# AI News Digest - 2026-05-25

1. Researchers from the University of Maryland, Google, Meta, and other institutions used Claude Code to discover an AI scaling control algorithm that cut compute by about 70% compared with standard self-consistency while matching accuracy, and the search reportedly cost $40 and took 160 minutes.

2. ByteDance reported that its Seed study found a 7B long-multimodal model answered questions on long, image-heavy documents more reliably than much larger models and learned to locate relevant passages by answering questions rather than transcribing pages, even for documents four times longer than those seen during training.

3. Anthropic was reported to be likely to continue supplying its Claude models to the NSA despite being flagged as a supply-chain risk by the Pentagon, with coverage noting intelligence agencies lacked Nvidia's latest Grace Blackwell chips and that an "any lawful use" clause was not part of the deal.

4. Microsoft Copilot was reported to produce spurious country-specific differences when given identical datasets with different country labels, delivering detailed stereotypes instead of accurate results, prompting researchers to advise against leaving model selection on default in Copilot, Gemini, and other AI tools.

5. Deepseek made a 75 percent discount on its V4‑Pro model permanent, pricing input tokens at $0.435 per million and reporting output-token pricing at least 34 times cheaper than GPT‑5.5, which created a substantial pricing gap for token-intensive agent systems.

# References

1. [https://the-decoder.com/researchers-let-claude-code-discover-ai-scaling-algorithms-that-humans-probably-wouldnt-have-designed/](https://the-decoder.com/researchers-let-claude-code-discover-ai-scaling-algorithms-that-humans-probably-wouldnt-have-designed/)

[Researchers let Claude Code discover AI scaling algorithms that humans probably wouldn't have designed](https://the-decoder.com/researchers-let-claude-code-discover-ai-scaling-algorithms-that-humans-probably-wouldnt-have-designed/)

1. [https://the-decoder.com/bytedance-study-finds-that-asking-lmms-questions-beats-making-it-transcribe-text-for-long-document-training/](https://the-decoder.com/bytedance-study-finds-that-asking-lmms-questions-beats-making-it-transcribe-text-for-long-document-training/)

[ByteDance study finds that asking LMMs questions beats making it transcribe text for long document training](https://the-decoder.com/bytedance-study-finds-that-asking-lmms-questions-beats-making-it-transcribe-text-for-long-document-training/)

1. [https://the-decoder.com/anthropic-may-keep-supplying-claude-to-the-nsa-despite-being-flagged-as-a-supply-chain-risk-by-the-pentagon/](https://the-decoder.com/anthropic-may-keep-supplying-claude-to-the-nsa-despite-being-flagged-as-a-supply-chain-risk-by-the-pentagon/)

[Anthropic may keep supplying Claude to the NSA despite being flagged as a supply chain risk by the Pentagon](https://the-decoder.com/anthropic-may-keep-supplying-claude-to-the-nsa-despite-being-flagged-as-a-supply-chain-risk-by-the-pentagon/)

1. [https://the-decoder.com/why-you-shouldnt-leave-model-selection-on-default-in-copilot-gemini-and-other-ai-tools/](https://the-decoder.com/why-you-shouldnt-leave-model-selection-on-default-in-copilot-gemini-and-other-ai-tools/)

[Why you shouldn't leave model selection on default in Copilot, Gemini and other AI tools](https://the-decoder.com/why-you-shouldnt-leave-model-selection-on-default-in-copilot-gemini-and-other-ai-tools/)

1. [https://the-decoder.com/deepseek-makes-its-75-percent-discount-permanent-pricing-output-tokens-at-least-34x-below-gpt-5-5/](https://the-decoder.com/deepseek-makes-its-75-percent-discount-permanent-pricing-output-tokens-at-least-34x-below-gpt-5-5/)

[Deepseek makes its 75 percent discount permanent, pricing output tokens at least 34x below GPT-5.5](https://the-decoder.com/deepseek-makes-its-75-percent-discount-permanent-pricing-output-tokens-at-least-34x-below-gpt-5-5/)

For the site tree, see the [root Markdown](https://ixtj.dev/.md).