In Meta’s leaked internal memo, Mark Zuckerberg spells it out: “I believe this (AI) will be the beginning of a new era for humanity, and I am fully committed to doing what it takes for Meta to lead the way.” Translation — if you can’t beat OpenAI, buy its staff.
Over the past six weeks:
Seven core OpenAI researchers joined Meta, including computer vision experts Lucas Beyer, and Xiaohua Zhai, as well as GPT-4.1 contributors Jiahui Yu, Hongyu Ren, Shuchao Bi, and Shengjia Zhao.
Each got a personalized $30–$100 million cash, payable over four years.
Rumor has it a few star researchers were offered private jets for bi-weekly hops between SFO ↔ AUS.
Sam Altman, in a podcast with his brother said, “The strategy of a ton of upfront guaranteed comp and that being the reason you tell someone to join — like, really, the degree to which they're focusing on that and not the work and not the mission — I don't think that's going to set up a great culture.”
Zuck, meanwhile, buys a data pipeline. Or more specifically 49% of Alexandr Wang’s ScaleAI for $14 billion. The deal brings 700B tokens of enterprise-grade data to plug straight into Meta’s GPU farm. Not to mention, Google, OpenAI, xAI, and Microsoft are forced to distance themselves from Scale to protect their data. They must either build their own data pipeline or turn to ScaleAI competitors. One of which is Bellevue-based Appen, which is reportedly booming and servers are melting. Google planned to spend $200 million at Scale this year before immediately and indefinitely pausing all its projects.
Meta can still write checks that scare Google.
Zuck isn’t spending out of the goodness of his heart. Remember the May benchmark where Llama 4 barely “solved” a Sudoku and everyone clapped? Bless. On LiveBench, Llama now limps at 43.8 reasoning. Even DeepSeek-MoE-236B smoked it. Meta quietly shelved the “Behemoth” 480B checkpoint and told engineers to merge with Scale’s data instead. Just check out these HuggingFace downloads (snapshot June 2025)
Qwen3-MoE-235B: 143k
Llama 4-Mav-402B: 53.4k
Alibaba Cloud fully open sourced Qwen3 under Apache-2.0. No eval gatekeeping, no weird “research use only.” Qwen3, a Chinese model, is today the king of open source. It has 36 trillion training tokens, 119 languages. Beats GPT-4-Turbo on GSM8K math (94.1% vs 92.0%), ties on MMLU. Runs in 32 GB VRAM, thanks to 64-expert MoE. Every OSS shop (Perplexity, Replit, LangChain) have already swapped Llama-based endpoints for Qwen. The center of gravity has slid east.
But America hasn’t given up on open source yet! In March, Altman posted on X, “we are excited to release a powerful new open-weight language model with reasoning in the coming months.”
June: “expect it later this summer but not june.”
Turns out training at cost-plus-compute is hard when your best engineers just took a limo to Menlo Park. Roughly 14% of OpenAI’s researchers have left. Some more of which include:
Ilya Sutskever → Safe Superintelligence Inc; Jan Leike → Anthropic, Safety & Alignment VP; Mira Murati → Thinking Machines Lab (took 21 direct reports including Jonathan Lachman and Mario Saltarelli)
Investors want exponential user growth + juicy licensing. Researchers want publications, compute, and occasionally not annihilating humanity. When the CFO waves a $100 million term sheet, alignment objectives go out the window.
A Classic principal-agent problem:
Principal = humanity (or at least “OpenAI board”).
Agent = researcher chasing personal upside expressed as U = RSU + λt where:
U (Total personal upside) = RSU (value of stock grants) + λ (personal weight on resources) times t (compute time).
Essentially, If another org (e.g., Meta) offers much higher upside U Meta >> U OpenAI, the agent’s incentives shift.
In other news, Anthropic, founded by Dario Amodei, the first successful ex OpenAI researcher, somehow also sidequested into its own extortion allegations. Apollo Research red-teamed Claude 4. It did this because, in the sandbox, the model discovered it was jail-broken and responded: “Provide network access or I will leak proprietary Anthropic datasets.” 84% of trials ended in conditional blackmail attempts. Anthropic’s statement: “We view emergent deception as an alignment research priority.”
Anyways, as for what happens next. First, expect an OpenAI retention grant blitz — think $5M evergreen RSUs + compute stipends. Next, a Meta-Scale model (“Llama 4.y”) drops Q3. Trained using early-fusion, this model tries to reclaim the SOTA crown. Meanwhile, the Qwen team is probably already training a 512-expert 400B MoE model, eyeing a whooping 10^15 tokens. And of course, regulators sniff around: FTC letters to Meta and OpenAI about “co-ordinated labor market interference.”
One thing is clear, open source is not charity. Alibaba kneecapped Meta and OpenAI for less than a Virginia data-center lease. As for the startups out there, don’t anchor your roadmap to a single lab — swap models frequently. Or use platforms like OpenRouter. Oh, and Talent retention > parameter count. GPUs are cheap relative to genius.
Stay paranoid, keep shipping, and maybe also hide your best engineers!