D.A.D. joke of the day: My boss asked if I'd been using AI to write my reports. I said, "Great question! Here's a comprehensive breakdown of my workflow."
What's New
AI developments from the last 24 hours
Anthropic Tells Investors AI May Soon Cross Its Own Threshold for Speeding Up AI Research
AI Safety & Risk AI Business & Money Anthropic
In an updated IPO filing on Monday, Anthropic told investors that AI may soon be dramatically speeding up its own development. Its internal estimates and expectations for future models "indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI,'" the company wrote. That's one of the capability thresholds in its own safety rules, which AI writer Andrew Curran summarized as a year's progress equal to two years at the 2018-2024 pace. Anthropic says no current model has crossed it.
The filing also admits the company may not see the moment coming: in the past it has not identified a significant acceleration until after it happened. Anthropic says it is considering rule changes to require human review and approval of some research decisions as Claude does more of that work, and draws a firm line: "We do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible." It warns that such safeguards, like the limited release of its Mythos model, could slow it down against competitors that don't adopt them.
Why it matters: For years, AI that helps build better AI was a thought experiment. Now a leading lab is telling investors, in a legal filing where misleading statements carry liability, that it may be close. It's a striking disclosure a week after Anthropic's first filing laid out its risks (D.A.D., October 27), and it lands alongside today's MIT research showing AI can edit other AI's inner workings directly. The open question is whether keeping humans in charge holds up once competitors move faster.
Sources: Anthropic S-1/A (SEC) · Andrew Curran on X · Theo Bearman on X
Critic Urges SEC to Block Anthropic's IPO Over Nearly $900 Billion in Commitments
AI Business & Money Regulation & Policy Anthropic
Anthropic's finances look strong on the surface. Footnotes tell a bleaker story, argues Matt Scherer, co-author of a recent report on protecting people when the AI bubble bursts. In an op-ed for Tech Policy Press, he writes that Anthropic's disclosures reveal nearly $900 billion in spending commitments, most stretching into the 2030s. Its 2026 results haven't been audited, he notes, and third-quarter figures suggest its revenue growth "slowed to a crawl" over the summer. He wants the SEC to block the planned November IPO. If it goes ahead, he warns, "workers, retail investors, and retirees will increasingly be left holding the bag when Anthropic's bills come due."
Why it matters: Anthropic's IPO is the first real public test of whether AI's giant compute bills can be paid. Scherer's case rests on numbers in Anthropic's own filing, so expect investors and regulators to press on the same footnotes. It's an advocacy piece, and the SEC rarely blocks an offering outright, but the question it raises applies to every AI lab planning to go public.
Sources: Tech Policy Press · Discuss on Hacker News
How a Tiny Tweak Made a Free Chinese Model Six Times Better at Math
AI Safety & Risk Science & Discovery OpenAI
Every AI model is, underneath, billions of numbers called weights. Normally nobody edits them directly: they're far too opaque for humans, so companies set them by training, a slow, data-hungry and expensive process. Now researchers at MIT say OpenAI's GPT-6 Astra can read and rewrite those numbers itself.
The results, posted in a blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, are striking. Looking only at a small network's raw weights, Astra could tell which video game it had been trained to play. It wrote game-playing networks from scratch that beat conventionally trained ones that had 100 million steps of practice. And asked to make a small Alibaba model, Qwen3.5-0.8B, better at math, it found the flaw that kept the model stuck in endless reasoning loops and fixed it by changing just 0.016% of the weights. The model's score on a standard math test jumped from 7.9% to 48.1%, with no retraining, and its ability to follow instructions roughly tripled.
There are limits. Astra couldn't make a different small model, one without the looping flaw, any better at math: its edits could remove an obstacle hiding an ability, the authors write, but not add one. Most tests used small networks. And the work is a blog post, not yet peer-reviewed or reproduced. "Waiting for replication but this is a pretty big deal," wrote Wharton professor Ethan Mollick.
