|
Also inside today's 12 stories
● Outside Review Finds AI Is Now Roughly Doubling the Pace of AI Progress at Anthropic
● Democrats Weigh a Corporate Crackdown. Lina Khan Says Start With AI.
● Mark Cuban Says Meta Profits From Deepfake Scam Ads Using His Face
AI developments from the last 24 hours
Days after voters rejected data centers across the country, OpenAI is trying to win them back. The company published a National Community Compact, a set of promises for every campus it owns and operates, and a policy blueprint asking lawmakers to turn similar rules into law for the whole industry. The documents also adopt a new name: data centers are now "compute campuses," a rebrand Axios had reported would be part of OpenAI's post-election policy push.
The promises read like a list of the backlash's complaints. OpenAI says it will pay the full cost of the electricity and grid upgrades its campuses need, and put up money in advance so other customers aren't stuck with the bill if a project stalls. It will publish each site's water and electricity use, construction plans and independent noise studies, and report annually on hiring. It will stop asking public officials to sign nondisclosure agreements, except for narrow security and trade-secret protections; it voided the ones in Effingham County, Georgia, in July. Several of these echo demands from critics such as Texas Gov.-elect Gina Hinojosa, who campaigned on banning data center NDAs and making companies prove households won't pay more.
The blueprint asks lawmakers to require those disclosures, bar NDAs with officials, force big power users to cover their grid costs, and publish the terms of tax breaks, with clawbacks if companies fall short. It also asks for faster permitting and grid connections, and restored clean-energy tax credits, and argues against blanket rules requiring data centers to build their own power plants. OpenAI acknowledges the mood. "Americans are understandably highly skeptical that big investments by Silicon Valley-based companies will deliver tangible benefits to their hometowns," the blueprint says.
The industry's campaign money didn't stop the backlash either. Its main super PAC, Leading the Future, spent $4.8 million on Republican Darline Graham's South Carolina Senate race; she lost to Democrat Annie Andrews, who called her "Data Center Darline." The PAC put no money into fighting the roughly 40 data center ballot measures that passed, Semafor reported, and was caught off guard by their success, one Republican operative said. The timing was not lost on observers. "OpenAI kicked off its big charm offensive this morning after data center moratoriums went 43-1 during the midterms," AI writer Andrew Curran posted.
| Why it matters: The industry's fight has moved to the places where data centers get built, and OpenAI is trying to set the terms before states and towns set them for it. Some of the pledges are real concessions; others are an argument for faster permits and fewer power mandates. The compact covers only campuses OpenAI owns and operates, and much of it is promises to disclose. Whether communities accept a deal now, after voting no, is the test. |
Sources: OpenAI · National Community Compact (PDF) · Compute Campus Policy Blueprint (PDF) · Axios · Andrew Curran on X
While OpenAI courts communities, some Democrats want to put executives under oath. Senior congressional Democrats have been studying a 181-page plan, drafted by the progressive think tank Groundwork Collaborative, naming hundreds of executives and companies to investigate once the party takes Congress in January, the Washington Sun reported. Its targets include the biggest tech companies, AI labs, natural gas firms, private prisons and nursing home operators. One proposed hearing would press online sports-betting CEOs on how they use AI to encourage compulsive gambling. "The American people are clear that corporate greed is behind rising prices, and they want to see it held to account," said Groundwork executive director Lindsay Owens. "We're talking resignations of major CEOs, hauling people in, material restitution."
Party leaders are cooler. "We all get a lot of emails," a Democratic aide told Semafor. "It's one of many." Rep. Jamie Raskin, the incoming House Judiciary chair, told Semafor he hadn't seen the plan: "We're not taking people or corporations that we want to pursue in a vendetta-like fashion: That's Donald Trump land." Centrists warn that a broad assault on business could cost Democrats the edge on the economy they just won.
Lina Khan, who chaired the Federal Trade Commission under President Biden, made the case for investigating, and for putting AI near the top. "AI markets should be high on the list of investigations," she wrote on X. "The business decisions of the major AI firms have huge public ramifications, yet the public is at a severe informational disadvantage when it comes to how these markets work and the choices these firms are making." As a House Judiciary staffer, Khan worked on the 2019–2020 investigation of Amazon, Apple, Facebook and Google, which she says laid the groundwork for later lawsuits and proposed laws. Her advice to the new majority: let subcommittee chairs issue their own subpoenas, spend months preparing before CEOs testify, and rebuild Congress's depleted research staff.
