Article URL: https://doublepulsar.com/anthropic-hacked-to-deliver-remote-access-trojan-to-claude-developer-laptops-worldwide-1148a92baae2 Comments URL: https://news.ycombinator.com/item?id=50302187 Points: 17 # Comments: 3
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Running AI models on a company's own hardware means the model operates on a machine the business owns or controls rather than on a provider's servers. The main limit is graphics card memory, since an 8 billion parameter model needs about 4.9GB and a 70 billion parameter model about 43GB. This setup suits recurring overnight jobs and sensitive records, but it requires ongoing updates, security, monitoring, and hardware ownership. For that reason, most small service firms are advised to begin with cloud-based tools.
A developer building a model-upgrade pipeline for a .NET 10 support assistant with Microsoft.Extensions.AI shares lessons learned. The assistant has tools that move money or create work for staff, so the stakes of a shadow deployment are higher than for a plain chatbot. The article warns that model, prompt, and option changes often ship together, which obscures what actually caused behavior shifts. It also describes a case where a supposedly passive shadow model executed its own refund tool call, resulting in a duplicate refund. The author proposes fixes such as fingerprinting releases, running shadow candidates with declaration-only tools, and avoiding stale references in background comparison jobs.
An analysis of 15 models' published MMLU-Pro predictions found that when they answer incorrectly, they select the longest wrong option more often than chance would predict. The effect is strongest on the STEM-website biology and psychology questions, where it is larger than on questions derived from MMLU. Three weaker models scored below a simple longest-option heuristic that ignores the question entirely on those 656 questions. Even the strongest models showed the bias on STEM-website items, though their error counts were small. The researchers noted the lean is most visible where answer options are prose rather than numeric, since numeric options often tie in length.
Due diligence begins after a letter of intent is signed and typically lasts 45 to 90 days, during which buyers verify the claims a seller has made. Revenue quality is usually where price reductions originate, more so than the codebase itself, with issues like messy data, missing paperwork, and heavy reliance on a few large customers being common culprits. Buyers commonly request monthly financial statements, customer-level MRR data, and churn and retention analysis, and they often rebuild these figures from raw billing exports. A practical step for founders is to prepare a month-by-month MRR rollforward covering new, expansion, contraction, and churn, tied to billing and accounting records. Further scrutiny typically covers code ownership and IP assignments, open-source licence risks, security practices, and infrastructure documentation.
Discussion (0)
Log in to join the discussion and vote.
Log in