Why it matters: Small, free-to-download models like Alibaba's Qwen are what many schools, hospitals, newsrooms and small businesses can actually afford to run. Today, making one better at a specific job means costly retraining. If a targeted edit can do it instead, capable AI gets cheaper and reaches far more institutions, changing the economics that now favor the few companies that can afford giant training runs. Safety researchers might audit a model's insides for hidden flaws, such as planted backdoors, before it ever runs. That's still an aspiration: Astra couldn't spot a deliberately sycophantic model from its weights alone. It also raises a harder question the authors name themselves: AI that can inspect and improve AI, including its own weights, is a step toward AI that improves itself, a milestone safety researchers have long watched for.
Sources: Takehi et al., MIT CSAIL · Rikiya Takehi on X · Ethan Mollick on X
OpenAI Pledges Its Data Centers Will Power Down When the Grid Is Strained
Chips, Data Centers & Energy OpenAI
OpenAI launched what it calls the Grid Protectors Program, a national effort to have its future data centers cut their electricity use when the grid is under strain. The first test is its planned campus in Effingham County, Georgia. Its deal with Georgia Power includes up to 1,000 megawatts of flexible demand. That's equal to the average electricity use of roughly 700,000 households, which OpenAI says is more than Washington, D.C., and Boston combined. OpenAI says it will pay the full infrastructure and electric-service costs to serve the Georgia campus. "Families and small businesses should not be asked to subsidize a compute campus's electricity needs," the company wrote. Elsewhere, the details are still to come: OpenAI calls the program an ambition, says arrangements will differ by location, and promises more commitments as they're finalized.
Why it matters: Electricity bills are becoming a political problem for AI, and OpenAI is answering the charge that data centers make everyone else pay more. Its argument is that a campus that uses power when the grid has room and backs off at peak times helps pay for grid upgrades without forcing new ones. The Georgia deal is the one concrete test. Whether the pledge holds elsewhere depends on terms OpenAI hasn't released yet.
Anthropic's AI Found a Critical Linux Flaw. Ubuntu Still Hasn't Patched It.
A critical flaw in the Linux kernel, present since 2016 and exploitable remotely, sits on virtually every Linux server. An Anthropic AI model found it, CSO Online reports, and Anthropic says it built a working exploit that gives an attacker full control of a server. Debian patched it on October 16. Arch and Alpine followed. But as of Monday, every supported version of Ubuntu, including the long-term releases businesses rely on, was still listed as vulnerable, a developer wrote in a widely shared post. That's nine days after the flaw was rated 9.8 out of 10 for severity, and six days after the working exploit. The developer took their Ubuntu servers offline rather than wait.
Why it matters: AI is now finding serious flaws faster than the companies selling business Linux are fixing them. Ubuntu runs a large share of corporate cloud servers, and it's made by Canonical, a company that sells security subscriptions. Volunteer-run Linux versions shipped the fix first. Any business on Ubuntu stays exposed until the patch lands.
Sources: chalmovsky.com · CSO Online · Discuss on Hacker News
Reddit Says Google's AI Search Is Choking Off Its Traffic. Its Stock Had Its Worst Day Ever.
Google's AI answers are starting to cost the websites that feed them. Reddit told investors last week that searches sending people its way dropped by more than a million visits a week on average in October. CEO Steve Huffman put the blame on Google's AI Overviews and AI Mode. "The reality is, AI Mode does not drive traffic," he said. Reddit warned its daily users could fall in the fourth quarter if the trend continues, and gave a revenue outlook below Wall Street's hopes. Its shares plunged 29% on Friday, its worst day as a public company, CNBC reports, even though third-quarter results beat expectations. Reddit says it will test requiring people to log in, Search Engine Journal reports.
Why it matters: This is the other side of Google's AI search squeeze. Google said its AI answers earn it less money than its old results (D.A.D., October 30). For the sites that supply the answers, the hit is bigger: the visitors stop coming. If even Reddit, among the web's most-searched sources, can't hold onto Google traffic, publishers, review sites and anyone who depends on search should plan for less of it.
Sources: CNBC · Search Engine Journal · eMarketer
What's in the Lab
New announcements from major AI labs
OpenAI Quadruples Bug Bounties as It Turns Its Own AI on Its Defenses
OpenAI is more than quadrupling what it pays outside researchers who find security flaws. A critical bug in ChatGPT, Codex or its API now earns up to $30,000. Exceptional findings earn up to $100,000, or as much as $172,000 with bonuses. A one-click account takeover that used to pay $6,500 would now pay over $51,000. OpenAI is also running another internal security sprint with the firm Cathedral, using Codex to find and patch its own weaknesses. And it has committed $1 billion in subsidized AI access over six months to "Daybreak" customers, organizations that defend critical infrastructure.
