Sunday, 7 June 2026

The gap between what AI companies claim and what actually works is widening fast, and I think we're starting to see the consequences in real time.

Take Microsoft's situation. They've released new MAI models at Build 2026 with considerable fanfare, but early testing suggests they're not living up to the hype. GitHub's troubles compound the problem. You can have the best AI research in the world, but if your products don't deliver tangible value to developers and enterprises, it doesn't matter much. Microsoft has the resources to iterate, of course, but there's a pattern here worth noting: the company that once dominated developer tooling is now scrambling to prove its AI strategy is more than marketing momentum.

By the way, contrast that with what Alibaba is attempting. Qwen3.7-Plus represents a genuinely interesting direction—taking multimodal AI and actually turning it into an agent capable of visual perception, GUI operation, and coding within a single loop. That's not flashy, but it's practical. The question isn't whether it's "revolutionary" but whether it actually solves problems for people building applications. That's where my skepticism kicks in: the real test isn't the benchmark, it's whether businesses adopt it at scale.

What's catching my attention, though, is the tension surfacing between capability claims and risk management. Anthropic is being unusually candid as it approaches a reported trillion-dollar IPO—warning that AI systems could soon achieve recursive self-improvement and that we risk losing control. I respect the honesty, but there's an uncomfortable contradiction embedded here. You're raising capital at unprecedented valuations while simultaneously raising alarms about existential risks. Markets and caution usually don't coexist well. Either the risk is real and priced in, or it's being downplayed for investor appetite. I'm not sure which concerns me more.

The robotics and video generation spaces tell a different story. Chinese humanoid robots are proliferating but remain largely performative—the demand and scale needed for real mass production just isn't there yet. Video generation, meanwhile, is maturing in the opposite direction. Kling, Gemini, and others have moved beyond unpredictable outputs; director-level control is becoming standard. That's meaningful progress.

What strikes me is that the winners in AI aren't always the ones with the biggest announcements. Google's work on research assistants—tools that help scientists generate hypotheses and analyze data—feels understated but genuinely useful. World Labs raising $1B on a spatial-intelligence platform suggests investors still believe in infrastructure plays, even if consumer-facing AI products are hitting friction.

The real question going forward: which of these narratives actually matter in two years? That's what I'll be watching.