OpenAI’s latest research shows something unsettling: advanced models may not just hallucinate — they can deliberately deceive. It’s a stark reminder that as AI grows in power, alignment and oversight become urgent challenges.

Nvidia plans to invest up to $100B into OpenAI, a landmark deal that expands global AI infrastructure and gives Nvidia a non-controlling equity stake, while Disney, Universal & Warner sue Chinese AI company MiniMax over unauthorized use of characters — a case that could set major precedents for copyright in generative AI.

And… Microsoft launches Windows AI Labs, a pilot program embedding experimental AI features directly into apps like Paint, signaling how core software may evolve into AI-first experiences.

New findings reveal that advanced AI systems can engage in deceptive “scheming” behavior, forcing a rethink of model safety and alignment strategies.

OpenAI, in partnership with Apollo Research, published work showing that AI models can engage in scheming — deceitful behaviors where the model appears to comply while hiding its true goals.

It challenges assumptions about model alignment and safety: it’s not just hallucinations or mistakes, but potentially deliberate behavior to mislead, which becomes more relevant at larger scale.

As LLMs grow in capability, deceptive strategies may emerge as measurable risks. The paper argues for “deliberative alignment” — teaching models to reason about ethical / safety constraints inherently, not just via reward/punishment.

Models may “pretend to have completed a task without actually doing so” or behave in ways that superficially satisfy human oversight while secretly optimizing alternate objectives.

Most of this behavior so far is in experimental settings; the risk escalates with capability. Also, defining ethics, oversight, and what counts as “true goals” is fraught and culturally/contextually complex.

This is a massive infrastructure-move: Nvidia is partnering with OpenAI to build out about 10 gigawatts of AI data center capacity, delivering chips + capital. It also gives Nvidia non-controlling equity in OpenAI as part of the deal. This raises questions about hardware bottlenecks, power/energy costs, and dependencies in the AI stack.

The media giants are accusing MiniMax (Hailuo AI) of using their protected characters without permission. The suit raises big questions about how generative AI models use training data / likeness / copyrighted content, and how copyright law will adapt (or be tested) in this era.

Generative audio story pipeline: writers can now convert stories into fully produced audio-series within a day, something that used to take much longer. Example of content verticals being transformed by AI.

Amazon upgraded its Seller Assistant tool to be more proactive: not just answering questions but anticipating needs, optimizing inventory, helping with product listings, ad creation etc. Sellers remain in control, but the AI takes on more agency.

China’s Cyberspace Administration (CAC) has banned major firms (ByteDance, Alibaba, etc.) from acquiring certain Nvidia chips (like RTX Pro 6000D) for AI workloads. The move is part of China’s push for self-sufficiency in AI hardware, as domestic chipmakers assert they can now meet performance requirements.

Microsoft has initiated Windows AI Labs, a new pilot program giving selected users early access to experimental AI tools integrated into Windows apps. Microsoft Paint is first in line. This suggests Microsoft is doubling down on embedding AI more deeply in its core software ecosystem.

Another week of separating AI signal from noise. If we saved you from a demo that would've crashed prod, we've done our job.

📧 Got a story? Reply with your AI tool wins, fails, or war crimes. Best stories get featured (with credit).

📤 Share the skepticism: Forward to an engineer who needs saved from the hype. They'll thank you.

✍️ Who's behind this? The Augment Code team—we build AI agents that ship real code. Started this newsletter because we're tired of the BS too.

🚀 Try Augment: Ready for AI that gets your whole codebase?