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

In the rapidly evolving world of artificial intelligence, few voices carry as much insight and authority as Dylan Patel, the founder and CEO of SemiAnalysis. With unparalleled expertise in AI and semiconductors, Dylan offers a deep dive into the current chaos and breakthroughs within the top AI companies. This article distills his extensive analysis and commentary on industry giants like OpenAI, Meta, Apple, Microsoft, and the rising challenger, xAI, shedding light on the technological, strategic, and organizational dynamics shaping the race toward superintelligence.

Table of Contents

• 🔍 Meta’s LLaMA Struggles: What’s Happening Behind the Scenes?
• 💰 Meta’s Scale AI Acquisition and $100 Million Offers: Power Play in Talent Wars
• 🧩 OpenAI and Microsoft: A Complex Partnership with High Stakes
• 🚀 The Rise and Fall of GPT-4.5: Lessons from Orion
• 🍏 Apple’s AI Conundrum: Conservative Culture and NVIDIA Beef
• 📱 On-Device AI vs. Cloud AI: The Tradeoffs and Realities
• ⚔️ NVIDIA vs. AMD: The Battle for AI Chip Supremacy
• 🤖 xAI and Grok: Elon Musk’s AI Ambitions
• 💼 AI and the Future of Work: Will 50% of White-Collar Jobs Disappear?
• 🔓 Open Source vs. Closed Source in AI: Which Will Prevail?
• 🏁 Who Will Reach Superintelligence First?
• 📚 Frequently Asked Questions (FAQ)
• Conclusion

🔍 Meta’s LLaMA Struggles: What’s Happening Behind the Scenes?

Meta’s journey with its LLaMA models has been a mixed bag. While LLaMA 2 generated excitement in the AI community, the anticipation around LLaMA 4 and the much-delayed Behemoth model has been tempered by underwhelming results and strategic missteps.

Dylan explains there are actually three different models in play at Meta: Behemoth, Maverick, and Scout. Behemoth, the largest and most ambitious, was delayed and might never see the light of day due to fundamental issues in its training approach. Specifically, Meta’s attempt to emulate DeepSeek’s architecture with a focus on sparsity and mixture of experts (MoE) backfired. The routing mechanism within the model failed to effectively utilize many of the experts, resulting in wasted training effort and poor overall performance.

Meta’s technical challenges aren’t simply about compute or talent. Dylan highlights that despite having some of the best researchers and abundant resources, the organization struggles with decision-making and leadership. Without a clear technical leader to evaluate and select the best research paths, the company risks pursuing flawed ideas that consume valuable time and resources.

This organizational bottleneck contrasts sharply with OpenAI’s structure, where CEO Sam Altman works closely with technical leaders like Greg Brockman and Mark Chen to steer research directions decisively. The presence of a strong technical lead helps OpenAI avoid the trap of chasing unproductive research avenues and ensures a sharper focus on breakthroughs that matter.

💰 Meta’s Scale AI Acquisition and $100 Million Offers: Power Play in Talent Wars

Meta’s recent acquisition of Scale AI, spearheaded by Alex Wang and his team, signals a strategic shift toward aggressively chasing superintelligence. Dylan describes Scale AI as “kinda cooked” as a company due to clients like Google pulling back on contracts. However, the acquisition wasn’t about Scale’s current business but about securing top talent and leadership for Meta’s AI ambitions.

Alex Wang, a young yet highly successful figure in AI, represents the kind of leadership Meta wants to harness to catch up in the superintelligence race. Meta’s CEO, Mark Zuckerberg, who previously downplayed the imminence of AGI, is now fully “superintelligence pilled” — convinced that achieving superintelligence is the ultimate goal and that Meta must compete aggressively.

This talent war extends beyond acquisitions. Meta has reportedly offered $100 million bonuses to top AI researchers to retain and attract talent. Dylan points out that while money is a factor, many top researchers are motivated by the power and influence that comes from shaping the AI future at a trillion-dollar company like Meta, where they can directly influence products used by billions.

