Think Differently. Solve the Impossible.

A conversation with Manas Talukdar (Senior Industry Leader in Enterprise AI and Data Infrastructure

Artificial intelligence has become the most overhyped and misunderstood technology story of our time. Some people talk about it like a coming super-being that will outthink humanity, take away jobs, and quietly decide the fate of civilization. I sat down with Manas Talukdar, an industry leader in data infrastructure for AI, to pull AI out of the fog and put it back on solid ground: what it is, what it isn’t, and what it means for your work, your privacy, and your ability to stay relevant.

Manas has built mission-critical AI systems across sectors, bootstrapped multiple products from ideation to prototype, holds multiple patents in data processing and machine learning, and works closely with AI labs on specialized “ground truth” training data—the kind that actually makes models smarter in specific domains. (As noted in the show: none of this reflects any employer endorsement—past or present.) Here’s what we uncovered.

AI Myths, Truths, and the Future of Work
 

1) The myth: “AI is intelligent like a human”

A major theme of our conversation was simple: what most people call “AI” today is not human-like intelligence. Manas drew a clear line between:

• AGI (Artificial General Intelligence): human-like intelligence that can reason from unknown information, adapt, and generate insight in new situations.
• Today’s mainstream AI (LLMs and most ML systems): models trained on huge amounts of data that generate outputs based on learned patterns—powerful, but not truly “reasoning” the way humans do.

That distinction matters because the public debate often assumes AI is already near-human, when in reality we’re mostly dealing with advanced pattern engines. A helpful mental model Manas offered: today’s LLMs are probabilistic systems. They behave as an ultra-scaled-up version of a “type-ahead” autocomplete, trained on massive bodies of text, with transformer architectures that make them far more sophisticated than a basic prediction algorithm.

The result can look like intelligence, especially when the answer is well-written. Well-written isn’t the same thing as “true,” and it isn’t the same as “reasoned.”

2) What AI really does well: synthesis and format

At one point I said what many people feel: using ChatGPT can feel like “a super Google.” Manas sharpened that: with search, you might read 10 pages, then you do the reasoning and synthesis. With LLMs, the model has already absorbed patterns across tons of sources and can combine disparate material into a clean, well-formed response. That’s the big productivity leap:

• Search engines gives you fragments
• LLMs give you a coherent narrative

The warning is built into the benefit. It can generate a coherent narrative easily, and people assume it must be right.

3) The truth about training data: AI can only learn what we feed it

We talked through a concrete example: a model that detects broken bones from X-rays. The process Manas described is a good reality-check:

1. You collect many examples (X-rays at a doctors office).
2. Humans label them,  literally drawing boxes around the fracture area, thereby creating “ground truths”.
3. The model learns patterns from those labeled examples.
4. The model can then generalize to similar images— within limits.  If the training set is incomplete, the model’s output will be incomplete. This is where a lot of the “AI will replace humans” fear collapses into something more accurate: AI is dependent on humans to define the truth it learns from.

That doesn’t mean AI can’t outperform humans in narrow areas. Manas pointed out that models can sometimes detect signals humans miss, especially when trained on massive datasets beyond what a human brain can hold. The operating mode is still: data → pattern learning → probabilistic output. Not magic. Not consciousness.

4) “New knowledge” is still a hard problem

One of the most important sections of the conversation was about something the industry still struggles with: How do you get AI systems to reliably respond to brand-new information?

If the model wasn’t trained on it, it may not know it. The “new knowledge” challenge becomes: how do we add knowledge safely, accurately, and quickly. That’s where we discussed platforms prompting “experts” to answer questions. I described how I’d received what looked like genuine human questions—about coaching, climbing, songwriting— to realize later they were AI-generated prompts designed to harvest expert responses for training or evaluation. Manas confirmed this general pattern exists: systems can use expert answers for evaluation and further training. That leads to a fair question:

Should people be cautious about answering “expert prompts” online?

Manas’s view was clear: transparency is essential. If an organization is collecting answers to train models, it should clearly state that intent. Otherwise, you get a quiet value transfer: humans provide expertise and labor, models absorb it, and the humans get little or no credit, control, or compensation.

5) Privacy: the real issue isn’t “AI,” it’s responsibility

We moved into privacy using a practical frame: How do we get good data while keeping people safe and private? Manas explained that privacy risk often comes down to how companies handle PII (personally identifiable information) and sensitive attributes. The core idea is PII should be anonymized before training, and safeguards can be implemented at multiple points:

1. Before training (mask/anonymize the data)
2. On output (mask sensitive details when the model responds)
3. On prompting (intercept requests that ask for PII and refuse)

In other words, privacy isn’t solved by AI. it’s solved, or violated, by the systems and policies surrounding it. If you’re in a minority group, you often want to be represented in datasets so systems don’t ignore you. You don’t, however, want sensitive personal details to become casually retrievable. That’s the challenge: inclusion without exposure.

