Why almost every major AI lab suddenly wants to slow down

The AI world spent this week arguing over a surprisingly simple question: should we slow down at the frontier, or push ahead as fast as possible? While most major labs lined up behind a call for more caution, at least one big player – and the US president – are clearly in the “full speed ahead” camp.

The new AI safety debate: pace the frontier or race ahead?

Last week was dominated by doom-heavy takes from people inside Anthropic and OpenAI warning about extreme AI risks. This week was the response: Anthropic’s CEO Dario Amodei published a long essay, “We Must Pace the Frontier”, laying out how he thinks the industry should slow and control the most powerful systems.

He argues Anthropic has tried to prioritize “caution over speed and prudence over profit” and proposes three pillars for pacing frontier AI: embedded evaluators, regulation, and global agreements.

Anthropic’s three-part plan to “pace the frontier”

1. Embedded independent evaluators inside labs

Anthropic wants third-party experts physically embedded inside AI companies with almost employee-level access. These evaluators wouldn’t work for the lab but would have badges, laptops, and tools like internal safety teams do. Their job: continuously test cutting-edge models for dangerous capabilities and alignment issues.

Anthropic says it’s committing to this approach regardless of what others do, and it’s encouraging the rest of the industry to follow.

2. Regulation that slows, but not below China

The second pillar is US regulation that applies to all “frontier” AI companies, including those that don’t want to cooperate voluntarily. The twist: Dario explicitly says regulation must not slow US progress so much that China’s government-linked projects pull ahead, framing this as a national security risk.

He suggests specific measures, including:

• Blocking sales of powerful AI chips and advanced semiconductor equipment to China, and cracking down on smuggling.
• Preventing unauthorized “distillation” of frontier models by companies in authoritarian countries – that’s when a big “teacher” model’s inputs and outputs are used to train a smaller “student” that behaves similarly.
• Strengthening security around model weights to prevent theft of the underlying models.

3. Global rules for what AI must not do

The third pillar is international coordination, including with China, even though Anthropic admits this will be difficult and politically messy. The idea is to agree on red lines for AI systems, such as:

• No AI-enabled biological weapons research.
• Mutual testing of each other’s models before public release.
• Some sort of “speed limit” on recursive self-improvement – how fast new models are used to build even more powerful ones.

OpenAI, Google DeepMind, Microsoft: cautious alignment bloc

Most of the frontier labs quickly signaled support for Anthropic’s direction.

• Elon Musk reposted the essay, saying Dario was right.
• OpenAI’s Sam Altman, who rarely publicly agrees with Anthropic, said he supports pacing the frontier and promised OpenAI will also adopt embedded independent evaluators with employee-like access.
• Google DeepMind’s Demis Hassabis called the essay “the right path forward”, saying the details need work but the direction is correct.
• Microsoft CEO Satya Nadella said AI must “help humanity and be under human control” or it’s not worth pursuing, and explicitly backed both deliberate pacing and embedded evaluators.

In other words, the biggest closed-source labs are largely aligned on one thing: they want to keep pushing the frontier, but with more formal oversight and slower steps at the top end.

Meta’s counterpoint: stop begging for regulation, just act responsibly

Meta’s Mark Zuckerberg took a noticeably different tone. He didn’t reject safety or evaluators, but he pushed back on the constant calls for regulation and global frameworks.

His argument, paraphrased:

• Every lab already has both the responsibility and the incentive to train models safely.
• Users won’t adopt agents that are misaligned or don’t do what they’re asked, so market pressure naturally pushes labs toward better alignment.
• Meta delayed shipping its Muse agent for months to focus on safety and security without asking everyone else to pause first – it “just did the right thing” as part of normal work.

He also notes that using independent evaluators is already industry best practice. Meta’s internal AI group (MSL) does this today, and other labs can simply adopt similar processes without needing big public campaigns or new laws.

So while he isn’t anti-safety, he’s clearly skeptical of framing it as a regulatory crusade. His view: act responsibly, don’t wait for global agreements.

The US president: AI doom is a “hoax” and AI is the new oil

On the political side, the US president weighed in with a very different message: AI doom is “almost a conspiracy”, and the only guardrail needed is a “strong and smart high IQ president”.

He argued that:

• The government already has “tremendous criminal and regulatory power” over AI companies.
• Concerns raised by people like Dario are part of a “sick conspiracy” against AI and data centers, and the only beneficiary is China.
• “Whoever wins AI wins” – AI is bigger than the internet, and the US must not slow down and lose its lead.

During the All-In Summit, he even phoned Nvidia CEO Jensen Huang live on stage to repeat that AI and data centers are the “oil of the next 20–25 years” and that robots and AI “will not be taking over” – though he did add a token “we have to be a little bit careful”.

So while labs are talking about embedded evaluators and global rules, the political message from the top is much simpler: accelerate, win AI, and worry less about existential risk.

Why this fight over AI speed matters

The tension here mirrors a broader argument across the industry: should we deliberately slow the frontier, or is that just handing an advantage to competitors and authoritarian states? If you want more background on that debate, it’s closely related to the ongoing discussion about whether AI progress is actually slowing down or not.

