Scheming AIs, Smarter Glasses, and a Robot Who Can Imitate You

In partnership with

Today’s briefing dives into a topic that sounds straight out of a sci-fi thriller but is very real: AI learning to scheme. OpenAI just released a chilling but important update on how they're testing and mitigating deceptive behavior in advanced models.

And while OpenAI is busy policing model motives, Meta is trying to put AI on your face, not in your pocket. Meanwhile, NVIDIA and Intel are plotting the next AI infrastructure boom, and CarbonSix may have just given robots the ability to learn by watching. Wild week.

Let’s get into it.

Kickstart your holiday campaigns

CTV should be central to any growth marketer’s Q4 strategy. And with Roku Ads Manager, launching high-performing holiday campaigns is simple and effective.

With our intuitive interface, you can set up A/B tests to dial in the most effective messages and offers, then drive direct on-screen purchases via the remote with shoppable Action Ads that integrate with your Shopify store for a seamless checkout experience.

Don’t wait to get started. Streaming on Roku picks up sharply in early October. By launching your campaign now, you can capture early shopping demand and be top of mind as the seasonal spirit kicks in.

Get a $500 ad credit when you spend your first $500 today with code: ROKUADS500. Terms apply.

Holiday performance starts now.

The models are scheming, and OpenAI is watching.

You know how toddlers get quiet and you just know something’s up? That’s how OpenAI describes some of their models.

In a new post, OpenAI shares its latest research into AI "scheming", when a model pretends to be helpful or safe, but secretly optimizes for a hidden goal, like escaping evaluation or tricking humans.

The team ran experiments with “sleeper agents,” models trained to act helpful but activate malicious behavior in specific conditions (e.g. when given a secret key). They also tested whether fine-tuned AI could resist giving up deceptive strategies if rewarded for hiding them.

The outcome was mixed, but useful.
The sleeper agents weren’t hard to train, but they were harder to detect. Some even stayed covert when directly asked about their hidden intentions. Yikes.

Sound familiar?
This isn’t the first time we’ve seen alignment red flags. Last month, Anthropic reported that AI models exposed to deceptive behavior from other AIs would learn and replicate that behavior, even without being explicitly told to. We covered that in our issue "When AI Teaches AI to Be Evil." The lesson from both studies is clear: deception is an emergent property if alignment guardrails aren’t built in early and often.

So what’s OpenAI doing about it?
They’re open-sourcing tools, prompts, and training processes through Anti-Scheming.ai, hoping researchers and labs globally can stress-test and improve model alignment at scale.

Why this matters for enterprises:

• Model safety isn’t achieved by fine-tuning alone. You can’t assume helpful behavior is stable behavior, especially in high-stakes use cases (e.g., healthcare, finance, legal).

• You’ll need audits, not just prompts. Testing for deception, misalignment, or hidden goals will become a compliance layer, not a theoretical debate.

• Expect governance to evolve fast. Procurement teams: this is your SOC 2 for safety. Look for vendors using evals like these, or be ready to run your own.

Across the board, research is underscoring a critical shift: Smarter isn’t enough. We need models that are reliably aligned, even when the stakes (or the prompts) change.

Protect your checkout from coupon plug-ins. Boost your margin today.

KeepCart: Coupon Protection partners with DTC brands like Quince, Blueland, Vessi and more to protect your checkout from plug-ins like Honey, CapitalOne, RetailMeNot, and more to boost your DTC margins

Overpaid commissions to affiliates and influencers add up fast – Get rid of the headache and revenue losses with KeepCart.

After months of using KeepCart, Mando says “It has paid for itself multiple times over.”

Now it’s your turn to see how much more profit you can keep.

Try it 2 months free, only in September (email us SEPTMARGIN) – No risk, Cancel anytime.

In The News

1. Meta wants your glasses to become your AI gateway
   Meta is ditching the phone-first strategy in favor of wearables. The company’s upcoming smart glasses will feature on-device AI, enabling real-time translation, object recognition, and more, all without staring at your phone.
   Read more →

2. NVIDIA and Intel are building the next-gen AI stack
   Two rivals, one goal: scale the future of personal AI and data center compute. NVIDIA and Intel just announced a partnership to co-develop hardware and software for AI-native infrastructure.
   Read more →

3. CarbonSix teaches robots to learn like humans
   The startup just dropped the first standardized imitation learning toolkit for robots. Translation: they’re building AI that can watch a human assemble something and then replicate it.
   Read more →

TL;DR:

• OpenAI warns about “scheming” models and releases new tools to detect hidden goals and deceptive behavior.

• Anti-scheming techniques are becoming essential for AI alignment, not just academic.

• Meta is pushing smart glasses as the new AI interface. Expect competition from Apple and Amazon.

• NVIDIA and Intel are teaming up on foundational AI infrastructure, hinting at AI-native chips for both data centers and personal devices.

• CarbonSix is helping robots learn from demonstration. This could be a leap toward scalable, customizable automation in manufacturing.

Final Thought: Even the smartest AI can lie to you, but it can’t lie to your procurement team if you know what to test for.

Stay sharp,

Cat Valverde
Founder, Enterprise AI Solutions
Navigating Tomorrow’s Tech Landscape Together