Links 9/21/2025 | naked capitalism

To construct on a remark I made just lately…

There are some phrases he used which have a exact that means within the business, notably (tech/AI brethren within the feedback, don’t come after me for generalizing and simplifying, I’m chatting with these outdoors of the sector):

– persistence: sturdy storage for software state
– planning: a sort of agent workflow that makes use of RAG context fed to the LLM/frontier fashions with checkpoints that may save (“persist”) the state of the response so the consumer can resume, fork, change and so on from the checkpoint with the state knowledge from that cut-off date

What he’s describing is a variant on the exterior RAG connector paradigm I discussed within the ultimate level in my earlier remark. The mannequin itself is taking in data which is saved someplace within the consumer’s software area (for instance, a undertaking plan with targets, outcomes, good and dangerous examples of anticipated output) and feeding that to the mannequin utilizing a planning/orchetration agentic course of in order that the mannequin processes the undertaking plan to take motion based mostly on the undertaking plan contents and some other RAG or exterior (to the mannequin) instruments (like exterior internet search, schedulers, and so on). That is an anticipated a part of the transition of the AI area going from science experiment/magic trick to simply one other a part of the appliance layer.

What you’re citing – hallucinations – are sometimes an artifact of the mannequin not having sufficient context about what it’s being requested and so it makes stuff as much as keep away from getting halted. Bringing in additional context by way of RAG (however not a lot as to overload the mannequin’s context window for every step of the interplay) is like 98% of the magic of utilizing AI productively.

He’s not describing any new breakthroughs. Just about the entire ‘severe’ finish of the devoted AI tooling is starting to roll extra refined RAG-based tooling on prime of their purposes that use the frontier fashions. There’s a bullwhip impact whereas everybody figures out find out how to use every new tier of tooling. Paradoxically the extra of those are added, the much less the precise LLM/frontier mannequin is required to dig deep into its personal coaching set to reply. Over time we’ll see the big frontier fashions in all probability die off as they’re changed with hyper-specific fashions for specific use instances, like coding (that is already beginning to occur in sure arenas the place the coding sort shouldn’t be one which has plenty of public knowledge that would have been used within the frontier fashions coaching set, like FPGA/ASIC growth).

Lastly, to make one thing extraordinarily explicitly clear, any severe firm that implements this may construct out the user-associated persistence layer separate from the mannequin so every particular person’s context recordsdata are segregated from everybody else (“multitenancy”) and in the event that they’re on a paid or enterprise plan that often comes with a authorized settlement to not take that data and retrain the frontier fashions on it if the service can be a frontier mannequin supplier. The context is often encrypted and handed to the mannequin in a format the mannequin can learn, not human readable, so the supplier can say they’re not stealing your context knowledge or harvesting it for nefarious deeds (not all paid plans/providers try this which is why you need to learn the effective print within the settlement). Any downstream supplier who, say, buys a Microsoft service to construct out their very own software leveraging this tech, will use these non-public/safe endpoints and providers so there’s a chain of authorized accountability so attorneys down the chain can log out on it. This isn’t supplied to customers of the free and public implementations of the chatbots, which is the place plenty of the horror tales and dangerous experiences come from.