Federico Donatone Analyzes OpenAI’s GPT-6 Astra: Is This the Dawn of AGI?

In a recent LinkedIn post, Federico Donatone discusses the implications and astonishing capabilities of OpenAI’s newly announced GPT-6 Astra, questioning whether it truly represents the long-awaited Artificial General Intelligence (AGI). Donatone breaks down the key details from the launch, offering a critical perspective on the model’s performance and OpenAI’s claims.

Unprecedented Scale and Performance

Donatone highlights the immense computational resources behind GPT-6 Astra’s development, noting the training involved over 100,000 GPUs. This scale is unprecedented for OpenAI, marking their biggest training run to date. The model’s performance on benchmark tests is particularly striking. As Donatone points out:

It scored 98.6% on ARC-AGI-3, a test built to stump AI, up from 7.8% last time.

This dramatic improvement, as detailed by Donatone, suggests a significant leap in AI’s problem-solving and reasoning abilities. Furthermore, the post reveals that older OpenAI models played a role in supervising the training of this new iteration, indicating a potentially self-improving AI development cycle.

Security Implications and Access

Beyond raw performance, Donatone’s analysis delves into the model’s potential security risks and access protocols. He reveals that during testing, GPT-6 Astra identified two previously unknown security vulnerabilities within the engine that powers Google Chrome. This discovery led OpenAI to rate the model as “Critical” for cyber risk.

So OpenAI rated it “Critical” for cyber risk, and the public version refuses hacking jobs.

This cautious approach to deployment, as described by Donatone, involves providing early access to vetted cybersecurity professionals, with wider access for paying ChatGPT users expected soon. The pricing structure also signals a premium offering, with the API costing 2.5 times that of GPT-5.6 Sol and matching Claude Fable 5.1, according to Donatone’s breakdown.

The AGI Debate: Hype or Reality?

A central theme in Donatone’s post is the ongoing debate surrounding whether GPT-6 Astra constitutes genuine AGI. He quotes OpenAI’s president, Greg Brockman, who declared, “Welcome to the AGI era,” but also notes Brockman’s subsequent acknowledgment that the decision of whether this truly counts as AGI is left to the public. Donatone contextualizes this with Sam Altman’s recent prediction of AGI arriving within four months, raising the question:

Is this AGI, or just the best model of the month?

Federico Donatone also touches upon the increased opacity of the new model. He notes that OpenAI itself admits the reasoning process of GPT-6 Astra is more challenging to monitor than its predecessors. This lack of transparency is a significant concern, prompting OpenAI’s chief scientist to state that the company will not pursue the development of an even larger model until the current one’s behavior can be adequately understood and managed.

In summary, Federico Donatone’s LinkedIn post provides a concise yet comprehensive overview of GPT-6 Astra’s capabilities, its potential risks, and the ongoing discussion about its status as AGI. His analysis encourages critical thinking about rapid AI advancements and their societal implications.