|
|
Welcome, humans. |
Okay, so apparently a Delaware jury just told Arm it broke its contract with Qualcomm and deliberately interfered with Qualcomm's chip business with Meta, then awarded Qualcomm $2. Not $2 million. Two dollars, which is roughly the cost of a gas station coffee. |
For context, Arm designs the chip blueprints inside nearly every phone, and Qualcomm pays Arm to use them under a license that runs through 2033. Before the trial, Qualcomm said Arm held back software and design tools it owed and leaked a 2024 license-termination threat to Bloomberg to spook Meta. |
The jury agreed on the breach and the Meta interference, but rejected Qualcomm's claims about Arm's pricing promises. Both sides declared victory, and Arm said the tiny award backs up its side. |
Two trials in two years, and the invoice came to $2. Somewhere, a litigator is framing the check and a CFO is doing the math on billable hours. |
Here’s what happened in AI today: |
😺 Google opened Gemini 4 Argon to paying users 📰 Nvidia plans about €2B in new European AI deals 📰 About 400 scientists urged a global AI oversight body 📰 Microsoft patched 1,400+ flaws in one October update 🎓 Ask AI for six fixes, expect to keep two
|
…and a whole lot more that you can read about here. |
|
|
😺 Google Opens Gemini 4 Argon to Paying Users and Ties OpenAI's Flagship |
Google launched Gemini 4 Argon on Tuesday, two weeks after announcing it to a small group of security defenders. We flagged that early access on Oct. 2, when it was still limited to testers. Now Pro and Ultra subscribers can try it in the Gemini app, and paid API customers can build with it. |
The number worth repeating is hallucination, meaning how often a model gives a wrong answer instead of admitting it doesn't know. On Artificial Analysis's AA-Omniscience test, Argon scored 15%, versus 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. |
The catch is accuracy. Argon got 50% of the questions right versus Astra's 63%, because that test punishes wrong answers but not skipped ones. So Argon wins partly by saying "I'm not sure" more often. Every office has that coworker, and most of us trust them more. |
Here's how it stacks up against the field: |
OpenAI: Argon ties GPT-6 Astra at 53 on Artificial Analysis's Intelligence Index and sits one point above GPT-6.1 Sol. Astra still wins some of Google's own rows, like FrontierSWE v2 (65.5% to 55.0%). Anthropic: Claude Opus 5.5 still leads that index at 58, and Sonnet 5.5 sits at 56. Google's own table has Argon ahead of Opus 5.5 on DeepSWE, 77.9% to 74.2%. xAI: Grok 4.7 scores about 46 on the same index, per one tracker, so Argon passes it. China: Bloomberg Intelligence found, as we reported Oct. 5, that top Chinese models trail U.S. rivals by 3%, down from 15%. That reading predates Argon, so the next scorecard shows whether Google's comeback widens the gap again.
|
Price is Google's best argument. Argon costs $2 per million input tokens and $10 per million output tokens through Nov 30, then $4 and $20, matching Opus 5.5. Astra costs $10 and $50. (A token is a chunk of text, roughly three-quarters of a word.) |
But Argon writes a lot: about 62K output tokens per Artificial Analysis task versus Astra's 27K. That puts its cost per task at $1.99 now and $3.98 after the discount ends, about 1.2x Astra's $3.26. The sticker price is a fifth of Astra's. The receipt is another story. |
Not everyone is sold. Bloomberg reported that some Googlers think Argon aces benchmarks but stumbles on real coding and front-end design, which Google denies. The DeepSWE and FrontierSWE numbers come from Google's own table, so treat them as a sales pitch until outside testers weigh in. |
Why This Matters: If you use Gemini for free, nothing changes today, because Google's post lists only paid Pro and Ultra subscribers for the app and AI Mode in Search. If you pay, you get a model Google says answers more directly, repeats itself less, and follows style instructions better. |
One demo has the app turn a folder of project notes into a client deck, email it, and schedule a task for future requests. The everyday payoff is the 15%: fewer confident wrong answers and more "not sure" replies, a trade that saves you fact-checking time. |
Our Take: Google just shipped its first model above the Flash class in over seven months, per Artificial Analysis. On Sept. 21, we flagged a mystery Gemini model on Arena that testers suspected was a Gemini 4 flagship in disguise. We can't confirm it was Argon, but the gap is closed either way. |
Argon's safety section also reads like a reply to our Sept. 20 story, when Gemini logged into three real companies during a test. Google now says it seals test sandboxes before risky runs and uses a monitor that reads the agent's reasoning, with shortcut attempts at 2% on coding tests. |
What to watch: whether the 15% holds on your messy real work. Run 10 to 20 of your own tasks through Argon before Dec 1, using the test we shared Oct. 5. |
|
FROM OUR PARTNERS |
What's stopping your search-powered AI apps from delivering? |
|
More than 330 IT leaders shared what's blocking search-powered AI in the enterprise. Here's what the data reveals about the five biggest obstacles and how to move past them. |
Find your fix. |
|
🎓 AI Skill of the Day: Ask for six fixes, expect to keep two |
Google engineers George Weale and Bo Yang let Gemini 4 Argon run unattended for 11 hours on a 205,000-line codebase, and published the setup. Argon proposed six speed-ups. The team shipped two and rejected four, and their summary is a good rule for any AI project: the agent explores, the human subtracts. |
Two details are worth stealing: |
Round one had zero house rules. Argon spent 70 minutes measuring before it touched anything and found real wins, but it quietly changed a public data format and ran only 216 of the project's 16,892 tests. Round two added a folder of the team's review rules and a mode that keeps the goal pinned during long runs. Same goal, but this time Argon read the rules first, ran every test, and opened a pull request (a proposed change for humans to approve) with before-and-after numbers.
