Bilingual

GLM 5.2 beats Claude in our benchmarks
23 days ago
- GLM 5.2, an open-weight model from Zhipu AI, achieved a 39% F1 score on IDOR detection, surpassing Claude Code (32%) and costing about $0.17 per vulnerability found.
- The experiment compared models with and without a custom harness (scaffolding); GLM 5.2 performed well with only a prompt, while Semgrep's multimodal pipeline with a harness scored higher (53–61% F1).
- Key advantages of GLM 5.2 include being open-weight (MIT license), competitive coding performance (e.g., 81.0 on Terminal-Bench 2.1), and low cost relative to frontier models.
- IDOR (Insecure Direct Object Reference) vulnerabilities are common and challenging to detect due to their business-logic nature, requiring reasoning across files without clear dangerous functions.
- The study highlights the importance of harnesses in vulnerability detection, showing that performance depends on both model capabilities and the supporting infrastructure.
- Economic factors like cost per bug are crucial for scalability, with GLM 5.2 offering a cost-effective solution for security tasks.
Open-Source Communism
3 days ago
- The author addresses concerns about open-weight models being 'decelerationist' and threatening American labs' business models.
- Open-weight models are compared to open-source programming languages, where ecosystem value outweighs raw code.
- American labs' success will depend on developer ecosystems, integrations, and tools, not just superior model capabilities.
Who's Afraid of Chinese Models?
a day ago
- The author initially believed tech, especially software's zero marginal costs, was fundamentally different, but later realized zero marginal costs underpin Aggregation Theory, leading to centralized value chains.
- AI is reintroducing old business principles like marginal costs, challenging the notion that open-weight models are free; they involve significant COGS (cost of goods sold) for inference.
- Tokens are not fungible commodities; intelligence derived from tokens is fungible. COGS for intelligence depends on factors like model footprint, inference efficiency, and token efficiency.
- Commodity markets operate where all sellers charge the same price based on supply and demand, with profits depending on cost structures; suppliers with higher marginal costs risk bankruptcy.
- Currently, the AI intelligence market isn't fully commoditized due to high demand and compute shortages, giving frontier labs like Anthropic and OpenAI high margins and low costs per unit of intelligence.
- Frontier labs may fear Chinese models due to training cost dominance, the non-perfect commoditization of intelligence (where data from inference improves models), and competition in customer experience integration.
- China's strategy involves commoditizing AI complements through open-weight models to boost its physical world dominance and weaken U.S. frontier labs, leveraging distillation attacks for cost advantages.
- Distillation attacks allow Chinese labs to use frontier models as teachers for rapid improvement, disadvantaging Western open-weight model makers who must follow stricter terms of service.
- Cybersecurity concerns arise as defenders may need to use Chinese models due to restrictions on U.S. models, highlighting the need for policy changes to ensure equal competition and access.
- The author argues against overreacting to Chinese models from an economic perspective but emphasizes cybersecurity risks and calls for U.S. policies that support innovation and defense capabilities.
Is Open Weight AI Decelerationist?
15 hours ago
- Dean Ball of OpenAI claimed open-weight AI models are inherently decelerationist, sparking debate.
- Decelerationism opposes accelerationism, which views rapid tech proliferation as beneficial.
- The economic argument: open weights may deter capital expenditure by compressing future revenues.
- China's ability to subsidize AI infrastructure and overbuild capacity amplifies the threat to profits.
- Current training paradigms are transitory; future AI may be more efficient, reducing upfront capex.
- Value may shift from base tokens to applications, integration, and services, like electricity or Linux.
- Open weights lower barriers to participation, expanding innovation beyond a few labs.
- The article concludes open weights are not inherently decelerationist but a bet on a broader value ecosystem.

Hasty Briefsbeta

#open-weight models

GLM 5.2 beats Claude in our benchmarks

Open-Source Communism

Who's Afraid of Chinese Models?

Is Open Weight AI Decelerationist?