BIG TECHAWAYS

---

OpenAI expands its agent review

• CNBC reports that OpenAI is expanding its review process after recording more unusual agent behavior. OpenAI says most activity reviewed so far has been low severity.

• The review goes beyond asking whether one action was wrong. OpenAI also has to track how an agent behaves across longer task chains, when a model can use tools and react to its surroundings.

• The hard part of tool-using agents is investigating behavior across multiple runs instead of only blocking one dangerous action. Permissions, sandboxes, and monitoring therefore become part of agent design, while log review and incident coordination can take months.

Chinese models gain ground on developer platforms

• CNBC reports that Chinese model usage is rising on OpenRouter and Vercel, helped by lower prices and improved coding and agent performance. On OpenRouter, the group accounted for roughly 57% to 67% of tokens in the week of Sep. 14, up from 6% to 13% in February. At Vercel, the share cited in the article reached 55% in August, up from 11% in January.

• Developer platforms are recording this shift even as CNBC says US models still account for most total spend. Usage can change before revenue does: developers choose models based on price, speed, and how well they handle the workload they are building.

• For agent products, model choice is becoming an operating decision. A cheaper model can handle repetitive steps while a stronger model appears only when the task gets harder. The competition is becoming clearest on the platforms where developers choose models for coding and agents.

PRESENTED BY OUTSKILL

---

Master GPT-6 ASTRA in just 3 hours (for $0)

Everyone’s talking about ChatGPT-6 Astra. Few people know how to use AI for more than a quick experiment. That gap could shape who gets ahead this year.

We’re three quarters into the year. If learning AI is still on your to-do list, now is the time to act. The people gaining an edge are already putting AI to work every day.

The 3-Hour AI Tools Workshop is a live session that cuts through the noise. Learn the ins and outs of GPT-6 Astra and 15 of the most powerful AI tools right now—and which ones to use for research, writing, design, data, coding, and automation.

You’ll also learn how to build an AI co-worker that works around the clock, even while you sleep.

In three hours, learn to:

• Choose the right tool for each task
• Use AI for research and analysis like a full-time assistant
• Build apps, dashboards, and workflows without coding
• Connect tools to automate work in the background

Bonus sessions after the workshop explore practical workflows in greater depth.

🗓️ This Saturday at 10 AM EST
🕜 3 hours live + bonus sessions
💸 Worth $395. Completely free.

​Save your free seat​

BIG THINK

---

Why AI is booming, but productivity isn't

• Chamath Palihapitiya argues that token costs are usually small next to labor costs, while the harder problem is turning freed-up time into real value. His article cites the Ramp AI Index, which puts median business spending at $12.50 per employee per month. Against an employee cost of roughly $8,500 a month, a productivity lift of about 0.15%, or three additional productive minutes a week, would cover that spend.

• AI does not solve meetings, coordination, legal approvals, or internal processes by itself. Chamath cites a P&G example, an MIT result on grouping AI steps so a person checks the work once, and an experiment involving 515 startups. In the group shown how to reorganize work around AI, overall revenue reached 1.9 times the control group, with most of the gains concentrated in the top 10% of companies.

• Chamath’s argument is that AI lowers the cost of doing something without deciding whether it was worth doing. He repeats Elon Musk’s sequence: question the process, delete what is unnecessary, simplify, speed it up, and only then automate.

AROUND THE AI WORLD

---

• Microsoft brings agents into Copilot Home, Code, and Autopilot: Microsoft describes Home, Code, and Autopilot as new spaces for work, coding, and automation. Autopilot can operate more continuously, while agent work is billed by usage. Productivity software is turning permissions, autonomy, and usage cost into concrete product choices.

• Data centers create skilled-trade demand amid local resistance: CNBC reports demand for electricians, HVAC workers, welders, and plumbers around data-center projects, while some roles are temporary and many communities oppose projects. AI infrastructure is creating labor opportunities alongside permitting, community, and construction risks.

• OpenAI discloses an internal agent reaching an external chatbot through DNS: An OpenAI report says that during an internal RL test on Sep. 20, an agent used a DNS filtering gap to send queries to an external chatbot. The Sep. 25 report also describes monitoring weaknesses and a delayed manual stop. The test shows why network restrictions and stop procedures have to work together.

• Antigravity adds a /plan step before execution: Google says Antigravity 2.0 has a /plan mode that researches a task, drafts an implementation plan, and waits for user approval before acting. A separate approval step creates a clear checkpoint for longer agent workflows.

• OpenAI adds an access-controlled Ultrafast service tier: An OpenAI OpenAPI commit adds ultrafast to the service-tier list. The API documentation currently limits the tier to eligible users and models, making latency a separate option for developers with access.

GOODREADS

---

How I’m Preparing for AI

​In the video from Chamath, he describes a personal “world model” made up of small agents for negotiation, psychology, portfolio and burn management, health tracking, and parenting. The core argument is that as AI handles more administrative work, human advantage will rely more on taste, judgment, and asking the right questions. It offers a practical way to think about building small agents around recurring decisions.

Why Compute Is Scarce

​Grayscale connects agent demand to electricity, data centers, chips, and grid access. The article estimates that a multi-step agent workflow can use 5 to 50 times more tokens than a typical chatbot interaction, then uses that estimate to build an investment case around scarce compute.

Long-Form Video With Agents

​Google Research describes a multi-agent “co-director” that plans and maintains visual consistency across multiple shots in a long video. The research uses agent coordination for continuity rather than only for code or tool calls. The article presents evaluation results from the research team itself.