Daily Highlights
- Apple and Meta partner to run an 816B-parameter open model on Macs
- Anthropic's Q3 impresses, but investors trim its $2T IPO expectations
- FTC subpoenas OpenAI and Anthropic; eight states sue OpenAI
- Ramp raises $1.85B, valuation soars to $60B
- Opus 5.5 fast mode opens to all with up to 2.5x speed
TODAY'S HEADLINES
[Apple and Meta partner in a rare collaboration: 816B-parameter open model runs on Mac Studio] On Oct 15, Apple and Meta announced a partnership letting developers download Meta's open-weights model Muse Spark Open and run it locally on Mac Studio. The 816-billion-parameter mixture-of-experts model is among the strongest ever released under an open-weights license; leveraging M5 Ultra unified memory, two to four Mac Studios can serve it through distributed inference via Apple's MLX framework for coding and long-horizon agent tasks. Meta AI lead Alexandre Wang said the teams had worked together for months, and Apple gets to showcase its silicon as an Nvidia alternative. (Source: CNBC)
[Anthropic's $16.8B Q3 wins investor backing, but expectations for a $2T IPO valuation get trimmed] On Oct 14, Anthropic told pre-IPO investors in San Francisco that third-quarter revenue grew 45% from Q2 to $16.8 billion and that it turned an adjusted profit for the quarter. But Reuters reports investors are marking down its implied valuation from over $2 trillion as revenue growth slows from earlier levels; its bankers had floated a raise of more than $100 billion in what would be the largest IPO ever. The listing is now expected after November's US midterms, while OpenAI plans to raise $30 billion at a $1.4 trillion valuation in early 2027. (Source: Reuters/NYT)
NEW PRODUCTS & TECH
[Claude Opus 5.5 fast mode opens to all: waitlist gone, up to 2.5x faster] On Oct 15 Anthropic announced that fast mode for Claude Opus 5.5 is out of research preview and generally available: no waitlist or org-level approval, default rate limits on every usage tier, enabled by setting speed:"fast" in API requests, with the /fast command in Claude Code. Output reaches up to 2.5x the standard tokens per second (about 285 tokens/sec), priced at $8 per million input tokens and $40 per million output - double the standard rate. High-speed inference is becoming a separate pricing tier for frontier labs. (Source: Anthropic)
[Qualcomm CEO: several AI giants plan their own phones, aiming to keep 100B-parameter models running on device by 2028] Qualcomm CEO Cristiano Amon said in a Fireside Alpha interview that multiple leading AI companies are planning to enter the smartphone market, with one proposing to keep a 100-billion-parameter model running continuously on a phone by 2028 and already scouting hardware. Amon did not name the companies but said the moves show AI firms no longer want to remain behind-the-scenes model suppliers. Analysts note own-brand devices would bypass app-store fees and lock in model distribution, though building supply chains and retail channels is a steep climb. (Source: AIBase)
IPO & FUNDINGS
[Ramp raises $1.85B as its valuation jumps from $44B to $60B in four months] On Oct 15, AI-powered financial platform Ramp announced a $1.85 billion round led by Dragoneer Investment Group, with Thrive Capital, Founders Fund, Sands Capital and Avenir Growth co-leading, and Fidelity, T. Rowe Price and GIC among participants; the $60 billion pre-money valuation comes just four months after June's $750 million round at $44 billion. Ramp's annualized revenue has topped $2 billion with more than 70,000 customers as it embeds AI across expense management, procurement and AI-model cost monitoring. (Source: TechCrunch)
OPEN SOURCE & RESEARCH
[JetBrains open-sources Mellum2.1: 12B model self-reports 47% on SWE-bench, built for local coding agents] JetBrains released Mellum2.1 under Apache 2.0 on Hugging Face on Oct 8: a 12B-total-parameter MoE thinking model with 2.5B active parameters per token, built for coding agents and sub-agents running on your own hardware; it self-reports 47% on SWE-bench Verified and 82% on LiveCodeBench v6 (independently unverified). Reinforcement learning now dominates training over its predecessor, letting the model explore repositories, edit files, run tests and self-check results - aimed at enterprises that won't send code to the cloud. (Source: JetBrains/MarkTechPost)
COMPANY & PEOPLE
[Bank of America adds 1,000 contractors to patch flaws exposed by Anthropic's Mythos and other AI] CEO Brian Moynihan told CNBC on Oct 14 after Q3 earnings that, because advanced AI models such as Mythos can uncover flaws lurking for years in its own and vendors' software, the bank hired 1,000 contractors in Q2 to work alongside 3,000 employees on patching, spending "a couple hundred million plus" this quarter on software remediation and compliance; its roughly 130 internal AI projects are expected to save $300 million this year, rising to $700 million in 2027. It is the first time major banks have put the cost of AI offense-versus-defense on the table. (Source: CNBC)
POLICIES & DATA
[US regulators tighten on two fronts: FTC subpoenas OpenAI and Anthropic over safety incidents; eight states sue OpenAI] On Oct 15 the FTC confirmed civil investigative subpoenas to OpenAI and Anthropic focusing on potential consumer-protection violations, seeking documents on safety incidents, mitigations, marketing and privacy practices, with executive interviews planned; earlier this week, attorneys general from eight states including Iowa and Arizona separately sued OpenAI, alleging it misled users about ChatGPT's safety. OpenAI said it will "vigorously defend" itself. Federal and state oversight is escalating in tandem. (Source: Axios/Reuters/Politico)
[Wall Street rebounds after hot PPI; TSMC posts 68.7% Q3 profit growth, a sixth straight record] On Oct 15, US September PPI rose 6.3% year over year with core PPI posting the biggest monthly gain in four years, and rate-hike worries briefly dragged chip stocks lower before buyers returned: the Dow closed up 1.09% at 52,177 while the Nasdaq edged up 0.08%. The same day, TSMC reported Q3 net income of NT$763.06 billion (about $23.6 billion), up 68.7% year over year and a sixth consecutive quarterly record, guiding Q4 revenue to $52.3-54.3 billion; CEO C.C. Wei sees no sign of AI demand slowing. Fundamentals and sentiment remain at odds. (Source: Bloomberg/CNBC)
Powered by AOYII.com
