Open-Source AI

Today's Stories

1. 1

   AT&T builds AI strategy around open models and token economics

   AT&T is adopting open-source AI models as a core part of its strategy, focusing on what the company calls 'tokenomics'—managing costs through token consumption (the unit by which AI models charge for processing text). The telecom is moving away from reliance on proprietary closed models. Open models offer AT&T greater cost control, vendor independence, and the ability to run AI workloads on its own infrastructure rather than paying per-token fees to outside providers. This approach allows large enterprises to optimize their AI spending as token costs remain a significant operational expense.

   How AT&T's token-efficiency strategy influences its broader cloud and edge computing footprint, and whether other large telecom or enterprise players follow a similar model-agnostic approach to AI adoption.

2. 2

   AI helps J.P. Morgan catch fraud as attempted attacks rise 20% yearly

   J.P. Morgan's Commercial & Investment Bank reports attempted fraud volume rose more than 20% every year for the last 5 years, driven partly by open-source AI making it easier for fraudsters to generate polished emails and impersonate contacts at scale. The bank counters by using AI models trained on millions of transactions to generate risk scores that flag suspicious payments for review before release. Open-source AI is a double-edged sword—while it enables more sophisticated fraud, J.P. Morgan's fraud detection AI can identify subtle anomalies and sophisticated attacks that traditional tools like linear regression and decision trees cannot. The bank processes $12 trillion in payments each day, so even small improvements in detection precision reduce both fraud losses and false-positive alert fatigue that slows legitimate transactions.

   J.P. Morgan emphasizes that AI alone is not enough; strong internal controls and authentication technology remain critical to preventing fraud. Joel Kalamba, business analysis director at J.P. Morgan, notes that "technology combined with strong internal controls is what keeps clients safeguarded," signaling that the bank's defense strategy depends on both new AI and time-tested operational discipline.

3. 3

   Laptop AI model matches cloud performance, just slower

   Qwen3.8-27B, a model that runs on a laptop, ranks #1 of 135 models on Artificial Analysis's Intelligence Index with a score of 52, outperforming GLM-5.2 (the state-of-the-art open-source model from Z.ai, at 753b parameters) which scored 51—despite being roughly 28 times smaller. Small, locally-run models achieve the same answer quality as cloud-based models like DeepSeek V4, but use different reasoning strategies. Smaller models must reason more from first principles rather than rely on memorized knowledge, making them viable for anyone with a laptop instead of cloud access.

   Speed trade-offs are significant. In a benchmark of 25 venture-capital tasks, Qwen3.8-27B and DeepSeek V4-Flash both delivered quality scores of 8.0 out of 9, but DeepSeek answered in 1.1 seconds on average while Qwen3.8-27B took 7.2 seconds; Qwen3.6-35B took 10.0 seconds.

4. 4

   Snowflake adds smart routing to cut AI costs per task

   Snowflake launched dynamic model routing in Cortex AI Gateway, which automatically directs requests to the lowest-cost model capable of completing each task. The company also expanded its open model portfolio to include DeepSeek-V4-Flash 0731 (available in private preview) and GLM-5.3 (coming soon in private preview), both self-hostable within Snowflake's governed data environment. Most organizations default every AI request to their most powerful (and most expensive) model, even when simpler tasks don't need that capability. Snowflake's routing layer cuts this waste by matching task complexity to model cost—internal testing showed up to three times greater token efficiency on data pipeline workloads and approximately 25% fewer tokens on coding tasks while maintaining comparable quality. This approach lets companies reduce AI spend without rebuilding applications or manually coding routing logic.

   DeepSeek-V4-Flash 0731 scores 74.4% on ADE-bench (Snowflake's internal data engineering benchmark), outperforming the leading proprietary model tested; GLM-5.3 is noted for low token footprint and solid data engineering accuracy. Both models run inside Snowflake's security perimeter alongside governed data, so organizations don't route data through third-party APIs.

5. 5

   Alibaba launches laptop AI model as open-source race heats up

   Alibaba has released an open-source AI model designed to run directly on laptops, just ahead of a similar move by Meta. The release underscores a shift in competition toward on-device AI that users can download and run on their own hardware. Open-source models are no longer just cheaper alternatives to proprietary AI from Silicon Valley labs — they are now leading the race on the newest competitive frontier: AI that runs locally on consumer devices rather than relying on cloud servers. This means faster inference, greater privacy, and reduced dependence on big tech platforms' servers.

   The timing and capabilities of both Alibaba's and Meta's laptop-optimized models will likely shape how quickly businesses and consumers adopt on-device AI, and whether open-source or proprietary solutions dominate the personal-computer AI space.

6. 6

   Amodei says AI centralizes by nature; open models just shift power to chip owners

   Anthropic CEO Dario Amodei entered a public debate on X with investors Gavin Baker and David Sacks, and Meta researcher Yann LeCun over whether AI should be concentrated among a few regulated companies or spread widely through open models. Amodei argued that open models do not solve concentration—they simply shift power to whoever controls the most computing power and AI chips, because AI centralizes by its very structure due to scaling laws. The dispute cuts to the core of AI regulation strategy. Amodei contends that regulation can rein in corporate power and help ordinary citizens when good institutions anchor power in ideas rather than individuals, and that Anthropic deliberately designs proposals to slow down leading labs and favor smaller ones (citing California's SB53 law as proof). His critics counter that this amounts to regulatory capture—that Anthropic has hired several former Biden administration AI officials and is building a large government apparatus while proposing a federal agency to review top models before release, which they say would create delays and disadvantage the US against China.

   Baker noted that almost every major company except Anthropic has signed Jensen Huang's letter backing open models, signaling that the industry consensus may be shifting away from Amodei's concentrated approach.