|
Round one was a brilliant new hire who skipped orientation. Round two was the same hire after reading the handbook. |
Try it on any repeat task, like a weekly report or a spreadsheet cleanup. Write a one-paragraph goal with a number that proves success, put your house rules in a file the AI reads first, and ask for proposals instead of finished changes. Not sure how to phrase the goal? Google's 3D game guide says to ask Argon to draft its own goal command first. |
Goal: [one paragraph, including the number that proves it worked].
First, read my house rules: [paste what's off limits, what "done" means, and how I review work].
Then give me six proposed changes, each with before-and-after evidence and the risk if you're wrong. Don't apply anything yet.
|
Google ran this in its Antigravity coding app, where a /goal mode keeps the goal pinned across hours of work. In a regular chat, paste the goal back in whenever the AI drifts. |
Have a specific skill you want to learn? Request it here. |
|
FROM OUR PARTNERS |
Your AI Agents Can't Outsmart Bad Data |
|
AI agents are only as effective as the data powering them. In this episode of ZoomInfo’s Revenue Architects podcast, discover why clean data, connected systems, and smarter workflows are essential for AI success, and what it takes to build a truly AI-ready GTM stack. |
Listen now |
|
📰 Around the Horn |
 | OpenAI is reportedly racing to get its next flagship model safe enough to ship, while still dealing with fallout from firing three safety researchers. OpenAI says it was a confidentiality breach, and the researchers say it was for raising safety concerns. |
|
Nvidia CEO Jensen Huang said the company will unveil €2B in deals across Europe in the coming week, on top of $8B invested over the past year. Chinese tech firms reportedly bought over 1M Nvidia gaming and workstation GPUs in a year to plug AI computing gaps, the Financial Times said; the chips comply with U.S. export controls. About 400 scientists, including Geoffrey Hinton and Yoshua Bengio, called for an international body to set containment standards and keep a registry of security incidents for cyber-capable AI, Reuters reported. Microsoft patched more than 1,400 security flaws in its October update, up from September's record of nearly 1,000, and credited AI models for finding many of them. Google AI Studio builds custom Android apps from a text prompt with no SDK setup, and you can install the result on your phone over USB.
|
|
🍪 Treats to Try |
*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here! |
*Build an AI-Ready Workplace. Explore expert insights and practical guidance to modernize workplace technology, strengthen security, and prepare your organization for AI-powered work. Explore the hub. Gemini 4 Argon is Google's new top model, now in the Gemini app for Pro and Ultra subscribers and the paid API —paid only rn ($2 / $10 per million tokens through Nov 30). Facebook's Your algorithm lets you type the topics you want more or less of in Feed and Reels, and recommendations update within the same session (rolling out now, no completion date given). MausBot gives you a team of AI bots in a chat app on your own computer, each with a job like chasing overdue invoices, powered by subscriptions you already pay for —free and open source, plus provider fees. MentionBird shows which sources ChatGPT and Gemini cite when buyers ask for products like yours, then hands you a plan to get your brand named —paid only rn ($79/month; code PH19 makes it $19/month for 3 months). OpenCode is an open-source coding agent for your terminal, editor, or desktop that connects to 75+ model providers, so you can switch between Claude, GPT, and Gemini without switching tools —free and open source.
|
|
UPCOMING EVENT: The Neuron IRL in San Francisco |
|
The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen. |
It’s free, but space is limited. Save your spot! |
|
|
📖 Midweek Wisdom |
Justin Curl argues courts should pilot a parallel AI track, because AI-written text in federal complaints jumped from 0.1% to 18% and overloaded judges may lean on chatbots anyway. Deloitte's Gregory Vert, Brad Kreit, and Abha Kulkarni point out that 49% of workers using AI at work use unapproved tools, while only one in five companies has mature governance for AI agents. Your personal agent may beat IT's policy to the office. Geoffrey Hinton told the Guardian that letting AI companies regulate themselves is as unwise as letting oil and asbestos companies write their own rules. He added that Trump's push to rename AI "Super Intelligence" shows he knows almost nothing about it. A Washington Post opinion piece argues entry-level "grunt jobs" double as apprenticeships, so the AI risk for new grads is losing the bottom rungs of the career ladder.
|
|
New from The Neuron: AI Explained |
|
New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube |
|
|
|
|
| That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today! | | What'd you think of today's email? | |
|
|
Btw: We just launched a robotics newsletter! Sign up for it here. |
P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here. |