Why it matters: OpenAI is betting that AI tools, not just human analysts, are now needed to keep up with AI-assisted attackers. That shift will shape how every company budgets for cybersecurity.
Sources: OpenAI (bug bounties) · OpenAI (security sprint)
OpenAI Is Grading ChatGPT on Life Admin: Childcare, Budgets, Government Forms
OpenAI built an internal benchmark called LifeVal to measure how well ChatGPT handles everyday practical tasks—planning childcare, navigating government services, budgeting, grocery shopping, building workout routines—rather than abstract reasoning or coding problems. The tasks are drawn from real time-use surveys and a demographic survey of nearly 4,000 people. OpenAI says it's used LifeVal to guide a year of model improvements, and claims recent upgrades to browser use, computer control, and app connectors are pushing ChatGPT from a chat window toward something closer to a true personal assistant. No comparative scores were disclosed.
Why it matters: If OpenAI is explicitly optimizing for life-admin tasks rather than just benchmarks and coding tests, expect ChatGPT to get noticeably better at the chores you'd normally hand to an assistant—not just the ones you'd hand to an engineer.
What's in Academe
New papers on AI and its effects from researchers
One Economist Used AI to Recheck 108 Studies. A Feared Flaw Turned Out to Be Rare.
For years, economists have worried that one of their workhorse tools for measuring what policies do can give misleading answers, even pointing the wrong direction. The technique is called two-way fixed effects. Peter Hull, a Brown University economist, used AI to replicate 1,346 analyses from 108 published papers and compare them with newer methods. In a working paper, he reports that the results mostly agreed. The two approaches pointed in opposite directions in 10.8% to 14.5% of comparisons. After accounting for statistical noise, Hull puts the chance of a statistically significant result pointing the wrong way at just 0.6% to 2.1%.
Why it matters: Checking 108 papers by hand would take a research team years. AI let one economist do it, and Hull says it shows how AI can support large-scale checks of research methods. Policymakers lean on this kind of research, so faster auditing of it matters well beyond economics.
Source: Brown University (NBER working paper)
Study: Mental-Health Chats With AI Reveal More Than Users Say Outright
Privacy & Surveillance Kids, Teens & Mental Health
Many people now use chatbots for informal mental-health support. A new study asks what those conversations give away. Researchers led by Synthia Qia Wang built 240 realistic mental-health conversations from a nationally representative health survey. They checked them against real counseling transcripts and found people disclosed about as much overall. Then they tested what an AI could work out about each user from the conversation alone. Its first guess at a personal attribute was right 76.5% of the time, and 94.3% of the time within three guesses. All but one of the 240 conversations exposed at least one sensitive mental-health detail. Nearly half (45%) exposed all six that the researchers tracked. The authors call for chat designs that make these inference risks visible to users.
Why it matters: A chatbot doesn't need you to name a condition to work it out. For anyone using AI as a sounding board, and for employers or health apps that offer it, the privacy question isn't only what users type. It's what the AI can deduce. Privacy policies built around "information you provide" may miss most of it.
Mixing Experts to Red-Team AI Risks Finds Blind Spots, Study Finds
Researchers ran a pilot workshop testing "violet teaming," a governance exercise that combines red-teaming (finding how a technology could be misused) with blue-teaming (designing defenses) across technical, security, operational, and policy experts simultaneously. Using a hypothetical biodefense-relevant AI scenario, participants worked through three stages—how the capability develops, how it could be misused, and how it would actually be deployed. The paper reports qualitative findings only: the mixed-expert format surfaced misuse risks that purely technical reviews missed and pointed to non-technical fixes, like policy or operational controls, rather than just software patches.
Why it matters: As AI models edge closer to genuinely dangerous capabilities in biology and chemistry, this suggests effective oversight may require mixing policy and security experts into technical safety reviews rather than leaving AI risk assessment to engineers alone.
What's On The Pod
Some new podcast episodes
How I AI — How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma)