top of page
Search

Voice of AI: 72 hours between "AGI era" and "nobody is prepared" ⏱️

On September 3, OpenAI President Greg Brockman opened the launch livestream with the words, " Welcome to the AGI era ." On September 6, the company's chief scientist wrote that it was a time that demanded extreme caution and that he was concerned no one was prepared for the consequences. In between lie three days and a model that OpenAI itself, for the first time, classifies as critical for cybersecurity. Both are true simultaneously. That is the news.



Three days between two truths

On September 3rd, OpenAI released GPT-6 Astra and simultaneously revealed that it is the first proprietary model to achieve the "Critical" level of cyber capabilities under its in-house Preparedness Framework. In practical terms, this means that with the appropriate tools and access, the model can find previously unknown vulnerabilities and develop ways to exploit them in well-protected systems without human intervention.


Tested without production protection, it achieved 100 percent on the company's proprietary exploit benchmark, compared to 78.5 percent for its predecessor, and discovered two previously unknown zero-day vulnerabilities during testing. Access will be rolled out gradually and is disabled by default for enterprise users.

Three days later, Jakub Pachocki, Chief Scientist at OpenAI, published an essay on the company website titled "An Alien Mind." His central argument: In his current opinion, no laboratory has solved alignment and monitoring to such an extent that responsible scaling at maximum speed can be justified for much longer . He expects and hopes that voluntary slowdowns will become commonplace until common safety thresholds are established. And he is concerned that no one is prepared for the consequences of a continued rapid increase in machine intelligence. Sam Altman shared the text and called it important.

This is not a contradiction in the communications department. This is an honest description of the situation.


The number that explains both

It's in the datasheet for the same model, and it's the most important number this week.

The controllability of the thought process has improved significantly, from 16.1 percent in the predecessor to 60.9 percent in Astra. However, the monitoring capability has plummeted in the same model. If Astra is instructed to circumvent oversight, the pure chain-of-thought monitor detects less than 11 percent of cases. The predecessor achieved nearly 100 percent. This is due to an architecture with recurrent depth, whose internal traces are more difficult to read.

We are better at controlling these systems, but less able to observe them. Pachocki identifies three reasons for this: Thought processes are increasingly intertwined with communication, which must be monitored. Models are becoming better at reflecting on and manipulating their own thought processes. And improved pretraining makes them smarter, even without verbalized thought.

This has an inconvenient consequence for your company. If monitoring a model from within becomes more difficult, control must come from the outside: defined permissions, logged calls, shutdown conditions with designated responsibilities. None of this is elegant, but all of it is verifiable. That's precisely what ARGUS , our Technology Due Diligence, is all about.


An important point to note: All performance figures mentioned come from OpenAI itself, and according to the company, the evaluations were run at maximum computing power. This doesn't make them incorrect. It makes them manufacturer specifications.


302 companies from a single university

From San Francisco back to Vancouver, where this series originated.

On August 31, Innovation UBC reported that the University of British Columbia has now spun off 302 companies, 15 of them in the current year alone. Together, these companies have raised more than $11 billion in capital, created over 2,800 jobs, and, according to the university's estimates, generated around $13 billion in revenue.

The interesting part isn't the number. It's the leverage behind it. Eight of the 15 new spin-offs used the UBC Express License Agreement, a standardized, ready-to-sign license agreement that was revised in June 2026 to make the process more efficient and founder-friendly. The university explicitly attributes the record year to this.

Not a billion-dollar program. A better contract.


That's the key takeaway from four weeks: Structure beats subsidies. And structure costs almost nothing, except the willingness to seriously examine your own processes. That's precisely where the AI Readiness Check at readiness.oakai.de comes in: not by asking which tool you need, but by asking what's preventing you from using it within your organization.


What the weeks in Canada have shown

Three figures have driven this series of events. Around 600 citizen petitions that postponed a data center hearing beyond a local election. 400 megawatts of data center power for two years, for which 15 applicants submitted bids totaling around 800 megawatts. 64 internationally appointed leading researchers, 48 of whom come from the USA and exactly one from Germany.

Each time, the core issue was the same: not whether there is enough technology, but whether there is enough judgment to distribute it.

I came here with one theory and I'm leaving with another. My assumption was that Canada simply had the better prerequisites: plenty of electricity, plenty of land, plenty of capital. What I saw instead is less spectacular and considerably more uncomfortable: it's primarily the processes. A simplified licensing agreement that accelerates spin-offs. A public procurement process that makes scarcity visible instead of merely managing it. An appointment practice that recruits internationally without having to debate it fundamentally every time. None of this is expensive, and none of it is tied to Canada. That's precisely why I'm pursuing this further. Ralph Schwehr


And the 400 megawatts?

In the last issue, I promised that we would know today who won the contract. We don't know.

BC Hydro continues to cite mid-September as the target timeframe. The deadline for withdrawing without forfeiting bid security expired on September 1st, but no decision has been made yet. I will provide an update as soon as the commitments are received. Procedures that seriously address scarcity take longer than those that ignore it.


💡 Key findings

  • Controllability and monitoring diverge. Control over the thought process increases from 16.1 to 60.9 percent, while the detection rate for conscious circumvention drops below 11 percent.

  • The call for slowing down comes from the center, not the periphery. OpenAI's chief scientist, three days after the firm's most powerful model.

  • External control replaces internal observability. Permissions, logs, shutdown conditions. Inelegant, but verifiable.

  • Structure beats subsidies. 302 spin-offs from one university, and the accelerator was a simplified licensing agreement, not a funding program.

  • Manufacturer specifications remain manufacturer specifications. All performance figures for Astra are from OpenAI, measured at maximum computing power.

🔧 To try: three steps for this week

  1. Select a process that takes too long in your organization and count the number of approvals involved. Don't evaluate, just count.

  2. For each agentic system, a shutdown condition should be recorded in writing: which signal, who is observing, who is allowed to stop.

  3. When asking about provider figures, inquire about the measurement setup. What settings are used, what protection mechanisms are employed, and against which baseline?


📚 Sources


Conclusion

My time in Canada was over, and yet the most important news came from San Francisco. That's no coincidence, it's the punchline.

Because both stories are about the same thing. In Vancouver, it becomes clear how much dynamism is created when friction is removed and scarcity is openly distributed. In San Francisco, it becomes clear what happens when capability grows faster than the ability to observe it. Both are questions of process, not technology.


For your company, this means: Waiting for an industry slowdown is not a strategy. Building your own capacity for control is. And the gap we in Europe need to close is not a knowledge gap. It's a decision gap.


Where does your company stand? Start with the AI Readiness Check at readiness.oakai.de or write to me at info@oakai.de .

All services: oakai.de/services

CLARITY INSTEAD OF HYPE. OAK AI



 
 
 

Comments

OAKAI®

E-Mail: info@oakai.de

NEWSLETTER

Sign up for the OAKAI newsletter.

  • LinkedIn

© 2024 OAKAI®

Imprint

Data protection

Thanks for subscribing!

bottom of page