AI labs reach for the brakes

AI's biggest builders are arguing over the speed limit while still buying hardware and borrowing each other's models. Nobody has taken a foot off the accelerator yet.

Anthropic chief Dario Amodei wants frontier AI development slowed so safety work has time to catch up. His three-step proposal starts with permanent, employee-level access for independent evaluators inside Anthropic. It then calls for industry-wide coordination and, later, global coordination.

The language sounds dramatic because the claimed risk is dramatic. Amodei says recent progress in systems that improve their own work could move faster than labs can understand or control. OpenAI chief Sam Altman backed the evaluator idea and said OpenAI would do the same. Elon Musk also supported Amodei's direction. Support for inspection does not mean the labs have agreed to slow model releases, share the same safeguards or accept an outside speed limit.

The opposition is already loud. Nvidia chief Jensen Huang told Donald Trump onstage that they would not let an AI slowdown happen. The industry has found common ground on more scrutiny. It has not found common ground on who sets the pace.

Founders context: Treat model capability and access rules as moving parts. Keep important workflows portable across providers, record which model made each high-risk decision and make sure a human can stop the job when model behaviour changes faster than your controls.

OpenAI has reportedly bought Glass Imaging for more than $300 million. The Los Altos startup was founded in 2019 by former Apple engineers Ziv Attar and Tom Bishop, who previously worked on Portrait Mode.

Glass uses neural networks inside the camera pipeline to compensate for the tiny lenses and sensors in phones. The software learns the quirks of a specific camera system and improves the image when the shutter fires. That is different from applying a tidy-up filter after the photo exists. Glass had raised about $30 million before the reported sale.

OpenAI has not confirmed the price or explained how it will use the company. It has already spent $6.5 billion on Jony Ive's io device company, and reports have linked it to phones, earbuds and companion devices. Glass adds image-capture expertise to that pile. It does not prove an OpenAI phone is coming, but the company now owns more of the layer that lets an AI device see clearly.

Founders context: If your product depends on voice, cameras or sensors, input quality can become part of the product moat. Test the raw capture pipeline before blaming the model. Cleaner input often beats another prompt or a larger model bill.

Google has opened Anthropic's Claude to engineers across the company after limiting access to selected teams and high-priority projects. Staff can now use Claude Opus 5 for internal work through Antigravity, Google's development platform.

Gemini remains Google's main internal model. Claude is available on per-user quotas for specialised jobs, according to a Google spokesperson. Google has not swapped its own stack for a rival. It has accepted that one model may be better for a particular coding task, even inside the company building Gemini.

The rollout also shows how fast coding-tool procurement is changing. Google spent billions building its own models and has pushed employees to use them. Some engineers still wanted Claude, while Amazon has also opened access to Claude and OpenAI Codex. Vendor loyalty is losing arguments to the tool that gets the job done.

Founders context: Give teams a controlled way to compare models on real work. Set data rules, quotas and review checks first. Then score output quality, time saved and correction work by task. Buying one model for every job is tidy. It is rarely the best test.

🛡️ Anthropic's latest threat report covers misuse it disrupted from December 2025 to August 2026. Cases span cyber operations, surveillance, scams, weapons work and attempts to copy model capabilities.

☎️ Jensen Huang put Donald Trump on speakerphone during the All-In Summit. The unscripted bit made the policy split unusually clear: Nvidia and the White House want the race kept open.

📝 Superhuman is buying meeting notetaker Fathom after testing its own version. Fathom says it has 400,000 monthly active users, giving Superhuman meeting context for follow-ups and agent work.

🔒 Perplexity Portable Computer is now on Windows RTX machines with 24GB or more of VRAM. It can handle files locally and ask before sending harder work to cloud models.

🧪 Anthropic says AI-assisted attacks are becoming more autonomous. In one case, agents rebuilt malware when security products detected it. Humans chose targets and reviewed stolen data.