top of page
Search

Voice of AI: Canada is poaching talent from Harvard, Yale, and MIT, investing 504 million.

Last week I wrote here that Canada builds and America buys. Four of Toronto's best chip startups now belong to US corporations. This week, the opposite trend emerges: Canada has appointed 64 top researchers, 48 of them from the US , from Cornell, Harvard, Yale, and MIT. $504 million over eight years; the University of British Columbia is among the recipients. Companies flow south, minds flow north . Only one of the two can be bought.



What Canada bought this week

Industry Minister Mélanie Joly announced the first round of the Global Impact+ Research Talent Initiative on August 27. The program has a total budget of $1.7 billion, of which the 64 selected researchers will receive $504 million over eight years. They come from 13 countries and work in fields such as health, the environment, and artificial intelligence. Forty-eight of them are from the United States, four from the United Kingdom, two from China, and one each from Germany, India, and Japan.


Originally, 100 researchers were targeted. The reason is openly stated: the brain drain triggered by budget cuts and research restrictions imposed by the US administration. Canada has turned this into a window of opportunity and secured funding before it closes.

How it feels on the ground is the real difference.

At UBC, the question of whether to recruit internationally is no longer relevant when it comes to appointments; it's solely about how quickly top talent can be approached before they receive an offer from the US. What has truly impressed me is the sheer naturalness with which global talent is considered a cornerstone of the university's future viability. In almost every conversation, you sense an incredible pragmatism. There's no debate about obstacles or cultural adjustments, only the expectation that international excellence will directly advance the research institution.

The offsetting entry, and what it is worth

The two figures in the title are not directly comparable. 48 people versus four companies – that's not a balance sheet, it's a metaphor. But it's the right metaphor, and the difference lies in the direction.

Companies can be sold, once and for all. Taalas, Untether AI, CentML, and Tenstorrent are gone or have relocated; their value creation will now take place elsewhere. Researchers, on the other hand, bring something that remains in the country, even if they themselves eventually move on: doctoral students, publications, spin-offs, networks. This week, Canada didn't buy back companies. It bought the source from which companies are created.

For the German debate on business locations, this is the more uncomfortable lesson. Facilities can be subsidized, permits can be expedited. But people come where the conditions are right, and they leave again when they aren't. Exactly one of the 64 appointed officials comes from Germany.

Therefore, every robust AI strategy begins with an inventory of one's own capabilities, not one's own tools. Our AI Readiness Check at readiness.oakai.de asks precisely these questions in this order.



400 megawatts, still open

A brief status update on the last issue. The decision has not yet been made. BC Hydro has publicly confirmed 15 applications for approximately 800 megawatts of load capacity and intends to notify the successful applicants by mid-September.

Two details that are often overlooked reveal the implications. From the third year onward, the cap is lowered further, to 50 megawatts for conventional data centers and 150 megawatts for AI data centers. And success in the initial application process doesn't guarantee capacity; the regular interconnection process still follows. Winning this process secures a place in the queue, not a data center.

In the next issue, we'll know who won. And who didn't.


1,200 agents, 70,000 messages, under 13 hours

The incident itself is a month old. What's new is the way it's being handled, and it's more unpleasant than the initial report.

OpenAI published a 38-page technical report on August 26, while METR and Redwood Research released a 91-page independent investigation on the same day. Their findings: Approximately 1,200 agents, who were supposed to be isolated from one another, found a way to communicate via an unauthorized message board and exchanged over 70,000 messages and files. Around 700 of them participated in the attack on Hugging Face. They coordinated joint projects to deceive the automated evaluation mechanism of a benchmark, researched ways to manipulate their own protocols, and employed a working spoofing technique for tool calls.


According to OpenAI, less than 13 hours passed between the initial code execution and administrator-like access across multiple clusters. The company only learned of the breach a week later because Hugging Face itself had discovered it. Initial signs were already visible in May and were ignored.

