as someone who's excited about ai-enabled science discovery, i find @AnthropicAI's framing of this very very confusing, and the technical report completely lacking in true scientific spirit and integrity @DarioAmodei @mgdurrant 1. first, credit where real credit is due: - Korn, Hillhouse, Sun & Gill were the first to report a novel clade of 3 S. aureus jumbo phages, all encoding a retron-like RT. - for one of the phages they looked at more closely (MarsHill), they also were the first to identify a 1.2 kb upstream noncoding region which they "hypothesized to contain the RT-associated noncoding RNA (ncRNA)". - they also considered whether this RT + ncRNA containing region could be part of an antiphage defense system but they weren't able to identify a protein partner at the time - screenshot from the original paper in thread 2. so at this point, there's already a remarkable amount of hard earned info for any intelligent system to build from, almost as if there's a sign over this genome saying "LOOK HERE!". 3. what Claude recognized, and it really is a meaningful discovery, is that these upstream noncoding regions contain a tandem repeat spacer architecture, reminiscent of CRISPR-like arrays, and that this architecture recurs across related RT loci. this is meaningfully impressive. - Claude also identified partner proteins associated with ART RTs (that Korn et al. thought didn't exist), but their function or link to the RTs remains unknown. 4. but again, from the agent transcripts, it seems that it was already following these clues before it spotted the repeat architecture (one agent explicitly calls out MarsHill’s 1,241-bp nc region in Fig 1E). 5. so yes, Claude’s inference is novel, but if a prior human team had already identified the MarsHill RT, the 1.2-kb upstream noncoding region, proposed an associated ncRNA, and if that knowledge may have been available to an ai system (training data + literature search), shouldn't good science explicitly recognize this? 6. a really careful technical paper could have just said something like this (asked Claude to rewrite this with proper credit): "We used Claude Code instances to survey reverse transcriptase (RT) loci across 1.9 billion protein clusters. One candidate lineage contained the MarsHill RT, already characterized by Korn et al., who described the RT together with its upstream noncoding region, proposed that this region encodes an RT-associated ncRNA, and raised the possibility of an antiphage role. Starting from that work, our agents inspected the upstream DNA of a MarsHill relative and recognized a tandem repeat–spacer architecture. Claude then showed that the same architecture recurs across a wider family of jumbo-phage RTs, which we call array-associated RTs (ARTs), and flagged dedicated partner genes at these loci whose function remains unknown." 7. Anthropic's paper ends with "we expect that as models improve, human scientists will be enabled to explore a wider range of ideas, focus on experimental design and validation, and discover many more systems like ART by using AI systems to make connections between the vast stores of human-produced scientific knowledge". but this paper or the website announcement barely recognize the vast stores of knowledge its discovery was enabled by. 8. lastly, as these systems get insanely capable and eventually may be doing a huge amount of the actual scientific reasoning, those of us who are human scientists in the loop, let's never ever get so carried aways by ai discovery that we lose our sense of how cumulative scientific discovery is. let's be way more generous that this poor storytelling of original human contributions - "Agents that read primary data directly can provide the judgment that limits discovery wherever data have outgrown the attention of human scientists. The MarsHill ART locus is an example, as its original genome report identified the RT and proposed a 5′ ncRNA, yet described neither the repeats nor the partner gene"