POLICY

Will the US debate over open models bear fruit?

US tech companies are calling on American policymakers to support open models. But they may be swimming upstream, and they have one big naysayer.

On Friday, a group of 25 leading American tech companies led by Microsoft and Nvidia published an open letter titled "Open Weights and American AI Leadership," calling on policymakers in the US to:

• Avoid blanket bans on open-weight AI models 

• Take targeted legal and commercial steps against large-scale model distillation (see Kimi and Alibaba) and avoid sweeping restrictions on distillation for legitimate purposes

• Expand access to compute and other AI resources for researchers and startups to expand US innovation and competition in AI beyond a small handful of frontier labs

The main point of the open letter can be summed up by one quote from it: the idea is "keeping the frontier plural by avoiding premature restrictions on open models that stifle competition or drive innovation overseas."

By Monday night, the list of companies that signed the open letter had swelled to 133, and it included two of the three leading frontier labs in the US: OpenAI and Google. 

Conspicuously absent was Anthropic. But by the end of the day on Monday, CEO Dario Amodei had published a post titled "Our position on open-weights models" that:

• Asserted "Anthropic has never advocated for a ban on open-weights models"

• Expressed concern about the risks of AI capabilities falling into the hands of authoritarian governments, including China

• Raised the specter of AI models being misused to carry out cyberattacks and biological attacks

• Recommended that the most powerful models, closed and open, should be subject to mandatory safety testing and that testing should be based on a global, internationally agreed-upon standard

Where Amodei and the signers of the letter are the farthest apart is on the issue of cybersecurity. 

The letter wrote, "Just as open-source software demonstrated that transparency can be more secure than obscurity, AI safety may depend on giving more people the ability to test and strengthen the models on which society relies."

Amodei stated, "I don’t agree with the letter’s assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers. It seems at least as likely to me that the opposite will be true. For example, I worry that biology will have a strong attacker-defender asymmetry, where sufficiently capable models may be able to quickly weaponize pandemic-level viruses with widely available materials, whereas defense against these agents is a multi-year operational task in the best case." 

The other area worth highlighting is the economic incentives. Today, there has been so much investment in OpenAI, Anthropic, Google, and a host of other public companies and startups in the US, and those investors are deeply committed to profitable returns. That doesn't naturally lend itself to a long-term commitment to a cooperative environment, unless there's a crisis or another unifying factor. 

However, I was able to find someone to disagree with me on that. 

"Open models are extraordinarily important," Anastasios Angelopoulos, CEO of Arena, one of the signers of the open-letter, told The Deep View. "And the reason, fundamentally, is that companies want to own their own intelligence."

Angelopoulos said, "I actually view American open-source as being inevitable. That's why we signed the letter. The economic incentives for a huge American open-source company [are] too strong to ignore, because everyone's going to want to own their own intelligence, and everyone's going to want to be able to fine-tune their [own] large and small models. And nobody's going to want to do it based on an intelligence chain that relies on foreign imports."

We should focus less on who turns out to be right in this debate, because none of these are fundamentally original ideas. The greatest value from this open letter, like OpenAI's AI New Deal, the Pope's AI manifesto, and Demis Hassibis's AI safety plan is that it's creating societal dialogue around the kind of world we want to create in light of AI. This is a tacit acknowledgment that nothing is written in stone and there's no predetermined course of action that AI is going to follow. There are tons of choices, large and small, to be made in shaping a future that will be informed by AI capabilities. More dialogue will be needed at every level of society to ensure the best outcomes for the most people.

TOGETHER WITH AWS RE:INVENT

Experience hands-on building at re:Invent

Sitting still isn’t on the agenda. With over 2,200 sessions, 70% are hands-on at AWS re:Invent, returning to Las Vegas, November 30 - December 4.

• Test a service against a realistic workload before you commit to it in production

• Ask a service engineer about an edge-case constraint

• Access AWS certification discounts, with exam prep built into the schedule

• Talk through implementations with 400+ partners on the expo floor

Learn more and sign up by August 25th to save $1,200.

