Once Deployed, AI Models Stop Learning—Freshness Will Come from Outside the Parameters

Once an AI model goes live, its learning largely stalls — and the cost of a model going stale after taking on real work is high. On August 14, @coallaoh argued in a thread that over the next few years, the learning that keeps models current will happen mostly outside the parameters: post-training has a high barrier and risks damaging existing knowledge, so what most users can actually do is external engineering such as memory management, search/RAG, and tool use. A related article by Ben Lorica likewise points out that post-deployment learning stagnation is a current pain point, with continual learning techniques only landing in piecemeal fashion.

Confirmed

• @coallaoh's core claim: once AI takes on real work, people need not just better models but models that know last week's events, codebases, companies, and individuals, so the cost of a stale model is high
• On "why not just update the parameters": post-training is the first answer, but it is constrained by data, compute, and talent that most dependents lack; updating parameters can also break a model's existing knowledge
• The realistic path: the adaptation most people can do lies outside the parameters, including engineering, memory management, search and RAG, and tool use; most real progress over the past two years has happened in these areas
• Ben Lorica writes that models can keep improving during training via reinforcement learning, but after launch their learning largely stalls; faced with new tactics, edge cases, or user feedback, they typically get one-off patch fixes rather than lasting behavioral improvement

Unconfirmed

• The post relaying Ben Lorica's article is truncated, so the specific technologies and case details behind "continual learning landing only in fragments" cannot be verified from this batch of material

Why it matters

• Model freshness directly determines AI's reliability in production, especially for tasks involving recent events, private codebases, and individual context
• This view implies that engineering approaches like memory, RAG, and tool ecosystems may become the mainstream way to keep models fresh sooner than parameter updates — a contrast with the industry's "continual learning" research direction

2026-08-14 ~ 2026-08-14 · 5 related posts

• [source] Overcoming the Post-Deployment Stall: Continual Learning Arrives in Pieces — bigdata · 2026-08-14
• [source] AI models need to stay current; future learning will happen outside parameters — coallaoh · 2026-08-14
• Post-training works but is gated; updating parameters risks existing knowledge — coallaoh · 2026-08-14
• [source] Model adaptation happens outside parameters: engineering, memory, RAG, tools — coallaoh · 2026-08-14
• Bet: most learning keeping models current will happen outside parameters — coallaoh · 2026-08-14