Today's Highlights
- OpenAI lands first $11B at $1.25T valuation
- Anthropic IPO expectations cut below $1.5T
- Nasdaq enters correction; Nvidia drops 9.3%
- Anthropic unveils agent memory with "dreaming"
- Claude Code leaks 2,539 secrets to public GitHub
Headlines
[OpenAI closes $11B first tranche at $1.25T valuation] OpenAI closed the first $11 billion of a new round valuing it at $1.25 trillion on November 6, Bloomberg reports, with SoftBank, MGX, Coatue and T. Rowe Price in. The round may still reach $20 billion but falls short of the original $30B-at-$1.4T target, a sign of cooling sentiment after last month's AI rout. (Bloomberg/Reuters)
[Anthropic IPO valuation hopes slashed] Anthropic filed its S-1 for a Nasdaq debut targeted for the week of November 16 at ~$2 trillion, but investors have cut expectations below $1.5 trillion and Reuters says the roadshow timing is in question. Its S-1/A shows ~$895.4B in compute commitments and Q3 revenue of $16.96B with growth slowing from 144% to 47%. Its pricing will anchor sector valuations. (TMTPost/Reuters)
[US AI stocks enter correction] The Nasdaq fell ~5.4% on November 5, over 10% below its October 6 record — a formal correction. Nvidia dropped 9.3%; Philadelphia semis fell over 12% intraday. The trigger: a Democratic midterm sweep raising fears of tighter AI regulation and slower datacenter buildout, with top-rated datacenter bond yields up from under 5% to ~7% this year. (TMTPost)
Products & Tech
[Anthropic details three-layer agent memory with "dreaming"] A November 6 research post splits agent memory into three layers: salient notes written during sessions, salience judged by stronger models like Opus 5.5, and idle-time "dreaming" that consolidates knowledge — the first public description of Anthropic's sleep-time compute, improving long-horizon task performance. (Anthropic blog)
[Claude Code's "Don't Ask Again" leaked 2,539 secrets] Apiiro found Claude Code's "Yes, and don't ask again" option saves full approved commands — keys and tokens included — in plaintext to .claude/settings.local.json, which is not gitignored. Of 104,185 public copies on GitHub, 1,122 held real credentials: 2,539 leaked secrets. (The Hacker News)
[Google ADK 2.14 adds Teams integration] Released November 7, ADK 2.14.0 graduates 26 core modules and 16 toolset integrations, adds a TeamsRunner for Microsoft Teams and MultiRunConfig eval replays; long multi-turn workflows run 23% faster with 9% lower peak memory. (GitHub)
[Meta gives away Muse proactive memory] On November 6 Meta made Muse Wonder's "proactive memory" and "proactive focus" free for all users: 2.4x less forgetting of target info, 67% higher task completion under interruptions — locking in users while amplifying privacy risks. (TMTPost)
IPO & Fundings
[Meta prices $25B bond, AI risk premium widens] Meta priced ~$25 billion of dollar and euro bonds in early November; the dollar tranche drew ~$58B of orders, nearly 4x covered, but the longest note's spread jumped 73bp to 220. Q3 free cash flow was -$2.12B, its first negative since 2012, and S&P cut its outlook. (TMTPost)
[Chip shortage is back] Microchip fell 10% on November 7 after CEO Steve Sanghi said AI buyers "came in and took over the whole factory" at shared assembly houses, stretching industry lead times to 32 weeks and squeezing out auto and industrial customers. (247wallst)
Open Source & Research
[Recursive open-sources Ripple 65B] On November 6 chip-design lab Recursive released Ripple 65B, a dense 65B-parameter code model under Apache 2.0, after $32M of training spend. (AIFOD)
[Moonshot's Kimi K3.1 Pro Highspeed lands on AWS Bedrock] AWS added Moonshot AI's open-weight Kimi K3.1 Pro Highspeed to Amazon Bedrock on November 6, bringing a leading Chinese open-weight model into a global hyperscaler's managed catalog. (AIFOD)
[Thinking Machines unveils Inkling 2 Small] On November 5 Mira Murati's lab introduced Inkling 2 Small: 594B parameters trained on 17M agent rollouts, extending its open-weight-plus-customization strategy. (Medium)
Company & People
[Microsoft's AI Frontiers group shut down] On November 6 John Langford confirmed the ~50-person AI Frontiers team he co-founded — behind Autogen and the Fara computer-use agent — is "finished" after being folded into Microsoft AI in September. (Hunch.net)
[Anthropic-backed AI safety PAC wins 10 of 17 bets] Public First spent over $63M in the midterms; 10 of its 17 endorsed candidates won. Anthropic gave ~$20M-plus and CEO Dario Amodei $1M personally; the opposing Leading the Future PAC spent ~$100M with weaker results. (Bloomberg)
Policies & Data
[ECB weighs review of banks' AI datacenter loans] Bloomberg reported November 6 that the eurozone's top banking supervisor is considering a review — possibly with on-site inspections — of banks' AI datacenter lending, focused on cash-flow and collateral assumptions; Morgan Stanley sees ~$50B of securitized datacenter credit and a $1.5T funding gap. (Bloomberg/TMTPost)
[China: ~600 TWh of new annual power demand] At a November 6 NDRC briefing, the National Energy Administration said China will add ~600 billion kWh of demand per year in the coming years — three Shanghai-sized loads annually — with AI datacenters a core driver. (MyDrivers/Sina Tech)
Powered by AOYII.com
