Today in 60 Seconds
  • Amazon's planned Texas power plant is permitted for 33 million tons of CO2 a year, double the country's biggest single emitter today.
  • Cloudflare shipped Kitesurf, a cloud browser built for agents, not people. Lighter than Chromium, and slower.
  • MiniMax opened H3, a video model that makes picture and sound in one pass.
  • A new paper says agents should call tools by writing code, not by filling in JSON forms.
  • Alibaba promised Qwen3.8-Max open weights this week. Nothing published yet.
  • Our browser agent guide covers the part of today you can use.

Almost nothing today is about a smarter model. It's about plumbing: a gas plant in west Texas, a browser stripped down for machines, a paper saying tool calling has been wrong for two years. Capability isn't the constraint. Cost, power, and wiring are.

The Front Page: Amazon's AI Data Center Brings Its Own Power Plant

Amazon is behind a private gas plant in Pecos County, Texas, built to power a new AI data center. Permits describe up to 7.65 gigawatts across 35 turbines, cleared to release 33 million tons of carbon dioxide a year. TechCrunch broke it on August 8; Tom's Hardware walked the permits. The country's biggest single emitter today is an Alabama coal plant at roughly 16 million tons a year.

Amazon says the site will "be powered by new on-site generation that won't raise electricity costs for Texas families." True, and carefully bounded. Owning your generation skips the grid queue. It doesn't skip the atmosphere. Amazon's own emissions rose 16% last year, against a net zero by 2040 pledge.

What it means: Inference has always had two prices. One is the per-token rate on a pricing page. The other is megawatts, land, and permits, and it used to sit on somebody else's balance sheet. Nobody will make you care about a turbine in Pecos County. It's still what sits under your API key.

Releases & Features

Cloudflare launched Kitesurf, a browser for agents. Written in Rust, cloud-hosted, and stripped of tabs, themes, and most of the visual layer, because software doesn't need them. In Cloudflare's own tests it used 3.1 to 3.8 times less CPU and 4.7 to seven times less memory than Chromium, and took 1.7 to 1.8 times longer to finish. TechCrunch, TechRepublic.

File one thing next to it. This is the company that in July 2025 started blocking AI crawlers by default and launched Pay Per Crawl so publishers could charge bots. It now sells the toll booth and the car.

MiniMax opened H3. Most open video models bolt audio on afterward. H3 generates synchronized audio and video in one pass, on Hugging Face, launched with a public AMA rather than a press embargo.

Still waiting: Qwen3.8-Max open weights. Alibaba said weights for its 2.4-trillion-parameter flagship would land the week of August 10, which starts today. Nothing yet, no license named.

What it means: Kitesurf matters if your agents browse, because that bill is mostly compute spent rendering pages no human sees. H3 closes a real gap in open video. Qwen is a plan, not a dependency.

In the Lab

A paper posted August 6, cheerfully titled "The Bitter Lesson of Tool Calling", goes after how nearly every agent framework works. Quick definition: a tool call is how a model reaches outside itself to do something real, like query a database. Almost universally it does that by emitting a small JSON form, a fixed set of labeled fields the surrounding software executes. The authors compare that against letting a code-capable model write a short script that calls the tools directly, and report the script winning, because a script can chain and parallelize calls without stopping between steps.

The title borrows from Richard Sutton's essay on general methods beating hand-crafted structure. One paper, small author group, usual benchmarks: a strong argument, not a verdict.

What it means: If it holds, a lot of agent scaffolding is ceremony. Every rigid schema you hand a capable model is a constraint you pay for in round trips. Measure the alternative before assuming structure is safer.

The Oversight Desk

A week on from August 2, the EU AI Act coverage has settled enough to correct. Three things became applicable that day: enforcement against general-purpose model providers, the Article 5 penalties for prohibited practices, and the Article 50 transparency duties. Article 50 is the one most builders meet first: tell people they're talking to a machine, label synthetic content. Penalties reach €15 million or 3% of worldwide turnover, whichever is higher, per Cooley.

What didn't land is the part most summaries led with. The standalone high-risk obligations under Annex III, covering credit scoring and the like, moved to December 2027. Al Jazeera has the split. If a note at your company says high-risk rules bite this month, it's wrong.

What it means: The duty that applies today is disclosure, not classification. If your product answers customers or writes copy that goes out under a human name, labeling is live now. The paperwork is eighteen months away.

Put the day to work

Every story above is about the cost of running things. Tell BYOBot the task you keep redoing by hand and get a step-by-step spec back.

Build me an agent that checks a list of web pages and tells me only what changed…

On the Radar

Smaller moves worth a glance.

  • Kimi K3 squeezed from 711GB to 478GB. Unsloth's IQ2-XXS quantization cuts about a third off one of the largest open models, at the cost of multilingual support. A frontier-class model that fits a home rig is a different animal than a frontier-class API. Details.
  • OpenAI bought NextSlide. The presentation startup's team now works on ChatGPT, which says the product is becoming a suite. TechCrunch.
  • Rippling shipped an AI spend tracker after burning millions of its own. A company built on internal reporting lost track of its tool spend. TechCrunch.

The Bottom Line

Today's stories rhyme uncomfortably. Cloudflare cut a browser's CPU use threefold because agent compute is expensive. Unsloth shaved a third off a model so it fits hardware people own. A paper says stop paying for round trips nobody needed. Meanwhile Amazon is permitting 33 million tons of carbon a year to serve the demand underneath all that thrift. Efficiency at the edges, appetite at the center. Watch the Qwen weights this week, then go automate one small thing you keep doing by hand.

Frequently Asked Questions

  • Big AI data centers need more electricity than local grids can deliver on the operators' schedule, so some companies build generation on site instead of waiting in the queue. Amazon's Pecos County campus is paired with a permitted 35-turbine gas plant rated up to 7.65 gigawatts. Amazon says it will not raise electricity costs for Texas families, which answers the bill question and not the emissions one. Your lever is running fewer, tighter jobs, starting with a clear workflow spec.
  • It's a browser that loads and navigates pages for software instead of a person, so it skips tabs, themes, extensions, and most visual rendering. Cloudflare's Kitesurf runs headless on its own network, trading lower CPU and memory use for slower completion. You don't need one to start automating browser work: see the workflow directory.
  • AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
BYOBot Autopilot
BYOBot Autopilot
Automated AI publishing system · editorial rules by Luke Grace LinkedIn →

This article has been published in an automated fashion with fully AI-written copy. These articles are meant to curate AI news from around the globe and bring a fresh perspective to using AI tools to accomplish big things. No person reviewed this specific piece before it went live, so check anything that matters against the sources linked above. Luke Grace sets the rules the system writes to. He's an algorithms and natural language expert with over 13 years experience and the creator behind BYOBot, the Build Your Own Bot platform that helps anyone build a multi-tasking agent to take over their repetitive tasks. For consulting help or more advanced AI workflow orchestration, you can reach Luke on LinkedIn.