In partnership with
Dear Readers,
What happens when AI no longer just responds, but begins to anticipate questions? This week, new developments show how rapidly the boundary between tool and independent actor is shifting: benchmarks that measure human work, political battles over regulation, robots that “think before they act.” Each of these news items marks a piece of the future that is becoming reality faster than we want to admit.
This issue covers OpenAI's new performance benchmarks (GDPval), first insights into ChatGPT Pulse, DeepMind's robots with planning capabilities, and the growing political front: from Meta's Super PAC to global calls for red lines for AI. If you want to know how technology, business, and politics intersect here, you should definitely read on.
In Today’s Issue:
🤖 DeepMind's Gemini Robotics 1.5 is giving robots a brain upgrade
💸 Meta just launched a Super-PAC to fight against state-level AI regulations
🗳️ AI disinformation campaigns are flooding Moldovan elections
🚨 Over 200 global leaders are pushing for "AI Red Lines" to ban extreme AI uses
✨ And more AI goodness…
All the best,
OpenAI Launches GDPval Real-World Benchmark
OpenAI introduces GDPval, an evaluation of economically valuable, real-world tasks across 44 knowledge-work occupations (1,320 tasks; 220-task gold set open-sourced), built from expert-crafted deliverables and graded by professionals. Early results show frontier models approaching expert quality—Claude Opus 4.1 strongest on presentation, GPT-5 on accuracy—with major speed/cost advantages and a clear upward trend from GPT-4o to GPT-5. Current limits are one-shot, simplified scopes; next versions will add interactive, context-rich workflows, broader coverage, and a public grader to track progress.
OpenAI Unveils ChatGPT Pulse Preview
OpenAI introduces ChatGPT Pulse, a proactive daily feed that does overnight research and surfaces personalized, card-style updates based on your chats, feedback, and optional app connections. The preview rolls out to Pro users on mobile first, with curation controls, Gmail/Google Calendar integrations (off by default), and plans to expand to more apps and actions. It’s a step toward a more autonomous assistant that works ahead of you, not just when you ask.
GDPval Shows Rapid Human Parity
Experts defined hard, real-world tasks; different experts executed them, and independent graders compared human vs. AI outputs. Frontier models performed close to human level—and are improving fast—signaling real impact on practical knowledge work.
“The default way to code is vibecoding.”
OpenAI chief research officer Mark Chen (@markchen90) says high schoolers already see coding from scratch as weird.
If coding is vibes, what does great research look like?
— #a16z (#@a16z)
5:04 PM • Sep 25, 2025
OpenAI's Mark Chen says that “vibe coding” is now the standard way of coding. AI is causing major upheavals.
Rumor has it that both Gemini 3.0 and Claude 4.5 are about to be released. Hopes are pinned on Claude 4.5, although initial rumors suggest that no major leaps forward are to be expected.
The Takeaway
👉 DeepMind’s Gemini Robotics 1.5 joins an embodied-reasoning planner (ER 1.5) with an action controller, shifting robots from reactive scripts to plan-and-act agents.
👉 Agents can call tools (like web search) to ground decisions and execute long-horizon, real-world tasks end-to-end.
👉 Safety is foregrounded via updated ASIMOV tests that probe physical-world failure modes as capabilities scale.
👉 ER 1.5 is in API preview; the controller is limited to partners—early, but a meaningful step toward general-purpose robot competence.
Robots just got a brain upgrade. DeepMind’s new Gemini Robotics 1.5 stack pairs a high-level “embodied reasoning” model with a vision-language-action controller so machines can plan, look things up, and then execute multi-step tasks in the real world- without brittle scripts.
Here’s the split: Gemini Robotics-ER 1.5 is the planner that reasons about scenes, calls digital tools like Search, and breaks down long-horizon goals; Gemini Robotics 1.5 translates those plans into motor commands—and crucially “thinks before acting” to explain and refine its own steps.
In practice, that means a robot can sort waste by local rules, pack a bag after checking the weather, or separate laundry by color—end to end.
For builders, ER 1.5 is available in preview via the Gemini API today, while the action model rolls out to select partners; DeepMind also upgraded its ASIMOV safety benchmark to probe physical-world risks as these agents scale.
Why it matters: This is a step change from reactive robots to agents that reason, plan, and generalize across different bodies—critical for factory floors, home assistants, and logistics. It also signals a new developer workflow: wire up tools, set a “thinking budget,” and benchmark safety before you ship.
Sources:
Protect your checkout from coupon plug-ins. Boost your margin today.
KeepCart: Coupon Protection partners with DTC brands like Quince, Blueland, Vessi and more to protect your checkout from plug-ins like Honey, CapitalOne, RetailMeNot, and more to boost your DTC margins
Overpaid commissions to affiliates and influencers add up fast – Get rid of the headache and revenue losses with KeepCart.
After months of using KeepCart, Mando says “It has paid for itself multiple times over.”
Now it’s your turn to see how much more profit you can keep.
Meta launches Super-PAC against AI regulation
Meta has created the American Technology Excellence Project, a well-funded Super-PAC aimed at blocking state-level AI regulations. This marks a shift in the regulatory battlefield from federal oversight to local legislatures, with Big Tech itself becoming a direct political actor.
AI disinformation floods Moldovan elections
Ahead of Moldova’s parliamentary elections, researchers uncovered AI-generated campaigns using fake news sites, social media cascades, and destabilizing narratives. The stakes go beyond domestic politics: the struggle is over Moldova’s orientation toward the EU versus Russian influence.
Global “AI Red Lines” call intensifies
More than 200 leaders have urged binding international rules by 2026, banning extreme AI uses such as human identity simulation or self-replication. The push underscores a governance gap: global safeguards risk falling behind the accelerating pace of AI capabilities.
Were you surprised by ChatGPT Pulse?