6) The self-driving car dilemma: AI can’t “decide” morality without a moral framework

We tackled a classic scenario people love to use as evidence that AI will become morally superior, the case of when a self-driving car must choose between hitting an old man or a child. Some believe AI will know to hit the old man. That dilemma is hard even for humans, and the car wouldn’t have access to the hidden context people add later. For example, what if the old man was actually a scientist working to find a cure for the child who was in sick and going to die without the man’s help?

Moral dilemmas require a moral framework, not just pattern prediction. Even if we can formalize the framework, we  face a legal landmine: Who is liable—the car manufacturer, the model developer, the data provider, or someone else? Manas  referenced the broader tradition of “safe system rules” (including the spirit of Asimov’s Three Laws of Robotics) as an example of how humans try to encode safety, yet still struggle with edge cases. Bottom line: we’re not at “AI moral wisdom.” We’re at “humans arguing about how to encode human values into machines.”

7) “Sentient AI” isn’t here. and we’re not close

We clarified the word sentient—capable of thinking and feeling like a human. Manas connected sentience to AGI. If people mean sentient, reasoning intelligence, that’s basically the AGI category, and we’re not remotely there.

One of the most practical signs he mentioned: LLMs can become less accurate the further they go, especially in long outputs, because of their probabilistic nature. They can drift. That drift isn’t a sign of “thinking.” It’s a sign of pattern continuation.

8) The gold rush is training data, and it’s creating new work

AI labs are running out of free public data. Most major models have already been trained on broadly similar internet-scale corpora, which is why different LLMs can feel similar. The frontier is in specialized, high-quality training data. The “broken bone X-ray” example is not hypothetical, it’s the template.

Manas’s work involves helping AI labs get domain-specific labeled data—for cardiology, astrophysics, and other specialized areas—so models can improve where the internet doesn’t provide clean ground truth. This leads to a future-of-work insight most people miss: AI doesn’t only automate work—it also creates new categories of work:

• Expert labelers and reviewers (human-in-the-loop)
• Domain specialists who validate outputs
• Synthetic data evaluators
• Sata infrastructure engineers
• Safety, privacy, and governance roles

The employment question becomes less, “will AI take jobs?” and more, will you learn to use AI as leverage, or refuse to adapt. The fear is overstated, but people who don’t learn to leverage AI may be vulnerable in certain fields.

9) The “right” use of AI: make life easier, don’t replace meaning

We hit a cultural nerve with a meme that resonated with both of us: “I want AI to do my laundry and dishes. I don’t want it to write my article.” That captures a healthy boundary. Let AI reduce the logistical burden of life; Don’t outsource the parts of being human that create meaning—art, voice, love, originality

We talked about writing: AI can provide scaffolding (structure, ideas); Copying and pasting a love letter is a kind of self-erasure. That’s the theme in one sentence: Use AI as a tool, not as a replacement for your agency. The best use is:

• “Give me an outline.”
• “Help me find the right tone.”
• “Offer a few phrasing options.”
• Then write it in your words.

10) Advice for students: foundations, projects, and scrappy entry points

For people entering the field, Manas emphasized:

• Build fundamentals: linear algebra, ML, deep learning
• If you’re pivoting later: structured online specializations (Coursera-style paths)
• Do personal projects
• Look for early-stage opportunities, even volunteer roles, to get real experience
• Expect a tougher entry-level market: companies are favoring experienced hires right now, but doors still open through startups and applied work

The “meta-advice” he ended on is timeless. If you love it enough to keep learning and building, you’ll find a path.

The practical conclusion: How to think about AI without falling for the hype

If you take nothing else from this episode, take this:

1. Today’s AI is not sentient and not AGI.
2. LLMs are powerful synthesizers. They are also probabilistic and can drift.
3. Training data is the real engine, and it needs humans.
4. Privacy and safety are design responsibilities, not automatic guarantees.
5. AI will shift work toward new opportunities for people who adapt.
6. The best role for AI is enabling humans, especially in the parts of life that matter.

AI won’t be “smarter than you” in the way people fear. It can outperform you in narrow tasks, scale your output, and amplify your creativity. Stay in the driver’s seat. That may be the real future of work. It's not humans versus AI. It's humans who can think clearly with AI, versus humans who cannot.

Editor’s Note: This article is based on my podcast interview with Manas Talukdar, published in August, 2024. The ideas discussed here originate from that conversation. The structure, emphasis, and commentary are my own. Any errors or interpretations should be attributed to me, not to Manas Talukdar. 

Watch or listen to Artificial Intelligence With Senior Industry Leader Manas Talukdar: