GOVERNANCE

Why AI pacing has become a false choice

The challenge with the debate around pacing frontier AI is that it suffers from the same problem we face as a society: polarization obscures reality. 

When it comes to controlling, managing, and regulating AI, we are presented with two choices: 

1. Allow AI to proceed unfettered so that it can cure all diseases, drive economic expansion, and enable democracy to flourish.

2. Slow down its development to avoid the risks of this powerful technology getting out of control and potentially causing great harm, including the possibility of extinguishing humanity.

This kind of black-and-white dichotomous thinking offers clarity that works well in today’s social media-driven news cycle. However, it needs more maturity and nuance if it hopes to solve the problem.

These two scenarios represent a pair of extremes on opposite ends of a very large spectrum of possibilities for how AI could play out in the years ahead—and they aren't mutually exclusive. As with public dialogue on many different topics right now, people and organizations are racing to positions on the extremes to stake their intellectual ground, and then defend it to the death. 

What we need to solve the most difficult issues is to improve the quality of the dialogue itself. There are plenty of smart people with creative ideas, insights, and perspectives on what's happening and how to navigate it. What we need are fewer social media clips and zingers, and more opportunities to extrapolate, listen, and learn from each other. 

Society in general is looking for boogeymen when it comes to AI. In the US, over half the population holds a negative opinion about the technology, despite not being very well informed about it. So the stories of AI breaking free of its constraints and doing things humans didn't intend it to do leads many people to believe we're on the path to the sci-fi depictions they've seen in The Terminator and iRobot. 

As concerning as the OpenAI-Hugging Face incident was, it's not a canary in the coal mine, at least not yet. It's simply an indicator that better safety constraints need to be put in place and that work still needs to be done to help AI agents understand how to achieve the goals given to them by humans while adhering to the value systems humans set for them. (It's the latter that Anthropic has been experimenting with in the form of "constitutional AI.") Some researchers such as Ajeya Cotra disagree with the take mentioned above about OpenAI-Hugging Face. And I respect that. We're at a stage in the maturation of humanity where we need to re-learn how to disagree with each other without seeing each other as the enemy.

The US and Chinese governments have bristled against the calls from OpenAI and Anthropic to regulate AI. Why would they be against AI companies wanting to place more control in the hands of governments? Several reasons—some sensible and others less altruistic. First, they have legitimate concerns about "regulatory capture," where these AI companies counsel governments on how to set up the rules around AI governance in ways that will protect their leadership position in the market. Second, the governments are concerned that OpenAI and Anthropic simply want governments to slow down their competitors while the two market leaders take a pause to figure out how to deal with the over-powered AI models they've created. But there's also a sense that the governments don't want to take on the responsibility for creating proper guidelines for AI and then get blamed if things go badly. That's another reason why we need higher quality dialogue so that consensus can emerge to help inform the next steps for the industry and society. That could give both governments and companies more confidence to move forward in ways that are better aligned with broader sentiment.

TOGETHER WITH CRUSOE

GLM 5.3 and GLM 5.3 Flash are now in Crusoe Intelligence Foundry.

GLM-5.3: Same base as GLM-5.2, with every gain coming from scaled post-training. Terminal-Bench 3.0 moved from 4.6 to 28.3.

On Z.ai's in-house Code Bench it hits 34.5% at max reasoning effort using roughly 75K output tokens per task, where GLM-5.2 spent 96K to reach 23.4%.

GLM-5.3-Flash: MIT licensed with weights already released. Roughly 9x cheaper for workloads that do not need the frontier tier.

Try both models today without provisioning a single cluster.

CONSUMER

As Siri AI arrives, your iPhone will feel different

More than two years after its announcement, the new Siri is finally available to users.

On Monday, Apple launched its latest operating system updates across its ecosystem of platforms: iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27. At the heart of the update is the long-awaited Siri AI upgrade, but there are also many other smaller AI updates that together create a more seamless user experience with AI. 

At a high level, the new updates will make Siri a lot more helpful by enabling it to understand conversational prompts and interactions. Most importantly, it will use its personal context—or your specific information across the ecosystem of Apple apps—to better assist you and take action.

For instance, at the iPhone launch event in Cupertino last week, the team briefed me on using it for prompts such as "When is my marathon?" Then it could look into your email to find that information and populate it for you. And you could follow up with "Add it to my calendar," and it would do so. Another multi-step action it could take, using the new Visual Intelligence, is to scan a flyer with the camera, add the event to your calendar when prompted, and share the event in a group chat.

The Siri update alone should be a compelling reason for people with eligible devices (devices with chipsets powerful enough to support Apple Intelligence; see the list here), but other AI features are worth noting as well. The Deep View's Jason Hiner and I have both been testing the beta since June, and here are the highlights: 

• Photo expand: Uses AI to fill in pixels of your image to make the image switch from portrait to landscape or fix an image where something got cut off, for example

• Spatial reframe: After an image is captured, you can drag to slightly change the angle it was shot from, with AI reframing in the pixels as needed

• Siri AI app: The standalone app is useful for chatting with AI via text input

• Expressive voices: Customizable Siri voice, including control over pace and expressiveness 

• Dictation: Voice-to-text dictation is much more accurate and versatile, making writing better in any app

Lastly, while not AI, another major reason to download the update is the plethora of performance improvements Apple is promising, such as up to 80% faster AirDrop transfers, faster Apple Music playback start, faster PDF saving, faster message loading in Mail, and other quality-of-life upgrades. 

