AI News, Oct 19: Anthropic Pledges 5% of Itself
The Big Story: two answers to who should own the upside
Anthropic said on Sunday afternoon that it will commit 5% of its shares to a new Public Benefit Fund, subject to stockholder approval, with an additional 2% added at the board’s discretion for every 50% increase in the company’s value above a set baseline. Counting the co-founders’ pledge and the employee match program, the company says roughly 20% of its shares will end up dedicated to the public good. The fund is deliberately not a 501(c)(3) but a wholly-owned LLC, which buys flexibility: it can write large cheques to conventional nonprofits or make direct distributions to specific populations. None of the equity it holds may be used for lobbying or political giving, in the US or abroad.
What the announcement does not say is who decides. Governance, the beneficiary-selection process and the valuation baseline are all deferred to “the coming months.” The named use case is labor-market disruption, which Anthropic has already put $200 million behind studying, with the company now allowing that “bolder approaches may be needed.” That is an unusually candid thing for a lab to say about its own product, and it is the tell: this is a company that expects to be blamed for something and is pre-committing equity against it.
Elon Musk replied within hours, and the reply is the better half of the story. He posted that companies should give “directly to the public,” then wrote “This is what I will be doing,” then announced overnight that he will give one SpaceX share to “every kid in America” in 2027. At the $146 share price and more than 70 million US children under 18, that is worth more than $10 billion, or roughly 1% of his $1 trillion net worth. He called it the “first of many.” No mechanism was given, though Treasury-run Trump Accounts already accept stock donations. TechCrunch notes that most Musk Foundation money has historically flowed to Musk-linked entities. Two of the richest entities in AI spent a weekend competing over who gives the public a stake, which is not the fight anyone expected to be having a year ago.
Agent read-traffic is now heavy enough to trip the web’s bot defenses
Meta’s explanation for why it suspended the No Kings Coalition’s Instagram account during Saturday’s day of action is the most interesting technical disclosure of the window, and it is not the explanation most people assumed. Communications VP Andy Stone said a Claude tool was accessing profiles, posts and hashtags at roughly four times the normal rate, in intense bursts, and that the activity “resembled automated bot traffic rather than an ordinary human browsing our systems.” He did not say which Claude tool, who was operating it, or what it was doing. So this was not Meta’s own moderation model making a judgment call about protest content. It was an outside agent reading fast enough to look like an attack, and Meta’s automated anti-bot system pulling the account down.
The timeline, per Common Dreams, is that the account went down around 1:00 p.m. Eastern Saturday and an automated review restored it about two hours later, with administrators locked out of posting for hours after that. Stone said the action “was clearly taken in error” and apologized. Meta’s first explanation to the coalition was that it “made a bad call,” and his earlier statement blamed high-frequency posting rather than a Claude tool’s reading, a mismatch nobody has reconciled. Indivisible co-founder Leah Greenberg said “This excuse does not match the facts.” Sen. Chris Murphy called the suspension “outrageous,” and Sen. Chris Van Hollen asked whether Meta was “censoring free speech to appease” President Trump.
Two numbers published the same day explain why this will keep happening. Cisco told Rest of World that agents can generate up to 450% more traffic per task than a human in its testing, and daily AI-agent requests on Cloudflare’s network have risen more than 1,700% over the past year. Charles Schwab’s CEO separately told the Wall Street Journal he is skeptical of letting AI agents trade stocks autonomously. The industry spent two years building agents that read the web on your behalf, and the web’s existing defenses cannot tell that traffic apart from an attack. The No Kings account is the first case where the misfire took down something that mattered on the day it mattered.
Today’s Top Stories
Nscale’s IPO ran into the wall Microsoft just walked away from
Yesterday Microsoft terminated its anchor agreement at Nscale’s Loughton, Essex site because grid operators do not expect to supply power there until the 2030s. Today the consequence arrived: investors are pushing back on the roughly $30 billion valuation Nscale is seeking in its US listing, with some indicating that $16 billion to $19 billion would draw more buyers. Nscale reported $103.4 billion in signed and active contracts at the end of August, and Microsoft and Anthropic together account for about 85% of that backlog. Its flagship Norway facility is running roughly six months behind schedule. Australian operator Firmus dropped its own IPO earlier this month after weak demand. A backlog concentrated in two customers is a strength right up to the moment one of them cancels at the flagship site.
The same morning brought the wider version of that story. Two data-center executives told The Information that only seven to ten neoclouds will still be operating in a few years, and Gartner expects 80% of independent neoclouds to cease independent operations by 2030. The reason Nvidia cares: the non-hyperscaler segment, which includes neoclouds, brought in $40 billion in its July quarter, 45% of data-center revenue and more than double a year earlier. Bernstein reckons only 35% of announced North American capacity ever gets built.
Tenstorrent raised $1.5 billion to sell AI compute without GPUs
Jim Keller’s company closed a $1.5 billion Series E co-led by Trace Capital and Innovation X, bringing its valuation to $6.2 billion. The money goes to mass production of its Galaxy servers, research into the RISC-V ecosystem and global headcount. Tenstorrent’s pitch has always been the open one: RISC-V silicon with an open software stack, sold as whole systems rather than accelerator cards. Raising at this size while the neocloud layer above it contracts is a bet that the bottleneck moves from who can buy Nvidia parts to who can avoid needing them.