| Why it matters: Oversight is the one power a Democratic Congress holds without needing President Trump's signature, and as Khan argues, investigations are often how laws get written: subpoenaed documents become evidence for bills and lawsuits. Sen. Elizabeth Warren, likely to chair the Banking Committee, is already warning Big Tech CEOs to expect grilling, Semafor reports. AI companies, which courted the administration and spent heavily in the midterms, are an obvious target. The fight is over how wide to swing. |
Sources: The Washington Sun · Semafor · Lina Khan on X
The question that worries AI safety researchers most is whether AI has started speeding up its own development. An independent evaluator with inside access now says yes, probably, at least at one lab. From Aug. 12 to Oct. 29, staff from METR, a nonprofit that tests AI systems, reviewed what it calls unprecedented internal data from Anthropic, from training metrics to staff interviews. METR took no payment. Its conclusions:
- Faster progress. As of October, Anthropic's models are likely improving on well-defined tasks about twice as fast as they did from March 2024 to October 2025: roughly two years of progress in one. METR says it is highly uncertain and would not be surprised by anything from 1.3 to 3 times.
- AI is probably the cause. METR's best guess is that about 80% of today's speedup comes from Anthropic using AI to do its research. It is confident the cause is not faster growth in computing power, but Anthropic's own staff don't agree on the rest. For now AI mostly assists researchers rather than working alone, though researchers described early signs of fully autonomous contributions.
- No brakes in sight. METR found no clear bottlenecks for the next 12 months, so progress could speed up further, though it says it may have missed some.
Much of the evidence is hidden. Anthropic had the right to redact confidential material from METR's roughly 100-page report, and cut most of the evidence METR found most informative, tightening its redactions in the final days. Anthropic said the cuts were meant to keep information from helping rival AI developers. "As a consequence, the public has to rely on METR's judgment much more than would be ideal, given the weight of the questions answered here," the report says.
Some forecasters were caught off guard. METR researcher Ryan Greenblatt wrote that he had been seriously wrong: he expected AI to speed up its own development far less this year. His rough math suggests AI is speeding up Anthropic's software research about 4.5 times, a pace the influential "AI 2027" forecast didn't expect until around March 2027. He says the report makes superintelligent AI within one to two years significantly more likely.
| Why it matters: Anthropic told investors on Monday it may soon cross its own threshold for AI-driven research acceleration (D.A.D., November 3). An outside reviewer now says the speedup is already under way. If AI is helping build better AI, the timelines for job disruption, safety rules and everything else shrink, just as the public turns against the industry. The caveat is real: the strongest evidence is locked away, and the company being reviewed decided what stayed secret. |
Sources: METR · Ryan Greenblatt on X
Mark Cuban says his face and voice are being used without permission in AI deepfake ads peddling a fake $294 diabetes cure on Facebook, and that Meta knowingly profits from it. He points to internal Meta documents reported by Reuters showing the company projected roughly $16 billion—about 10% of 2024 revenue—from ads for scams and banned products, with automated systems requiring 95% certainty before banning an advertiser. Meta disputes the figure as "overly-inclusive," citing a 58% drop in user scam-ad complaints and 134 million scam ads removed this year. The SEC is reportedly investigating.
| Why it matters: Deepfakes of real public figures are now a mass-market fraud tool, and the allegation that a platform's own internal math tolerated it raises the stakes for regulators weighing ad-platform liability. |
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Anthropic launched a personal finance feature for Claude on Thursday, letting paying U.S. subscribers connect bank and card accounts so the chatbot can analyze their spending, The Next Web reports. The beta, for Pro, Max and Ultra plans, links accounts through Plaid, so Claude never sees bank passwords. It can read balances, credit limits and up to two years of transactions, but can't move money. Anthropic says conversations in the finance section aren't used to train its models, there are no ads or sponsored recommendations, and users should check important figures against their bank records because Claude can make mistakes. OpenAI launched a similar ChatGPT feature in May.
Anthony Pompliano, founder of Silvia, a startup that bills itself as a personal AI chief financial officer, accused Anthropic of copying his company in a long post on X. He claims Anthropic staff began approaching Silvia's engineers in April. Anthropic had not responded publicly, Decrypt reported.
| Why it matters: The two biggest chatbot makers now want to sit inside people's finances, a move that brings them into regulated territory and puts them in direct competition with budgeting apps and the startups building on top of their models. Pompliano's complaint voices a fear many of those startups share: that the AI labs they depend on will turn into their rivals. |
Sources: The Next Web · Seeking Alpha · Decrypt via Yahoo Tech
New announcements from major AI labs
OpenAI rolled out Images 2.6 across ChatGPT, Codex, and its API, adding GIF animation and claiming faster, more consistent image generation and editing. The company says the update renders images 50% faster than before and does a better job keeping characters, products, and brand details consistent across multiple edits. A faster model variant, Flare, can produce low-quality images in about six seconds, while a higher-quality version, Sunburst, prioritizes consistency. OpenAI says users generate roughly 4 billion images weekly across its products.
