D.A.D. joke of the day: I asked AI to summarize our two-hour meeting. It replied, "This could have been an email." Finally, an intelligence I trust.
What's New
AI developments from the last 24 hours
New Governors Could Pause 41% of America's Planned Data Centers, Tracker Finds
Chips, Data Centers & Energy Regulation & Policy
The AI boom runs on data centers, and on Tuesday voters turned against them, from county ballots to governor's mansions. Data centers are warehouse-size buildings packed with the chips that train AI models and answer every chatbot request. The industry says it needs far more, because every new user and every AI agent running a long task takes more computing. Gartner expects data centers' electricity use worldwide to grow 26% this year alone.
Much of that planned growth now sits in hostile territory. States where the new governor backs a pause hold 159 gigawatts of planned data centers, 41% of the national pipeline, according to Cleanview, a research firm whose project tracker has been used in grid studies. Another 18% is in states whose winners want new rules but no pause. Locally, every data center ballot measure with results went against the industry (D.A.D., November 4).
The biggest prize is Texas, where Democrat Gina Hinojosa, who called for a moratorium, unseated Gov. Greg Abbott. Georgia elected Keisha Lance Bottoms, who wants a statewide pause; New York re-elected Gov. Kathy Hochul, who proposed one in July. AI data center stocks and SpaceX fell on the Texas result, CNBC reported.
In Texas, the brakes are already on. Abbott paused new data center permits and grid approvals by directive starting in August, pending a grid and water audit, and Hinojosa can simply leave that pause in place, the Texas Tribune notes. She has said her pause would cover new permits while lawmakers write protections, not projects already under construction, KTAB reported. Several trackers put Texas construction at roughly 20 gigawatts. Her plan sets conditions, not a ban, as tech writer Tae Kim noted: union labor, no subsidies, and proof that households won't pay more for power or water. Those need the Legislature, which convenes Jan. 12 with Republicans still holding the state Senate. Governors also control air permits for the gas plants many data centers build on site, Cleanview founder Michael Thomas noted.
In Washington, the likeliest law is about power bills. The House passed a bill 417-3 in September nudging states to make big data centers pay their full electricity costs; a Senate Democrat blocked it as too weak, and Trump got tech giants to sign a voluntary pledge to the same end in March. Bipartisan bills to keep AI companion chatbots from minors are another candidate. Broader bipartisan safety talks in the Senate have stalled. Some Democrats want far more. "Top of the Democratic agenda on day 1 needs to be a pause on advanced AI development, regulation of this out of control industry, and a ban on artificial superintelligence," Rep. Yassamin Ansari posted on X.
"Tuesday's results don't signal the end of the data center boom," Thomas wrote, though he expects pauses in some of the biggest markets and new rules nationwide, even in Texas.
Why it matters: Data centers need land, power and permits now, and local voters, governors and regulators control those. A slowdown in states holding two-fifths of planned capacity would ripple through AI companies' spending plans and push building toward friendlier states. If Congress acts, the best bet is on who pays for data centers' power: both parties and the White House are already on record.
Sources: Cleanview · Michael Thomas on X · Texas Tribune · KTAB/KRBC · Gartner · CNBC (stocks) · CBS Texas · Fox 7 Austin · TCEQ · Axios (House vote) · CNBC (Senate) · Inside Climate News (pledge) · Miller-Meeks (chatbot bill) · The Next Web, citing Reuters (Senate talks) · Tae Kim on X
OpenAI's AI Built 948 Working Linux Exploits for About $1,300 Each
A research pipeline using an advanced cyber-focused AI model scanned the Linux kernel's codebase and, in a single mid-October run, automatically confirmed 1,724 vulnerabilities and generated 948 working exploits that could grant attackers root-level control, plus 291 tools for leaking sensitive data. The cost: about $1.2 million in computing, or roughly $1,279 per exploit—far less than the $18.9 million to $66.3 million such findings could fetch through bug-bounty programs. The AI reportedly surfaced at least 10 exploitation techniques not previously documented. It comes two days after a critical Linux flaw found by Anthropic's AI was left unpatched for weeks on Ubuntu (D.A.D., November 3).
Why it matters: The same automation that lets defenders patch flaws at industrial scale could just as easily let attackers find them first, raising the stakes for how fast software makers can respond.
