The Pros and Cons of Open-Source Models and the Promise of Geothermal Energy: A Dual Focus in the AI Industry

The Real Benefits and Risks at Play in Open-Source Models

According to reports cited by a Bilibili creator and a YouTube channel, referencing MIT Technology Review's The Download, AI company Anthropic recently released a report warning that a powerful new model from the Chinese company Z.ai could pose a significant cybersecurity threat. This assessment has pushed the debate over the pros and cons of open-source models back into the spotlight.

Supporters of open-source models have long emphasized their value in driving innovation, lowering barriers to entry, and promoting transparency. But Anthropic's report highlights a downside: when a sufficiently capable model is made public, malicious actors can access it too, using it to launch cyberattacks or generate harmful content. This is not a theoretical exercise—it is a real risk that can emerge once model capabilities cross a certain threshold.

The core of the debate is not whether open source is good or bad, but how to balance capability release with safety guardrails. Cross-referencing multiple sources, the industry has yet to reach a consensus—one side argues that openness is the necessary path to technological democratization, while the other advocates stricter release controls for frontier models.

Anthropic is an American AI safety company founded in 2021 by former OpenAI research executives, best known for its Claude series of large language models. The company has long positioned "AI safety" as its core brand identity, advocating for responsible release strategies for frontier models. Z.ai (the overseas brand of Zhipu AI) is an international product line launched in 2024 by the Chinese large-model company Zhipu, whose GLM series models perform close to the international frontier on multiple benchmarks. Anthropic's judgment that Z.ai's model poses a "cybersecurity threat" has drawn attention because it touches a sensitive point in the open-source community: when a company that positions itself on safety issues a warning about a competitor's open model, it is difficult for outsiders to fully determine whether the judgment is based on technical assessment or is mixed with commercial competitive considerations. This "both referee and player" role means the objectivity of such reports is often questioned.

Open-source models generally refer to AI models whose weights can be publicly downloaded and freely used and modified, as opposed to closed-source models accessible only via API. Supporters argue that open source accelerates innovation, lowers barriers for small and mid-sized developers, and facilitates security auditing; opponents point out that once weights are public, they cannot be recalled, and malicious actors can fine-tune models offline to bypass safety alignment and use them to generate malicious code, phishing content, or automated attack tools. Safety guardrails refer to filtering and restriction mechanisms at the model output level, including alignment during training (RLHF), inference-time moderation, and access controls. The core dilemma of open-source models is that guardrails typically reside on the server side, and once weights are released, guardrails become effectively useless. This has led the industry to seek compromises such as "open weights but restricted capabilities," including delayed releases, tiered licensing, or capability-threshold controls.

Why Geothermal Energy's Promise Is Heating Up Again

Alongside the debate over AI models, The Download also focused on the prospects for geothermal energy the same day. Geothermal has long been regarded as a niche energy source, but according to analysis cited by a YouTube channel and a Bilibili creator, new technological pathways are changing that impression.

Traditional geothermal development is constrained by geography, with plants only feasible in areas of frequent volcanic activity or thin crust. A new generation of enhanced geothermal systems aims to break through this limitation by artificially creating reservoirs to extract underground heat under a broader range of geological conditions. If costs can be reduced, geothermal energy could become a stable baseload power source, compensating for the intermittency of wind and solar.

Multiple independent sources take a cautiously optimistic view, believing that progress has been made in the technology validation stage, but commercial-scale deployment still faces uncertainties such as drilling costs, induced seismic risk, and long-term operations and maintenance.

Enhanced Geothermal Systems (EGS) is an engineering approach that artificially modifies underground rock formations to extract heat. Traditional geothermal power generation relies on naturally occurring high-temperature reservoirs, mostly limited to plate boundaries or volcanically active zones. EGS involves drilling into deep hot dry rock, using high-pressure water injection to create an artificial fracture network in the rock, and then extracting heated water or steam through another well to drive power generation. In theory, this approach could expand geothermal development to most regions of the world. Baseload power refers to power sources that can deliver stable output around the clock without weather fluctuations, complementing the intermittency of wind and solar. The main obstacles to EGS commercialization include: drilling depths often reaching several kilometers, leading to high costs; induced microseismicity from artificial reservoir creation; difficulty ensuring long-term connectivity of fracture networks; and long payback periods for upfront project investment.

A Dual-Track Technology Agenda

Placing open-source model governance and geothermal development in the same newsletter issue is no coincidence. Both involve a common question: is the pace of technological capability release outpacing society's ability to digest its consequences?

The debate over open-source AI models concerns the boundaries of information and capability diffusion, while the discussion of geothermal energy concerns the physical boundaries of energy infrastructure. The former iterates rapidly in digital space, while the latter advances slowly deep within rock formations, yet both equally test the judgment of policymakers, companies, and research institutions.

For now, neither side has easy answers. Open-source models will not stop being released simply because risks exist, and geothermal energy will not be abandoned because of its many challenges. What truly deserves attention is whether the industry can, while advancing, establish assessment and governance mechanisms that match.

Technology governance refers to institutional arrangements through which society guides and constrains the speed, direction, and consequences of technological development, covering laws and regulations, industry standards, ethical review, and international coordination. The article's question of whether "the pace of capability release exceeds society's capacity to digest" corresponds precisely to the Collingridge Dilemma in governance research: a technology is difficult to regulate effectively when its early impacts are unclear, but by the time those impacts become apparent, it is already deeply embedded in society and difficult to change. This dilemma manifests simultaneously in AI open source and geothermal development—the former has an extremely short governance window due to rapid iteration, while the latter adjusts slowly due to long infrastructure cycles. Whether the industry can establish assessment mechanisms matching the pace of technology is the shared open question in both discussions.