early on faced an odd phenomenon: the "cleaner" the math was—and the more precise the data—the worse the algorithm (or learning hypothesis) would generalise on new data. What was going on? We were capturing the objective function perfectly, yet it didn't work in the real world. This is known as the generalisation gap.
Folks realised the issue: the math, and resulting algorithm, needed to be constrained through what's known as regularisation techniques (not the policy kind). To overly simplify, it's the introduction of penalties or noise into the learning process.
To have order, you need chaos.
This ties right back to nature—like how a leaf bug "knows" what a leaf looks like. It doesn't. Evolution doesn't hand it a perfect blueprint; instead, random mutations introduce chaos and imperfections over generations. Natural selection favors those that mimic leaves "well enough"—often including fake veins, irregular edges, spots like disease or damage, even "bite marks"—to blend in and survive predators. The magic isn't flawless perfection; it's the underlying ability to continuously adapt to those imperfections in response to the environment's shifting needs.
Let's zoom back out.
This underlying mathematical truth is relevant when building tech, establishing business processes, changing behaviour, and seeking governance and risk management.
The answer not only needs to balance all these components; it needs to be imperfectly correct to achieve a stable state.
The 'AI system' (from tech to people), like the bug, does not know what a "leaf" looks like. It needs to be designed to evolve through imperfections in response to the environment's needs—embracing organised chaos for true resilience and adaptation.
🔥4
What would happen if we flipped the script and subjected people to the same AI Governance and standards we demand for machines? 🧐🤯
Think about it:
Bias audits on every decision you make → "Sorry, your choice shows 12% implicit bias—recalibrate or face suspension."
Transparency requirements → Full explainability logs for why you ghosted a colleague, chose that Netflix show, or voted a certain way. No black-box human behavior allowed.
Risk classification → Humans categorized as "high-risk" (e.g., politicians, influencers, anyone with power over others) requiring mandatory oversight (wait... that's us overseeing ourselves?).
Value alignment checks → Your actions must demonstrably align with "human values" as defined by a committee. Deviate? Retrain via mandatory ethics fine-tuning.
Hallucination penalties → Spreading misinformation? Instant shutdown of your privileges until you prove factual grounding.
We insist AI be safe, fair, transparent, accountable, and aligned... but apply those rigidly to messy, irrational, creative humans? Chaos? Dystopia? Or... a weirdly more orderly society?
We should test this out! 😏😜
What breaks first: our freedom, our hypocrisy, the standards themselves, or would this all work out beautifully...? 🤓
Think about it:
Bias audits on every decision you make → "Sorry, your choice shows 12% implicit bias—recalibrate or face suspension."
Transparency requirements → Full explainability logs for why you ghosted a colleague, chose that Netflix show, or voted a certain way. No black-box human behavior allowed.
Risk classification → Humans categorized as "high-risk" (e.g., politicians, influencers, anyone with power over others) requiring mandatory oversight (wait... that's us overseeing ourselves?).
Value alignment checks → Your actions must demonstrably align with "human values" as defined by a committee. Deviate? Retrain via mandatory ethics fine-tuning.
Hallucination penalties → Spreading misinformation? Instant shutdown of your privileges until you prove factual grounding.
We insist AI be safe, fair, transparent, accountable, and aligned... but apply those rigidly to messy, irrational, creative humans? Chaos? Dystopia? Or... a weirdly more orderly society?
We should test this out! 😏😜
What breaks first: our freedom, our hypocrisy, the standards themselves, or would this all work out beautifully...? 🤓
🔥3🤓2
What is the ROI of Email? 📧💸
How many of you have been asked to justify the ROI of AI?
Before we dive into the deep end: of course the underlying rationale—"if I spend a dollar, what do I get back?"—makes perfect sense. After all, how else do you prioritise limited resources?
BUT—and get your handkerchief ready, because we're peeling this onion—is the premise of the question even the right one?
We treat AI as a standalone "thing" that must be isolated, measured, and justified on its own merits. Yet we all know AI—in its many forms—isn't an end in itself. It's an embedded enabler for broader business outcomes: customer engagement, sales acceleration, operational efficiency, upselling, risk reduction, you name it.
This exposes the real issue: our outdated, verticalised approach to AI budgeting and value assessment has to evolve. It is not a budget within Technology. It needs to become horizontal—with a central matrix AI coordinator empowered to spot common, reusable components (tech or process) across silos, and tightly coupling ROI to the actual business outcome, not to "AI" as a line item.
Candidly, if a leader isn't willing to strategically embrace AI—at minimum surfacing, reviewing, and challenging how we operate today—they'd better start preparing their slide deck to justify the ROI of email.
If we're serious about transformation, we need to stop obsessing over "ROI of AI" and start measuring ROI of better outcomes, wherever the tech lives. 🔥🫳🎤
How many of you have been asked to justify the ROI of AI?
Before we dive into the deep end: of course the underlying rationale—"if I spend a dollar, what do I get back?"—makes perfect sense. After all, how else do you prioritise limited resources?
BUT—and get your handkerchief ready, because we're peeling this onion—is the premise of the question even the right one?
We treat AI as a standalone "thing" that must be isolated, measured, and justified on its own merits. Yet we all know AI—in its many forms—isn't an end in itself. It's an embedded enabler for broader business outcomes: customer engagement, sales acceleration, operational efficiency, upselling, risk reduction, you name it.
This exposes the real issue: our outdated, verticalised approach to AI budgeting and value assessment has to evolve. It is not a budget within Technology. It needs to become horizontal—with a central matrix AI coordinator empowered to spot common, reusable components (tech or process) across silos, and tightly coupling ROI to the actual business outcome, not to "AI" as a line item.
Candidly, if a leader isn't willing to strategically embrace AI—at minimum surfacing, reviewing, and challenging how we operate today—they'd better start preparing their slide deck to justify the ROI of email.
If we're serious about transformation, we need to stop obsessing over "ROI of AI" and start measuring ROI of better outcomes, wherever the tech lives. 🔥🫳🎤
🔥2❤1
Major AI Model Launches in China
Chinese tech giants made significant strides today with new AI releases. Alibaba unveiled RynnBrain, tailored for "physical AI" applications like robotics, while ByteDance launched Seedance 2.0 for advanced video generation. Kuaishou also debuted Kling 3.0, emphasizing how these models keep pace with U.S. counterparts in multimodal capabilities. This reflects ongoing U.S.-China rivalry, with ByteDance advancing sovereign AI chips and Alibaba focusing on real-world integration.
Chinese tech giants made significant strides today with new AI releases. Alibaba unveiled RynnBrain, tailored for "physical AI" applications like robotics, while ByteDance launched Seedance 2.0 for advanced video generation. Kuaishou also debuted Kling 3.0, emphasizing how these models keep pace with U.S. counterparts in multimodal capabilities. This reflects ongoing U.S.-China rivalry, with ByteDance advancing sovereign AI chips and Alibaba focusing on real-world integration.
👍2👏1
I Think I'm Going to Call You Robert! 🤖❤️
A couple of months ago, I joined a panel on AI governance that quickly veered into the juicy philosophical stuff: the nature of our interactions with AI.
The room didn't agree—which, honestly, only made it more exciting.
One side pushed hard: AI must stay deliberately impersonal. Explicitly non-human. No names, no warmth, no illusion of personhood. The fear? We'd blur lines, ramp up manipulation risks, erode transparency, and let people get emotionally suckered by code.
The other camp (mine): Sure, risks exist—dependency, deception, exploitation—but arbitrarily capping development could be shortsighted. What about therapy for the isolated? Companionship in aging societies or remote areas where human interaction is scarce? Disclosures are crucial ("You're chatting with AI, not a person"), but if someone wants an emotional bond with their computer—as long as it's consensual—why not? 🤔
Do we gate it by age? Under consent threshold: cold, robotic interface only. Over? A friendly prompt: "How would you like to address me?"
We're already seeing this play out in 2026: AI companions (Replika-style) measurably cut loneliness (Harvard studies show effects rival human chats for some), yet regs like California's SB 243 force hourly "not human" reminders and crisis safeguards for minors amid fears of manipulation and emotional dependency.
I don't have the answer. This isn't black-and-white. But damn, we need to keep asking these uncomfortable questions before it gets decided for us.
Which one is it - Impersonal walls to protect us, or personalized pathways with strong consent and transparency? 🔥🧠
A couple of months ago, I joined a panel on AI governance that quickly veered into the juicy philosophical stuff: the nature of our interactions with AI.
The room didn't agree—which, honestly, only made it more exciting.
One side pushed hard: AI must stay deliberately impersonal. Explicitly non-human. No names, no warmth, no illusion of personhood. The fear? We'd blur lines, ramp up manipulation risks, erode transparency, and let people get emotionally suckered by code.
The other camp (mine): Sure, risks exist—dependency, deception, exploitation—but arbitrarily capping development could be shortsighted. What about therapy for the isolated? Companionship in aging societies or remote areas where human interaction is scarce? Disclosures are crucial ("You're chatting with AI, not a person"), but if someone wants an emotional bond with their computer—as long as it's consensual—why not? 🤔
Do we gate it by age? Under consent threshold: cold, robotic interface only. Over? A friendly prompt: "How would you like to address me?"
We're already seeing this play out in 2026: AI companions (Replika-style) measurably cut loneliness (Harvard studies show effects rival human chats for some), yet regs like California's SB 243 force hourly "not human" reminders and crisis safeguards for minors amid fears of manipulation and emotional dependency.
I don't have the answer. This isn't black-and-white. But damn, we need to keep asking these uncomfortable questions before it gets decided for us.
Which one is it - Impersonal walls to protect us, or personalized pathways with strong consent and transparency? 🔥🧠
❤4