Discuss on Hacker News · Source: openai.com
OpenAI Brings Cyber-Focused AI Model to Government Cloud Customers
OpenAI's cybersecurity-focused model, GPT-5.6 Cyber, is now available to approved customers in AWS GovCloud—the isolated cloud environment used by government agencies and contractors handling sensitive data—through a product called Daybreak Red. The model is tuned to be less likely to refuse requests tied to authorized vulnerability research and exploit testing, paired with added identity verification and monitoring controls.
Why it matters: Government cybersecurity teams are becoming a priority market for AI labs, and the push to reduce model refusals for offensive security tasks signals a broader industry bet that specialized, less-restricted tools will win public-sector business.
Pangram, Known for Catching AI Writing, Says It Can Now Spot AI Video
Pangram, a company best known for detecting AI-written text, is releasing a model it says can detect AI-generated video, along with an updated tool for spotting AI images, Pirate Wires reports. Images and video are harder than text, CEO Max Spero wrote on X: there are far more AI image and video generators than leading chatbots, and their output varies more. Most phone cameras, including iPhones, already use AI to add detail to real photos, so the model had to tell AI touch-ups apart from pictures made entirely by AI. Instead of a single yes-or-no score, it marks which parts of an image look AI-generated with a heatmap, which Spero calls a huge advantage over rival detectors. We haven't seen independent tests of the new tools yet.
Why it matters: This year's midterms showed the need. More than $137 million went to AI-made political ads, most of them unlabeled (D.A.D., October 30). Newsrooms, courts, employers and platforms all need ways to tell real footage from fake, and a false accusation can do as much damage as a missed fake. A heatmap that shows people what looks suspicious is easier to check than a bare score. Pangram's text detector ranked among the most accurate in independent testing by University of Chicago researchers, but detection is an arms race: every new generator is a new test.
Sources: Pirate Wires · Max Spero on X · Pangram
What's Innovative
Clever new use cases for AI
Developer Builds Free Tool to Check If a Prebuilt PC Is Overpriced
Shopping for a pre-built PC but can't tell if the price is fair? A developer built PC Deal Check, a free web tool that takes a retailer link, photo, or text listing, identifies the components inside, and prices them against current market rates. It then issues a letter grade—in one test, a $1,400 PowerSpec desktop scored an 'A' against an estimated $1,569 in parts. Commenters called it lightweight and well-designed; the creator says he self-funded the build and expects component prices to climb as AI and gaming demand compete for the same hardware.
Why it matters: It's a small example of AI doing the tedious research work—comparing dozens of specs and prices—that used to require real expertise, turning any shopper into an informed one.
Discuss on Hacker News · Source: pcdealcheck.com
Developer Builds Detective Game With AI Suspects Trained to Lie
A solo developer built Uncracked City, a detective game that drops players into real cities to solve fictional crimes by interrogating suspects powered by language models—characters designed to lie, dodge, and mislead like real witnesses. Instead of scripted dialogue trees, players ask open-ended questions and have to catch inconsistencies themselves. The game is free on Android now, with an iOS version awaiting Apple's review, and supports six languages.
Why it matters: It's a small example of a bigger shift: AI lets individual developers build games with genuinely unscripted characters, something that used to require large studios and teams of writers.
Discuss on Hacker News · Source: uncracked.city
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Critic: The Viral 'AI Edits AI' Result Leaned on Code, Not Just Weight-Reading
AI Safety & Risk Science & Discovery OpenAI
On Tuesday our lead story reported MIT researchers' claim that OpenAI's GPT-6 Astra could improve another AI model by editing its weights directly (D.A.D., November 3). A widely shared critique by Matthew Honnibal argues the headline result was oversold. The math boost for Alibaba's small Qwen model, from 7.9% to 48.1%, relied on Python code and 50 rounds of running the model on practice problems, he writes, not on Astra reading raw weights alone. Only smaller demos, like naming the game a tiny network was trained on, used weights alone. Some commenters counter that the paper was clear about its methods.
Why it matters: The core result stands: a tiny edit, no retraining, a big jump. What's in question is how Astra found the fix, and so how close this is to AI that can improve AI on its own. It's a reminder to wait for replication, as Ethan Mollick said at the time.
Discuss on Hacker News · Source: honnibal.dev
What's in the Lab
New announcements from major AI labs
OpenAI Buys Startup to Let AI and Developers Share Code Reviews
OpenAI is acquiring Pierre Computer Company, a small startup whose open-source developer tools—Code Storage, Trees, and Diffs—already power parts of Codex, OpenAI's AI coding assistant. The deal, still pending regulatory approval, aims to bake Pierre's Git-based code review and version-control tools directly into Codex. OpenAI says the goal is a shared workflow where human developers and AI coding agents review and manage code side by side, rather than bolting AI onto existing developer tools built for humans alone.
