Cognitive warmup. China’s Zhipu AI (Z.ai) has an open-weight GLM-5.2 model which some researchers insist matches Anthropic’s controversial Mythos in certain cybersecurity and vulnerability identification tasks. It must be noted that while GLM lags compared to other models from Anthropic and OpenAI in more general tasks, this is a representation that Chinese AI models have systematically reduced the gap in average capabilities compared with other AI companies.
GLM-5.2 continues to rank among the 10 most-used AI models on OpenRouter’s LLM leaderboard, siting alongside models from Anthropic, Deepseek, Xiaomi and Tencent. In some benchmarking tests, according to cybersecurity company Semgrep, GLM-5.2 performed better than Anthropic’s Claude Opus 4.8 model (this was released in May). Depending on instructions and specificity, Opus 4.8 and GLM-5.2 can match Mythos in vulnerability-finding, according to the researchers. This must be worrying?
PREVIOUSLY, ON NEURAL DISPATCH
A SEMICONDUCTOR BREAKTHROUGH
IBM had made the world’s first sub-1 nanometer (nm) chip technology, specifically operating at the 0.7 nm or 7 angstrom node. This achievement represents a landmark moment for an industry that has been finding ways to work around physical limits of traditional chip scaling. Semiconductors play increasingly critical roles in everything from computing and household appliances to transportation systems and critical infrastructure, the ability to continue minimising transistor size while improving performance, had a broad impact.
At the heart of this is IBM’s entirely new transistor architecture known as “nanostack”, and architecture that becomes the first to stack components vertically like a skyscraper. IBM says this is a significant advancement from their own previously developed nanosheet used for many 3nm and 2nm chips. Here are some headline capabilities:
- 100 billion transistors onto a chip the size of a fingernail—that is twice the density of IBM’s 2nm node chip, which marked a major leap only a few years ago
- 50% higher performance or 70% greater energy efficiency, which plays the dual role of powerful computing with lower power requirements.
- IBM also says this methodology enables 40% scaling in SRAM (they say this is the biggest leap in a decade) which is crucial for AI workloads, cloud infrastructure, and next-generation electronic devices.
“IBM’s latest chip breakthrough marks a landmark moment in computing, pushing technology beyond the nanometer era to the scale of atoms. With our new nanostack architecture, we’re not just making smaller transistors, we’re reinventing how chips are built to deliver dramatically more power and energy efficiency,” says Jay Gambetta, Director of IBM Research and IBM Fellow.
This technology means it is possible to extend scaling below the 1nm node, advancing the semiconductor industry into an era of angstrom-level scaling where dimensions approach the size of individual atoms. While transistor nodes now refer more to a generation of manufacturing technology rather than an exact physical dimension, IBM’s 0.7 nm technology is a sure enough demonstration that continued scaling remains highly possible. According to IBM, this nanostack architecture establishes a semiconductor roadmap that projects at least a decade of future scaling.
GOOGLE’S PRAGMATIC APPROACH
Google’s recently published white paper titled “A Pragmatic Approach to AI Governance in America,” is a clear and cogent attempt to get the artificial intelligence regulation conversion on track. Google pitches two very clear things. First, a clear distinction between frontier models and widely-used AI applications. Secondly, what they call a “ pragmatic, evidence-based approach” for the overlaps between the two.
I find merit in Google’s call for that separation as the foundation for any regulation. AI is in both spheres, that is everyday chatbot and tool use, as well as extraordinary scientific discoveries. The two streams (and there are many sub streams, not to be ignored) cannot be regulated in the same way, at the same intensity either. Google’s call is one that points to a rather unhelpful dichotomy the AI space is currently grappling with—it is either draconian over-regulation that stifles progress, or inattentive regulation that inevitably endangers users. Google advocates for a “middle way”—a pragmatic, evidence-based approach tailored to the diverse realities of different AI systems.
The cornerstone of Google’s proposal is a bifurcated regulatory framework that actively distinguishes between “frontier AI” (the most advanced, highly capable models) and “widely-deployed AI applications” (everyday tools with lower, narrower capabilities). By avoiding a one-size-fits-all legislative blanket, Google argues that regulators can successfully target specific, identifiable real-world harms without fundamentally impeding the underlying computer science.
For Frontier AI, Google believes these are key:
- An independent regulatory organisation that can keep pace with fast-moving AI research and development.
- Scientific benchmarks for identifying frontier capabilities in the cyber and chemical, biological, radiological, and nuclear (CBRN) domains, complemented by clear safety and security standards for building, testing, and deploying the most advanced AI systems.
- Annual audits to demonstrate procedural, and ultimately substantive, compliance with safety standards, supported by model transparency and reporting requirements
For frontier AI, where the implications span systemic safety and national security, Google proposes the creation of a Frontier Regulatory Organisation (FARO)—an independent, federally overseen, and industry-backed entity. FARO would be tasked with establishing nimble safety standards and verifying independent, voluntary audits of the most advanced AI models. This model mirrors how other critical sectors are managed, providing a flexible framework that can keep pace with rapid algorithmic innovation.
“For AI applications enabled by models at lower levels of capability, the federal government does not need new regulatory regimes that duplicate or conflict with existing law. AI applications like chatbots raise social and consumer safety issues distinct from the national security issues raised by the most advanced frontier AI models. For these widespread applications the federal government can draw on, and in some cases amend, existing laws and rules to address real-world outputs and specific harms,” Google writes in their white paper.
Beyond model governance, the white paper addresses a broader ecosystem requirement for sustainable AI leadership. They broach the topic of public-private initiatives to scale America’s energy generation (this is a very uncomfortable topic at this time) and transmission grids with something similar to the “Eisenhower Highway Program”, importance of information integrity, urging regulators to mandate watermarking technologies like SynthID, and tamper-resistant cryptographic provenance standards like C2PA for generative AI services.
COST, VALUE AND SENSIBILITY
Uh-oh! AI is proving to be more expensive to run, than the humans it urged corporates to replace? Common sense hasn’t exactly been common since investors and boardrooms have become AI obsessed over the past couple of years. Anyway, don’t believe me, but believe the numbers.
By 2028, AI coding costs will overtake an average software developer’s salary, according to estimates by research firm Gartner. Core to this would be the extremely high costs of large language model (LLM) token consumption and an industry that largely works on consumption-based licensing models. Gartner warns that organisations are rapidly transitioning from initial experimentation to scaled deployment of AI coding agents, while vastly underestimating the financial impact of this rising token usage. Tokens—the fundamental units of data processed by generative AI models—directly dictate the cost of these software tools under new consumption-based pricing structures.
AI companies are of course smart. The subtle shift away from a more financially predictable seat-based licensing to more volatile (and therefore expensive) token based structure, is designed to earn them revenue. Not save a corporate money by replacing humans with AI. I wouldn’t put it past anyone to compound this issue with lack of transparency about token consumption calculation and a distinct inability to accurately plan budgets as well as track cost to result.
It wouldn’t be outlandish to say that most organisations lack the maturity and frameworks required to actively measure the cost of AI with actual business impact, apart from the boardroom excitement about being AI-first or whatever the terminology is.
