SUMM

Meta's recent models (Llama 4, Behemoth, Maverick, Scout) had mixed results, with Behemoth facing training issues and likely never releasing.
Organizational structure and lack of clear technical leadership hamper Meta's ability to select and pursue the best research paths.
Meta's acquisition of Scale AI was primarily to obtain its leadership (Alexander Wang and team) to drive Meta's superintelligence efforts.
A shift has taken place at Meta, with Mark Zuckerberg now focused on superintelligence rather than near-term AGI.
Meta has attempted to acquire or hire talent from various top AI companies (SSI, Thinking Machines, Perplexity), aiming for both product and AI research leaders.
Highly lucrative offers (up to and over $100 million, even $1 billion) have been made to attract top researchers; some have accepted, contradicting claims otherwise.

The term AGI has lost specific meaning; "superintelligence" is now the main industry narrative.
Ilia Sutskever's SSI company catalyzed the superintelligence branding, influencing other firms to adopt the focus.
Meta and other major tech companies are heavily recruiting for superintelligence projects, frequently trying (and often failing) to acquire or poach high-profile AI talent.
True believers in mission and power, rather than just money, motivate many top AI researchers to join or stay at elite firms.

Microsoft and OpenAI's partnership is complex, involving unique revenue and profit-sharing agreements without clear ownership stakes.
Microsoft previously had exclusivity for OpenAI's compute; this ended due to antitrust concerns and OpenAI's need for faster infrastructure expansion.
Microsoft retains rights to all OpenAI IP and profits up to an AGI milestone, maintaining significant leverage in the relationship.
OpenAI expects to remain unprofitable for several years, targeting massive future revenue as its AI scales.

GPT-4.5 (Orion) was envisioned as a huge leap (possibly GPT-5), but suffered from slow speed, high cost, and low real-world utility.
Overparameterization led to memorization rather than generalization due to insufficient data, despite vast compute used.
Infrastructure issues and a small but significant bug in training code also hindered model performance.
OpenAI found greater improvements in reasoning-focused architectures and synthetic data generation, making brute-force scaling less viable.

Apple is seen as conservative, favoring small early-stage acquisitions and maintaining a secretive culture that's unappealing to top AI researchers.
Difficulty attracting AI talent is compounded by Apple's historical aversion to Nvidia (stemming from supply chain disputes and legal threats).
Apple prefers on-device AI for privacy and latency, but this is limited by hardware constraints and less appealing to most users compared to free cloud-based AI.
Apple's AI strategy is now shifting more towards cloud, building large data centers powered by its own chips, while continuing to invest in edge/on-device AI for specific low-value use cases.

AMD's new chips offer some advantages, but Nvidia remains ahead in hardware and especially in software and ecosystem support (e.g., CUDA, PyTorch).
Nvidia's networking capabilities (e.g., NVLink) allow tightly integrated large-scale training, giving it an edge for inference and training workloads.
Nvidia has supported smaller AI cloud providers to reduce dependency on big clouds like Amazon and Google, but is now competing directly with them via its Lepton acquisition, creating industry tension.
AMD pursues aggressive sales tactics, including renting back GPUs from clouds to build goodwill and demonstrate viability; some clouds are adopting AMD more for diversification and pricing pressure.
For most workloads and for best-in-class models, Nvidia remains the preferred choice, though AMD is gaining modest share for certain use cases.

Elon Musk's XAI and the Grok series have attracted significant resources and talent, with massive compute investments.
Grok 3 was better than industry expectations; claims for Grok 4 remain unverified, though optimism is high.
XAI is leveraging unique real-time data access (e.g., from X/Twitter) for current events and controversial queries, but faces challenges with data quality.
Fundamentally, most leading labs—including XAI—are pursuing similar approaches: large-scale transformer pre-training with RL in verifiable domains.

There is concern about massive displacement of white-collar jobs; up to 50% could be affected.
Despite fears, overall human labor hours have been falling for decades, with quality of life increasing.
Automation will likely first sweep through tasks with long time horizons and high routine, with AI assistants eventually operating independently for days or weeks before human feedback.
Robotics automation lags digital/creative automation; creative freelance markets (e.g., graphic design) are affected earlier than manual labor.
The future likely involves fewer entry-level tech jobs—especially for junior software engineers—unless self-starting or upskilled; most growth will be in leveraging AI for more ambitious projects.

The U.S. risks losing in the open-source AI race unless firms like Meta significantly improve their research and engineering efforts.
Open-source dominance (e.g., by China) is largely opportunistic; major players will restrict openness once leading.
Closed-source models are expected to dominate significant AI economic value, possibly leading to concentration among a few firms.
There is hope for more distributed leadership, but the trend favors major closed-source players consolidating power.

OpenAI is seen as most likely to reach superintelligence first due to its track record of being first to key breakthroughs.
Anthropic is considered second, having improved its operational agility.
Meta, XAI, and Google are viewed as contenders but lag behind; Meta is expected to rise if it successfully recruits top talent.

Dylan Patel: GPT4.5's Flop, Grok 4, Meta's Poaching Spree, Apple's Failure, and Super Intelligence