Will OpenAI Reach ASI First? Dylan Patel’s Bold Prediction

01 Meta’s Struggle and DeepSeek Copycat Failures

Dylan Patel explains that Meta’s recent AI projects flopped because they hastily copied DeepSeek’s MoE architecture without proper training, resulting in expert modules that never received tokens and wasted GPU resources. Despite having top talent and compute, Meta’s models underperformed and wasted billions.

02 Why GPT‑4.5 Failed

Patel describes GPT‑4.5 (codenamed Orion) as overly parameterized and slow, offering little practical value. Training suffered from a minor PyTorch bug, excessive checkpoint restarts, and a mismatch between model size and token data, violating the optimal parameter‑to‑token ratio highlighted by DeepMind’s Scaling Laws.

03 OpenAI and Microsoft: Former Partners Now Diverging

OpenAI’s early reliance on Microsoft’s Azure gave it a rapid rise, but the partnership is complex: Microsoft holds most profit rights and IP usage without equity, while OpenAI seeks other cloud providers. Both companies now guard each other, and the balance may soon break as OpenAI pursues massive funding and Microsoft retains priority purchase rights.

04 Apple’s AI Missteps

Apple lags behind in AI due to conservative acquisitions, a lack of open‑source engagement, strained relations with Nvidia, and an insufficient AI research culture, limiting its ability to attract top talent. Although it promotes on‑device AI, Apple is secretly building large data centers and investing in custom AI accelerators, ultimately betting on cloud AI rather than local inference.

05 Who Will Reach ASI First?

When asked which company will achieve ASI first, Patel confidently names OpenAI, citing its leadership in pre‑training, inference, and multimodal capabilities. He ranks Anthropic second—strong but cautious—and places Google, Meta, and xAI in a close third, emphasizing that the race is as much about resources and willpower as it is about technology.