POLICY

Mythos returns, but AI may be permanently changed

Mythos and Fable are back. 

The US Commerce Department officially lifted its export controls on both Anthropic models on Tuesday, clearing the way for Claude users to regain access to the frontier lab's most powerful AI. 

In a letter to Anthropic seen by Reuters, Howard Lutnick, secretary of the US Commerce Department, said, "Anthropic has agreed to proactively detect and address security risks associated with the models; to work diligently with the US government on protocols and standards and releases for ​Mythos, Fable, and ​future models; and ⁠to inform the US government of any malicious activity."

The two models have been restricted since Jun 12, when the US imposed export controls that made it illegal for all foreign nationals, outside or inside the US, to access the technology. 

Anthropic has been warning about the potential cybersecurity implications of Mythos since early April, when it released it to a limited set of companies to help them prepare, as part of an initiative called Project Glasswing. 

US agencies were particularly concerned about Mythos being co-opted for military purposes by countries such as Iran, North Korea, Russia, China, and others. The speed at which Mythos uncovers software vulnerabilities and can potentially arm bad actors for cyberattacks was at the heart of the concern.

Washington hasn't only picked on Anthropic. It also kept OpenAI from broadly releasing its GPT-5.6 model, leading OpenAI CEO Sam Altman to call it "bad news" on X. Altman said, "At the request of the US government, it is launching today in limited preview instead of the open access launch we were planning." That was despite the fact that GPT-5.6 has tested even stronger than Mythos in some benchmarks. There's no word yet on when the government will clear GPT-5.6 for general use. 

But Washington's intervention with Mythos and GPT-5.6 may have set a new precedent for safety, audits, and oversight.

"If AI is going to be as impactful as all the major AI CEOs and some countries are saying, it should have the same audits and the same safety that other powerful technologies have," Chris Canal, CEO of AI eval vendor Equistamp, told The Deep View. "It deserves the same audits and safety standards as nuclear technology, [for example]."

Now the challenge is setting up a framework for pre-release model review, because no one wants a situation where AI labs release a model, the government intervenes, and then it gets re-released weeks later.

Perhaps the most notable outcome of the US government's standoff with Anthropic — and OpenAI now as well — is that it has resulted in a pause in promulgating these ever-powerful models. Such a pause to evaluate and prepare for the effects of the latest models had appeared highly unlikely. While leaders such as Anthropic CEO Dario Amodei have extolled the benefits of such a pause to assess safety, they've also made it clear they wouldn't do it unless other labs agree to a similar pause, because otherwise it wouldn't be effective. Will the staggered rollouts of Mythos and GPT-5.5 reset the expectations and norms for the way the most powerful frontier models are launched? If you are concerned about AI safety, that would be a win. The question would then shift to the who and the how of such a pause. 

TOGETHER WITH KALSHI

Trade The World Cup on Kalshi

The World Cup is the biggest sporting event on Earth.

Kalshi is a CFTC-regulated exchange where you can take a position on who lifts the trophy, backed by real-time market pricing.

Right now, Spain is priced at 10.7% to win it all. A $100 trade on Spain pays out $935 if they go all the way.

Trade $10, Claim $10

WORKFORCE

Forward Deployed Engineers come with a catch

Amazon is all in on the trendiest new job in AI: forward deployed engineers (FDEs).

On Wednesday, Amazon Web Services announced it is investing $1 billion to embed forward deployed engineers within its customer network. This network of engineers will help customers develop and deploy agents into their organizations, aiming to help them become "self-sufficient with AI" in days, the company said in its announcement. 

"Customers have moved past exploring what AI can do; they want to make it core to how they operate," Francessca Vasquez, VP of frontier AI engineering and services at AWS wrote in a blog post. "I have also heard loud and clear that many customers need expert AI engineers working directly with their teams to help them build and become AI-native organizations." 

As part of this new Amazon organization, engineers will work directly with a customer’s business, engineering, and security teams to help them build systems that integrate with their data and governance structures. 

• Rather than treating these deployments as "standalone projects," Vasquez said, the goal is to help these customers create lasting AI strategies, deployments and skills to "innovate independently." These agentic systems will run within a customer's existing AWS environment. 

• After the engineer is done, the organization is left not only with deployed agents, but with knowledge graphs, runbooks, architectural documentation and "trained internal champions" to keep the momentum going. The idea is for every new engagement with these AI systems to compound its knowledge and make it more useful. 

Amazon's investment is the latest indication that tech companies are betting on forward deployed engineers to solve the ROI problem in enterprise AI. This role, which essentially loans AI experts to a client's organization, has taken off in recent months, with job postings for forward deployed engineers up around 700% year over year, according to Business Insider. 

And Amazon's not the only one building its forces. In May, OpenAI launched the OpenAI Deployment Company, designed to help organizations build and deploy its technology using forward deployed engineers.

It makes sense that forward deployed engineers have taken off. AI is incredibly complex, and the barrier to entry can be daunting for many organizations. Forward deployed engineers can help remove those obstacles, and do so using the technology of whichever vendor they represent, creating a built-in user base. However, organizations that decide to take in forward deployed engineers have to remember that they come with a catch: These engineers represent the best interests of Amazon, OpenAI, or whatever other organization they come from. Working with them to integrate their tech into your systems can lead to unwanted lock-in. 

TOGETHER WITH LAMBDA

Cut your AI training costs by 25% or more

Most large-scale AI training runs use less than half the computing power they're paying for. Lambda's team found the root causes and built a reproducible framework that boosted efficiency by over 25%, without changing the model itself.

Lambda’s whitepaper shows you how to address:

• Memory inefficiencies silently inflating your costs

• Training configurations that aren't making full use of your hardware

• Bottlenecks that slow down GPU communication

Get the whitepaper.

HARDWARE

How Apple's decade-long bet on chips won AI

Walk into any of the frontier AI labs, and you'll find wall-to-wall Macs.

Decisions Apple made years ago, long before ChatGPT and the LLM era, are now paying big dividends.

In an exclusive conversation with The Deep View ahead of WWDC 2026 in June, Doug Brooks, senior product manager of Apple silicon, discussed how Apple's long-term investments in neural processing, unified memory, power-efficient computing, and hardware-software integration have positioned its devices for the AI era. 

This interview has been edited and curated for brevity and clarity.

Jason Hiner: Tell us where Apple silicon is today when it comes to AI.

Doug Brooks: We're seeing tremendous momentum with how people are using Apple products and Apple silicon, particularly in on-device AI workflows.... It really highlights the tent poles of Apple silicon. It's a very balanced architecture that provides CPU, GPU, unified memory, and the Neural Engine all contributing to performance across the chip. That's particularly important for these evolving agentic workflows.... It's not just about the GPU crunching on an LLM anymore. It's about the whole chip contributing to different parts of the task.

Jason Hiner: Apple was working on machine learning long before ChatGPT… What were the origins of that journey?

Doug Brooks: I think a lot of it starts with having really good foundational building blocks.... All of that was done with a relentless focus on power efficiency.... When we first introduced the Neural Engine, it was designed to solve very specific compute needs for machine learning.... So if you go back to the core of the question, the tent poles of Apple silicon, performance, efficiency, and unified memory, have played extremely well in the AI domain.

Jason Hiner: One of the trends we're seeing is developers wanting to run more AI locally because of privacy concerns, security requirements, and rising inference costs. Are you seeing that trend too?

Doug Brooks: Absolutely.... Why send your data, your code, or your intellectual property somewhere else if you can run it locally?... People are realizing they have incredibly powerful devices sitting right on their desks that are capable of handling a wide range of workloads.... I'm a big believer in using the right tool for the right job.... I think the hybrid approach is really interesting because agents can decide what needs to happen locally and what needs to happen in the cloud based on the workload.... We've talked publicly about the incredible demand for Mac minis and Mac Studios because they've become the platform for running so much of this [agentic] work.

Jason Hiner: How do you think about the future of on-device AI for iPhone and iPad?

Doug Brooks: Sometimes I think of it as transparent AI. There are so many AI features embedded throughout the operating system and throughout applications that don't necessarily jump out at you as AI. They're powered by AI under the hood, but they're just helping people get things done.... The work we've done with foundation models and exposing those capabilities to developers has allowed AI to become pervasive across applications in both small and large ways.

Read the rest of the interview with Apple's Doug Brooks.

LINKS

• Abu Dhabi's MGX raises $49 billion for AI-focused fund

• Palantir's Alex Karp calls military AI labs "effing insane" 

• Uber dismisses two leaders in AI data labeling business 

• Meta reportedly plans a cloud business for selling AI compute

• Open source AI firm Together AI raises $800 million

• ChatGPT: For Pro users, OpenAI has added new personal finance experience to securely connect financial accounts, see where their money is going, and query about your information. 

• Gemini Spark: Google's AI personal assistant is now available on MacOS for local tasks. 

• Google TabFM: Google has launched a zero-shot foundation model for tabular data. 

• xAI Voice Agent Builder: xAI has launched a no-code platform to create human-like voice agents with Grok Voice. 

• xAI: Member of Technical Staff - Inference

• Amazon: Applied Scientist, AWS Automated Reasoning

• Scale AI: Security Engineer, Product Security

• OpenAI: Agent Post-Training, Computer Use Research

GAMES

A QUICK POLL BEFORE YOU GO