[ishtanbul]

I work at F500 company and this is not my experience at all. We are rolling out claude to basically everyone in my division and people are doing pretty amazing things with it and the uptake is high. Building “skills” around specific workflows has been very useful. We have regular training where we meet to share “skills”. Its exciting. Its not all rainbows but the impact and growth potential is very clear to me. This blog seems to reflect a 2024-ish state of affairs.

[luco17]

Agree. It’s good and exciting. It is so strong at information retrieval. Much enterprise work is finding, categorising, slicing, summarising info.

Querying SQL, searching Sharepoint, AWS/Azure CLI, log analysis… etc. Amazing.

ABAC/RBAC, data labelling, scoped tokens, Purview all v helpful in managing risk.

[aakresearch]

What gives you basis to believe and accept that all of the above is exhaustive, complete and devoid of fakes? That "finding" finds ALL what exists and NONE of what does not exist? Ditto for categorising, summarising and labeling?

There is no fundamental guarantees provided by LLM architecture, so to assert anything of sorts would require manually executing the operation and comparing the results. And why not asserting the correctness is not negligence or fraud?

I am not saying it is impossible, but so far I have not found any explanations that would close this huge gap in logic for me. LLMs themselves "admit" the gap cannot be explained rationaly.