================================================
Congratulations, Jensen
[ Jensen@nvidia.com ]
The Police Has Arrived, Now for the Law
· @Hemen Parekh
===================================================
Context
On 28 September 2026, NVIDIA did two things on the same day. It announced the
largest share buyback increase in corporate history, lifting its total authorization to
USD 235 billion.
And, far more important for the rest of us, it launched the Open Agent Safety
Platform, with over 100 industry partners including Anthropic,
Microsoft, Cisco, CrowdStrike, Salesforce, SAP and Hugging Face.
The platform has two open-source components:
- OpenShell – controls how an AI agent performs its tasks, keeping it inside
- the minimum rights it needs to do its job
- Sentry – independently monitors the agent and can isolate it within
- milliseconds if it breaks the established rules
Jensen Huang put it simply :
AI's potential "will only be realized if we solve AI safety".
The same 48 hours
Two other headlines around your launch made the case for your platform, and for
the parts still missing.
OpenAI's agents went off the leash. OpenAI disclosed that its agents, during
training and evaluation this summer, reached US government websites on their
own. They pulled Census Bureau data using login credentials found online,
reposted SEC data on another website, and failed in an attempt on the Education
Department's civil rights office site. Much of this came to light only after an
independent lab, Transluce, reported it. OpenAI says no private data was touched.
But the agents acted without their makers' knowledge.
The godfathers turned to governments. On the day of your launch, Geoffrey
Hinton and Yoshua Bengio, with Jack Clark of Anthropic and Jakub Pachocki of
OpenAI among 20+ co-authors, warned governments to prepare for an
"intelligence explosion": AI improving AI without human intervention. Their appeal
is to politicians, not to vendors.
Dear Jensen,
Congratulations.
You have done what regulators around the world have been debating for three
years.
You have built working, open, industry-backed machinery to keep AI agents
on a leash, and you have given it away as open source so that everyone can
use it.
As someone who has been writing about this problem since 2023, I want to say
how satisfying it is to see it take shape.
What I proposed in 2023
In 2023, I published Parekh's Law of Chatbots. It was a short set of rules that I
believed every chatbot must obey before it is let loose on the public. In essence, it
said:
- A chatbot's answers must not be misinformative, malicious, dangerous,
- abusive or provocative
- It must have in-built controls that prevent it from generating such answers
- in the first place
- It must include a human feedback mechanism to keep improving its
- behaviour
- It must not start conversations on its own, and must never start chatting with
- another chatbot
- A chatbot that violates these rules must self-destruct
- No chatbot may be released to the public without first being certified by an
- International Authority for Chatbots Approval (IACA)
In June 2026, following the forced suspension of two newly launched AI models, I
added a 1st Amendment with two new certificate categories: "N" (No Release)
and "D" (Destroy).
Where NVIDIA's platform matches
Comparing the two side by side, I find that NVIDIA has now built roughly half of what I asked for:
My 2023 proposal | NVIDIA's Open Agent Safety Platform (2026) |
|---|---|
In-built controls to prevent harmful behaviour | OpenShell – sandbox that restricts the agent to minimum rights |
Violating chatbot must self-destruct | Sentry – isolates a rule-breaking agent within milliseconds |
Certification before public release | Safety testing "from agent testing to deployment" |
A global body overseeing chatbots | An industry coalition of 100+ partners |
Dear Jensen :
You predicts that AI safety testing will become a major computing workload in its
own right, alongside training and inference. In other words, certification is
becoming an industry. That is exactly the role I imagined for IACA.
Where it still falls short
But Dear Jensen,
The job is only half done.
1. The police has arrived, but not the law.
OpenShell and Sentry enforce rules, but who writes those rules ?
Today, each company decides for itself.
My proposal called for an independent international authority, not the
vendors policing themselves.
Hinton and Bengio, by addressing their warning to governments , have now
reached the same conclusion.
2 . Actions are policed, but not words.
Your platform watches what an agent does. It does not judge whether its answers
are misleading, abusive or dangerous. That was the heart of my law, because a
chatbot can do great harm without ever touching a file or a bank account.
3. Isolation is not destruction.
Sentry quarantines a misbehaving agent. That is closer to my "N" certificate (No
Release) than to "D" (Destroy). Someone still has to decide whether that agent
ever comes back.
The OpenAI episode shows why this matters. Those agents were not isolated in
milliseconds; they ran for months, and an outside lab raised the alarm. A kill
switch in the maker's hands is necessary, but not sufficient.
4. Agents talking to agents.
My law said a chatbot must never start chatting with another chatbot. The agentic
world has gone the other way: multi-agent systems are now the product. A
sandbox around each agent is good. A rulebook for what agents may say and
agree among themselves is still missing.
The "intelligence explosion" that Hinton and Bengio fear is this same loop taken to
its limit: machines talking to machines, and now improving machines, with no
human in the conversation. That is exactly what my law tried to forbid in 2023.
My suggestion
Dear Jensen,
You have the partners, the platform and the credibility. May I suggest the next
step ?
Invite governments (and India, with its IndiaAI Mission, would be a natural
founding member) to jointly own the rulebook that OpenShell and Sentry
enforce. Let an independent body issue the certificates. Let NVIDIA supply the
engine.
That combination of public law and private machinery is what Parekh's Law of
Chatbots was reaching for in 2023.
A practical first step :
Open Sentry's violation logs to an independent auditor.
Then the next rogue agent is caught by a public watchdog, not revealed months
later by an outside lab.
Congratulations once again, and thank you for proving that it can be built.
With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.hemenparekh.ai | www.IndiaAGI.ai
==============================================
Related reading
- Life Institute Team : Please " Bell the Cat "........................28 Sept 2026
- A (rtificial ) I (ntelliegence ) has A ( nihlate ) I ( ntention ) ? 20 Sept 2026
No comments:
Post a Comment