Why promise the moon on open AI when your track record screams ‘open-ish’?
Meta’s fresh hire, Alexandr Wang — the 19-year-old dropout who built Scale AI into a data-labeling behemoth — now runs its Meta Superintelligence Labs. He’s vowing to unleash a batch of the company’s AI models under an open-source license. Sounds democratizing, right? Especially from a firm that’s already dropped Llama, PyTorch, and React into the wild.
But hold up. Open-source heavyweights aren’t popping champagne. They’re squinting at the fine print — or lack thereof. No release date. Vague details. And Meta’s history of ‘community licenses’ that cramp commercial use.
Look, Meta’s open pedigree is legit. PyTorch powers half the ML world today. Llama’s ecosystem exploded, with 189,719 commits in 2024 alone — 71,018 from outsiders, per their engineering blog. That’s not chump change; it’s a thriving bazaar.
“Our open source codebases grew at an impressive pace, reaching 189,719 total commits in just one year – Meta Engineering.”
Yet the pendulum swings. Early Facebook openness morphed into walled gardens. Llama 3? Open weights, sure, but training data? Opaque. Restrictions on big-scale training? Sneaky guardrails.
Will Meta’s Alexandr Wang Deliver Real Open-Source AI?
Wang’s no rookie. Scale AI labels data for the world’s hungriest LLMs — OpenAI, Google, you name it. Joining Meta in June 2025, he’s got Zuckerberg’s ear on superintelligence. The pitch: Flood the zone with U.S.-built models for engineers everywhere. Counter DeepSeek’s Chinese surge, maybe.
Skeptics like Professor Amanda Brock, CEO of OpenUK, aren’t convinced. She’s seen Meta’s playbook.
“We need to understand what Meta is really planning to do here and what the company means by saying it will open-source the technologies,” Brock says. “If it’s a re-hash of the commercially restricted ‘Llama Community license’, then it’s not open-source according to any rational person’s understanding of the term.”
Bingo. That ‘community license’ caps users at 700 million monthly actives. Fine for tinkerers, poison for enterprises. Brock nails it: Open-source isn’t a buzzword; it’s OSI-approved licenses like Apache or MIT. No asterisks.
Jason Corso, AI prof at Michigan and Voxel51 co-founder, chimes in. Meta leads on open weights — those models spurred real innovation. But weights without full training recipes? Blind spots galore. Risks for misuse, provenance headaches.
“This creates risks for both Meta and model adopters, and it will be interesting to see how Meta addresses this problem differently,” says Corso.
Here’s my take, the one you won’t find in the press release: This reeks of Meta’s consumer-seeding playbook, straight out of 2011’s Open Compute Project. Back then, they open-sourced data center designs to lock in hardware ecosystems. Today? Flood consumers with ‘free’ models via Facebook, Instagram, WhatsApp. Anthropic and OpenAI chase enterprise bucks and gov contracts. Meta? Billions of eyeballs daily. Prediction: These models seed consumer AI dominance, making devs dependent on Meta’s stack — long-term moat disguised as generosity.
Data backs the strategy. Llama downloads hit 100 million in months. PyTorch? 70% of top ML papers use it. Wang’s move amplifies that flywheel.
But why now? Timing’s everything. OpenAI’s o1 secrecy fuels backlash. China’s models like DeepSeek-V3 crush benchmarks openly. U.S. export controls choke closed rivals. Meta positions as the patriotic open alternative — while Zuckerberg dodges safety suits.
Why Do Open-Source Leaders Doubt Meta’s AI Giveaway?
It’s the ‘openish’ specter. Delayed features. Stripped capabilities. Enterprises love flexibility, but governance nightmares follow — security, bias audits, all on you.
Vangala (an analyst quoted elsewhere) puts it sharp: Lower barriers, shape standards, breed infrastructure lock-in. Trade control for influence. Smart, if you’re Meta.
Wang’s calculus? Scale’s $14B valuation came from labeling closed models. Now flipping the script. But Zuckerberg’s developer heart — yeah, that guy’s still coding — clashes with boardroom caution. One lawsuit, and openness evaporates.
Zoom out to market dynamics. Open weights democratize inference, not training. Fine for fine-tunes, useless for from-scratch rivals. Meta keeps the real superintelligence crown.
Historical parallel? Think Linux vs. Solaris. Sun open-sourced Solaris late, half-baked. Linux won because Torvalds meant it. Meta’s Llama flirtations echo that — momentum, but no kernel-level commitment.
Enterprises? They’ll grab it for cost savings. 90% cheaper inference than GPT-4o. But provenance voids kill regulated use — finance, health. Devs rejoice short-term; compliance teams groan.
Wang could change that. Scale’s data moat means he knows clean pipelines. If he pushes full recipes, OSI-compliant? Game on. But bet on open-weights-plus. Safer for Meta.
Consumer angle seals it. Ina Fried at Axios: Wang eyes consumers while rivals enterprise. Billions of users tweaking models in feeds? Viral lock-in.
Bold call: By 2027, Meta’s consumer AI share triples, forcing OpenAI to ‘open’ scraps. Wang’s the catalyst — if he delivers.
The flip side. Fully open? Accelerationism risks. Unfettered models in wild hands. Meta’s guardrails might be mercy.
Still, pressure builds. Brock’s right — define terms. Wang, prove it.
🧬 Related Insights
- Read more: 17-Point AI Performance Gap from Bad Instructions — And the Tool Fixing It
- Read more: Gateway API on Kind: Local Testing Without the Hassle
Frequently Asked Questions
What models is Meta planning to open-source?
No specifics yet, but from Meta Superintelligence Labs under Alexandr Wang — likely advanced LLMs rivaling Llama 4.
Is Meta’s open-source AI truly free to use commercially?
Doubtful; expect Llama-style restrictions unless Wang pushes pure OSI licenses.
Why hire Alexandr Wang for this?
Scale AI expertise in data for top models; aims to counter closed rivals like OpenAI.
