What changed
- Anthropic raised $50B at a $900B pre-money valuation with ~$45B ARR (FT, May 8). Anthropic Snapshot updated.
- Anthropic-Akamai $1.8B seven-year compute deal added. Anthropic
- Jack Clark's Import AI 455 assigns 60%+ probability to recursive AI R&D by end-2028. Jack Clark reinforced; AGI Timelines gets an Aggressive-2027 row.
- OpenAI launched Trusted Contact (consumer-AI mental-health alerting) and GPT-5.5-Cyber (limited preview to critical-infrastructure defenders). OpenAI
- LIFT AI Act introduced (Schiff D-CA / Rounds R-SD, May 4 — bipartisan K-12 AI literacy bill, OpenAI/Google/Microsoft endorsed). New page: LIFT AI Act (Literacy in Future Technologies Artificial Intelligence Act).
- IMF financial-stability blog named Claude Mythos Preview and GPT-5.5-Cyber as macro-financial risk vectors — first IMF designation of frontier models as systemic-financial risk.
- Apple settled the Siri AI class action for $250M + Apple-Intel preliminary chip-making agreement. Apple
- Isomorphic Labs raised $2B+ from Thrive Capital + Alphabet. Isomorphic Labs
- Microsoft 2018 internal emails released in Musk v. Altman (and OpenAI / Microsoft / Brockman) — Nadella/Zander/Scott/Horvitz pre-investment skepticism.
- Anthropic Opus 4.7 prompting guide (31 pages) surfaced — documents a sharp behavioral shift (literal-instruction-following, fewer emojis, fewer tool calls). Claude Opus 4.7
- Epoch AI median estimate: 660,000 H100-equivalents smuggled into China through 2025 (range 290K–1.6M). Chip Smuggling and Export-Control Evasion gets first numeric end-to-end estimate.
What it connects to
The dominant thread is Anthropic's commercial-frontier divergence from the political-frontier: a $50B raise + $1.8B Akamai deal + $200B Google-cloud commitment + Anthropic-SpaceX Colossus 1 deal land in the same 30-day window where Anthropic is excluded from the Pentagon's classified-network cohort and where the Trump admin is reportedly drafting two parallel pre-release-vetting EOs (one Anthropic-friendly broad model-review track, one cybersecurity-only track that omits mandatory testing — see AI Pre-Release Vetting subsection 1.5). The May 8 Bloomberg report about the narrower track contradicts the May 5 Politico framing of the broader EO; the wiki now flags both. This is a live tension worth tracking — it instances the broader pattern in AI Safety Frameworks Compared: RSP, Preparedness, NIST RMF, and Safety Cases of capability-threshold-vs-procurement-gate as competing governance modes.
Jack Clark's Import AI 455 strengthens the Recursive Self-Improvement (RSI) → AGI Timelines chain (Aggressive-2027 row) but more importantly it's the first piece of insider frontier-lab forecasting since Situational Awareness: The Decade Ahead to use first-party Anthropic capability data rather than scaling-law extrapolation. It belongs in conversation with AI Pre-Release Vetting because if a 60%+ probability of recursive AI R&D by end-2028 is correct, the procurement-driven vetting model is on a much shorter clock than its proponents assume.
The IMF Mythos designation is a category change for Claude Mythos Preview: until now Mythos was discussed as a safety-policy object (will it trigger an EO? will it be deployed in the Pentagon?). The IMF designation makes it a macro-financial-risk object — the same conceptual move that took CDOs from "obscure financial instrument" to "stress-test category" after 2008. This connects to AI and Cybersecurity and to the Mozilla-Firefox-271-bugs item from the May 8 cycle in a way the wiki has not yet fully exploited.
One question worth sitting with
If the Trump admin ships two AI executive orders in the next 90 days — a broad model-review EO that resembles a "FDA for frontier models" and a narrower cybersecurity-only EO that omits mandatory pre-release testing — which one is the real policy and which one is theater? The wiki currently has evidence pulling both ways (AI Pre-Release Vetting sources_count = 11) but no clean read on the political economy of which constituency wins when the two tracks conflict.
Predictions touched
- Aschenbrenner Situational Awareness timeline (predicts AGI by 2027): reinforced by Clark's Import AI 455 (60%+ recursive-AI by end-2028 is broadly consistent). Status: ACTIVE, no resolution date passed.
- Anthropic RSP v3.1 capability thresholds as the de-facto pre-deployment standard: weakened by Bloomberg May 8 report of narrower cybersecurity-only EO that omits mandatory testing. Status: ACTIVE, contested-leaning. Watch the next 90 days.
- (No predictions resolved this cycle.
analysis/track-record.mdnot yet populated — it will fill as## Predictionssections get added to high-traffic pages and resolution dates pass.)