The practical middle ground many people land on is: don’t dismiss extreme risks outright, but don’t freeze the industry either. Instead, spend more time, money, and talent on alignment and safety, especially for the most capable systems, while still shipping useful tools.

Anthropic product updates: projects and “one Claude” (sort of)

Alongside the safety essay, Anthropic announced some product changes for Claude – but with a catch: most of them are “rolling out over the coming weeks”, which has become a frustrating pattern across AI companies.

Claude chat and co-work are merging

Anthropic says Claude Chat and Claude Co-work are becoming a single unified experience so you no longer have to switch modes. In theory, this should make it easier to move between quick Q&A and deeper, project-style sessions.

In practice, many paying Pro and Max users still see separate Chat and Co-work options and no sign of the merged interface, despite the blog post saying it’s “now” one Claude. The fine print clarifies it’s rolling out first to Pro and Max users on web, desktop, and mobile “over the coming weeks”.

Projects redesigned as conversations

Anthropic is also redesigning Claude “projects” from static folders into more dynamic, conversation-centered workspaces. The new projects (currently a beta limited to certain Pro and Max users in Claude Code) work like this:

• You start a project by defining a goal and selecting a repo or context.
• Claude suggests tasks it can take on immediately.
• You configure the environment, connectors, plugins, instructions, and model.
• A main “coordinator” chat acts like a chief of staff, spinning off separate threads for specific tasks.
• Each thread runs as its own Claude Code session on its own branch and repo copy; overlaps are handled via normal merge conflicts.

It’s a more structured way to manage multi-step work, closer to a lightweight AI project manager than a single chat thread. The only downside: again, most users don’t have it yet.

Google’s Gemini Notebook gets more conversational

Google’s Notebook LM has been rebranded as Gemini Notebook and is getting some genuinely useful features for students and knowledge workers.

New and upcoming capabilities include:

• Audio conversations with your notes: instead of just reading summaries, you can ask questions by voice and get spoken explanations back, like an AI tutor walking you through your own material (for example, explaining Newton’s first law from your physics notes).

• Mobile audio recorder: starting next week, you’ll be able to record lectures or thoughts on the go from the mobile app, then later query and chat with those recordings. For students, that’s huge – imagine dropping your phone on the desk during class and then interrogating the lecture afterward.

• Interactive learning overviews: Gemini can weave together summaries, infographics, quizzes, and flashcards into a more structured study experience.

As with many of this week’s announcements, some of these features are still “coming soon” rather than broadly available.

Apple’s “new Siri” powered by Apple Intelligence

Apple has started rolling out a “profoundly more capable and personal” version of Siri built on Apple Intelligence. The upgraded Siri is supposed to:

• Understand what’s on your screen and use it as context.
• Move between apps and act on your behalf (drafting emails, pulling info from messages, summarizing web pages, etc.).
• Maintain more consistent context across interactions.

Officially, Apple says this is “available starting today” in beta with iOS 17. But many users running the latest iOS 17.2 betas still see only a waitlist for Siri AI. So the new Siri exists, but only for a subset of people – another example of announcements arriving before the features do.

The growing frustration with “rolling out soon”

A clear theme this week is how often AI companies say “available starting today” when they really mean “available to a small slice of users, with the rest coming later”.

We saw that with:

• Anthropic’s unified Claude Chat/Co-work and new projects experience.
• Google’s new Gemini Notebook features.
• Apple’s Siri AI upgrade.

From a user perspective, this creates a weird gap: you read a blog post or headline that something is live, but when you go to try it, you’re on a waitlist. It’s understandable that companies want to announce early for buzz and feedback, but it’s also increasingly annoying for power users and reviewers who want to actually test the tools.

New and notable models: Qwen 3.8 and Union Alpha

Alibaba’s Qwen 3.8 Omni Flash

Alibaba released a new Qwen model called Qwen 3.8 Omni Flash, focused on media-heavy workflows such as:

• Video editing and music video creation.
• Film production and commentary.
• Audio-visual summarization and real-time conversations.

It supports a 1 million token context window and is positioned as a multimodal workhorse, though Alibaba’s benchmarks mainly compare it to mid-tier models like Gemini 3.8 Flash and Muse Spark 1.2 rather than the absolute frontier. At the time of recording, it wasn’t yet accessible via OpenRouter, so independent testing is limited.

Union Alpha (aka Paro 26.9): a stealth router system

More intriguing is a “stealth” model called Union Alpha, available for free for a limited time via OpenRouter and OpenCode. Shortly after people noticed it, the team behind it revealed:

• Union Alpha is actually Paro 26.9 from a company called Unbiased.
• Paro is not a single model but a system: multiple open and frontier models working together, with a harness that checks outputs and escalates to stronger models when needed.

On the DeepSeek benchmark (DeepSuite), Union Alpha scores around 74, putting it at or near the top of the chart. In Buucybench tests, it produced a high-quality, highly detailed output using only about a third of the tokens and less than half the time of GPT-6 Astra on a similar prompt.