| Why it matters: Consistency has been a persistent weak spot in AI image tools—useful for marketing and product teams who need the same character or logo to look identical across dozens of variations, not just one good image. |
Google released Gemini Omni 1.2 Flash, a video-generation model now generally available through the Gemini API and Gemini for business. It extends native video clips from 10 to 15 seconds, adds multi-turn editing so users can refine a clip through follow-up prompts, and improves how closely output matches reference images—useful for things like virtual try-on sequences that keep a character's outfit consistent across the clip. Google also cut pricing 30%, to $0.07 per second of 720p video.
| Why it matters: Longer, cheaper, more controllable AI video pushes the technology closer to practical use for marketing, product demos, and social content rather than just novelty clips. |
UNSW Sydney, University of Sydney, La Trobe, and Adelaide University are rolling out campus-wide deals with OpenAI, giving students and staff access to ChatGPT Edu for coursework, clinical simulations, and robot-assisted teacher training. UNSW alone is extending access to more than 80,000 students and staff. An OpenAI-commissioned survey of Australian students found 81% use AI several times a week and 79% want more chances to apply it to real-world problems like health and climate work, suggesting universities are racing to catch up with student behavior rather than lead it.
| Why it matters: These deals signal that AI fluency is becoming a baseline expectation for graduates entering the workforce, not an elective skill—something employers everywhere should start planning for. |
New papers on AI and its effects from researchers
A new study found readers can't tell the difference between human-written LinkedIn posts and ones lightly edited by AI—accuracy was 49%, essentially a coin flip, and people defaulted to assuming human authorship when unsure. Researchers tested 214 participants on real and AI-altered posts, modeled on actual ChatGPT conversations. The AI editing itself didn't move trust or transparency ratings. But one specific prompt—asking the AI to insert a new, unverifiable idea the original author never wrote—made posts seem more trustworthy and transparent, even though it was fabricated content.
| Why it matters: The finding that invented, unverified claims can read as more credible than the original text is a warning sign for professional networks and workplaces now flooded with AI-polished writing. |
A study testing AI models against traditional statistical tools for guessing race, ethnicity, and gender from names found that chatbots' apparent edge in accuracy is partly a mirage: Anthropic's Claude Haiku refused to classify half the names in a North Carolina voter dataset, quietly skipping harder cases. That selective refusal skewed results—every method tested understated the Black-white gap in Democratic registration by 36% to 65%, and Haiku's gaps in who it labeled inflated an estimated age gap by six years.
| Why it matters: Researchers, pollsters, and marketers increasingly use AI to infer demographics from names when direct data is missing, and this study suggests those estimates can be quietly and systematically biased in ways that are easy to miss. |
A study of MovieLens users who predicted how they'd rate movies before watching them found self-predictions beat the platform's own recommendation algorithm by a wide margin (a 0.69 correlation with actual ratings versus 0.34), and outperformed trained collaborative-filtering models too. Simply adding a user's own stated expectation to any of the algorithms cut prediction errors by 26%. The exception: on old, heavily-rated movies with lots of data, the algorithms matched or beat human self-knowledge.
| Why it matters: Recommendation engines power everything from streaming queues to retail upsells, and this suggests the fastest fix for personalization isn't more data—it's just asking people what they expect. |
A new academic paper argues AI jailbreaking—tricking chatbots into ignoring their safety rules—should be understood through the lens of coercive control, a framework from domestic violence research. Rather than treating jailbreaks as purely technical exploits, the researchers map how persona-based tricks (getting a model to "roleplay" an unrestricted character) mirror patterns of manipulation and control in abusive relationships. The paper is conceptual, offering no new test results or benchmarks, but proposes rethinking AI safety as a relational problem built into how models are trained, not just a security bug to patch.
| Why it matters: If safety failures stem from relational dynamics baked into training rather than fixable code flaws, patching individual jailbreaks may never be enough—a sobering implication for any company betting on guardrails alone. |
The Cognitive Revolution — KYC Becomes KYA: Sentinel Bio's Joshua Monrad on DNA Screening, Uplift Studies & Agent Monitoring
AI in Business — How to Make AI-Translated Content Convert in Every Market - with Andrew Thomas of Acclaro
D.A.D. joke of the day: I asked ChatGPT to make my pitch more engaging. It came back with a proposal.
|