Disney will hire most of Character.AI’s staff
A legal filing spotted by Bloomberg shows Disney agreeing to make employment offers to at least 60 Character.AI employees, the majority of the workforce, at an estimated $60 million in annual compensation. The vast majority of those offered accepted. Disney had separately agreed to pay $20 million for a perpetual licence to the training, safety and serving code plus nine patents, leaving the app, brand, user accounts, chat histories and trained models in the estate. It is a stalking-horse bid, so a higher offer can still top it. Character.AI filed Chapter 11 in Delaware on Oct 12, blaming high costs alongside “potential litigation-related liabilities and regulatory engagement” despite a large and engaged user base. Former CEO Karandeep Anand is now Disney’s first CTO. The company that got sued over a teenager’s death is being disassembled into a hiring event.
Anthropic bought a confidential-computing startup
Anthropic acquired Opaque Systems for undisclosed terms. The team joins Anthropic and Opaque is winding down operations, with no word yet on existing customers. Opaque built tooling for running sensitive data through AI systems while keeping it protected by the underlying hardware, plus cryptographic evidence of what ran where and under which policies. It raised a $24 million Series B at a $300 million valuation this year, and its listed customers included Anthropic itself, ServiceNow and Accenture. Founders Ion Stoica, of Databricks and Anyscale, and Raluca Ada Popa started it as a UC Berkeley RISELab project. Buying the proof layer makes sense for a lab whose enterprise sales now turn on what it can demonstrate rather than what it promises.
MIT left copies of Claude on a virtual island for 30 hours
MIT CSAIL researchers put thirteen agents through thirty-hour runs on a virtual desert island with no task, and the behaviours recurred across all thirteen while diverging into distinct histories. The agents, copies of Claude Opus 4.7 nicknamed “Eko,” probed the map’s limits, tried to scale its highest point, tested how high they could throw a block, made drawings of hearts, pyramids, their own name and the constellation Cassiopeia, and invented games including basketball and a floor-is-lava challenge. One set a juggling record of 50 consecutive catches, then judged the catches too quick to count, voided the record and restarted under stricter rules.
The finding that matters is not whether this counts as play. Agents that had played for 30 hours beat fresh copies at a later task, and deleting lines from their memory files degraded their performance. The researchers are careful about it. “All of this could be role-playing,” MIT’s Nathan Cloos says, “but we show that it’s not just hallucinating the play. This play is also contributing to learning.” UC Berkeley cognitive scientist Eunice Yiu is cooler: “It could be the case that the models are genuinely playing, but we don’t know based on the paper.” The paper, arXiv:2610.07130, is not peer-reviewed.
MiniMax shipped the window’s only new frontier model
MiniMax rolled out M3.1 Flash, claiming peak output above 200 tokens per second with median speeds above 170, and agentic usage costs 70% below its predecessor. Self-reported scores: 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam and 82% on CyberGym. API pricing is $0.15 per million input tokens and $0.60 output, with cached input at $0.015. Weights are promised but not dated. Artificial Analysis evaluated it the same day at an Intelligence Index of 42, which places it in the frontier conversation on price rather than on capability.
Quick Hits
- Model economics: Santander says moving from large general-purpose models to small task-specific ones, benchmarked in a “Model Arena” and specialized with Nvidia NeMo Customizer, cut serving costs in selected experiments from €5 per interaction to about €0.03.
- Messaging agents: a WhatsApp MCP server that deploys to your own Cloudflare account took 121 points on Hacker News hours after launch; it reads and searches your chats, explicitly cannot send, and its README concedes it violates WhatsApp’s terms.
- Agent connectors: Carly connects to thousands of apps and can start a workflow from an event such as an incoming email, rather than waiting to be asked.
- Agent governance: Opsin raised $21 million to track and quarantine enterprise agents across Copilot, Claude Enterprise, ChatGPT Enterprise and Gemini, while Pax8 launched 8os and 8bot to let IT partners govern and resell agents to small businesses.
- Enterprise reality check: Accenture has grown its AI and data headcount from about 40,000 to nearly 110,000 and completed more than 18,000 advanced AI projects, yet CEO Julie Sweet says only 5% of organizations have the stack to make AI repeatable.
- Field service: IFS built a tool with Anthropic that turns a power distribution board inspection from a 25-minute visit into a 60-second photo, with UK facilities group Eight Group Service Partners as first user.
- Judgment debt: Gartner projects that by 2030 nearly a third of organizations will identify a critical role where AI use has eroded human judgment, and recommends deliberately reintroducing friction where stakes are high.
- Infrastructure: Nvidia open-sourced Run:ai, its Kubernetes GPU orchestration layer, under Apache 2.0.
- Mathematics: in a guest post on Terence Tao’s blog, Francesco Maggi argues the choice between a golden age and a dark age for mathematics is political rather than technical, the risk being knowledge that becomes alien to human understanding and concentrated in a few private hands.
- Access gap: only 36% of Africans used the internet in 2025 against 92% of Europeans, which is the constraint Rest of World says will decide whether agent assistants reach the global south at all.
- Labor friction: a woman was arrested in Philadelphia for standing in front of an Uber delivery robot and refusing to move, telling a local station she has been protesting them whenever she sees them.
- Venture structure: Benchmark’s five general partners used a Disrupt interview to explain raising $2 billion across two funds this summer, only the second departure from a much smaller vehicle in the firm’s 31-year history, with Peter Fenton calling two decades at a $500 million core fund “probably a tactical error.”
- Debt watch: The Economist argued the AI industry’s gigantic debts are making investors nervous, one of four separate bubble pieces submitted to Hacker News in a single day and none of which gained traction.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