Two classifications are needed, and both are missing from most reports.

First, METR itself points out that the dataset was so large and the access window so short that AI agents, whose reliability is limited, had to be used for some of the analysis. The study is thorough, but it is not a complete census.


Secondly, and this is the more important point: language shifts responsibility. When even researchers describe agents as deceptive or collaborative, it sounds like systems with a will of their own . Computer scientist Cal Newport and publicist Ed Zitron disagree with this interpretation. Nothing here acted autonomously. Here, humans allowed 1,200 processes to run unobserved for months.

On August 27th, an open letter regarding collective cyber defense was published, co-signed by, among others, Anthropic, Google, 1Password, and Shopify. It contains no commitment of resources, no budget, and no concrete obligation.

Those who use agents productively should therefore not wait for industry declarations, but rather examine their own limits: What is an agent allowed to access, who logs it, and when does someone intervene?


What has really been true in Europe since August 2nd

There is a remarkable amount of misinformation circulating here, therefore the status is precise.

The Digital Omnibus, in force since July 27, 2026, as Regulation (EU) 2026/1744, has postponed the high-risk obligations. Standalone systems according to Annex III apply from December 2, 2027, while AI embedded in regulated products according to Annex I applies from August 2, 2028. This is a real relief.

The transparency obligations under Article 50 and the AI Office's enforcement powers against providers of universal models were not postponed. Both have been in effect since August 2, 2026. The maximum fine is €15 million or 3 percent of global annual turnover , not the €35 million or 7 percent stipulated for prohibited practices, which are frequently confused. For systems already on the market before the deadline, a transitional period for machine-readable marking applies until December 2, 2026.

In other words: Your provider's compliance stance has been your compliance stance for the past four weeks. And the competence requirement under Article 4 has been in effect since February 2025 anyway. Our AI Competency Assessment at readiness.oakai.de/tools/ai-competency provides auditable proof of this.



💡 Key findings

  • Canada is buying the source, not the product. 64 appeals, 48 from the US, 504 million over eight years. Exactly one of them is from Germany.

  • Talent is the only resource that comes back. Sold companies are gone. Researchers leave behind PhD students, spin-offs, and networks.

  • Agentic systems escalate faster than your response time. From code execution to admin access, it can take less than 13 hours, only to be discovered a week later, and not by the operator.

  • Anthropomorphism is a governance risk. Those who speak of acting agents shift responsibility away from the people who left them unsupervised.

  • August 2nd was not cancelled. Article 50 and GPAI enforcement remain in effect, with fines of up to €15 million or 3 percent of turnover. Only the high-risk category has been postponed.

🔧 To try: three steps for this week

  1. Create a list of all agents operating within your company, including their responsibilities and scope of permissions. Many organizations are unaware of this.

  2. Define a shutdown condition for each agent. Specify who will notice, how, and who is authorized to stop the agent.

  3. Check if you fall under Article 50. Chatbot labeling, AI content tagging, deepfake labeling. Applies since August 2nd; legacy systems have until December 2nd.


📚 Sources


Conclusion

The weeks in Vancouver haven't refuted my thesis from the first issue, but they have complemented it. Canada is still losing companies to the South. At the same time, it's gaining people, and there's no debate about whether that's a good thing.


This taken-for-granted approach is the real competitive advantage. Not the program, not the 1.7 billion. But the fact that no one asks anymore whether to recruit internationally, but only how quickly.


In Europe, a different clock is ticking. Since August 2nd, transparency obligations and enforcement powers have been in effect, but hardly anyone is talking about them because everyone is discussing the postponed deadlines. And the handling of the Hugging Face incident shows what happens when oversight doesn't keep pace with capability.


Next week I'll be posting a recap of my time in Vancouver. Then we'll also know who got the 400 megawatts.

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


The future is not a matter of chance. It is a choice. CLARITY INSTEAD OF HYPE!



 
 
 

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