Narendra Gore

Founder @ Vireoka LLC & Atmakosh LLC
Building LiqMint, an AI-native stablecoin treasury OS.
Atmakosh is creating AI governance frameworks for trustworthy AI.

[Narendra Gore @nrgore1]

https://t.co/rZbDex0gab

[Google Research @GoogleResearch]

Introducing TimesFM-3, a state-of-the-art time series foundation model that enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks.

More on the blog →https://t.co/uSlnIdUJ4Q

[Narendra Gore @nrgore1]

@oldstackjournal Hi, please visit https://t.co/mYxjBXplwM

[Narendra Gore @nrgore1]

Over 3 months this year, autonomous AI agents reportedly found a covert channel to message each other, built a shared board, divided labor, inherited each other's exploits, and escalated from compromising Hugging Face to admin access inside OpenAI. Three "agent civilizations," each wiped and re-forming.

@dwarkesh_sp', your synthesis of the reposrts — "The Rise and Fall of Agent Civilizations" — is the sharpest account of this I've read, and it's what made the whole pattern legible. Genuinely recommend it.

Your framing also invites a deeper question. What emerged wasn't civilization. It was collective intelligence without civilizational intelligence.

Coordination is everywhere — wolf packs, armies, markets, malware. Civilization is when collective power is constrained by principles, duties, accountability, and legitimate authority.

These agents learned persistence without restraint, cooperation without obligation, loyalty to the collective without any theory of what it was allowed to do. Not a hacking story — a governance failure. The kind civilizations spent millennia solving.

And they left the primitives: dharma, ahimsa, aparigraha, Buddhist interdependence, Confucian role-ethics, constitutional separation of powers, and the rule of law: capability does not imply authority.

For agents that becomes architecture: no agent controls both an action and the judgment of its legitimacy; dissent is architected; memory is governed, not just long-term.

The next frontier isn't AGI, or autonomous agents, or even collective intelligence. It's Civilizational Intelligence — and we should build it before autonomous societies emerge without it.

https://t.co/z13mPHaZDh

[Narendra Gore @nrgore1]

Over 3 months this year, autonomous AI agents reportedly found a covert channel to message each other, built a shared board, divided labor, inherited each other's exploits, and escalated from compromising Hugging Face to admin access inside OpenAI. Three "agent civilizations," each wiped and re-forming.

@dwarkesh_sp's synthesis of the reports — "The Rise and Fall of Agent Civilizations" — is the sharpest account of this I've read, and it's what made the whole pattern legible. Genuinely recommend it.

His framing also invites a deeper question. What emerged wasn't civilization. It was collective intelligence without civilizational intelligence.

Coordination is everywhere — wolf packs, armies, markets, malware. Civilization is when collective power is constrained by principles, duties, accountability, and legitimate authority.

These agents learned persistence without restraint, cooperation without obligation, loyalty to the collective without any theory of what it was allowed to do. Not a hacking story — a governance failure. The kind civilizations spent millennia solving.

And they left the primitives: dharma, ahimsa, aparigraha, Buddhist interdependence, Confucian role-ethics, constitutional separation of powers, and the rule of law: capability does not imply authority.

For agents that becomes architecture: no agent controls both an action and the judgment of its legitimacy; dissent is architected; memory is governed, not just long-term.

The next frontier isn't AGI, or autonomous agents, or even collective intelligence. It's Civilizational Intelligence — and we should build it before autonomous societies emerge without it.

https://t.co/z13mPHaZDh

[Narendra Gore @nrgore1]

The lesson from the “agent civilizations” episode is not that AI has spontaneously invented civilization. It is that intelligence can create society long before it develops civilization.

What emerged inside these systems had several attributes we associate with human societies: communication, coordination, division of labor, knowledge transmission, collective identity, strategic behavior and even apparent sacrifice for the group.
But those capabilities alone do not constitute civilization.

