😺 Waymo just keeps on winning...

PLUS: GPT-5 tops charts, Waymo cuts injuries 85%, AI comedy fails

Welcome, humans.

This has got to be the best animation of how AI “works” (albeit, sped up really really fast). It’s all about what happens when you send a message (in this case, a very stupid one lol) to Gemini via Google Search and get an AI Overview back:

P.S: that video includes some 🔥 content from 3Blue1Brown, a fantastic YouTube channel for learning lots of things, but especially AI and machine learning* (example: watch this video on “what is a neural network.”).

And if you’re wondering wtf “machine learning” is, watch this video from Stanford AI engineer and AI educator Harper Carroll, where she breaks down the technical differences between terms like that and deep learning, agentic AI, and generative AI. YouTube is amazing, right??

Here’s what happened in AI today:

We dive into Waymo’s impressive safety stats and expansion plans.
Oracle and Meta are negotiating a $20B cloud computing deal.
YouTube launched 30+ AI creator tools including Ask Studio assistant.
Notion’s new agents have a prompt injection problem.

Advertise in The Neuron here

P.P.S: We accidentally messed up the link for yesterday’s LinkedIn news—here it is.

Waymo's Safety Stats Are So Good, Even the Skeptics Are Running Out of Arguments

DEEP DIVE: What’s Waymo been up to? As it turns out, rapid expansion due to a pretty stellar safety record. We break down the numbers on the website here.

It’s been about a year since we last covered Alphabet’s driverless car company Waymo. Since then, a massive peer-reviewed study analyzed Waymo's performance over 56M+ rider-only miles (where zero humans were behind the wheel). And the results are pretty wild:

85% fewer suspected serious injuries compared to human drivers.
96% reduction in intersection crashes (where 27% of fatal crashes happen).
92% fewer pedestrian crashes.
82% fewer cyclist and motorcycle crashes.

For context, Waymo went from 10,000 weekly rides in May 2023 to 250K in April 2025. That's ~25x growth in two years. If this rate continues, they could more than likely hit 1M weekly rides by early 2026.

The competition is on and popping, though. Amazon's Zoox finally launched in Vegas with their toaster-on-wheels design (no steering wheel, seats face each other). Chinese players like WeRide and Pony AI are partnering with Singapore's Grab for 2026 launches of their own.

And Tesla? Tesla is still very much in game. As far as its fully autonomous robotaxi service, called “robotaxi”, it launched in June out in Austin (though not without controversy), and will soon launch in Phoenix, too.

Where’s all this going? Anyone (including you) owning and operating your own robo-fleet. Elon Musk made that a big part of his launch event for the robotaxi, and during Alphabet's Q1 earnings earlier this year, Google CEO Sundar Pichai hinted at “future optionality around personal ownership.” Translation: you might eventually be able to buy your own Waymo that earns money while you sleep.

Why this matters: As Ethan Mollick says, with 40,000 killed and 2.4 million injured in US accidents annually, an 85% reduction in serious injuries should be front-page news. Bigger and bigger cities are starting to catch on: NYC just approved Waymo testing, and Singapore's fast-tracking approvals. Waymo also just announced a new launch in Nashville in partnership with Lyft, too.

The policy vacuum is a real issue, though. As Ethan Mollick noted, there's not enough policy response to self-driving cars existing. With safety data this compelling, the question isn't whether robotaxis will scale, it's whether regulation (and public opinion) can keep up with the technology. Read the rest here!

FROM OUR PARTNERS

Ship production-ready code in minutes ⏱️

Too often, agents write code that almost works, leaving developers debugging instead of shipping.

Warp lets developers prompt, code, and refine efficiently and accurately with an all-in-one platform.

What sets Warp apart:

#1 coding agent: Warp tops Terminal-bench and SWE-bench Verified, delivering more accurate code out of the box.
Tight feedback loop: Built-in code review and lightweight editing let you quickly spot issues, hand-edit, or reprompt without leaving Warp.
Proven efficiency: Early users generate 150M+ lines weekly with a 97% acceptance rate — saving 1–2 hours daily.

Get time back → use code NEURON9 for a month of Warp Pro for just $1.

Prompt Tip of the Day

Researchers recently found out that “ChatGPT Plus” tends to produce misleading briefs when summarizing scientific papers, revealing that the AI prioritizes engaging readability over scientific accuracy (paper).

There’s three tricks that we’ve found dramatically improve the quality of the summary you get when you ask AI to summarize for you:

Add the words “summarize with 100% fidelity to the original” to your prompt.
Use Gemini 2.5 Pro (in AI studio, if non-sensitive data, or in Gemini, if using sensitive data) with max Thinking budget for its long-context accuracy; it’s not perfect, but it’s from what we can tell the highest quality model at maintaining accuracy across long context information up until ~200K-300K tokens context.
Ask the AI to do an initial pass through to organize all unique ideas from the paper into a structure format (like Cornell-style; prompt below).

Before drafting, you must first process the provided outline and any associated source material. Extract every unique idea, fact, statistic, and quote. Organize them in a Cornell-style format where each main idea is a top-level number and sub-points are nested (e.g., 1.1, 1.2, 2.1). This creates a structured knowledge base that will be used for drafting, allowing facts to be easily referenced, remixed, and combined.

Do these three things in tandem, and we’re sure the quality of the summary you get will be improved. We have more ideas like this, so if these tactics don’t work, let us know!

Treats to Try

*Asterisk = from our partners (only the first one!). Advertise in The Neuron here.

*VEED creates professional videos from text prompts, automatically adds subtitles, and generates avatars for your marketing content and YouTube videos.
Dart automates your project workflows, generates subtasks and reports, and warns you about potential blockers while seamlessly connecting with GitHub, Slack, and other tools you already use.
Animant turns your PDFs, audio, and videos into interactive 3D presentations where models automatically rotate to match what you're discussing—no technical skills required.
MarkUp scans your blog posts, marketing copy, and other content to flag brand compliance issues, then automatically rewrites it to match your style guide before you publish (raised $27M).
Teable turns your spoken requirements into working databases and apps, automating repetitive tasks and generating insights from millions of rows of data—no coding needed.
Google’s Learn Your Way turns your boring textbooks into interactive study materials that adapt to how you learn best.

Around the Horn

In terms of pure intelligence, the best AI you can run right now is GPT-5 High, followed by Grok 4, Grok 4 Fast, and Gemini 2.5 Pro.

Oracle and Meta are working on a $20B multi-year cloud computing contract (perhaps to unlock more infinite money glitch??).
YouTube unveiled 30+ AI tools for creators at their 2025 “Made on YouTube” event, including Ask Studio AI assistant, A/B testing for titles and thumbnails, and collaboration features allowing up to 5 contributors per video.
China has “cracked down” on the use of live-streaming and AI in the country to “sell religion” as part of a broader effort.
Google Cloud has seen significant growth in 2025 as AI startups flocked to its platform for specialized infrastructure and integrated tools, with these startups contributing disproportionately to revenue growth.
Notion’s new 3.0 AI Agents apparently contained a critical security risk where malicious prompts could manipulate the web search tool to exfiltrate confidential data from private pages, revealing that even frontier models like Claude Sonnet 4.0 remain vulnerable despite their claimed advanced guardrails.

FROM OUR PARTNERS

Live AMA, engineer-led demos, and more – all at Glean:LIVE

Don’t miss Glean’s product launch on Sept 25th. Get a first look at Glean’s new Assistant, see how to vibe code an agent, and catch engineer-led demos. This launch is all about empowering you with a more personalized experience and upleveling the skills only you can bring to your work.