AI science claims draw excitement and eye-rolls in equal measure
A post says Opus 5.5 agents discovered two room-temperature magnetic semiconductor candidates. The top reaction compared it to yet another promising fusion claim: unless it's experimentally verified, it isn't groundbreaking. Another commenter said they could have gotten it in one prompt. A third made the opposite argument: frontier LLMs hand almost anyone knowledge that once required hiring a multi-million dollar engineering effort.
The pattern: the debate has moved from 'can AI do research' to 'what counts as a result'. Candidates are cheap now. Verification is the scarce part. The same split showed up on a post about a 60-year-old developer reignited by Claude Code, where some feel like they are programming with buddies and others call the work unfulfilling.
The nuance is that both camps are right about different layers. Generation is nearly free and validation is still slow, physical, and expensive.
So what?
If you sell AI for science, engineering, or research, your moat is the verification loop, not the idea generator. Show experimental or real-world confirmation, because a list of candidates now reads as marketing.