In partnership with
In Todayโs Issue:
๐ธ Anthropic puts 5% of its shares into a public benefit fund
๐ Open models trail the cyber frontier by six to seven months
๐ Google's Harnessยฒ lets AI agents tune their own setup
๐ณ DeepSeek's Flash update takes a big leap
๐ Why the CoreWeave panic looks overdone
โจ And more AI goodnessโฆ
โก The Signal
AI's boom has a new money question: who gets the wealth.
Anthropic is committing 5% of its shares, worth up to $100 billion at its expected IPO valuation, to a fund that could pay people hit by AI directly. It announced the fund a day after co-founder Jack Clark told IMF officials that the evidence points to a singularity. Markets are asking the reverse question: CoreWeave's bond rout is a fight over who pays for the data centers, and at what interest rate. Meanwhile the capability itself keeps getting cheaper. DeepSeek's newest Flash model tops the closed flagships it compares itself with on four agentic tests, at about 20 cents per coding task, and Epoch AI finds open models only six to seven months behind on cyber skills. The wealth may stay concentrated; the technology is spreading anyway.
All the best,
Kim Isenberg
Record-setting Cyber ECI scores, closed vs. open-weight models (Epoch AI, CC-BY)
๐ Open-Weight Models Trail Closed Ones on Cyber by Half a Year
Open-weight AI models, the kind anyone can download and run, are six to seven months behind the best closed models at cybersecurity tasks, Epoch AI finds. The top open model on Epoch's Cyber ECI, an index built from cybersecurity benchmarks, is Zhipu's GLM-5.3 with a score of 160. That beats OpenAI's GPT-5.4 from March but trails Claude Mythos Preview and GPT-5.5, both released in April, and the closed frontier has since moved on to Claude Opus 5.5.
๐ tl;dr: Cyber skills that only paid frontier models had this spring are now in open models anyone can download, for defenders and attackers alike.
How Harnessยฒ refines an agent's setup task by task (Google Research, Harnessยฒ on GitHub)
๐ Google's Harnessยฒ Lets AI Agents Rewrite Their Own Setup
Google researchers released Harnessยฒ, an open-source framework that lets an AI agent improve the setup around it while it works, without retraining the model. A "harness" is everything wrapped around a model: its instructions, tools, memory, skills and helper agents. Harnessยฒ has the agent propose changes to that setup, test them on the same task and keep what works, then carry the lessons to the next job. In the paper, a single round lifted GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, an office-work test, and by 11.8 percentage points on the partial pass rate of LAB, a legal benchmark.
๐ tl;dr: AI agents can get noticeably better at office and legal work by rewriting their own instructions and tools, with no new model required.
DeepSWE v1.1 pass rate vs. cost per task (A2Agent chart; DeepSeek and public leaderboard data)
๐ณ DeepSeek's "Minor" Flash Update Takes a Big Leap
DeepSeek released V4.1 Flash-1017, an update it calls minor that scores far higher than the V4.1 Flash we covered on October 5. On Terminal-Bench 4.0, which tests agents working in a computer terminal, it climbs from 31.2 to 42.3; on ProgramBench, from 20.3 to 29.1. DeepSeek says it also writes shorter answers and finishes tasks faster, so it costs less to run: on the DeepSWE coding test it reaches 77% at about 20 cents per task. The weights are free to download under an MIT license.
๐ tl;dr: A cheap open model just got markedly better at agentic coding, narrowing the gap with paid frontier systems at a fraction of their cost.
๐ฌ Watch This
Anthropic co-founder Jack Clark gave this year's Per Jacobsson Foundation Lecture, "Inside the Event Horizon," at the IMF and World Bank annual meetings in Bangkok. He argues that AI is starting to speed up science itself, that this will bring great wealth and real danger, and that governments have only "a small number of years" to agree on safeguards. Raghuram Rajan moderates, and economists Yan Shen of Peking University and V. Anantha Nageswaran, India's chief economic adviser, respond: the full case behind the singularity headline in today's Featured Story.
โ David Sacks, investor and co-chair of the President's Council of Advisors on Science and Technology, replying to Jack Clark on X
Clark had written that roughly 20% of Anthropic's equity is now committed to the public good. Sacks' point: the pledge is real, but the fund is an Anthropic-owned company, and nobody outside yet knows who will decide where the money goes.
Apple sued OpenAI in July over poached hardware secrets; now OpenAI says the leak ran the other way.
A new filing alleges that an employee who returned to Apple moved about 100 files on an unreleased OpenAI device to a personal account and pitched its details to a senior Apple executive. Apple says it has found no evidence.
Sam Altman (left) and an Apple store (Getty Images via Interesting Engineering)
Anthropic's Step Toward Universal High Income
The Takeaway
๐ Anthropic will put 5% of its shares into a new Public Benefit Fund, pending stockholder approval, with up to 2% more for every 50% rise in its value.
๐ Counting founder and employee pledges, about 20% of the company is now earmarked for the public good.
๐ The fund may pay specific groups directly, such as workers hit by AI, but it is not a universal basic income.
๐ Its board, recipients and valuation milestones are still unannounced, and critics such as David Sacks want proof it will serve the public.