GOVERNANCE

Microsoft rolls out Mythos-armor for enterprises

Frontier AI is now powerful enough to escape the confines of its creators. Tech companies are fighting fire with fire. 

On Monday, Microsoft introduced MAI-Cyber-1-Flash, a multi-agent tool for identifying and remediating vulnerabilities. The tech giant claims that the model offers powerful cyber capabilities at half the cost of leading models. 

"As the cost of finding a flaw collapses, the old model of security, where you scan occasionally and patch eventually, is now obsolete," Mustafa Suleyman, Microsoft AI CEO, and Hayete Gallot, EVP of Microsoft Security, said in a joint blog post. 

Microsoft said that MAI-Cyber-1-Flash was built specifically to find difficult vulnerabilities in complex code bases. The model is integrated with MDASH, the agentic bug-hunting system that the company announced at Microsoft Build in early June. 

• The model was honed by the "largest security estate" in the world, the company claims, and beats out Anthropic's Mythos, Google's Gemini and OpenAI's GPT-5.5 Cyber and 5.6 Sol on the cybersecurity benchmarking system CyberGym.

• Additionally, because token cost has increasingly become a limiting factor in cyber defense, MAI-Cyber-1-Flash was designed to handle 90% of all tasks, while MDASH uses heavier-duty models to handle the exceptionally hard work, noting that it relies on GPT-5.4 for those tasks.

• Previously, MDASH relied on a number of external models, with its most efficient mix being GPT-5.4, GPT-5.4 mini, and GPT-5.3-Codex. Now using MAI-Cyber-1-Flash to cover most of this ground, the costs are 50% cheaper than previous offerings, Microsoft said. 

To centralize these capabilities, Microsoft also launched Perception, an agentic security system to continuously monitor, patch, and close new threats, which will utilize MAI-Cyber-1-Flash for both security workflows and work beyond the scope of software bug hunting. 

"If we’re to unlock the true benefits of AI, we must first build outstanding cyber models that help all of us harden the software the world runs on," Suleyman and Gallot said in the post. 

Microsoft's new offering comes at a pivotal time in cybersecurity. OpenAI's accidental breach of Hugging Face last week while testing its powerful new models marked one of the first instances of frontier AI going rogue, sending shockwaves through the AI industry and serving as a wake-up call that enterprises are not ready for the cybersecurity risks that the latest AI models are poised to unleash. Any solutions that could help enterprises increase security are likely to be welcome. For example, last week The Deep View covered Cisco launching SLMs to help bolster enterprise security.

AI has proven itself to be incredible at spotting software vulnerabilities. A report from the US National Vulnerabilities Database found that the number of flaws discovered in popular tech products is on par to roughly double this year, largely due to AI's superior bug-hunting capabilities. But vulnerability spotting itself isn't inherently a good thing, and these capabilities can easily be used by malicious actors to find exploits for cyber attacks. It's why the industry has no choice but to fight powerful AI with powerful AI. Microsoft is well positioned to champion this, and investing its resources into powerful AI for security could turn into a winning strategy. Clearly, Cisco sees the opportunity there as well. Additionally, since it has arguably the largest installed base of enterprise software, Microsoft already has the trust and relationships to deploy AI security solutions broadly. If those solutions can help companies cut down on token costs, that's an added bonus. 

TOGETHER WITH GENERAL ASSEMBLY

25% of AI initiatives get scaled back within a year. Bad output is the top reason why.

A quarter of leaders have had to reverse or scale back an AI initiative in the past year. 41% said it was because the AI's output wasn't as good as what people produced. 36% said their data wasn't ready.

23% said their team simply didn't have the skills to make it work.

GA's AI For Leaders training closes the skills gap behind these rollbacks, right alongside data readiness and workflow resistance.

See how the training closes that gap.

