Cool - thank you for the updates on the progress here and feedback from
participants.
LGTM to experiment an additional 6 milestones/3 months, M154 to M159
inclusive.
On 8/21/26 1:09 p.m., Isaac Ahouma wrote:
Quick correction: the link to the explainer in my previous email was
broken. The correct link is:
https://github.com/webmachinelearning/prompt-api#configuration-of-sampling-modes.
On Friday, August 21, 2026 at 10:06:28 AM UTC-7 Isaac Ahouma wrote:
Hi Mike,
Thanks for checking in! Here is a quick update on where we stand
across those areas:
Milestones:
As noted in the intent, we are looking to extend the experiment
through Chrome 159. This aligns with the standard extension policy
and gives developers sufficient time to test the updated
categorical presets.
Draft Spec & TAG Review:
The draft specification
<https://webmachinelearning.github.io/prompt-api/> and explainer
<http://LanguageModelSamplingMode> have been updated to reflect
the new |AILanguageModelSamplingMode| enum structure. TAG review
is currently pending.
Outreach for feedback & Signals:
While official vendor signals are still pending, we've gathered
helpful qualitative feedback from the community. This input
surfaced the large gaps in our initial presets, which directly
motivated this update.
On the quantitative side, our latest developer survey data shows:
* 45.8% of developers actively tune parameters (vs. 39.2% who
rely on defaults).
* Among those who tune parameters, 54.5% are favorable to using
semantic presets, while 24.5% prefer raw parameter access.
Note that this demand for tuning largely reflects broader
developer habits across LLMs in general, rather than Built-in AI
specifically. If we need even more signal during this extension,
we plan to tap into WEC/partnerships or reach out to the EPP
mailing list.
WPT tests & Eval framework:
Web Platform Tests currently cover the API shape and surface.
However, for evaluating the actual model outputs (the effects of
sampling parameters), our ultimate destination is the work in
progress with other browser vendors on shared use-case benchmarks.
We are using Web AI Studio as an initial stepping stone toward
that broader evaluation framework.
Let me know if you need any additional details!
Cheers,
Isaac
On Thu, Aug 20, 2026 at 10:18 AM Mike Taylor
<[email protected]> wrote:
Hi there,
Can you clarify what milestones you are looking to experiment
on, and any progress on the following since the initial
experiment?
Draft spec
TAG review
Signals requests
Outreach for feedback from the spec community
WPT tests (or some kind of open eval framework)
thanks,
Mike
On 8/19/26 12:37 p.m., 'Isaac Ahouma' via blink-dev wrote:
*Contact emails *[email protected], [email protected]
Explainerhttps://github.com/webmachinelearning/prompt-api#sampling-parameters
<https://github.com/webmachinelearning/prompt-api#sampling-parameters>
Specificationhttps://webmachinelearning.github.io/prompt-api/
<https://webmachinelearning.github.io/prompt-api/>
SummaryThe Prompt API Sampling Parameters allow developers to
control the output variety of the built-in AI language model.
Instead of exposing raw numerical parameters (e.g. topK and
temperature) which can behave inconsistently across different
underlying model families and versions, this feature
introduces a categorical samplingMode enum. This allows the
browser to handle the heavy lifting of mapping semantic
presets to optimal raw parameters for a specific underlying
model, providing developers with the necessary granularity to
tune responses while maintaining cross-browser
interoperability.enum AILanguageModelSamplingMode
{"most-predictable", // For strict consistency/factual
extraction"predictable", // For highly focused
outputs"slightly-predictable", // For moderately focused,
consistent outputs"balanced", // The default state for
standard prompting"slightly-creative", // For moderately
varied, expressive outputs"creative", // For tasks favoring
variety over strict facts"most-creative" // For maximum token
diversity and brainstorming};
Blink componentBlink>AI>Prompt
Web Feature IDhttps://webstatus.dev/features/languagemodel
<https://webstatus.dev/features/languagemodel>
TAG review statusPending
Link to previous “Intent to Experiment” blink-dev
discussionhttps://groups.google.com/a/chromium.org/g/blink-dev/c/4KvH5XEBYtE
<https://groups.google.com/a/chromium.org/g/blink-dev/c/4KvH5XEBYtE>
Goals for experimentationOur primary goal during this
extension is to gather real-world usage data on the newly
expanded categorical presets to see which modes developers
gravitate toward most. We will use this data and developer
feedback to conduct more concerted medium-term mode
evaluations. Specifically, we want to evaluate developer
adoption of this expanded spectrum, and validate that the new
granularity effectively covers the previously identified dead
zones.
Experimental timelineThe extended experiment will continue
through Chrome 159.
Reason this experiment is being extendedWe are extending this
experiment because we are actively iterating on the API
surface based on developer feedback. During this trial, we
received developer feedback requesting predictable space
granularity to cover dead zones while restoring the
balancedpreset to the API default parameters. We have
expanded the preset enum values to cover these dead zones,
and we need the extended timeline to give developers
sufficient time to integrate and test these specific changes.
Interoperability and Compatibility RisksThe original raw
parameters were excluded from the initial Prompt API launch
due to cross-browser interoperability concerns. By refining
the categorical sampling modes to provide better coverage of
the predictable space based on developer feedback, we
maintain the cross-browser interoperability benefits of
semantic presets while offering the necessary granularity
developers requested.WebView application risksNone
Ongoing technical constraintsNoneDebuggabilityIt is possible
that giving DevTools more insight into the nondeterministic
states of the model, e.g. random seeds, could help with
debugging. See related discussion at
https://github.com/webmachinelearning/prompt-api/issues/9.
Will this feature be supported on all six Blink platforms
(Windows, Mac, Linux, ChromeOS, Android, and Android
WebView)?No, the Prompt API currently supports Windows, Mac,
Linux, and ChromeOS.
Is this feature fully tested by web-platform-tests?No; while
the API shape is fully tested, automated testing of sampling
parameter effects on probabilistic model response is not
readily feasible; instead we conduct rounds of evaluations on
configuration updates.
Flag name on about://flagsprompt-api-sampling-mode
Finch feature nameAIPromptAPIParams
Requires code in //chrome?True
Tracking bughttps://crbug.com/502214118
Launch
bughttps://launch.corp.google.com/launch/4463387<https://launch.corp.google.com/launch/4463387>Link
to entry on the Chrome Platform
Status:https://chromestatus.com/feature/6325545693478912
<https://chromestatus.com/feature/6325545693478912>
--
You received this message because you are subscribed to the
Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from
it, send an email to [email protected].
To view this discussion visit
https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAFk4SLZXUY8FQTQQrtLHkzLnkxkgJ7XXFBhDEAQSvy6XiYKgaQ%40mail.gmail.com
<https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAFk4SLZXUY8FQTQQrtLHkzLnkxkgJ7XXFBhDEAQSvy6XiYKgaQ%40mail.gmail.com?utm_medium=email&utm_source=footer>.
--
You received this message because you are subscribed to the Google Groups
"blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion visit
https://groups.google.com/a/chromium.org/d/msgid/blink-dev/7c4940db-bacc-4a3c-8035-dcc6cf069097%40chromium.org.