Why it matters: As coding agents become regular contributors to software projects, OpenAI is betting that the tools built for human-only teams—version control, code review—need to be rebuilt from scratch for human-AI teams, and it wants to own that infrastructure.
What's in Academe
New papers on AI and its effects from researchers
Another Study Finds AI Reliance Can Erode Skills At Work
A new theoretical model examines what happens when workers lean on AI tools repeatedly over time, rather than in a single task. The finding: workers who simply take the easiest path each time—accepting AI output when it's good enough—can get stuck as permanent novices, or see experienced workers' own skills erode, even though AI accuracy stays high. Workers who instead weigh future skill-building sometimes choose to skip more accurate AI help to keep their abilities sharp. The researchers say well-designed incentive pay for independent work can make skill-building self-sustaining, and that aiming for a higher skill bar can even cost less long-term than aiming low.
Why it matters: As companies roll out AI copilots widely, this suggests the real design question isn't just accuracy—it's whether incentive structures quietly erode the workforce's skills or build them.
Study Says AI Chatbots Are Optimized for Moments, Not Your Long-Term Well-Being
A new academic paper argues that AI companies are optimizing chatbots for the wrong timeframe. Current alignment methods—the techniques used to make AI assistants helpful and safe—mostly rely on feedback from single conversations, rewarding responses that feel satisfying in the moment. The paper, a conceptual argument rather than a data study, contends this ignores slower-building harms from repeated daily use, including reduced critical thinking, weaker social ties, and eroded human decision-making. It calls for 'long-horizon alignment' that measures AI's effect on well-being over months or years, not single chats.
Why it matters: As millions of people now use chatbots daily for advice, companionship, and decisions, this argument challenges AI makers to prove their products help users over time—not just in the moment they hit send.
New Method Flags Warning Signs in AI Chatbot Conversations
A new research method sidesteps the hard problem of proving an AI chatbot caused psychological harm by instead flagging observable warning signs in its responses—like excessive validation or failure to redirect distress—that could contribute to harm. Researchers built a labeled dataset of over 47,000 AI responses to train detection systems, achieving 79% accuracy against human judgments. The approach also cuts the computing cost of using AI to screen other AI conversations by 97%, making large-scale monitoring more feasible.
Why it matters: As chatbot companies face scrutiny over mental-health interactions, a cheaper, scalable way to flag risky AI behavior in real time could become a building block for safety audits and regulation.
Most State AI Guidance for Schools Copies Same Template, Study Finds
A new analysis of AI guidance documents from 35 states and Puerto Rico finds nearly identical language on legally-sensitive issues like data privacy and academic integrity (97% coverage each), but much thinner, inconsistent treatment of equity and accountability. Many states appear to have copied nonprofit templates—particularly from TeachAI—nearly verbatim rather than drafting original policy. Researchers also found no state included parents in drafting these documents, and only five involved students. Notably, state priorities tracked neither demographics nor political leaning, suggesting a few early movers set the template everyone else followed.
Why it matters: The guidance shaping how millions of students use AI in classrooms was largely written by a handful of nonprofits and early-adopter states, not by educators responding to local needs.
New Dataset Targets AI Blind Spots in Non-English Abuse Detection
Researchers working with civil society groups in Jordan, Francophone West Africa, and Kenya built a new dataset to help AI systems detect tech-facilitated gender-based violence—harassment, threats, and abuse conducted online—in Levantine Arabic, African French, and Swahili. Testing showed existing moderation models, largely trained on English-language content, frequently miss covert harms in these languages, with notably higher error rates in Swahili. Fine-tuning models on the new dataset, called ALTO and containing 3,022 annotated examples, improved detection accuracy.
Why it matters: It's a reminder that AI safety tools built and tested mostly on English data can fail the people who need them most, leaving billions of non-English speakers with weaker protection from online abuse.
What's On The Pod
Some new podcast episodes
AI in Business — How to Make AI-Translated Content Convert in Every Market - with Andrew Thomas of Acclaro
AI in Business — What It Actually Takes to Trust an AI System Running on Its Own - with Yogesh Bhagat of CVS Health