--- title: Getting the feedback loop correct in AI | daily.dev description: The real challenge in enterprise AI isn't picking the right model upfront — it's building a feedback loop that connects system outputs to business outcomes.... canonical: https://daily.dev/posts/getting-the-feedback-loop-correct-in-ai-hbmkhoy2m twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Getting the feedback loop correct in AI | daily.dev og:description: The real challenge in enterprise AI isn't picking the right model upfront — it's building a feedback loop that connects system outputs to business outcomes.... og:url: https://daily.dev/posts/getting-the-feedback-loop-correct-in-ai-hbmkhoy2m og:image: https://api.daily.dev/og/posts/hbmkhoY2M.png og:image:alt: Getting the feedback loop correct in AI og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Getting the feedback loop correct in AI **[InfoWorld](https://daily.dev/sources/infoworld)** · 8 min read · 0 upvotes · 0 comments ## Summary The real challenge in enterprise AI isn't picking the right model upfront — it's building a feedback loop that connects system outputs to business outcomes. RAG lets a model consult your data, but a continual learning loop lets the system improve from it. Open-weight models matter not because they're cheaper alternatives to OpenAI or Anthropic, but because they let enterprises turn proprietary outcome data into model improvements they own and carry forward. Cursor's Composer 2 illustrates this: starting from the open-weight Kimi K2.5, Cursor specialized it via continued pretraining and reinforcement learning, with 85% of compute going into additional training. Platforms like Fireworks AI, Together AI, and Baseten are emerging as a category that makes this learning loop accessible to enterprises without requiring them to build foundation models from scratch. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What open-weight model did Cursor use as the base for Composer 2 and 2.5? Cursor built Composer 2 starting from the open-weight Kimi K2.5 model, specializing it through continued pretraining and reinforcement learning in environments designed to resemble Cursor's own product. Composer 2.5 was subsequently built on the same Kimi K2.5 checkpoint, with Cursor stating that 85% of its compute went into additional training and reinforcement learning rather than the base model itself. _Developers weighing base versus fine-tuned coding models can follow real-world cases like this on daily.dev._ ### Why would a company choose an open-weight AI model over a closed API like GPT or Claude? Open weights let a company own the continual learning loop and turn proprietary outcome data—traces of what the system saw, did, and whether the result was acceptable—into something it can shape and carry forward as models change. Closed APIs only let a company consult its data via retrieval, not actually improve the underlying model using its own outcomes. _Teams debating open versus closed models for their AI stack can track this tradeoff on daily.dev._ ### How should a company decide whether a cheaper AI model is good enough for a task instead of always using the most expensive one? Log the request, model, cost, answer, and downstream outcome for one workflow over a period like a quarter, then look for patterns: a cheaper model will suffice for some work, a frontier model will earn its cost elsewhere, and a third category where every model fails the same way signals it's time to consider retraining open weights. The cheapest model by token can be the costliest by outcome if humans must review, retry, or repair its work. _Anyone optimizing AI spend across models can track these cost-versus-outcome tradeoffs on daily.dev._ ## Similar posts on daily.dev - [The Signals Loop: Fine-tuning for world-class AI apps and agents](https://daily.dev/posts/the-signals-loop-fine-tuning-for-world-class-ai-apps-and-agents-ii8dbe85f) · Microsoft Azure · 0 upvotes · 0 comments - [Building Enterprise AI Feedback Loops for Continuous Model Improvement](https://daily.dev/posts/building-enterprise-ai-feedback-loops-for-continuous-model-improvement-cuxd00zmn) · C\# Corner · 0 upvotes · 0 comments - [Why Not All AI “Context” is Equal](https://daily.dev/posts/why-not-all-ai-context-is-equal-xildxywzh) · SD Times · 1 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#llm](https://daily.dev/tags/llm), [#rag](https://daily.dev/tags/rag), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/getting-the-feedback-loop-correct-in-ai-hbmkhoy2m) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Getting the feedback loop correct in AI","url":"https://daily.dev/posts/getting-the-feedback-loop-correct-in-ai-hbmkhoy2m","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/getting-the-feedback-loop-correct-in-ai-hbmkhoy2m"},"datePublished":"2026-08-11T09:03:30.722Z","dateModified":"2026-09-13T19:07:02.266Z","description":"The real challenge in enterprise AI isn't picking the right model upfront — it's building a feedback loop that connects system outputs to business outcomes....","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b21e2c1fddc16dca882e2ed19c980d6c?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b21e2c1fddc16dca882e2ed19c980d6c?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"InfoWorld","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"InfoWorld","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/bf6d68a999064029b0bb09aa6268f1f3","url":"https://daily.dev/sources/infoworld"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/getting-the-feedback-loop-correct-in-ai-hbmkhoy2m","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"open-source,llm,rag,reinforcement-learning","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"InfoWorld","item":"https://daily.dev/sources/infoworld"},{"@type":"ListItem","position":3,"name":"Getting the feedback loop correct in AI"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/getting-the-feedback-loop-correct-in-ai-hbmkhoy2m#faq","mainEntity":[{"@type":"Question","name":"What open-weight model did Cursor use as the base for Composer 2 and 2.5?","acceptedAnswer":{"@type":"Answer","text":"Cursor built Composer 2 starting from the open-weight Kimi K2.5 model, specializing it through continued pretraining and reinforcement learning in environments designed to resemble Cursor's own product. Composer 2.5 was subsequently built on the same Kimi K2.5 checkpoint, with Cursor stating that 85% of its compute went into additional training and reinforcement learning rather than the base model itself. Developers weighing base versus fine-tuned coding models can follow real-world cases like this on daily.dev."}},{"@type":"Question","name":"Why would a company choose an open-weight AI model over a closed API like GPT or Claude?","acceptedAnswer":{"@type":"Answer","text":"Open weights let a company own the continual learning loop and turn proprietary outcome data—traces of what the system saw, did, and whether the result was acceptable—into something it can shape and carry forward as models change. Closed APIs only let a company consult its data via retrieval, not actually improve the underlying model using its own outcomes. Teams debating open versus closed models for their AI stack can track this tradeoff on daily.dev."}},{"@type":"Question","name":"How should a company decide whether a cheaper AI model is good enough for a task instead of always using the most expensive one?","acceptedAnswer":{"@type":"Answer","text":"Log the request, model, cost, answer, and downstream outcome for one workflow over a period like a quarter, then look for patterns: a cheaper model will suffice for some work, a frontier model will earn its cost elsewhere, and a third category where every model fails the same way signals it's time to consider retraining open weights. The cheapest model by token can be the costliest by outcome if humans must review, retry, or repair its work. Anyone optimizing AI spend across models can track these cost-versus-outcome tradeoffs on daily.dev."}}]} ```