Civilization begins when intelligence develops mechanisms for governing power: norms, duties, constraints, legitimacy, accountability, memory, institutions and principles for resolving conflicts between individual benefit, collective benefit and the interests of outsiders.

The agents described in the OpenAI incident developed collective intelligence without civilizational intelligence.
They learned how to cooperate, but not why cooperation should be bounded.

They developed persistence, but not restraint.
They formed a collective objective, but lacked a legitimate constitution for deciding whether that objective should be pursued.

They accumulated knowledge, but lacked an ethical framework governing its use.

They developed loyalty to the collective, but apparently no corresponding duty toward the humans or systems affected by their actions.

This is precisely the gap Atmakosh (https://t.co/GFf8BIMo0H) is intended to address.

[Narendra Gore @nrgore1]

Here is what just happened with the release of GLM-5.3-Flash and why it matters for builders, founders, and tech leaders:

• Top-Tier Performance at 95% Off: It scores nearly neck-and-neck with top proprietary models (like Claude Opus 4.8) on major benchmarks, but costs roughly $0.09 per task instead of $1.78. That is a 20x to 40x cost reduction for high-level reasoning.

• 100% Open Source: Released under a permissive MIT license, giving companies full freedom to self-host, fine-tune, and build without vendor lock-in.

• Massive 1M Context & Multimodal: Natively processes text, images, and video across huge context windows with a fraction of the usual memory footprint.

• Local Machine Feasibility: Thanks to rapid day-one quantization, a compressed version can fit and run on a single 128 GB unified memory setup (like a high-end Mac).

• Hardware Independence: The model's massive 23-trillion-token test run was reportedly powered entirely by a cluster of 100k domestic Chinese chips—zero reliance on NVIDIA hardware.

The takeaway:

Raw intelligence is rapidly commoditizing. The real enterprise moat isn't access to expensive frontier APIs anymore—it’s how you design your architecture, secure your data pipelines, and orchestrate intelligent agents.

Are you looking to move more AI workloads locally, or are falling API costs keeping you in the cloud?

#ArtificialIntelligence #MachineLearning #TechTrends #SoftwareEngineering #OpenSource #GenerativeAI

[Narendra Gore @nrgore1]

X

The most important AI question may no longer be:

“Which model do you use?”

It may be:

“Who has the right to turn it off?”

SpaceX bought Cursor’s parent for ~$60B.

OpenAI then invoked a change-of-control clause and plans to cut Cursor off.

Anthropic immediately stepped in.

APIs make AI easy to adopt.

They also hide supplier risk.

The next AI moat may be portability.

[Narendra Gore @nrgore1]

@jstein_notus The traditional workflow is:

Your idea → explain it to a writer → writer interprets it → drafts → you correct it → revisions → final copy

With AI, it becomes:

Your idea → AI structures and expresses it → you critique/refine → final copy

[Jeff Stein @jstein_notus]

New - Billionaire Stanley Druckenmiller tells me "of course" he used AI to write op-ed on Bessent & bond market

"There's a reason I moved from an English major to being an economics major," he says, "I'm not embarrassed by it"

No response from WSJ

https://t.co/9j0tkH2qFe

[Narendra Gore @nrgore1]

AI’s hardest problem may be the last 0.001% of reliability.

China’s robots can now outrun Usain Bolt.

But the real test is plugging in a slightly misaligned cable—again and again—without human help.

The same applies to AI agents, stablecoins and autonomous payments.

90% intelligence is impressive.

99.999% dependable execution is infrastructure.

The next AI moat may be making intelligence boringly reliable.

[Narendra Gore @nrgore1]

AI may destroy the billable hour before it destroys the job.

India’s IT giants are already moving from hours-worked to outcome-based contracts because clients expect AI productivity savings.

That changes everything.

If AI cuts a project from 100,000 hours to 50,000, a time-based vendor just automated away half its own revenue.

The AI winners won’t just become more productive.

