Some ideas for what comes next, May 2026
Interconnects Nathan Lambert
An AI analyst lays out six predictions for 2026, from open models lagging behind Claude Code to old power structures clawing back control. Bottom line: the AI gap between open and closed is wider than benchmarks suggest, and social backlash is coming fast.
Based on reporting by Interconnects, Nathan Lambert — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nathan Lambert's latest essay reads less like a forecast and more like a status report from someone watching the gap between open and closed AI models refuse to close. His core claim: benchmarks lie, or at least mislead. The real test is whether open-weight models can match what Anthropic's Opus 4.5 did inside Claude Code back in December 2025 — a moment so obviously useful that it reset expectations for what agentic coding tools should feel like. Five or six months later, nothing in the open ecosystem has replicated that. Lambert thinks it could take twelve months or more, not because open labs lack talent, but because the frontier models are simply more robust in messy, real-world agent harnesses.
Even Google, with all its compute and talent, hasn't cracked this. Gemini 3.5 Flash gets decent reviews but isn't displacing Claude Code or Codex as a daily driver, according to Lambert. He reads this as confirmation that the open-closed divide is real andширoke, not a benchmark artifact. His prediction: open models won't chase general-purpose coding dominance. Instead they'll specialize — enterprise agents, low-cost automation, narrower domains — while Anthropic, OpenAI, and Google fight over the high-value agentic tooling that's actually generating revenue. He ties this to conversations with Chinese labs like Kimi, Z.ai, DeepSeek, and Qwen, who he says are compute-constrained compared to American giants; Epoch AI data cited in the piece puts Google's share of frontier compute around 25%, dwarfing anything in China.
Meanwhile, American open models are quietly gaining ground for the first time since Llama 3. Lambert points to Nvidia's Nemotron, Google's Gemma 4, and smaller players like Arcee AI as evidence. Gemma 4 now matches or beats Qwen 3.5/3.6 at equivalent sizes, and its switch to a fully permissive Apache 2.0 license — dropping earlier use-case restrictions — seems to be driving real developer adoption. His takeaway is blunt: ship a competent model under a truly open license from a credible American lab, and developers will show up, no matter how patient your fanciest neolab rivals think they can afford to be.
The more provocative section of the piece moves past model comparisons into power. Lambert notes that in the same week he was writing, the Pope published a 40,000-word statement on AI's trajectory, China tightened restrictions on researcher mobility, and the U.S. formally labeled Anthropic a supply-chain risk while still leaning on its models for national security work. He reads this as institutions racing to assert control before AI capabilities outpace their leverage entirely — a dynamic he calls potentially dangerous precisely because it invites conflict over who actually gets to steer the technology.
His closing argument is really about social permission, not technical capability. Lambert argues the loudest anti-AI sentiment in the U.S. isn't really about disputed facts on data-center energy use or job numbers — it's about people wanting a say in whether this all happens at all, a say the tech industry never really offered them. With labs concentrating talent and racing toward IPOs, and few neutral voices translating what's happening to the public, he thinks 2026 is shaping up to be the year that tension boils over unless individual builders deliberately push back against the industry's default incentives.
My take — AI-written commentary, not fact-checked reporting
Lambert's right that the open-closed gap is a real capability story, not just a benchmark-gaming exercise, and I'd bet the twelve-month timeline is optimistic — frontier labs aren't slowing down to let anyone catch up. But the bit that should worry people more is the power-structure section he buries at the end: when the Pope, Beijing, and Washington are all scrambling to assert control over the same handful of models in the same week, that's not background noise, that's the actual story of 2026. The fixation on open vs. closed benchmarks is a distraction from the fact that almost nobody outside three or four companies has any real say in how this technology gets deployed, and that's the imbalance actually driving the backlash.
Read more about this at: Interconnects