# AI News Digest - 2026-03-01

1. OpenAI signed a deal with the Pentagon to provide classified AI networks shortly after Anthropic was barred from federal agencies; Anthropic was labeled a supply chain risk by the Pentagon and said it would challenge that designation in court.

2. OpenAI promised Canada tighter safety protocols after ChatGPT flagged a shooter's violent chats but did not notify police; the company had blocked the suspect's account but failed to inform authorities.

3. Perplexity open-sourced two new text embedding models that matched or exceeded Google's and Alibaba's offerings while operating at a fraction of the usual memory cost.

4. Researchers at Apple, Stanford, and the University of Washington found that common HTML extractors pulled substantially different content from the same web pages, causing large portions of the internet to be omitted from language model training data.

5. The Decoder reported that frontier LLMs, including models such as GPT‑5.2 and Claude 4.6, lost up to 33% accuracy over the course of long conversations.

# References

1. [https://the-decoder.com/openai-signs-pentagon-deal-for-classified-ai-networks-hours-after-anthropic-gets-banned-from-federal-agencies/](https://the-decoder.com/openai-signs-pentagon-deal-for-classified-ai-networks-hours-after-anthropic-gets-banned-from-federal-agencies/)

[OpenAI signs Pentagon deal for classified AI networks hours after Anthropic gets banned from federal agencies](https://the-decoder.com/openai-signs-pentagon-deal-for-classified-ai-networks-hours-after-anthropic-gets-banned-from-federal-agencies/)

1. [https://the-decoder.com/openai-promises-canada-tighter-safety-protocols-after-chatgpt-flagged-a-shooters-violent-chats-but-never-called-police/](https://the-decoder.com/openai-promises-canada-tighter-safety-protocols-after-chatgpt-flagged-a-shooters-violent-chats-but-never-called-police/)

[OpenAI promises Canada tighter safety protocols after ChatGPT flagged a shooter's violent chats but never called police](https://the-decoder.com/openai-promises-canada-tighter-safety-protocols-after-chatgpt-flagged-a-shooters-violent-chats-but-never-called-police/)

1. [https://the-decoder.com/perplexity-open-sources-embedding-models-that-match-google-and-alibaba-at-a-fraction-of-the-memory-cost/](https://the-decoder.com/perplexity-open-sources-embedding-models-that-match-google-and-alibaba-at-a-fraction-of-the-memory-cost/)

[Perplexity open-sources embedding models that match Google and Alibaba at a fraction of the memory cost](https://the-decoder.com/perplexity-open-sources-embedding-models-that-match-google-and-alibaba-at-a-fraction-of-the-memory-cost/)

1. [https://the-decoder.com/current-language-model-training-leaves-large-parts-of-the-internet-on-the-table/](https://the-decoder.com/current-language-model-training-leaves-large-parts-of-the-internet-on-the-table/)

[Current language model training leaves large parts of the internet on the table](https://the-decoder.com/current-language-model-training-leaves-large-parts-of-the-internet-on-the-table/)

1. [https://the-decoder.com/even-frontier-llms-from-gpt-5-onward-lose-up-to-33-accuracy-when-you-chat-too-long/](https://the-decoder.com/even-frontier-llms-from-gpt-5-onward-lose-up-to-33-accuracy-when-you-chat-too-long/)

[Even frontier LLMs from GPT-5 onward lose up to 33% accuracy when you chat too long](https://the-decoder.com/even-frontier-llms-from-gpt-5-onward-lose-up-to-33-accuracy-when-you-chat-too-long/)

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