Open-Weight Models Beat GPT and Claude in Financial Document Evaluation

Hedge fund Bridgewater and Thinking Machines Lab have jointly released a report on how AI models perform on financial documents. The conclusion is intriguing: fine-tuned open-weight models outperformed top-tier API models such as GPT and Claude at a fraction of the cost.

Why did the big-name models lose? The reason is that "the correct answers were not public." Finance is full of ground-truth data that has never been published, so it never made it into the pretraining data of general-purpose models. As a result, smaller models with specialized fine-tuning gained the upper hand.

The findings suggest that, beyond finance, an open-weight model plus proprietary data is a powerful strategy in any specialized domain. That said, the evaluation rests on a specific dataset, and how to assess general-purpose performance remains an open question.

AI Data Centers' Water Usage Far Exceeds Reported Figures

A Wall Street Journal investigation has found that the actual water usage of AI data centers significantly exceeds what major tech companies report. Cooling water consumption driven by AI inference and training is reportedly far greater than the numbers companies disclose.

Accurate accounting of water usage is an issue that bears directly on regional water resource planning. For companies concentrating data centers in dry regions, water depletion is becoming a genuine business risk. This is a moment that calls for greater transparency in reporting.

Running GLM5.2 on Five RTX Pro 6000s — LocalLLaMA's "Expensive Journey"

A user in Reddit's LocalLLaMA community has posted a full rundown of a build that runs GLM5.2 on five RTX Pro 6000s and one RTX 5090.

The build started with two 5090s, but PCIe bandwidth constraints and VRAM shortages led to GPU after GPU being added. It eventually landed on a Threadripper Pro 9975WX paired with a WRX90 motherboard carrying five Pro 6000s.

The result is a satisfaction level of "98–99%" — alongside self-deprecating remarks like "it won't pay for itself unless I run it for over 10 years" and "please, whatever you do, don't copy this." The post chronicles struggles unique to local-inference diehards, from GPU heat problems to additional power supplies.

Also on LocalLLaMA, a user running DeepSeek V4 Pro on an RTX PRO 6000 Max-Q with an Epyc 9374F reported hitting 192 tokens/s prompt processing and 11.7 tokens/s generation with a custom optimized branch of llama.cpp. As a fresh benchmark of a large-scale build handling a 794GB GGUF, it is well worth a look.

Surface Evolver Bench — A New Benchmark That Tests LLMs Through Physics Simulation

A unique LLM benchmark called "Surface Evolver Bench" has been announced on Reddit. The task: have LLMs write data files for Surface Evolver, a liquid surface modeling tool released in 1992. The design evaluates a natural agent loop (consult the documentation → implement → run → debug).

GPT5.5 took the top spot overall, while GLM5.2 posted the highest score among open models. It is a meaningful benchmark that puts both coding ability and iterative improvement to the test in the specialized domain of physics simulation.

---