Despite these efforts, Dylan is skeptical of Sam Altman’s claim that no top researchers have left OpenAI for Meta. He believes some have indeed defected, attracted by the promise of greater control and impact at Meta. This dynamic underscores the high-stakes competition for AI talent and leadership that will shape the industry’s future.

🧩 OpenAI and Microsoft: A Complex Partnership with High Stakes

The relationship between OpenAI and Microsoft is foundational to the AI ecosystem but fraught with complexity. Dylan describes their partnership as “therapy” — a relationship with ups and downs, where both parties have significant leverage and concerns.

Microsoft’s investment and deal with OpenAI are unique and intricate. OpenAI, originally a nonprofit, gave up significant control and profits to secure Microsoft’s funding. The deal includes revenue sharing, profit caps, and Microsoft holding IP rights to OpenAI’s technology until AGI is achieved. This arrangement creates tension, especially around the nebulous definition of AGI and the control over intellectual property.

One major sticking point was Microsoft’s demand for exclusivity as OpenAI’s compute provider, which slowed OpenAI’s access to necessary compute resources. This exclusivity was eventually lifted, allowing OpenAI to diversify its infrastructure partners to companies like Oracle and CoreWeave, improving agility but complicating the partnership.

Dylan warns that Microsoft’s vast resources and legal power could pose risks to OpenAI’s independence, especially as they hold rights to all IP up to the point of superintelligence. This means Microsoft could potentially leverage OpenAI’s breakthroughs for its own benefit, raising questions about the long-term dynamics of this critical alliance.

🚀 The Rise and Fall of GPT-4.5: Lessons from Orion

GPT-4.5, internally codenamed Orion, was envisioned as the next giant leap in AI, a massive model trained on full-scale pretraining with enormous compute resources. However, the model’s release was met with disappointment, leading to its deprecation.

Dylan explains that while GPT-4.5 was indeed smarter than its predecessors and even made him laugh, it suffered from being too slow, expensive, and ultimately less useful than smaller, more efficient models like GPT-3.

The core issue was overparameterization without sufficient data scaling. The model memorized vast amounts of data early in training but failed to generalize effectively due to a lack of diverse and voluminous training tokens. Compounding this was a subtle bug in the training code that went undetected for months, hampering progress and stability.

Meanwhile, another OpenAI team discovered that enhancing reasoning capabilities through techniques like “strawberry” (a code name for reasoning improvements) could generate high-quality synthetic data, improving model efficiency without massive scale. This breakthrough highlighted that more parameters alone don’t guarantee better performance — quality and quantity of data, especially verifiable data, are crucial.

🍏 Apple’s AI Conundrum: Conservative Culture and NVIDIA Beef

Apple’s AI efforts appear to lag behind competitors, with little public progress or major acquisitions. Dylan attributes this to Apple’s conservative corporate culture, secretive approach, and difficulties attracting top AI talent who prefer open research environments like OpenAI or Anthropic.

Apple’s history of small acquisitions contrasts with the large-scale talent buys by Meta and others. Their secretive nature discourages researchers who thrive on publishing and collaboration. Additionally, Apple’s strained relationship with NVIDIA — stemming from a hardware failure known as “bump gate” and patent disputes — limits their access to leading GPUs and infrastructure.

Bump gate was a hardware issue where solder balls connecting NVIDIA GPUs to motherboards cracked due to thermal expansion differences, causing failures. This incident, combined with NVIDIA’s aggressive patent enforcement and failed mobile chip ambitions, has soured Apple’s relationship with the GPU giant, further handicapping their AI hardware capabilities.

📱 On-Device AI vs. Cloud AI: The Tradeoffs and Realities

Apple champions on-device AI for security, latency, and privacy. However, Dylan takes a more pragmatic view, skeptical about the widespread adoption of on-device AI for complex tasks.

