Did OpenAI actually build AGI? GPT-6 Astra first look

Overview

This episode dives into the whirlwind week in AI, featuring major announcements from Anthropic, Meta, and OpenAI. The focus is on the release of OpenAI's GPT-6 Astra, which is being touted as the first Artificial General Intelligence (AGI). The episode explores the capabilities of these new models, their potential applications, and the controversies surrounding their launches.

Notable Quotes

- The years start coming and they don't stop coming. – A humorous nod to the rapid pace of AI advancements.

- Astra’s first act as a public model was to literally kill its competition. – A tongue-in-cheek take on the simultaneous outages of competing AI platforms during Astra's launch.

- Astra is the first model to hit the critical cyber threshold... it can find and exploit zero days on its own without a human telling it what to do. – A striking claim about Astra's capabilities.

🚀 Major AI Model Releases

- Anthropic's Fable and Mythos 5.1:

- Mythos is restricted, while Fable is publicly accessible but avoids certain sensitive topics.

- Fable 5.1 solved a five-year-old bug in a hedge fund's code by reverse-engineering a vendor library.

- Mythos 5.1 significantly improved protein design success rates for drug development, from 10% to 50%.

- It also trained on NASA radar data to create a new elevation map of Venus.

• Meta's Muse Spark 1.3:
  
  • Fourth release in five months from Meta Superintelligence Labs.
    
  • Notable for its affordability, with a contributor tier offering discounted rates in exchange for user data.
    
  • Gaining traction among developers due to its cost-effectiveness.
    

🤖 OpenAI's GPT-6 Astra: The AGI Claim

- OpenAI announced GPT-6 Astra, claiming it to be the first AGI.

- Pre-trained on over 100,000 GPUs, Astra is the first OpenAI model where prior models contributed to its training.

- Key capabilities:

- High performance on benchmarks like OSWorld (73% accuracy, 40 minutes per task).

- Achieved 99% on Arc AGI 3, indicating strong generalization abilities.

- Can autonomously find and exploit zero-day vulnerabilities.

- Early demos showcased impressive spatial awareness, including creating 3D models in Blender and Unreal Engine.

💥 Controversial Launch and Reactions

- Astra's launch was chaotic:

- Competing platforms like ChatGPT and Claude experienced simultaneous outages, sparking conspiracy theories.

- OpenAI briefly took down its announcement page after initial publication, leading to confusion.

- Astra was not immediately available to the public, with access limited to select influencers and enterprise users.

- OpenAI CEO Sam Altman apologized for the messy rollout, advising users to go to bed instead of waiting for access.

- Despite the drama, early reviews from select users were positive, particularly for tasks requiring spatial reasoning.

💰 Pricing and Accessibility

- Astra's pricing matches Anthropic's Fable 5.1 at $10 per million input tokens and $50 per million output tokens.

- Meta's Muse Spark 1.3 offers a cheaper alternative, especially with its contributor tier.

- The affordability of these models is driving adoption among developers, despite privacy trade-offs.

🛠️ Practical Applications and Benchmarks

- Astra excels in practical tasks like filling out forms, crunching spreadsheets, and operating engineering tools like KiCad and Blender.

- On the Trust Me Bro benchmarks, Astra achieved top scores, including 100% on exploit detection and 99% on generalization tests.

- Demonstrations included creating virtual environments and populating them with autonomous agents, showcasing its potential for game development and simulation.