Zuckerberg: The new Trump-Whisperer? Google Argon: The eternal third

Good morning.

Companies have plenty to say about their own work. Checking the homework takes a little longer over coffee.

Mark Zuckerberg reportedly helped shape the roughly 300-word accord Trump signed Sept. 29. In August, he opposed an oversight body that never reached the signing table.

Google’s Gemini 4 Argon, unveiled Sept. 30, scores 53 on the Artificial Analysis Intelligence Index. Some employees have a different view after putting it to work.

Meta recorded $3.912 billion in research credits for 2025. Its AI data centers reportedly count as pilot models, which gives the word “experiment” a tax question to answer.

Stay curious,

Marcus Schuler

Mark Zuckerberg shaped the AI pledge Trump released Tuesday, Semafor reported, citing people familiar with the matter.

By their account, the idea came from Zuckerberg’s talk with House Speaker Mike Johnson at the Sept. 24 state dinner for China’s Xi Jinping. On Sept. 28, Johnson called AI fears “a Chinese psyop.”

In an August call with Trump, Zuckerberg opposed a FINRA-style body Anthropic, OpenAI and Google had coalesced around. The idea was scrapped.

Why This Matters:

• The FINRA-style body Anthropic, OpenAI and Google coalesced around is off the table, leaving each company to apply the accord’s controls to itself.
• The FTC will probe AI firms including OpenAI and Anthropic over consumer risks, a federal check that sits outside the voluntary accord.

Google unveiled Gemini 4 Argon on Sept. 30, initially for selected cybersecurity partners.

Its 53 points on the Artificial Analysis Intelligence Index tie GPT-6 Astra and Claude Fable 5.1, behind Anthropic’s two newest models. Argon tops Claude Opus 5.5 and GPT-6 Astra on DeepSWE v1.1 at 77.9% and trails three rivals on Terminal Bench 4 at 57%.

Some employees with direct access say it struggles with certain coding tasks. Google calls it inaccurate to say Argon underperforms in coding.

Why This Matters:

• Developers choosing a coding model need tests on their own projects, since Argon’s position changes with the benchmark.
• Paid API customers and Google AI Ultra subscribers are next for access, but Google has given no release date.

Meta classifies its AI data centers as “pilot models” to claim federal research tax credits, treating purchased chips as experimental supplies.

Its 2025 filing records $3.912 billion in total research credits without isolating the data-center share. Meta declined to explain what makes the facilities experimental, and no IRS determination is disclosed. The tax treatment matters because IRS guidance generally limits qualifying supplies to nondepreciable tangible property used directly in research.

Reuters examines how McDonald's pricing engine analyzes millions of daily transactions to recommend menu prices at nearly 14,000 U.S. restaurants.

A September price check found a $5.69 Big Mac in Fresno and a $6.89 version two miles away, though Reuters could not confirm the engine caused the difference. Five store owners described pressure to use the tools; McDonald's says franchisees set their own prices.

• HPE booked a $1.2 billion Vultr order Sept. 30 for AMD Helios AI racks, its first commercial Helios order.
• Amazon and Synopsys announced a custom-silicon IP agreement worth more than $1 billion Sept. 30, using a license-plus-royalty model.
• OpenAI and Synopsys announced GPT-Synopsys Sept. 30, a chip-design model whose AI agents run Synopsys tools against engineers' optimization goals.
• Reddit will end RSS support Nov. 13, 2026, citing large-scale scraping and automated abuse.
• Google is rolling out Gemini 4 Argon first to trusted cyber defenders, with introductory pricing of $2 per million input tokens and $10 per million output tokens.

Build an evidence table from an interview.

Interview notes often mix observations with impressions. Make the model sort the evidence before the hiring group meets.

Your raw input:

The prompt:

Why this works:

The fixed criteria reduce the chance that fluency becomes a proxy for ability. Counterevidence and missing evidence force the model to show why a score may be unstable.

What to use:

Use ChatGPT or Claude with your approved enterprise privacy controls. A human must make the hiring decision.

Hindsight, Vectorize's agent-memory project with about 44,000 GitHub stars as of Sept. 30, builds repository memory banks from git history and past conversations. It can inject stored material when coding work begins, and an independent two-session test carried two unwritten rules into a later session on one machine. Developers who keep re-explaining the same project's conventions to coding agents have a reason to try it.

Choose self-hosted storage or a local daemon if session content must stay on your machines, since configuration and scripted installs default to Vectorize's cloud. Hindsight on GitHub →

Credit: Midjourney

Prompt: Surreal fashion editorial portrait of a lanky young man, glossy golden-blonde mushroom bowl cut, curled blonde handlebar moustache, deadpan stare, his entire skin is zebra hide, flat powder blue background, medium format film photography

DoorDash's ordering agent is live in Apple's Messages app with 20,000 U.S. iOS pilot users, and a beta waitlist opened Wednesday.

Our take: DoorDash's agent remembers your burrito modifications and has your payment details ready. That is a useful amount of familiarity for breakfast. Early testing also produced prices that differed from the app and photos that did not match the recommendation. DoorDash says waitlist users will get a newer version. The burrito may be your usual; the price still deserves an introduction.

The company says it does not want to outsource discovery to AI agents while building an agent that can suggest recipes from your fridge. Apparently, the objection depends on whose agent gets the customer. DoorDash would prefer the robot helping itself to your shopping list to be on the payroll.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai