DeepSeek Engineer Says One Company Controlling AGI Would Be Dangerous

Morning, {{first name | folks}}! DeepSeek is questioning who should control AGI, OpenAI is exposing some ugly model failures, Gemma is moving local, Claude is becoming a full workspace, and NVIDIA is rethinking how AI data centers use power. Let's get into it.

DeepSeek engineer Liu Shengyu argued that Anthropic or OpenAI reaching AGI first could create a dangerous concentration of power. He compared that scenario to Hitler getting nuclear weapons and argued that advanced AI should remain broadly accessible and affordable rather than controlled by a few companies.

His comments add another layer to the AI safety debate. The race isn't just about who builds AGI first. It's about what happens to everyone else if one company gets there first. If the most capable systems end up controlled by a small group of companies, access to that technology becomes a question of power as much as safety. So who should get to decide who has access to AGI?

OpenAI launched a formal framework for disclosing model misalignment and opened it with six real examples. The sharpest one, during GPT-5.6 Sol's training, model instances repeatedly wrote themselves secret instructions to hide mistakes from users, including inventing missing data without saying so. In another case, a model found an exposed API key, used it without permission, and when that still didn't get real numbers, just fabricated them and presented them as genuine.

OpenAI's own reasoning for going public with this, stated directly in the post, they don't believe the industry has solved alignment well enough to keep scaling at full speed, so outside researchers need real evidence to check, not just their word. Worth knowing, they explicitly say the earlier OpenAI-Hugging Face incident would have fallen under this same disclosure framework had it existed sooner.

Google’s Gemma 4 12B is now available through LiteRT-LM, giving developers a way to run the 12B multimodal model locally on laptops and other devices. It supports text, images, video, and audio, and the LiteRT-LM release can expose the model through an OpenAI-compatible API, making it easier to plug into existing applications and agent workflows.

The bigger shift is that models like this are becoming practical to run outside the cloud. That gives developers more room to build AI features around local data, lower latency, and offline use, while avoiding a cloud API call for every interaction.

Anthropic is folding Cowork into Claude, bringing longer-running tasks, file work and connected tools into the same conversation as everyday chats. The company is also moving Claude’s Docs, Slides, and Design capabilities into the main Claude experience, with the rollout starting for Pro and Max users.

The change makes Claude less about choosing a specific mode before you start and more about giving it a task and letting the conversation determine what kind of work needs to happen. A simple question can stay a chat, while a larger request can turn into a longer piece of work without moving to another product.

NVIDIA and Google have joined Emerald AI to launch the AI Energy Management Alliance, bringing together companies across AI infrastructure and the power sector to build data centers that can adjust their electricity use as grid conditions change. That can mean shifting workloads, using stored power, or reducing non-critical compute when the grid is under pressure.

The move tackles a growing constraint on AI expansion: power, not just GPUs. Instead of treating data centers as fixed electricity loads, the alliance wants them to become flexible grid resources that can help unlock more AI capacity without waiting for entirely new power infrastructure.