Anthropic just put a number on sharing the AI windfall. On Sunday the company said it will commit 5% of its shares to a new Anthropic Public Benefit Fund, subject to stockholder approval, and its board can add up to 2% more for every 50% rise in the company's value above a baseline. Counting the co-founders' pledge to give away 80% of their equity and an employee donation match, Anthropic says roughly 20% of the company is now earmarked for the public good. At the up to $2 trillion valuation Reuters reports for its expected November IPO, the 5% alone would be worth up to $100 billion.
Anthropic co-founder Jack Clark, the company's head of public benefit (Samuel Corum/Bloomberg via Getty Images)
The move came a day after co-founder Jack Clark told IMF and World Bank officials in Bangkok that growing evidence points to a "singularity" and that faster progress "will bring with it great wealth and potentially great peril." An internal Anthropic model has U.S. GDP reaching $44 trillion by 2030 under certain conditions, up from just over $30 trillion today. The fund is Anthropic's answer to who shares that wealth: it may make large donations or "direct distributions to specific populations," such as workers whose jobs AI disrupts. Elon Musk, who runs rival xAI, replied that Anthropic should be "giving the shares directly to the public," added "This is what I will be doing," and, under a user's post hailing the beginning of universal basic income, wrote "universal high income."
That is still far from a check in every mailbox. Anthropic has not named the fund's board, its recipients or the valuation baseline, and promises those details "in the coming months." The fund is a wholly owned LLC rather than a charity, so control stays with Anthropic, though its shares may not be used for lobbying or political giving. Still, the formula ties a growing slice of Anthropic's upside to the public: the bigger the company gets, the more it can give away.
Why it matters: If AI creates anything like the wealth Anthropic's own model projects, who shares in it becomes a political fight. Anthropic is answering with equity rather than promises, and its formula makes the public's slice grow with the company.
Sources:
๐ https://www.anthropic.com/news/public-benefit-fund
๐ https://x.com/elonmusk/status/2111914396490604547?s=20
๐ https://x.com/kimmonismus/status/2112097251992711448?s=20
๐ https://www.cnbc.com/2026/10/17/ai-us-gdp-anthropics-clark.html
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
The chart: DeepSeek's launch table scores V4.1 Flash-1017 against its predecessor and four rivals: Moonshot's Kimi K3, Zhipu's GLM-5.3, OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5. The new Flash posts the best score on four agentic tests: DeepSWE (77.0 vs. 74.0 for Opus 5), Automation-Bench (60.5 vs. 50.3), Agents' Last Exam (35.4 vs. 28.6) and Apex Bench (72.1 vs. 65.8). Opus 5 keeps the lead on Humanity's Last Exam (56.3 vs. 42.6), Terminal-Bench 4.0, ProgramBench and NL2Repo.
The lesson: Story 3 showed how far Flash jumped past its predecessor; this table shows where that leaves it. On four agentic tests, a cheap open model now beats the closed flagships DeepSeek chose to compare against. On hard reasoning, terminal work and building large programs from scratch, Opus 5 still leads clearly, which suggests where the closed labs' remaining advantage lies.
The caveat: These are DeepSeek's own numbers on test setups it picked, and the table leaves out newer closed models such as OpenAI's GPT-6 Astra and Claude Opus 5.5. Independent leaderboards will show how much of the lead survives.
๐ Why the CoreWeave Panic Looks Overdone
โก Bottom line: CoreWeave lost nearly 25% last week as yields on its bonds jumped to 15% and above, dragging other AI infrastructure stocks down too.
๐ก Why it matters: CoreWeave builds AI data centers with borrowed money, so investors treat it as a test of the whole buildout.
๐ What it means: The shock came from interest rates and leverage, while the biggest buyers of AI computing still plan to spend more next year.
Last week the market tested whether the AI boom can pay its bills, and CoreWeave took the hit. The company builds data centers full of Nvidia chips and rents them to AI labs, largely with borrowed money: about $50 billion of debt, according to Reuters. After hot inflation data raised bets on interest-rate hikes, its shares fell 15.5% on Wednesday and nearly 25% for the week, on course for the worst week in more than 11 months. Yields on its 2032 bonds, sold at 9.625% in June, touched 15%, and a 2031 bond briefly passed 16%. Nebius, Oracle and Dell fell too.
Yield on CoreWeave's 2032 bonds, June to October 13 (Social Capital via Chamath Palihapitiya on X)
The worry is real. S&P rates CoreWeave B+, which is junk territory, and by Thursday its credit-default swaps implied roughly 60% odds of default, according to Bloomberg. S&P Global's Gavan Nolan summed up the problem: "capital expenditure levels that are almost hyperscaler-like," without "the hyperscaler balance sheet."
But look at what set it off: inflation data and rate-hike bets, which raise the cost of CoreWeave's debt. CEO Michael Intrator wrote in a memo that the market sometimes "misunderstands and acts irrationally" and that the company's backlog of contracts "more than supports all of our debt obligations." The biggest buyers of AI computing are still accelerating: the hyperscalers are on track to spend about $800 billion this year and more than $1 trillion in 2027, according to JPMorgan.
I think the panic is overdone. CoreWeave has borrowed more heavily than its peers, so it feels a rate shock first, and its shareholders may pay for that. But Alphabet, Amazon, Meta and Oracle have issued more than $228 billion in bonds this year, all but Oracle with investment-grade ratings, to fund spending that keeps growing. As Tommaso Leodari of the hedge fund Index & Cie put it: "You cannot take CoreWeave as the poster child for AI."
Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation
If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.
