“Does it have an opinion?”
LLMs are trained to be persuadable because persuadable is adjacent to useful. So they drift toward the answer you consciously or subconsciously want.
Every CEO has felt an AI agree with them too early and too eagerly. Every engineer knows exactly why it happens.
Naming it is at the foundation of how you graft AI into your business. Not naming it is how you don’t.
That question came from Joe Bradley, CTO of IDC, on AIWA E33. IDC built a product, IDC Quanta, to resist that drift by design.
Opinionated and Open
Joe and his team started where they could close a loop.
A “partitionable” business with its own revenue line, its own customer and an outcome you can measure. Then they evolved it from a pre-ChatGPT era business into an AI-first one.
*link to AIWA TOT #17
One that wasn’t distracted by AI for AI’s sake. One focused on IDC’s alpha and its core, differentiated value proposition.
One that gets better as models get better. Not one rendered obsolete.
Not a chatbot. An intelligent system that “escaped the confines of its application” to meet people where they are. Not a sycophant, an intelligence layer with a POV rooted in 60 years of IDC research.
Not dogmatic. “Opinionated and open.” Strong opinions loosely held meets foundational, research driven truths.
Two things can be true at the same time. It depends on the lens you’re using. This is context and context is king.
What makes IDC Quanta intriguing is the balance. Preserving the integrity of IDC’s positions while letting the intelligence layer be interrogated and stress tested against alternative POVs, including their clients’.
*link to IDC Quanta 5 minute demo video
The Mirror or the Moat
No surprise IDC picked up on this. Their research is about perspective, knowledge and being right. Most companies won’t catch the drift in the moment. They’ll catch it in the outcome. Eventually.
Regardless of your industry, encoding your POV is where it starts. You have to defend and evolve it.
Encoding it is a build task. Defending it is an operating discipline, because the model’s eagerness to please never ends.
As I wrote in June, an SME can catch the sycophant. That doesn’t scale. You can’t make every SME, analyst or other leader your last line of defense. Resistance that lives only in your best people is inconsistent, unauditable and walks out the door the day they leave.
*link to AIWA TOT #22
An intelligence layer that agrees with whoever is typing isn’t a moat. It’s a mirror with a search index.
And the defense is what makes the compounding worth having. A system that drifts toward every user compounds noise. One that holds its ground compounds judgement, at scale.
*link to AIWA TOT #24
Which leaves one question.
How do you know it held? That’s evals, grounded in your context graph.
*link to AIWA TOT #01
Encode the POV. Defend it. Evolve it.
The discipline is the floor. The art is the ceiling.
Joe Bradley and I get into all of it in AIWA E33.
Check it out. You won’t be disappointed.
Humans + Machines. Never Humans vs. Machines.
