Open-Source AI

Today's Stories

1. 1

   AT&T cuts Anthropic spending with open-source AI models

   AT&T is deploying open-source AI models to reduce its reliance on Anthropic's commercial services and lower its spending on the AI company's products. Enterprise customers are increasingly exploring cost-saving alternatives to premium AI vendors, signaling that organizations can manage expenses by mixing commercial and freely available models — a shift that may pressure AI vendors to justify premium pricing.

   The specific models AT&T has chosen and the degree to which this strategy reduces Anthropic spending will indicate whether enterprises view open-source options as genuine substitutes for commercial AI services.

2. 2

   AI Revolution Shifts to Real-World Problems as Job Fears Ease

   U.S. software jobs have risen over the past year, and the 12-month moving average of workers in computer and mathematical occupations hit an all-time high in July 2026. Youth unemployment (ages 20–24) fell to 7.1% in July 2026, down from an average of 8.3% in 2025. Meanwhile, concerns have shifted from job losses to AI sovereignty—fears that companies will become vulnerable to cyberattacks as open-weight models (many from China) narrow the gap with closed-weight frontier models. Companies are increasingly worried about depending too heavily on third-party AI models. According to the Foundation for American Innovation, 56% of business owners and C-suite executives are concerned that competitors will benefit from AI models trained on their organization's proprietary data, workflows, or institutional knowledge. Most executives remain dissatisfied with their return on AI spend in real-world use cases. The next phase of AI requires building on proprietary, domain-specific data—manufacturing, clinical, material, and physical data that often doesn't yet exist—and protecting that data will require stronger AI sovereignty safeguards.

   Hundreds of companies and organizations, including Amazon, Google, and OpenAI, have signed the Open Weights and American AI Leadership letter, backing open-weight models as part of a strong AI ecosystem. Companies such as Palantir and Nvidia are building a middle-layer to help enterprises connect to AI models without exposing proprietary data. The body does not forecast a popping of an AI bubble, but signals that the AI revolution is entering a new phase focused on solving difficult real-world problems.

3. 3

   GLM 5.3 shows parameter count obsolete; scaling moves to data and compute

   Prof Jie Tang of Z.ai announced that parameter count alone no longer determines model capability. GLM 5.3's improvements came solely from reinforcement learning on long-horizon environments—the same base model as GLM 5.2 but enhanced via roughly one month of extra RL. Tang identified five scaling knobs beyond parameters, including MoE sparsity with new XA-YB notation, and argued that advanced skills like finding software vulnerabilities require long causal chains (20+ inference steps) rather than raw parameter size. The industry has long used parameter count as shorthand for model strength, but this shift signals that compute efficiency, data quality, and post-training strategy now matter far more. For businesses evaluating or building AI systems, it means comparing models by parameter count alone is misleading—reasoning depth, training recipe, and the environments used to refine behavior are what determine real-world performance on complex tasks.

   GLM 5.3 ranks #2 on Terminal Bench, #3 on Legal Bench, and #6 on Skills Bench among open-weight models. Several other labs are already adopting similar post-training approaches: Ornith-1.5 released in 9B dense, 35B MoE, and 397B MoE variants with end-to-end self-improvement; Microsoft's Agent Lightning moves Qwen3.5-9B on SWE-Bench Verified from 41.8% to 56.4% using modest compute and ~6K training examples.

4. 4

   Snowflake Routes AI Tasks to Cheaper Models, Adds DeepSeek and GLM

   Snowflake announced dynamic model routing in Cortex AI Gateway, which automatically directs each AI request to the lowest-cost model capable of completing the task, plus added DeepSeek-V4-Flash 0731 and GLM-5.3 (coming soon) to its open model portfolio. Early tests show up to 3× better token efficiency on data pipelines and ~25% fewer tokens on coding tasks—meaning organizations can cut AI inference spending without rebuilding applications or sacrificing output quality. As more capable, cheaper models become available, the routing system improves continuously without manual reconfiguration.

   DeepSeek-V4-Flash 0731 scores 74.4% on ADE-bench (outperforming the proprietary model Snowflake tested); GLM-5.3 is in private preview, designed for self-hosting with low token footprint. Both run inside Snowflake's security boundary, keeping data governed.

5. 5

   Nvidia, AMD vie for US open-model lead as China dominates scale

   Hugging Face's summer 2026 open model report found that Chinese labs have continued to lead in frontier open-model parameter scale since 2026, while competition in the US has shifted from model labs like Meta and Google to chipmakers including Nvidia and AMD. The report signals a structural shift in who is driving open AI development in the US—away from software-focused research labs toward hardware makers. This reflects changing incentives and resources in the open-model ecosystem, with chipmakers now positioned as key players alongside traditional AI researchers.

   The balance between Chinese frontier open-model scale dominance and US chipmaker competition may reshape investment and partnership strategies in the open-source AI sector.