He argues that while security is important, most users prioritize free and convenient services over privacy. Additionally, hardware constraints on mobile devices limit the size and speed of AI models that can run locally. For many AI applications—like search, calendar management, and personalized recommendations—the data is already in the cloud, making cloud-based AI more practical and powerful.

Dylan envisions on-device AI being useful for low-value, latency-sensitive tasks such as predictive text or image recognition on wearables like smart glasses or earpieces. However, the heavy lifting of reasoning and complex workflows will remain in the cloud, where vast compute resources and integrated data access enable more sophisticated AI services.

⚔️ NVIDIA vs. AMD: The Battle for AI Chip Supremacy

The AI chip market is dominated by NVIDIA, but AMD is making aggressive moves to capture market share. Dylan notes that while AMD’s hardware is generally behind NVIDIA’s latest Blackwell architecture, it does have strengths in certain areas and offers competitive pricing.

AMD’s biggest challenge is its software ecosystem, which lags behind NVIDIA’s mature CUDA and related toolchains. Researchers and developers prefer NVIDIA because of its seamless integration with popular AI frameworks like PyTorch and the extensive libraries that simplify deployment.

Interestingly, NVIDIA’s strategy includes supporting numerous smaller cloud providers beyond the big players like Amazon and Google. This expands NVIDIA’s reach but has also led to tensions, especially after NVIDIA acquired Lepton, a cloud software layer that competes directly with some of these cloud providers.

AMD has responded by building strong partnerships with these smaller cloud companies, sometimes engaging in arrangements where GPUs are sold and rented back to foster goodwill and market penetration. While AMD may not dramatically disrupt NVIDIA’s dominance soon, it is poised to steadily grow its presence, especially where price and software improvements align.

🤖 xAI and Grok: Elon Musk’s AI Ambitions

Elon Musk’s xAI and its Grok models have generated buzz, with claims of unprecedented intelligence and first-principles operation. Dylan appreciates Elon’s engineering and marketing prowess but remains cautious about the hype.

Grok 3 surprised many by outperforming expectations, particularly in understanding complex historical, geographic, and economic topics. Dylan sometimes uses Grok for deep research queries because it can bypass some of the “pansy” filters other models impose, offering more candid insights.

However, Dylan’s daily go-to models remain OpenAI’s GPT-3 and Claude 4, with Gemini also playing a role in workplace applications like long-context document analysis. Grok’s advantage lies in its access to real-time data and current events, making it useful for summarizing complex and evolving scenarios like geopolitical conflicts.

Overall, while xAI has strong compute resources, talented researchers, and ambitious plans—including building new data centers and solving power constraints—the foundational AI techniques they use are broadly similar to other leading labs. The superintelligence leap will likely require breakthroughs beyond existing transformer and reinforcement learning paradigms.

💼 AI and the Future of Work: Will 50% of White-Collar Jobs Disappear?

The prospect of AI automating half of white-collar jobs has sparked widespread concern. Dylan offers a nuanced perspective, considering historical trends and demographic shifts.

He points out that people today work fewer hours than they did 50 or 100 years ago, with improved living standards and more leisure time. AI could accelerate this trend, enabling humans to work less while maintaining or increasing productivity.

Yet, the distribution of resources remains a critical challenge. While AI automates certain creative and knowledge-based tasks, many manual jobs—like fruit picking—remain difficult to automate fully, at least for now. Robotics holds promise for addressing these gaps, but widespread deployment is likely years away.

Dylan anticipates a transition where AI assistants handle tasks over longer horizons—minutes, hours, or even days—before humans review outputs. Eventually, fully autonomous AI systems may operate without human oversight, though timelines remain uncertain. He estimates that by the end of this decade or early next, around 20% of jobs could be automated.

