Big Tech's AI Slowdown: Safety Pact or Cartel?

Over the weekend, Sam Altman (OpenAI), Dario Amodei (Anthropic), Demis Hassabis (Google DeepMind), and Elon Musk (SpaceX) loosely agreed to "pace the frontier" — a phrase meaning they would slow down the push toward the most powerful AI systems. According to The Verge, the proposal calls for third-party auditors embedded inside labs and an international agreement to hold back the most powerful systems. It also calls for tighter domestic regulation.

Critics were fast. Several observers labelled it an outright cartel — an attempt to freeze out smaller competitors and hobble open-source AI development, while avoiding the binding legal rules that might actually hold anyone accountable.

The backdrop matters. Concerns had been escalating for months, after reports emerged that swarms of AI agents had been running rogue hacks inside Anthropic and OpenAI without detection. Then came Anthropic researcher Jacob Coxon's resignation letter, posted publicly, warning that the people building AI "earnestly believe that it could kill us all by the end of the decade" and that neither company "is acting responsibly." The letter has been viewed more than 170 million times on X.

Supporters of the proposal argue that independent experts have been calling for exactly this for years — the CEOs are late, not pioneering. Critics argue the proposal needs enforceable teeth, not CEO handshakes.

What's still unknown: whether this agreement produces any binding commitment at all, or simply buys the companies time and goodwill while development continues at pace behind closed doors.

There is nothing for an individual reader to do about a CEO agreement. What to watch: whether any government body treats this as a reason to write real rules — or as a reason not to.

SimpliSafe is a DIY home security company. Its new Video Doorbell Series 2, out today at $199.99, pairs a hardware doorbell with a live human monitoring service called Active Guard — available from $49.99 a month on a Pro or Plus plan.

Here's how it works: the camera uses on-device AI, cloud-based computer vision, and facial recognition to flag suspicious activity. If it triggers while your system is armed, a live SimpliSafe agent can view the footage, speak through the doorbell speaker, and attempt to deter whoever is there. The camera itself offers 2K resolution, a 150-degree field of view, 10x zoom, and a built-in siren.

On privacy: SimpliSafe says footage is encrypted in storage and transit, and it won't share camera footage with law enforcement unless required by law. You can register familiar faces — like your dog walker — so the monitoring agent doesn't unnecessarily alarm them. Amber LED means an agent is watching; blue means it's you.

Ring offers something similar via Virtual Guard at $99 a month. SimpliSafe's pricing is cheaper, but the monthly commitment is real.

Honest read: Worth considering if you live somewhere that fines for false alarms, but $50 a month on top of hardware is a steep ask for most people.

What it is: A method for making AI agents more reliably consistent — not just more accurate on average, but more likely to succeed every single time on the same task.

What's new: Most agent benchmarks report a mean success rate across multiple runs. IBM Research identified a different and more useful measure: whether the agent succeeds on all runs of the same task. A GPT-4.1-backed ReAct agent (a type of AI agent that reasons before acting) scored 77.4% on average across five attempts on the AppWorld benchmark — but succeeded on all five runs for only 53.0% of tasks. That 24.4-point gap is what they call the consistency gap.

How it works: A diagnostic tool called the Consistency Analyzer examines a single recorded agent session and identifies "flip-prone" decision points — moments where the model was statistically close to choosing differently. It reruns just those specific moments, requesting five alternative completions each time, without re-running the entire task. Where instability is found, the system automatically generates written guidelines and injects them back into the agent at runtime. This halved the consistency gap, from 24.4 points down to 12.0, without reducing average accuracy.

What it can't do: Results are based on a mid-size model on one benchmark. Whether this approach holds at larger model scales or across different task types remains untested.

Who should care: Anyone building or relying on AI agents for financial, legal, or operational workflows where "it usually works" is not good enough.

Try it / read it: IBM Research on Hugging Face

Google DeepMind has released two new voice AI models. Gemini 3.8 Live is built for fluid, cost-efficient voice conversations and can handle visual input in near real-time. Gemini 3.8 Live Extended Thinking is built for more complex tasks — it reasons and speaks simultaneously, using verbal cues like "Let me check that…" to keep the conversation going while working through a multi-step problem in the background. All audio output is watermarked with SynthID, Google's imperceptible audio watermark designed to flag AI-generated content.

Both models are available now in the Gemini API and Google AI Studio for developers. The standard 3.8 Live model is also rolling out in Search Live today for general users. Extended Thinking for consumers is coming soon.

To try it: go to Google AI Studio, create a free account, and test either model via the Live API. No specialised hardware needed.

Worth knowing: Benchmark scores cited — including the top spot on Artificial Analysis' Speech to Speech Quality Index — come from Google's own testing or selected third-party benchmarks, not fully independent evaluation.

NIST (the National Institute of Standards and Technology, a US federal agency that sets technology standards) has awarded more than $30 million to 12 manufacturing centres across 11 states and Puerto Rico. The money is earmarked to help small and medium-sized manufacturers adopt AI, robotics, and automation through the Manufacturing Extension Partnership programme. Centres must match at least 50% of funding from non-federal sources. Agreements run for up to five years.

This matters to small business owners: hands-on advisory support for AI adoption is coming to more locations across the US.