Re: PSA: Automatic Minutes Generation

[Eric Rescorla <[e-mail]>]

On Wed, Nov 5, 2025 at 6:23 AM Phillip Hallam-Baker <[e-mail]>
wrote:

> Does the datatracker also track the result of shows of hands? I was
> scribing yesterday and it was a bit of a pain juggling multiple screens. I
> was pretty sure that the information was being captured but since I didn't
> know for sure had to do it.
>
> A point I am rather skeptical about is the ability of an LLM to
> accurately capture the conclusions from mic interactions which frequently
> start off with a hazy question, receive a hazy answer and gradually
> converge on a common understanding. In those cases, it is the common
> understanding that we want to go into the minutes, not necessarily how we
> got there.
>

It's reasonable to be skeptical of this, but we don't have to discuss it in
a vacuum, because I've generated 10+ IETF's worth of historical minutes,
plus the minutes I'm generating for this meeting, so people can judge for
themselves how good a job they do.

As should be clear, it's not like I'm generating these minutes by hand, so
if you think they're bad and don't want to use them, I'm not offended, and
I doubt Gemini will be either.

-Ekr



> On Tue, Nov 4, 2025 at 3:24 PM Orie <[e-mail]> wrote:
>
>> Hi,
>>
>> I filed https://github.com/ietf-tools/datatracker/issues/9881 as a
>> proposal to integrate autominutes into datatracker directly.
>>
>> I think it's important to enable corrections and clarifications in
>> minutes, especially to support direct links to related resources.
>>
>> Regards,
>>
>> OS
>>
>> On Tue, Nov 4, 2025 at 11:39 AM Lemmons, Chris <Chris_Lemmons=
>> [e-mail]> wrote:
>>
>>> Just a usage report… I found the real-ish time notes very helpful after
>>> taking notes for MOPS. I took notes as normal for the session, but after
>>> the session when the auto-minutes were available, I double-checked against
>>> them. I found that the auto-minutes had slightly mischaracterized some
>>> conversations, but not any worse than a human does. It did, however, make
>>> *different* mistakes than I did, so I was able to correct my notes with
>>> the things it captured effectively. It was very useful that the
>>> auto-minutes were available very quickly after the session while my memory
>>> was still fresh.
>>>
>>> Many thanks for your efforts here, they are very much appreciated.
>>>
>>> *From: *Eric Rescorla <[e-mail]>
>>> *Date: *Monday, November 3, 2025 at 3:32 PM
>>> *To: *Working Group Chairs <[e-mail]>
>>> *Subject: *Re: PSA: Automatic Minutes Generation
>>>
>>> Update:
>>> I am generating live-ish minutes for IETF 124 at:
>>> https://ietfminutes.org/minutes/ietf124/index.html
>>> <https://urldefense.com/v3/__https://ietfminutes.org/minutes/ietf124/index.html__;!!CQl3mcHX2A!Cp64CfaSh9BVeM9KKwGj5sFMR2ueryvxZh8WRGSaEZpko5keJ9ZPkFiqE3WPWX12o7SsF5aNhAEzPw$>
>>>
>>> It takes some time for Meetecho to produce the transcripts, so they're
>>> not available right after the meeting, but as a reference point Monday
>>> Session I is now available.
>>>
>>> -Ekr
>>>
>>>
>>> On Sat, Oct 11, 2025 at 7:02 PM Eric Rescorla <[e-mail]> wrote:
>>>
>>> Hi Chairs,
>>>
>>> Like many of you, I've long been unsatisfied with the situation around
>>> IETF minute taking, in which we conscript someone into taking minutes,
>>> thus preventing them from fully participating, and often with mediocre
>>> results.
>>>
>>> In an attempt to partly address this situation, I've put together an
>>> automatic minutes generator at (https://ietfminutes.org/
>>> <https://urldefense.com/v3/__https://ietfminutes.org/__;!!CQl3mcHX2A!Cp64CfaSh9BVeM9KKwGj5sFMR2ueryvxZh8WRGSaEZpko5keJ9ZPkFiqE3WPWX12o7SsF5by2XDQvg$>).
>>> The site
>>> takes as input the Meetecho transcript and produces AI-generated
>>> minutes. I've gone back and made minutes through IETF 112 (November
>>> 2021) and I plan to produce minutes for IETF 124 (Montreal) once
>>> transcripts are available, shortly after the meeting.
>>>
>>> Hopefully others will find this useful.
>>>
>>> -Ekr
>>>
>>> P.S. I'm not a WG chair, so you should make sure to keep
>>> me on CC if you want a response.
>>>
>>>