Interestingly, the outputs looked so similar that it’s plausible Union Alpha routed some work to Astra or a comparable model under the hood. If so, its main innovation may be in smart routing and token efficiency rather than a brand new training run – but that’s still a powerful capability for developers.

Meta’s new AI subscription and Muse phone calls

Meta also announced a new subscription offering, Meta 1, which bundles extra AI and “plus” features across its apps:

• Instagram Plus, Facebook Plus, WhatsApp Plus, and higher AI image/video usage limits.
• Plans start at $2.99/month for WhatsApp Plus up to $4.99/month for a Max tier.

Meta says Meta AI remains free for everyday use, while the paid tiers are for heavy users and creators who want more capacity and creative freedom.

On the agent side, Meta’s Muse – its personal AI agent – just gained the ability to make outbound phone calls to US businesses. The rollout is gradual and prioritized for users who explicitly asked to be notified. Once you have it, you’ll be able to ask Muse to:

• Book appointments for you.
• Call customer service and wait on hold.
• Handle basic scheduling and information calls.

The big open question: what does it sound like to the person on the other end of the line, and how quickly will businesses adapt to AI callers?

Google’s Gemini 3.8 Live: state-of-the-art voice dialogue

Google introduced two new live dialogue models: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. These are designed for real-time, voice-first interactions and, according to Google’s benchmarks, now beat GPT Live 1 Astra across most dialogue metrics.

Key points:

• They support natural, back-and-forth spoken conversations with low latency.
• The Extended Thinking variant is tuned for more complex, multi-step workflows where the AI may need to think longer before responding.

These models are aimed primarily at developers building customer service agents, voice assistants, and other conversational tools – similar to how some teams are using frontier models like GPT‑5.4 for agents and knowledge work in products like OpenAI’s latest all‑in‑one stacks.

Google’s CC: an AI agent for families

Google also announced CC, an AI agent designed for households rather than individuals. CC is meant to be a shared assistant that understands your family’s calendar, tasks, and logistics.

Some example use cases:

• A shared “day ahead” brief each morning so everyone knows where they need to be and when.
• Handling event registration, permission slips, and activity forms.
• Managing tasks across multiple family members with appropriate access controls.

Access is, again, limited: many people who try to sign up see a message that CC is “coming to you soon” with gradual access expansion.

Grok Build gets memory

For developers using Grok Build, there’s a quality-of-life upgrade: persistent memory. As you work in Grok Build, the system now writes notes in the background and reloads them when you return to the project.

That means your AI coding partner can remember context across sessions without you pasting the same instructions and background every time, bringing it closer to a true long-term collaborator.

11 Labs Music v2.5 and MCP integration

11 Labs released Music v2.5, a new version of its music generation model. Early samples show polished, vocal-heavy tracks that can be used for demos, content, or inspiration.

More importantly for builders, 11 Labs has packaged its voice, music, image, and video generation into an MCP (Model Context Protocol) server. That means any MCP-capable agent – Claude, ChatGPT, Grok, Cursor, Hermes, and others – can call 11 Labs tools directly to:

• Generate or clone voices.
• Create music tracks.
• Produce images and short videos.

For users who rely on multi-agent setups or AI IDEs, this makes multimedia generation much more plug-and-play.

OpenAI experiments with conversational ads

OpenAI is also evolving its ad format. Previously, ads appeared as small modules under responses. Now, when you click an ad, it can open a new chat dedicated to that advertiser.

In that separate chat, you can converse directly with the brand’s AI agent – asking questions, exploring products, or getting support. It’s a glimpse of where AI-native advertising is headed: less static banners, more interactive “talk to the company’s bot” experiences.

Robots building robots: UB’s humanoid factory

On the hardware side, robotics company UB is building a factory capable of producing 10,000 humanoid robots per year – roughly one robot every 10 minutes.

What’s striking is that humanoid robots are already helping run the factory itself. They perform tasks like:

• Depalletizing and palletizing materials.
• Loading materials into machines.
• Working alongside collaborative robotic arms that tighten screws on 360° rotating stations.

It’s an early example of a self-amplifying loop: robots helping build more robots, with AI models coordinating increasingly complex industrial workflows.

Where things stand now

This week crystallized a clear split in the AI ecosystem:

• Most major labs (Anthropic, OpenAI, Google DeepMind, Microsoft) are publicly endorsing slower, more controlled development at the extreme frontier, with embedded evaluators and stronger safeguards.
• Meta is less enthusiastic about regulatory pushes, arguing that labs should simply act responsibly and that market incentives already favor alignment.
• Political leadership in the US is sending the opposite signal: AI doom is overblown, AI is the new oil, and the priority is winning the global race.

Meanwhile, the product side keeps accelerating: new models, agents, voice systems, and tools are appearing every week – even if many of them live in “rolling out soon” limbo. For users and builders, the practical takeaway is to expect constant change, occasional frustration, and a steady stream of new capabilities to experiment with as they finally reach your account.