HARDWARE

5 new AI glasses that will test the form factor

Samsung is about to enter the smart glasses market this fall, but a lot of other competitors are also jumping in.

Despite AI exploding in popularity, using it is still mostly limited to laptop and phone screens. But AI wearables that give models access to your environment via cameras, microphones, and sensors are seeing a wave of new entrants. Smart glasses are leading the pack, thanks to a proven form factor and their ability to capture your exact point of view.

New players will be entering the market soon, and here's a list of the ones to watch (for more details about each one, see my full article). 

• Google and Samsung's Intelligent Eyewear: AI smartglasses with speaker, camera, and microphones | designed with Warby Parker and Gentle Monster | Gemini | 9-hour battery | ship in the fall 

• XREAL Project Aura: XR glasses | 70-degree field of view in the lenses | Android XR | Tethered battery pack | 96 grams | Snapdragon Reality Elite chipset | available to reserve | cost is targeted around $1,500 before tax | ship in the fall

• Snap Specs: AR glasses | 51-degree field of view | no tethered battery or compute pack | 132 to 136 grams | available for pre-order | $2,195 | ship in the fall 

• Halliday G2: Dual micro-LED projectors | 4-mic array | 12 hours of typical use | speakers | under 50 grams | focused on meetings | $599 | ship in September 

• Brilliant Labs Halo Glasses: Full-color display made possible by a tiny optical module | speakers | 14-hour battery life | 40 grams | conversational AI agent called Noa | focused on privacy | available for pre-order | $399 | ship in August 

Beyond the confirmed smart glasses above, there are rumored ones from major players. Apple will reportedly announce its smart glasses at WWDC 2027 next summer. Meta is hosting its annual conference, Meta Connect, on September 23-24. Mark Zuckerberg has teased that new glasses are coming at the event.

Bloomberg Apple watcher Mark Gurman has reported that the tech giant is targeting a late 2027 release of glasses without in-lens displays and in the $200 to $500 range. Apple could leave out cameras or keep them from recording photos and video. Gurman also said Apple will likely build its own frames, which will come in unique colors and frame styles and have a focus on privacy to overcome Meta Ray-Ban's reputation for clandestine recording.

The smart glasses sector is booming, with many more glasses launching in the upcoming months. Despite the increase in supply, there have yet to be any smart glasses that are really pushing the envelope in design. The closest is Snap Specs, which is tackling the very difficult task of making AR glasses with no additional battery packs or tethering. However, the price and form factor aren't where they need to be to justify the purchase for regular consumers. The inclusion of in-lens displays in a way that doesn't compromise comfort, cost, or aesthetics is what's needed to stand out in the market in a meaningful way, and perhaps spur on mass adoption. We know Meta and Google are working on it, and we have to believe Apple has it on its long-term roadmap. But miniaturized display technology may still need to improve for several generations before it's ready to power a mass market product.

LINKS

• Nvidia forms the Open Secure AI alliances with CrowdStrike, and Dell 

• Meta reportedly secured secret Louisiana data center deal

• Apple smart glasses face privacy hurdle, causing launch delay

• WSJ examines young adults using AI to script social, romantic exchanges

• Multiverse Computing raises $570 million at $1.7 billion valuation

• Nvidia, OpenAI in talks for $250 billion data center financing

• Chinese AI companies lack clear monetization strategy despite model gains

• MoonEP: Moonshot open-sourced its high-performance communication library for distributed MoE workloads

• Cohere: North Automations lets users use plain language to design automated workflows

• Kimi K3: Open weights are now available 

• Grok 4.5: The latest model from xAI, built for coding, agentic tasks, and knowledge work.

• Distyl: AI Engineer 

• Kargo: AI Engineer 

• Conduit: AI Engineer

• Campfire: AI Engineer 

GAMES

POLL RESULTS

Do you think tech giants' spending on AI compute will pay off?

Yes (39%)
Somewhat (36%)
No (22%)
Other (3%)