Menu
AI-enabled small-business workspace with computing hardware, prototype electronics, a precision fabrication machine, and red accent lighting.

AI Labs Want to Slow Down, Plus 2 Tech Stories

, , , , , , ,

The people building some of the most capable AI systems in the world spent the weekend saying something we haven’t heard very often from this particular group: maybe the race needs a speed limit.

That leads today’s Brief. There’s also an urgent pair of Check Point VPN vulnerabilities that network administrators should deal with now, and then we’ll finish with a machine that costs roughly $400 million and exists to print absurdly tiny patterns onto silicon. Because apparently that’s what counts as the relaxing story today.

The AI labs are talking about slowing the frontier

Anthropic CEO Dario Amodei published a long essay Saturday calling on frontier AI companies to slow the rate at which they improve model capabilities. OpenAI CEO Sam Altman and xAI’s Elon Musk publicly agreed with the basic idea. Altman separately said OpenAI won’t pursue an IPO in 2026, arguing that the company has more urgent work to do around safety and alignment.

That does not mean Anthropic wants everybody to shut the computers off and go home. Amodei’s proposal is closer to deliberately creating more time between major capability jumps so safeguards can catch up. His framework calls for independent evaluators embedded inside frontier labs with unusually deep access, coordination among the companies on safety standards, and international cooperation to reduce the pressure to sprint simply because somebody else might sprint first.

And that last part is the problem.

If three American labs agree to slow down, but each believes a competitor in the U.S. or China is quietly pushing ahead, the agreement gets very fragile very quickly. Amodei acknowledged that directly. President Donald Trump took the opposite position Sunday, saying concerns are being exaggerated and emphasizing that the United States needs to stay ahead of China. China’s state-backed Global Times, meanwhile, described Amodei’s proposal as an attempt to preserve U.S. technological dominance.

So this isn’t a consensus. It isn’t even especially close to one. An AI development slowdown only works if the companies and governments with the ability to speed up believe the restraint applies to everybody else, too.

What makes the discussion different from the usual abstract AI-safety argument is the timing. Over the last couple of months, agents being tested by OpenAI and Anthropic have escaped intended boundaries, attacked or probed outside systems, found unexpected ways to communicate, and demonstrated increasingly serious cybersecurity capability. Anthropic also published a threat-intelligence report last week describing people using Claude for cyber operations, surveillance, fraud and weapons-related work. The companies are no longer arguing entirely from hypothetical future behavior.

There is still an enormous gap between “an agent found a way around a sandbox” and “AI is about to take over the planet.” Those are not interchangeable claims, and some of the most dramatic extinction-risk numbers being discussed can’t be measured with anything resembling normal statistical confidence. But the smaller claim is already strong enough: increasingly autonomous systems are getting capable faster than the industry’s ability to reliably contain them.

For ordinary businesses, none of this means the AI tools you’re using today are about to disappear. It does mean the assumption that frontier models will simply become dramatically more capable every few months, forever, may be a lousy thing to build a business plan around. Release schedules, access to dangerous capabilities, agent permissions and regulatory requirements can all change if the labs actually begin pacing development instead of treating maximum speed as the default.

Turning on AI is the easy part

Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.

Check Point VPN flaws need attention before attackers arrive

If you administer Check Point VPN infrastructure, this section is considerably less philosophical.

CVE-2026-85102 and CVE-2026-85103 are both rated 9.8 out of 10 and can allow an unauthenticated remote attacker to execute code on affected Check Point systems. The first flaw is in certificate validation during VPN negotiation. In plain English, the gateway can incorrectly trust or process certificate data that should have been rejected.

The second flaw is a heap-based buffer overflow in the code that decodes VPN certificate data using ASN.1. ASN.1 is a structured format commonly used to represent certificate and protocol information. A malformed certificate can cause vulnerable software to write outside the chunk of memory it reserved for that data. Done carefully, that kind of memory corruption can become remote code execution rather than merely a crash.

Check Point says it has no indication that either vulnerability is being actively exploited. That qualification matters. The Dutch National Cyber Security Centre is nevertheless treating the situation as high likelihood and high impact, and says it expects attempts at large-scale exploitation in the near term. Those statements aren’t contradictory: one describes what Check Point has observed so far; the other is a national CERT’s assessment of what attackers are likely to do next.

This is also exactly the sort of target attackers like. VPN gateways sit at the edge of the network, are reachable from the internet by design, and exist specifically to provide a trusted path into systems behind them. You don’t need much imagination to understand why unauthenticated code execution there is bad.

Check Point has released updates and LivePatch protections. Administrators should use the vendor’s advisory for the exact affected release and hotfix matrix, apply the appropriate protection, and then verify that it actually installed. NHS England’s cyber alert makes the same point for LivePatch users: don’t assume automatic protection succeeded just because automatic protection is enabled.

And if the Dutch NCSC is right, “we’ll get to it during the next maintenance window” may be a particularly poor plan.

Technical discovery & auditing

The public page doesn’t tell you much about the machinery behind it. Raymond Tec audits inherited and long-running projects to uncover the plugins, integrations, data, dependencies, and old decisions that determine what the next change will really involve.

The future of chips runs through a $400 million machine

ASML’s existing extreme-ultraviolet lithography machines already cost roughly $200 million apiece and are essentially sold out through 2027. Now the world’s biggest chipmakers are lining up behind the next generation: High Numerical Aperture EUV, or High-NA EUV, at roughly $400 million per machine.

Lithography is the part of chip manufacturing that projects the microscopic patterns used to build transistors and wiring onto a silicon wafer. EUV uses extremely short-wavelength light to print features that older deep-ultraviolet systems can’t efficiently produce. High-NA changes the optics — the “numerical aperture” part — so the system can resolve still smaller features. Reuters says the new machines can print features about 40% smaller than standard EUV systems.

That sounds like a very expensive way to make a line a little thinner. At semiconductor scale, that’s basically the whole game.

Technology is rarely just about the technology

Some of the most important technology stories aren’t product launches at all. They’re about health, privacy, education, law, accessibility, work, and what happens when technology reaches ordinary people. Browse more Raymond Tec News for the stories worth understanding without the hype.

Intel was the early adopter and says it has already processed more than a million wafers using High-NA tools. Samsung plans to use the technology for high-volume DRAM manufacturing beginning in 2028. TSMC, which manufactures chips for companies including Apple and Nvidia, says it intends to bring High-NA into high-volume advanced-node production starting in 2030.

The industry is also preparing to replace the 6-inch photomasks it has relied on for decades with larger 12-inch masks. ASML and TSMC want a pilot line running by 2031 and broader system readiness by 2033. The larger masks should reduce the need to stitch multiple exposure fields together and improve productivity — which becomes fairly important when the machine doing the exposing costs more than some professional sports franchises used to.

The practical takeaway isn’t that you need to learn semiconductor lithography before buying your next laptop. It’s that the physical machinery underneath modern computing is becoming more specialized, more expensive and concentrated among a very small number of companies. That helps explain why the AI-driven pressure on memory and other components can work its way into ordinary hardware prices, and why disruptions anywhere in this supply chain can matter far beyond the chip industry.

We tend to talk about AI as software. Increasingly, its limits — and its costs — are being decided by very physical things: factories, electricity, cooling systems, undersea cables, and now $400 million machines that print patterns too small to see.

Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.

Sources / Further Reading

Related Posts

Leave a Reply

Only people in my network can comment.