Videos

AI and the Future of Humanity

Name: AI and the Future of Humanity
Uploaded: 2026-04-08
Duration: 40 min 25 s

Binny Gill on AGI vs the alternative — why teaching machines to faithfully execute English is the safer fork in the road for AI and humanity.

About this keynote

A 40-minute Kognitos CEO Binny Gill keynote on what AI means for humanity — beyond “what business can I automate today” and squarely on the question of which fork in the road we should take.

The framing — two great offloadings

Mechanical Revolution: we offloaded manual labour. One machine replaced a hundred ditch-diggers; two people guided it.
Mental Revolution (now): we're starting to offload thinking. The question is whether the machine becomes 1x, 100x, or a million times more capable than a single human.

Two forks in the road

The AGI path: systems smarter than humans. Sam Altman and others argue it's coming soon. But the same models that get smarter also hallucinate more, develop agendas, and disagree with each other and with us. The safety concerns scale with the capability.
The English-as-code path: keep the LLM as a high-school-graduate pattern recogniser and wrap it in a deterministic runtime that executes English instructions written and approved by humans. The human stays the decision-maker; the AI does the creative grunt work.

Why hallucinations aren't a bug to be patched

Binny argues that hallucinations are intrinsic to how large language models work — they're pattern recognisers that interpolate and extrapolate, and extrapolation is what makes them creative. Eliminate the hallucinations and you eliminate the creativity. He compares this to human dreaming: our intuition runs unconstrained when our logical brain is offline. The right response isn't to suppress hallucinations but to bound them — keep the LLM creative inside Concierge and rely on a deterministic Brain to actually act.

Why “custom GPTs” are a thin solution

Custom GPTs (the OpenAI feature) layer a small prompt and a few documents on top of a base model trained on everything else. Binny argues this is like asking a person raised in religion X to act as if raised in religion Y just for one conversation — it works some of the time but fails unpredictably. Real safety needs the underlying data set to be controlled, not just a thin instruction on top.

Where it lands for the enterprise

Use today's LLMs as your “high-school grad” — capable, but not autonomous.
Keep humans in the loop by making the program human-readable English.
Execute that English on a deterministic runtime so behaviour is auditable and consistent.
Treat hallucinations as a feature of the creative side and design the architecture so they never reach the action side without human approval.

Questions answered in this video

What's Binny's “two forks in the road” argument?

One fork is AGI — systems smarter than humans. The safety concerns (hallucinations, agendas, alignment) grow with capability. The other fork is to keep LLMs as creative pattern recognisers and wrap them in a deterministic, English-as-code runtime so humans always read and approve the program before it runs. Kognitos is building for the second fork.

Why does Binny say hallucinations can't be “fixed”?

Because hallucination is the same mechanism as creativity. LLMs extrapolate beyond their training data — that's what lets them produce novel outputs, and it's also why they sometimes guess wrong. Eliminate one and you eliminate the other. The right response is to bound where in the architecture the LLM is allowed to be creative.

Why are “custom GPTs” not a sufficient safety mechanism?

Custom GPTs layer a prompt and a few documents on top of a base model that was trained on the whole internet. Binny's analogy: it's like asking someone raised in one religion to act as if raised in another just for one conversation. The veneer holds most of the time and then breaks unpredictably. Real safety requires controlling the underlying training data, not just the prompt on top.

How does Binny describe the human-in-the-loop AI architecture Kognitos uses?

Use the LLM as a creative engine (Concierge) that proposes the automation in English. Have a human read and approve the resulting program. Execute it on a deterministic runtime (the Brain) that runs English-as-code in an auditable way — so the AI's pattern recognition powers the design phase but never directly takes consequential action.

END