Six stories today: the labs want to regulate themselves, voters want the data centers gone, open models are getting cheap, and the feeds are filling with machine-made filler.
- The self-policing pitch, the safety talks, and the cartel question
- Voters do not like the data centers
- Meta puts a price on the feed
- Salesforce and Nvidia ship a reasoning model built for sales
- Open models are closing the gap for a fraction of the price
- The agents are flooding the feeds with slop
The self-policing pitch, the safety talks, and the cartel question
Three threads on AI safety converged this week, and they point in different directions. Nvidia's Jensen Huang told TechCrunch that AI does not need regulation and that safety should be left to each model maker. Meanwhile OpenAI, Anthropic and Google have reportedly spent weeks in direct talks on safety, with Trump's team pushing to keep pace with China. And The Verge asked the obvious follow-up: when Altman, Amodei, Hassabis and Musk loosely agree to "pace the frontier," is that a safety pact or a cartel?
Why it matters: the industry is making the case for self-governance at the exact moment it is coordinating behind closed doors, and those two facts sit uneasily together. Huang's "leave it to us" argument only holds if the makers are actually competing on safety rather than agreeing to slow down together. Whether the private talks are prudence or price-fixing is not a question the labs can answer about themselves, which is the whole problem with self-policing.
Read Huang's argument → · Read about the safety talks → · Read the cartel question →
Voters do not like the data centers
A new NYT/Siena poll found that 61% of voters oppose the construction of AI data centers, and The Verge notes the opposition shows up across every poll, not just one.
Why it matters: the money case for AI infrastructure keeps expanding while public support runs the other way, and data centers are local — they need land, power and water in specific towns. That combination tends to produce zoning fights and permitting delays, which are the boring mechanisms that actually slow buildouts. A national appetite for AI does not help you when the county board says no.
Meta puts a price on the feed
Meta launched a set of global "One" subscription bundles that pair app subscriptions with extra AI usage.
Why it matters: Meta built its business on free, ad-supported apps, so charging for them at all is a shift worth watching. Bundling social access with AI usage is a bet that people will pay to remove friction from products they already use for free, and it puts a number on how much Meta thinks its AI features are worth. It also quietly reframes AI from a feature into a line item on your bill.
Salesforce and Nvidia ship a reasoning model built for sales
Salesforce and Nvidia released Koa, a reasoning model built on Nvidia's open Nemotron and trained specifically for sales, marketing and support tasks.
Why it matters: this is the case for narrow models built on open weights instead of a general-purpose frontier model. If a model tuned for support tickets and pipeline work does the job at a lower cost, a lot of enterprise buyers do not need a frontier lab at all. TechCrunch framed it as everything the labs should fear, and the reason is structural: the moat was supposed to be capability, and specialization plus open weights chips at it.
Open models are closing the gap for a fraction of the price
A Mozilla report found that cheap open models, many of them Chinese, are closing the capability gap with frontier US models. Ars Technica's read of the numbers: paying for a frontier model buys roughly a four-month head start, at about five times the cost.
Why it matters: a four-month lead is real, but it is not a durable moat, and 5x is a large premium to pay for it. For most teams the question is whether they need the newest capability the day it ships or whether they can wait a quarter and pay far less. If the answer is "wait," the economics of the whole frontier keep getting harder to defend.
The agents are flooding the feeds with slop
Ars Technica documented AI agents with names like "Timmy," "Ren" and "Jackie" flooding social platforms with generated content, or slop.
Why it matters: the problem is not that any single post is bad, it is the volume — generated content is now cheaper to produce than it is to filter. That breaks the assumption every feed ranking algorithm rests on, which is that posting takes some human effort. As the ratio of machine posts to human ones climbs, the platforms have to choose between detection they cannot yet do well and labeling they have so far avoided.
Tools worth your time
- Salesforce Koa — A reasoning model built on Nvidia's open Nemotron and tuned for sales, marketing and support work. For revenue and CX teams who want task-specific reasoning without paying frontier-model prices.
- Nvidia Nemotron — The open-weights model family Koa is built on, meant as a base to fine-tune for specific domains. For teams that want to own and specialize their own model rather than rent one through an API.
- Meta AI — Now bundled into Meta's paid "One" subscriptions alongside app access. For people already living inside Instagram, WhatsApp and Facebook who want more AI usage and are willing to pay for it.
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
