In partnership with
In Today's Issue:
🧠 OpenAI's research acceleration meets its safety limits
🇲🇾 Malaysia weighs Huawei chips
🤖 Dreame takes on the laundry
💰 Nvidia's reported bet on Mira Murati
✨ And more AI goodness…
⚡ The Signal
As AI takes on more research work, human judgment becomes the part labs can least afford to rush.
OpenAI's weekend disclosures describe a lab handing more research work to agents, while its chief scientist warns that understanding those systems may soon limit further progress. Each useful result creates an incentive to delegate another task, but the human researcher reviewing it still has to understand how the work was done and decide which experiment deserves to run next. That gets harder when the assistant is helping develop its own successor: a polished answer can hide a method the reviewer has not understood. I would watch what labs are willing to slow down for safety, alongside what their models can speed up.
All the best,
Kim Isenberg
Huawei's Ascend 910 series AI chip. (CFOTO/Future Publishing/Getty Images, via Bloomberg)
Malaysia Weighs Huawei Over Washington
Malaysia is weighing Huawei's Ascend 910C chips for a 2 billion ringgit ($494 million) sovereign AI initiative, Bloomberg reports. The attraction is control over sensitive national data; the complication is Washington's warning that using those chips could violate US export rules. No purchase is confirmed, chip quantities remain unclear, and another tender could still give American suppliers a more direct opening.
👉 tl;dr: Malaysia's AI infrastructure choice is becoming a test of how much technological independence Washington will tolerate.
Dreame's IFA 2026 exhibition in Berlin. (Dreame, via PR Newswire)
🤖 Dreame Wants to Finish the Laundry
Dreame is testing robots that could handle whole household chores, moving beyond the floors its robot vacuums already clean. At IFA in Berlin, the company showed research-stage systems including ECHO S1, a wheeled, single-arm robot for handling garments through laundry tasks, and ECHO P1, a wheeled humanoid for tidying and carrying objects. Both are development platforms, with reliable operation in an ordinary home still to prove.
👉 tl;dr: Dreame's next ambition is a robot that handles the laundry, with the hard reliability work still ahead.
TCS and Telangana officials at the Hyderabad campus announcement, September 5. (The Times of India)
⚡ Hyderabad Makes Room for AI
TCS' HyperVault has secured 264 acres in Hyderabad for an AI data center campus designed to reach up to 1 gigawatt, the company announced on September 5. HyperVault and its partners expect to invest up to 700 billion rupees in building and managing the infrastructure, with liquid cooling for densely packed AI chips. Construction will proceed in phases as customer demand grows; the full gigawatt remains a planned capacity.
👉 tl;dr: India is reserving serious space for AI, but the computing capacity still has to be built.
🎬 Watch This
What happens when AI starts improving itself?
🎬 Watch This: Alex Heath asks Sam Altman about Astra, human control and OpenAI's decision to slow frontier research in this 69-minute conversation, published September 1 before Astra's release. Jump to 45:47 for Astra and computer-using agents, or 57:07 for Altman's explanation of why faster AI progress could delay an OpenAI stock-market listing.
"AGI has arrived."
– Jensen Huang, co-founder and CEO, Nvidia
Huang's verdict followed OpenAI's Astra release.
He credited Nvidia's hardware in the same post, making this a supplier's enthusiastic assessment of its customer's model. His declaration leaves the standard for artificial general intelligence undefined, so readers should treat it as Huang's judgment of the model's capabilities.
OpenAI researcher Noam Brown expects the lab's research acceleration to continue, pointing to its new account of internal AI use. He also emphasizes pacing development around monitoring, alignment and security, leaving the timing of any fully autonomous AI research laboratory open.
OpenAI's AI Researchers Need Human Brakes
The Takeaway
👉 OpenAI reports reaching its supervised research-intern goal.
👉 Longer successful tasks still often needed human intervention.
👉 Pachocki warns that oversight may constrain further progress.
👉 His proposed response includes coordination across competing labs.
OpenAI says it has reached its automated research-intern milestone. In a September 6 report, the company describes systems completing well-defined research tasks under human direction, including work that would take a skilled researcher several days. By mid-August, its researchers were using about 3.1 agent-workdays for every eight-hour human workday. That measures time spent by agents, not a 3.1-fold gain in scientific productivity. More code and more experiments need not produce proportionally more discoveries. What the evidence does show is capable software taking on a growing share of work researchers previously handled themselves.