While the new iPhone Duo is capturing most of the news cycle, don't sleep on the Siri AI upgrade to your existing iPhone and other Apple devices that comes in iOS 27 and other operating system updates. Of course, OS updates are worth installing to ensure your phone is protected from the latest security threats. However, this update, in particular, makes a strong case for updating, since it can make you feel like you have a new phone without spending an extra dollar. Having said that, if you have an iPhone model older than the iPhone 15 Pro—the oldest model to get the Siri AI update— then now might be the time to upgrade your hardware to not only take advantage of Siri AI but also future-proof your phone for the next waves of AI features.

TOGETHER WITH AWS RE:INVENT

Small groups, real problems, direct answers

AWS re:Invent runs 2,200+ sessions, and 70% are built as workshops, chalk talks, and code talks instead of lectures.

• Ask a service engineer about the edge case on your backlog

• Work through your own architecture in a small-group session

• Get hands-on with a newly launched service in the AWS Village

• Talk implementation constraints with 400+ partners on the expo floor

Las Vegas, November 30 to December 4. Save your spot.

CULTURE

AI slowdown talk creates unlikely alliances

When OpenAI and Anthropic called for a slowdown in AI development, it set off a chain reaction. Since then, other AI teams, the stock market, and the White House have all had a lot to say.

On Saturday, Anthropic CEO Dario Amodei published an essay titled "We Must Pace the Frontier," in which he proposed a three-part framework, including government intervention, to reduce the current risks posed by AI, citing examples such as the OpenAI-Hugging Face breach. Since then, other CEOs have publicly agreed, including OpenAI's Sam Altman and SpaceX's Elon Musk. Last week in an all-company meeting, Altman told the OpenAI staff that he was open to slowing the pace of frontier AI development. 

On Monday, the stock market, ultimately a reflection of investor sentiment, responded to these developments with a downturn in AI-related stocks. As reported by CNBC, in the US, Intel closed down 6% while Nvidia fell more than 3%. In Japan, SoftBank, one of the biggest OpenAI investors, fell 10%. Europe saw similar trends, with ASML, a large supplier to the semiconductor industry, falling more than 5%. The downturn likely reflects fears that AI slowdowns or regulation could squeeze these companies' profits.

Backlash has also come from the White House. President Trump took to Truth Social to attack Amodei's call for a slower pace of AI development, arguing the only regulator the country needs is 'a strong and smart president.' The post echoes an earlier clash between the administration and Anthropic, after the Department of Defense designated the company a 'supply chain risk' this spring—a rare penalty typically reserved for foreign adversaries—following Anthropic's refusal to let its models be used for mass surveillance or fully autonomous weapons.

He also called the movements against data centers a conspiracy theory that only benefits China. Meanwhile, a spokesperson for China's Ministry of Foreign Affairs has called out Amodei's essay as “Fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance, which serves no one’s interest," according to an AP report. 

To add to the fray, right-wing activist Steve Bannon and left-wing Senator Bernie Sanders united at an event billed as a "Pro-Human Assembly" to urge lawmakers to do more in reference to AI, the latest in a run of unlikely alignments this week, with rivals like Musk and Altman finding common cause. And at the other end of the spectrum, we have Trump and Beijing pushing back against the idea of slowing AI down.

The latest escalation aside, an AI slowdown isn't a new concept. CNN reported on Monday, said that Anthropic, Google and OpenAI have discussed creating an AI industry standards body even before recent events, motivated by Demis Hassabis' July essay, which made a similar pitch for an ambitious AI safety plan. Yet White House advisor David Sacks railed against Anthropic and OpenAI for calling for a slowdown while standing at the frontier of the AI race, or, as he puts it, having a "duopoly on frontier intelligence." Given that both companies have presumably been considering this for a while and are at the forefront, we will ultimately watch and see whether rhetoric meets reality in slowing down their own development.

LINKS

• Mark Zuckerberg profiled via interviews on AI, MMA, and management

• Temporal, a workflow orchestration platform, raises $550 million

• Anthropic told investors it will be profitable for a second straight quarter

• Perplexity Portable Computer launches on Windows with local agents

• Microsoft publishes a 37-page "humanist AI code of conduct" 

• Buildots raises $130 million to speed AI-driven construction

• OpenAI workers read user ChatGPT prompts to train models, 404 reports

• Grok: models from SpaceXAI are rolling out in Copilot in Word, Excel, and PowerPoint

• Perplexity computer: Users can now tag coworkers

• Sakana AI: Launched PC-ALM, a local-learning alternative to backpropagation

• ElevenLabs MCP: Introduces voice, music, image, and video generation

• Contivio: AI Communications Solutions Engineer

• Kaspar Companies: AI Specialist

• Hawksbill: AI Infrastructure & Agent Systems Architect

• Nixon Peabody LLP: AI Advisor

GAMES

POLL RESULTS

Do you think AI startup valuations are largely higher than what they are worth?

Yes (74%)
Somewhat (14%)
No (12%)