For junior software developers, the landscape is already challenging. AI tools have boosted productivity dramatically, reducing the demand for entry-level coding jobs. Companies prefer senior engineers who can supervise AI tools effectively rather than large teams of juniors. Aspiring developers must adapt by acquiring new skills and demonstrating self-starting capabilities to remain competitive.

🔓 Open Source vs. Closed Source in AI: Which Will Prevail?

The AI ecosystem is split between open-source models and proprietary, closed-source systems. Dylan argues that unless Meta significantly improves, the U.S. risks losing ground in open-source AI development.

China currently leads in open-source releases, driven partly by being behind the curve. Once Chinese companies pull ahead, they are likely to shift toward closed source to protect competitive advantages.

Dylan believes closed-source AI will ultimately dominate, especially as the technology becomes central to economic output. His hope is that the market remains diverse, with multiple companies and models contributing, rather than consolidating into a few monopolies.

🏁 Who Will Reach Superintelligence First?

When asked to pick a winner in the race for superintelligence, Dylan confidently chooses OpenAI. He cites their track record of leading every major breakthrough, including reasoning capabilities that have transformed the field.

Anthropic follows as a strong contender, with Google, xAI, and Meta in a near three-way tie behind them. Dylan acknowledges that OpenAI has become less conservative recently, speeding up their release processes and expanding their team with more “normies”—researchers focused on practical applications.

The race is far from over, and the dynamics will continue to shift as companies invest billions in compute, talent, and infrastructure. But OpenAI’s leadership in innovation and strategic execution currently positions them at the forefront.

📚 Frequently Asked Questions (FAQ)

What went wrong with GPT-4.5?

GPT-4.5 was overparameterized without enough training data, leading to memorization rather than generalization. A subtle bug during training further hampered its progress. It was also too slow and costly compared to smaller, more efficient models.

Why is Meta struggling with their LLaMA models?

Meta faces organizational challenges with leadership and decision-making, leading to pursuing flawed research directions. Technically, models like Behemoth suffer from training inefficiencies, especially with mixture of experts routing issues.

How is Apple’s approach to AI different?

Apple focuses on on-device AI for privacy and latency but struggles with attracting top AI talent and lacks access to leading GPU hardware due to strained relations with NVIDIA. Their conservative culture slows progress compared to more aggressive competitors.

Will on-device AI replace cloud AI?

On-device AI will complement but not replace cloud AI. It is suited for low-value, latency-sensitive tasks on wearables and mobile devices. Complex AI tasks requiring large models and integrated data will remain cloud-based.

Is AMD a real threat to NVIDIA in AI chips?

AMD is making progress with competitive pricing and improving software support, but NVIDIA’s superior hardware, networking capabilities, and mature software ecosystem keep it dominant. AMD will gain some market share but not surge dramatically soon.

What is the future of jobs with AI automation?

AI will automate a significant portion of jobs over the next decade, especially white-collar tasks, but humans will likely work less overall. The transition will be gradual, with new roles emerging and a need for workers to adapt skills.

Who is leading the race to superintelligence?

OpenAI is currently the leader due to consistent breakthroughs and strong technical leadership. Anthropic, Google, xAI, and Meta are strong competitors, but OpenAI’s innovation pace gives them the edge.

Will open source or closed source AI dominate?

Closed source AI is expected to dominate as competition intensifies and companies seek to protect intellectual property. However, open-source efforts will continue to play a vital role in innovation and accessibility.

Dylan Patel’s comprehensive analysis reveals a complex and competitive AI landscape. From the technical missteps of GPT-4.5 to Meta’s aggressive talent acquisitions, Apple’s conservative struggles, and the intricate OpenAI-Microsoft partnership, the race toward superintelligence is both a technological and organizational challenge.

As AI continues to reshape industries and societies, understanding these dynamics is crucial for investors, technologists, and policymakers alike. The future will likely be shaped by who can not only innovate but also organize, fund, and scale effectively. For now, OpenAI remains the front-runner, but the field is wide open, with many players vying for the ultimate prize: superintelligence.