AIMarketsNewsTech

The Sunday AI brief: Google chatbot carveouts, DraftKings elasticity, and data-center spend past housing

  • State chatbot-safety bills keep Google-sized carveouts that can spare the big chatbots
  • DraftKings trained an elasticity score that chased expected post-bonus losers
  • U.S. computer and data-center gear spend hit $752B, edging past $748B in housing investment
  • Jensen Huang called the lab slowdown chorus “ulterior” and said CEOs want out of existing laws
  • A Mandiant undercover analyst sat inside TeamPCP while the supply-chain worm hit OpenAI devices

At least 10 states this year advanced chatbot-safety bills with near-identical carveouts, exemption lines that can pull ChatGPT, Claude, Copilot, Gemini, Meta AI, Grok, and Alexa out of the child-protection rules the bills advertise, NPR reports. Katie McQue’s exclusive finds heavy tech-industry input in the drafting. Some 2026 enactments already include language that can spare the most popular chatbots.

Colorado’s Chatbot Safety Act excludes 12 categories, including developer tools, search-engine chatbots, voice assistants, and bots built into social media. A person familiar with the process says Healthier Colorado worked with Google.

Hawaii Rep. Trish La Chica refused a Google lobbyist draft; Act 248 later stripped those exemptions. Arizona Governor Katie Hobbs vetoed her state’s bill in June.

The exit doors keep showing up in bills sold as child protection. The statehouse carveouts are the quieter liability map.

Longer cut: State chatbot bills keep Google-sized exits in the draft.

DraftKings built a 2023 machine-learning system that ranked online casino players by how much they were expected to lose after a free bet or bonus, then steered promotions at the highest scores, the Daily Caller reports, attributing a New York Times investigation. Workers called the ranking an “elasticity” score. Higher meant chase with more perks.

Former analyst Jayden Butts told the paper the best “investment” looked like a problem gambler. A separate 2024 model meant to flag people sliding toward crisis was shelved. DraftKings disputed the framing and said promotions go to customers with “sustained, engaged use.”

The promo model learned addiction markers. The harm model got shelved.

Longer cut: DraftKings elasticity score chased expected post-bonus losers.

U.S. spending on computers and data-center gear hit $752 billion in the second quarter, edging past $748 billion spent building and improving homes, Fortune reports, citing Bureau of Economic Analysis figures flagged by San Francisco Fed VP Adam Shapiro on LinkedIn. Those BEA categories are information-processing equipment and residential fixed investment.

Housing spend was down 18% from an early-2021 peak. Computer and data-center gear rose 51% over that span. An NBC News poll found 64% of registered voters less likely to support a candidate favoring a data center in their community.

The AI buildout crossed the residential line in the government books. Capital spending is still racing while the neighborhood vote turns cold.

Longer cut: U.S. data-center hardware spend tops housing investment.

Nvidia CEO Jensen Huang, in a CBS Sunday Morning interview covered by Business Insider, accused frontier-lab leaders of wanting something other than new oversight. Quote: “They’re actually not asking for more laws. They’re asking to be relieved of the laws we do have.” Follow-up: “They must be doing it for ulterior reasons.”

The Verge, on the same CBS cluster, has Huang claiming a “0% chance” AI ends the world and calling scaring people “unnecessary” and “irresponsible.” He said Amodei and Altman slowdown talk is “not grounded in science.”

The chip seller reframed last week’s safety chorus as a bid for liability relief. Existing product-liability and cybersecurity rules are the ones he wants applied, not waived.

Before Australian police arrested two alleged leaders last month, the hacker crew TeamPCP poisoned open-source packages, stole developer accounts, and automated the loop with a worm known as Mini Shai-Hulud, Ars Technica reports. Google now says a Mandiant undercover analyst was invited into the group’s inner chat almost from day one.

Public reporting names victims including GitHub, Mercor, OpenAI employee devices, and the European Commission. Compromises named this spring include Trivy, LiteLLM, Checkmarx, TanStack, and Mistral AI. The arrests of Ruben Ian Thomson and Louis Michael Gaebler are alleged charges, not convictions.

Mandiant watched from almost day one. Victims still got hit. Labs sold agent safety while a supply-chain worm ran through the same software stack.

Leave a Reply

Your email address will not be published. Required fields are marked *