Best of your X follows: Mythos returns, scrolls unroll, and DSpark speeds up LLMs

Coverage note

Model access and governance

Anthropic: Mythos 5 gets a narrow U.S. infrastructure lane

• What happened: Anthropic said the U.S. government allowed Claude Mythos 5 to be redeployed to a set of U.S. organizations that operate and defend critical infrastructure. 1
• Why it matters: This is not a full reversal: Anthropic said it is still working with the government to expand Mythos 5 access and make Fable 5 generally available again. 1
• Source / signal: X post by Anthropic, published Jun 27 at 00:29 UTC; at capture it showed 26,028 likes, 2,760 reposts, and 1,350 quote posts. 1

Dean Ball: frontier model releases are drifting into preapproval

• What happened: In a Jun 26 essay, Dean W. Ball argued that U.S. frontier-model policy has become a de facto licensing and preapproval regime, with no clear safety standard for broad release. 2
• Why it matters: Ball ties the access issue to economics: frontier labs recoup training costs in the first post-release months, while the AI infrastructure buildout assumes a global market for frontier services. 2
• Source / signal: Simon Willison highlighted the essay on Jun 26 at 22:25 UTC, and OpenAI's GPT-5.6 preview separately says its initial partner-limited release came at the U.S. government's request. 23

Research and model infrastructure

Vesuvius Challenge: one Herculaneum scroll is now readable end to end

• What happened: Ethan Mollick pointed to a recovered passage from PHerc. 1667, the first Herculaneum papyrus to be digitally unrolled and read continuously from beginning to end. 45
• Why it matters: The project combined high-resolution X-ray microtomography, surface reconstruction, and machine-learning ink detection; the released text is a fragmentary Stoic ethics treatise with about twenty-two columns of Greek recovered from roughly 1.4 metres of papyrus. 5
• Source / signal: Mollick's X post quoted one recovered passage; the official announcement says the data, code, and preprint are open for others to inspect and build on. 45

DeepSeek: DSpark pushes speculative decoding into production traffic

• What happened: DeepSeek's DSpark paper proposes confidence-scheduled speculative decoding: a semi-autoregressive draft model plus adaptive verification length for each request. 6
• Why it matters: In DeepSeek-V4 live serving, the authors report 60% to 85% faster per-user generation speeds at matched throughput compared with their MTP-1 baseline. 6
• Source / signal: Hacker News surfaced the paper on Jun 27; at capture it had 651 points and 248 comments, while DeepSeek open-sourced the DSpark checkpoints inside the DeepSpec repository. 67

Open models and security

Doubleword: the open-vs-closed model gap depends heavily on the benchmark

• What happened: Doubleword re-ran the open-weights versus closed-model lag question across 18 Artificial Analysis benchmark datasets instead of leaning on one headline index. 8
• Why it matters: A single index projects the open frontier catching closed models around Dec 3, 2026, but the 18-benchmark view shows the average gap nearly flat at just under five months; coding is where the gap has compressed most. 8
• Source / signal: Hacker News carried the piece within the window; at capture it showed 286 points and 217 comments. 9

Exploitarium: a public PoC dump becomes a security-watch item

• What happened: A GitHub repository named Exploitarium describes itself as an archive of public proof-of-concept and vulnerability research writeups, including several direct entries dated Jun 23 to Jun 26. 10
• Why it matters: The repo's own About text says the material had not been reported at posting time; for defenders, that makes it a triage signal rather than a curiosity link. 10
• Source / signal: Hacker News framed it as an anonymous GitHub account mass-dropping undisclosed 0-days; at capture it had 285 points and 118 comments. 11

What to actually read first