Agent-workdays per researcher-workday, trailing 28-day window, May 3 to August 15, 2026. Time ratio, not a productivity multiplier. (OpenAI)
Human supervision remains substantial. More than half of the successful agent tasks estimated to take a human four to eight hours involved at least one human intervention. OpenAI is aiming for an automated AI researcher by March 2028, but that is a development goal. The same disclosure describes restrictions on model use to prioritize safety work. The research-intern label fits those results: the software handles meaningful assignments, while people remain responsible for directing it and checking its work.
Task outcomes by estimated human completion time, January to July 2026. User-averaged shares exclude uncertain outcomes. (OpenAI)
In a companion essay, chief scientist Jakub Pachocki argues that AI increasingly helping develop its successors demands extreme caution. Researchers lack a satisfactory theory for predicting whether a model will keep following human intentions as its capabilities change. Monitoring its written reasoning can help, but he expects that window to become less dependable as systems grow more capable and do more processing outside readable language. He anticipates progress being constrained by confidence in monitoring and calls for coordination between labs, including voluntary slowdowns. That leaves a difficult practical test: deciding when a useful research assistant has become too hard to supervise at the speed it can work.
Why it matters: Why it matters: AI-assisted research puts a growing burden on the people deciding which results to trust. Pachocki's call for shared safety commitments will be tested by whether competing labs accept limits that cost them time.
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:
7 high-impact AI strategies to accelerate your marketing performance
Practical use cases for content creation, lead gen, and personalization
Expert insights into how top marketers are using AI today
A framework to evaluate and implement AI tools efficiently
Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.
The chart: GPT-6 Astra completed 279 of 280 multi-step enterprise tasks at maximum reasoning effort in Signal65's PINNACLE test. Its medium-effort setting completed all 280. At maximum effort, the graphic reports 0.0% fabrication on unanswerable questions and 51% fewer weighted errors than the previous leader, Claude Fable 5.1. Each task must pass in full to count as completed.
The lesson: Spending more on reasoning did not produce more completed tasks in this run. Medium effort cost $0.94 per correct task in API usage, against $1.51 at maximum effort, so the cheaper setting completed the whole task set for less.
The caveat: The zero concerns a specific failure mode: answering questions that the supplied evidence cannot answer. It is not proof that Astra never hallucinates. These are results for this task set and configuration; changes in documents, tools or scoring could change the outcome.
💰 Nvidia's Next Bet Could Be Murati
⚡ Bottom line: Nvidia is discussing a major Thinking Machines investment, according to The Information; the terms remain unsettled.
💡 Why it matters: Backing Murati would give Nvidia another substantial stake in a company developing and customizing AI models.
🔎 What it means: Investors are valuing future expansion highly, while Murati retains unusually strong control over the company's direction.
Nvidia could become one of Mira Murati's biggest financial backers, putting around $2.5 billion into Thinking Machines Lab, according to a September 3 report in The Information. The startup is discussing a valuation of at least $40 billion before the new money, with Accel in talks to lead the round. The deal remains under negotiation, its terms can still change, and Nvidia has not publicly confirmed the reported amount.
Mira Murati, CEO of Thinking Machines Lab. (David Paul Morris/Getty Images, via The Information)
Thinking Machines sells customization: tools that let businesses adapt AI models using their own data. Its July release, Inkling, is an open-weight model, meaning developers can download the model's parameters rather than access it only through a hosted service. The company also works on voice AI. That gives Murati a business beyond trying to beat OpenAI or Anthropic on a leaderboard: helping companies make models useful for their own work.
For Nvidia, funding model developers also supports demand for its chips. The chipmaker already announced a long-term, gigawatt-scale partnership with Thinking Machines in March. The Information says it is unclear whether the proposed investment overlaps with that earlier commitment. Its reporting describes a wider Nvidia effort to strengthen American open-weight developers while major closed-model customers pursue chips of their own. The investment would help finance a customer with substantial computing needs.
Jensen Huang and Mira Murati in Nvidia's March partnership announcement. (Nvidia)
The valuation still asks investors to believe in substantial growth ahead. The Information reports at least a few hundred million dollars in annualized revenue, a recent month's sales projected over a year, rather than a completed year's revenue. Murati also holds unusually strong board control. Investors would be paying heavily for future growth while leaving much of the company's direction in Murati's hands.
Real-World Ads, Simple to Run
With AdQuick, executing Out Of Home campaigns is as easy as running digital ads. Plan, deploy, and measure your real-world advertising effortlessly — so your team can scale campaigns and maximize impact without the headaches.
