I would like to propose a conversational AI architecture idea that may improve intent understanding, context retention, response latency, safety calibration, and the effective utilization of existing model capabilities without necessarily increasing model size.
1. The core problem
Current chat interfaces effectively define a user query as:
Query = final submitted text
But humans generate a large amount of meaningful information before pressing Send.
While composing a message, a user may:
- type and delete a phrase
- replace one interpretation with another
- strengthen or soften wording
- pause at a semantically important point
- repeatedly rewrite the same concept
- shift between serious and joking tones
- reject an earlier framing before choosing the final one
Almost all of this information is currently discarded.
The final message is therefore a compressed representation of the user’s intent, not necessarily the complete intent signal.
The simple question is:
Why should the final submitted string be treated as the entire query?
2. Draft-as-Context
The proposal is to optionally treat the composition trajectory as temporary context for the current query.
Conceptually:
drafting
→ edits / deletions / pauses / reformulations
→ final message
→ combined intent interpretation
→ response
→ draft context deleted
The model would not need to treat every keystroke as equally meaningful.
Instead, the system could derive transient semantic features such as:
rejected interpretation
semantic direction change
repeated concept
hesitation at meaningful phrase
tone shift
strengthening / softening
final selected framing
The important principle is:
Draft is context, not memory.
The composition process exists only to interpret the current message.
3. Speculative Intent
There is another opportunity hidden in the same mechanism.
Today, while a user spends 10–30 seconds writing a question, the model generally waits for Send before beginning inference.
Instead, during composition the system could maintain several provisional intent branches.
For example:
Possible intent A
Possible intent B
Possible intent C
Possible intent D
As the user continues typing, editing, deleting and reformulating, the probability of each branch changes.
The model could simultaneously prepare relevant context or reasoning for the most likely branches.
Then:
Send becomes less of a “start thinking” button and more of a “final branch selection” signal.
Conceptually:
USER TYPES
↓
temporary intent tree
↓
speculative context / reasoning preparation
↓
USER PRESSES SEND
↓
final text + composition trajectory
↓
select best intent branch
↓
generate response
↓
wipe transient draft state
This could potentially improve both accuracy and perceived latency with the same mechanism.
4. Privacy architecture
This does not need to become persistent keylogging.
A clean implementation could be explicitly opt-in.
Draft-as-Context: OFF
→ only submitted messages are interpreted
Draft-as-Context: ON
→ composition information may be used temporarily
→ only for the current turn
→ wiped after the response
→ not added to chat history
→ not added to memory
Where possible, raw drafting behavior could be processed locally and converted into transient intent features rather than permanently transmitting or storing the entire draft history.
The user should always control whether this mode is enabled.
5. Why this may be larger than a UX feature
At first glance this looks like a small interface improvement.
I think the deeper possibility is different.
A frontier model may already possess substantially more useful capability than the conversational interface allows the user to access.
Effective user-facing intelligence is not determined only by model capability.
A rough conceptual model is:
effective intelligence
≈
latent model capability
× quality of intent observation
× context continuity
× inference routing efficiency
If the model misunderstands the user’s intent, computation is spent on the wrong branch.
That creates secondary losses:
intent error
→ wrong reasoning branch
→ poor response
→ user correction
→ additional context
→ context pollution
→ more inference
→ slower conversation
Improving intent observation could therefore produce a nonlinear effect:
better intent recognition
→ fewer wrong branches
→ less wasted inference
→ fewer correction turns
→ cleaner context
→ better future responses
→ higher user trust
This means Draft-as-Context may be better understood as a latent-capability utilization layer, not merely a “humanizer.”
It does not necessarily make the underlying model larger.
It may allow more of the model that already exists to reach the output.
6. Last-mile Intelligence
As frontier models become increasingly capable, one of the major competitive differences may eventually become:
How efficiently can a system convert model capability into useful intelligence for a specific human?
Users ultimately do not experience parameter counts or benchmark tables directly.
They experience:
Did it understand me?
Did it remember what I actually meant?
Did I have to correct it?
Did it repeat an interpretation I already rejected?
Did it recognize my tone?
Did it answer the question I was actually trying to ask?
This suggests a possible future competitive layer:
Last-mile Intelligence
The final connection between the model’s latent capability and the individual user’s actual intent.
A slightly weaker model with much better intent observation could potentially feel substantially more intelligent than a stronger model with a lossy interface.
7. Even a failed first implementation would produce a useful research program
The strongest hypothesis is that Draft-as-Context + Speculative Intent could cause a large nonlinear improvement in perceived model intelligence.
But the idea does not need that strongest hypothesis to succeed.
If the complete drafting trajectory provides little benefit, the next question becomes:
Which parts of human intent formation contain predictive information beyond the final submitted text?
That creates several research directions:
Are semantic pauses useful?
Are deleted phrases useful?
Is the direction of revision more informative than deleted content?
Are rejected interpretations especially valuable?
Which tone transitions predict user intent?
How should transient intent states be represented?
When should speculative inference begin?
How does this generalize from text to speech, pauses, prosody and self-correction?
So there are multiple useful outcome levels:
Maximum:
large nonlinear capability-utilization gain
→ new conversational architecture / competitive advantage
Intermediate:
specific draft signals improve intent, context, latency or safety
→ useful component of future conversational systems
Minimum:
first implementation performs poorly
→ identifies a fundamental research problem:
how should AI observe the process by which humans form communicative intent?
The central research question survives even if the first engineering implementation does not.
8. Suggested evaluation
The cleanest experiment seems straightforward.
Use exactly the same underlying model and compare:
A: final-text-only interaction
B: final text
+ transient composition context
+ speculative intent preparation
Measure:
first-response task success
intent misunderstanding rate
number of correction turns
context drift
unnecessary refusal / over-cautious response rate
response latency
user preference
conversation completion time
The important comparison is not primarily a benchmark score.
It is whether the same model becomes substantially more effective for the human using it.
Core principle
**The final message is not necessarily the whole query.
The process of arriving at that message can also contain information about the query.**
And the broader hypothesis is:
Before spending all of our effort making the model stronger, we should measure how much of the strength we already have is being lost between the user and the model.
Draft-as-Context is one possible way to recover that lost information.
Speculative Intent is one possible way to use the recovered information before the user even presses Send.
Last-mile Intelligence is the broader problem this points toward.
Attribution
Original concept: Jae-Hyun Min
Working names: Draft-as-Context / Speculative Intent / Last-mile Intelligence
Structured and summarized with: ChatGPT, GPT-5.6 Sol