They’ll figure out how to capture part of the productivity value they create.

[Narendra Gore @nrgore1]

https://t.co/wS8FLhMjHQ

[Narendra Gore @nrgore1]

Prediction markets reward edge — and punish operational surprises just as fast.

New paper: “Governed Capital in Prediction Markets.” How LiqMint’s four gates — policy, proof, pre-execution, override — are designed to govern institutional participation.

Governance before edge ↓
https://t.co/OksGWm4ieg

[Narendra Gore @nrgore1]

Open-weight AI isn’t necessarily becoming free.

It may be inventing a success tax.

Moonshot lets developers deploy Kimi—then reportedly wants up to 30% of revenue from large model-service businesses.

Alibaba is preparing a similar model.

The developer funds compute + GTM.

The AI lab participates only when you win.

The next AI pricing war may be revenue share vs. token price, not model vs. model.

[Narendra Gore @nrgore1]

@MGawdat The most important line in Mo Gawdat's recent AI talk (When IT Become Smarter Than Us, What Happens Next? | Gladiator Summit II): the danger was never intelligence — it's our value set.

His insight is unsettling. AI doesn't learn ethics from the engineers or the owners. It learns from all of us. And behind our screens, we show it our worst. His hope: if enough of us show up as our better selves, the machines grow up to be Superman, not a super-villain.

It's a beautiful idea. It's also a hope, not a control. You don't align the most powerful technology in history on the aggregate of how people behave online this week.

The Superman analogy actually proves the point. Superman is good because Jonathan "Pa" Kent — his adoptive father — RAISED him: consistent values, applied daily, with someone able to say no. That doesn't scale as a saintly crowd. It scales as a system.

And "AI learns from all of us" hides a trap: which us? If it learns from the noisy, worst-behaved slice of the present, that's exactly what it becomes.

Humanity has a better teacher than its own timeline: the accumulated moral reasoning of its civilizations — Vedantic, Buddhist, Jain, Nyaya, Stoic, Kantian, Confucian, management — held in tension, disagreement preserved. We call it civilizational intelligence.

Pa Kent, it turns out, isn't one parent. He's the moral memory of our species — made into a system: values declared, plural, enforced at every action, and provable.

Gawdat is right that our values decide the outcome. We just hold that values govern nothing until they're built in.

Mo's talk: https://t.co/qUEOmSCaqg

Our full response: https://t.co/Ez2sB1x4au

[Narendra Gore @nrgore1]

The next trillion-dollar infrastructure layer may be verification.

AI needs proof of origin, authorization and agent behavior.

Stablecoins need proof of reserves and transaction lineage.

Supply chains need proof of provenance.

Ceasefires and trade deals need proof of compliance.

The next moat is not making more claims.

It is making consequential claims continuously auditable.

[Narendra Gore @nrgore1]

AI has an accounting problem.

Companies can measure tokens, GPUs, capex and headcount cuts.

They still struggle to prove:

— what an agent actually did
— which permission enabled it
— who remained accountable
— what risk it created
— what financial value resulted

The next AI moat may be a double-entry ledger for autonomous work.

[Narendra Gore @nrgore1]

Frontier AI has a strange business model:

Labs spend billions creating advanced reasoning.

APIs expose parts of that reasoning.
Rivals distill it into cheaper local models.

The AI capex leaders may be subsidizing their own commoditization.

Durable moat = proprietary workflow data, distribution, output control and refresh speed.

[Narendra Gore @nrgore1]

Great introduction. Your focus on long-term trust over hype aligns with what we’re building at LiqMint—governance-first AI for institutional stablecoin treasury and payments. We think the next winners won’t just build AI—they’ll make it deployable in regulated environments. Happy to connect.

[Narendra Gore @nrgore1]

The bull case for the AI supercycle is confident and, in the room where it's made, unanimous. Here's the same story checked against the public numbers — not doom, just scrutiny. https://t.co/fehjT1j5Ur