Hi Julien,
Thank you for bumping this thread.
I’m happy to share that we now have general consensus on the FSST Spec.
Many thanks to Micah, Gang, Andrew, Antoine, and others for their
thoughtful feedback and comments.

Our discussion with Alkis and Ryan during the last sync is also now
explicitly reflected in the spec: the amount of training data used is an
implementation choice.
We had initiated discussions with CWI, who are working on improvements that
may be released in the future. This does not block the specification from
moving forward, and can conduct a separate evaluation once the findings are
published.

(In parallel, I’ll evaluate OnPair16 and add findings and comparison with
FSST to the specification.)
At this stage, I would encourage everyone to take another look at the
specification whenever convenient. I’m happy to address any additional
comments or questions.

Thank you again to everyone for the thoughtful feedback and discussion.

Warm Regards,
Arnav

On Tue, Jul 28, 2026 at 4:35 AM Julien Le Dem <[email protected]> wrote:

> I wanted to bump this thread to discuss the next steps for evaluating FSST.
>
> My understanding is that we evaluated FSST with the original 8-bit codes
> and want to compare to variations with larger codes (and dictionaries) in
> particular 12 and 16 bits. Discussion with the CWI authors ongoing.
> Is that correct? What else do we need to do?
>
> Separately, someone brought OnPair16 <https://arxiv.org/abs/2508.02280> to
> my attention which has similar characteristics with slightly different
> trade-offs. I was wondering if someone on the list had experience with it
> and if we should compare it to FSST in the same evaluation.
>
> Best.
> Julien
>
>
>
> On Wed, Mar 11, 2026 at 6:29 AM Arnav Balyan <[email protected]>
> wrote:
>
> > Hi Micah and Gang,
> >
> > Thanks for taking a thorough look at the spec, really appreciate the well
> > thought of comments and feedback.
> > The comments are addressed and incorporated into the spec, could you pls
> > take another pass if time permits.
> >
> > Thanks again!
> >
> > Regards,
> > Arnav
> >
> > On Wed, Feb 4, 2026 at 11:30 PM Arnav Balyan <[email protected]>
> > wrote:
> >
> > > Hi team,
> > >
> > > Just wanted to share the spec design for FSST encoding based on
> previous
> > > discussions. I will add perf numbers to doc shortly.
> > >
> > > Spec Document:
> > >
> >
> https://docs.google.com/document/d/1Xg2b8HR19QnI3nhtQUDWZJhCLwJzW6y9tU1ziiLFZrM
> > >
> > > Would love to get your feedback, thoughts and suggestions.
> > >
> > > Warm Regards,
> > > Arnav
> > >
> >
>

Reply via email to