* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-16 18:36 ` Jonathan Corbet
@ 2026-07-16 19:53 ` Mauro Carvalho Chehab
2026-07-16 23:59 ` Theodore Tso
2026-07-16 20:05 ` Bart Van Assche
` (2 subsequent siblings)
3 siblings, 1 reply; 14+ messages in thread
From: Mauro Carvalho Chehab @ 2026-07-16 19:53 UTC (permalink / raw)
To: Jonathan Corbet; +Cc: Sasha Levin, ksummit
On Thu, 16 Jul 2026 12:36:53 -0600
Jonathan Corbet <corbet@lwn.net> wrote:
> Sasha Levin <sashal@kernel.org> writes:
>
> > On Thu, Jul 16, 2026 at 09:09:27AM -0600, Jonathan Corbet wrote:
> >>The use of LLMs in the development process appears to be a clear theme for
> >>the upcoming summit. On top of what others have already suggested, I think
> >>we may want to consider these questions:
> >>
> >>- Do we want to continue naming specific LLMs in the Assisted-by tags, or
> >> put something more generic? I *think* that this thread:
> >>
> >> https://lore.kernel.org/all/20260701-work-coding-assistants-v1-1-a20a94d1d606@kernel.org/
> >>
> >> reached a consensus that "Assisted-by: LLM" was better than what we
> >> require now, but it might be good to ratify that in this setting.
> >
> > So originally I've added the full name of the tool and LLM because there was
> > interest in a later audit of the tools to determine how useful (or useless)
> > some of the tools are.
> >
> > If those folks aren't interested in doing so anymore, then sure - we can drop
> > it.
> >
> > But... I find it difficult to see the point of having the tag if we do that.
>
> Folks like Greg have, in the recent past, said that it is useful even
> without specific product-name information:
>
> https://lore.kernel.org/all/2026070227-payroll-eradicate-8f66@gregkh/
I never saw any merit on having such tag: it doesn't help reviewers
and doesn't provide any useful information at the git history. The
only think they eventually allow is for someone to reject a patch
without actually looking on it.
What we need to ensure is that:
1. the patch is good;
2. author carefully reviewed/modified it and not just did vibe-coding;
3. the author understands the proposed code.
Unfortunately, (2) and (3)s easily said than done as technology avances,
though.
Yet, I recently saw a patch series that sounded to be produced by vibe
coding, probably generated by some proprietary paid model. That was something
that would very hardly be accepted by anyone, as LLM did lots of stupid
changes that are easily recognized as bad merge material. The final result was
also bad enough. So, for now, it is still easy to identify pure LLM generated
content.
> >>- There are many first-time contributors coming in with LLM-generated
> >> patches. At times, I could swear that every one of them is focused on
> >> documentation typos, but the truth of the matter is that they are
> >> reaching into subsystems all over the kernel. We have some brand-new
> >> contributors making significant changes to dozens of subsystems. An
> >> experienced developer would be hard-put to truly understand what those
> >> changes are doing; a newcomer is unlikely to have that understanding,
> >> and is unlikely to be around to fix eventual problems.
> >>
> >> Our maintainers are not scaling to handle this new flood, and I fear we
> >> are going to see some unfortunate things merged. One LLM-driven newcomer
> >> recently nearly succeeded in establishing himself as the maintainer of
> >> lib/. How do we hold the line against this stuff while remaining open to
> >> new developers?
> >
> > Shouldn't it be a merits question rather than a tools question?
> >
> > If the commits are correct, does it matter if they were written with
> > an LLM?
It does if the author doesn't understand the code and can't maintain
it without LLM. The bigger issue here is actually to allow people that
doesn't know how to code himself to become a maintainer.
> > we can insist more on supplying tests and demonstrating
> > correctness, something we seem to be doing quite rarely right now.
Perhaps we may need to have some interaction with the developers before
letting them to become new maintainers. This may allow checking if the
guy knows what he's doing or if he is only the man-in-the-middle.
> It's definitely a merit question. But we're not always all that good at
> determining whether a commit is correct, and we depend a lot on the
> contributor understanding their work and being around if something goes
> wrong with it. That is part of "merit" too. How confident are we of
> that merit when a brand-new developer makes significant changes to a
> dozen or more unrelated subsystems?
If a brand-new developer is touching lots of unrelated subsystems,
there is a high chance that it is vibe-coding.
>
> >>- Our process is becoming increasingly dependent on proprietary tools. We
> >> have done that before and, in 2005, it went pretty badly for us - and
> >> could have been worse. How do we prepare for the inevitable rugpull? I
> >> raised this last year, and it was largely brushed off, but I still think
> >> it's something we should be concerned about.
> >
> > Are we dependent on them, or do we just find them very useful? If
> > Claude/Codex/etc goes away next month, will it stall any of our processes?
> >
> > We have AI reviews, we have many AI tools that help both authors and
> > maintainers, but I don't think that any of them play an integral part of our
> > process.
>
> The related discussions have featured a number of maintainers talking
> about how much time Sashiko has saved them. I believe them. How long
> will it take until nobody does that level of patch review anymore? What
> will we do when the current round of corporate generosity ends and that
> tool goes away? Maybe I'm worrying too much, but this does seem, to me,
> like a possibility we should keep in mind.
This is a serious concern. It sounds risky to rely on that, as there's
no free lunch. We need to rely on something that can be managed in
an affordable way, prioritizing models that can run on affortable GPUs
and are open source.
Thanks,
Mauro
^ permalink raw reply [flat|nested] 14+ messages in thread* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-16 19:53 ` Mauro Carvalho Chehab
@ 2026-07-16 23:59 ` Theodore Tso
2026-07-17 0:58 ` Mauro Carvalho Chehab
0 siblings, 1 reply; 14+ messages in thread
From: Theodore Tso @ 2026-07-16 23:59 UTC (permalink / raw)
To: Mauro Carvalho Chehab; +Cc: Jonathan Corbet, Sasha Levin, ksummit
On Thu, Jul 16, 2026 at 09:53:42PM -0500, Mauro Carvalho Chehab wrote:
> > What
> > will we do when the current round of corporate generosity ends and that
> > tool goes away? Maybe I'm worrying too much, but this does seem, to me,
> > like a possibility we should keep in mind.
>
> This is a serious concern. It sounds risky to rely on that, as there's
> no free lunch. We need to rely on something that can be managed in
> an affordable way, prioritizing models that can run on affortable GPUs
> and are open source.
This gets tricky. We can divide ML models into a couple of
different classes:
1. Those that can fit on a mobile phone
2. Those that can fit on low-end developer machine (16GB-32GB
unified memory, or 16GB of VRAM in your GPU)
3. Those that fit in High-end developer machines (128GB unified
memory, such as could be found in a M5 Max Macbook Pro, a DGX
Spark, or an AMD Strix Halo)
4. Those that fit into one or more Enterprise servers with 8 H100
GPU's --- that is, frontier models.
Machines in category 3 run about $4k (on the low-end, without a
monitor) and go up from there. About six weeks ago, I invested in a
M5 Max Macbook Pro with 128GB, and it set me back $6,214 USD
(including tax). When Apple increased their prices due to the
DRAMpocalypse, for the first time, I've seen a computer *appreciate*
in value after being purchased --- by $1,400 USD. :-) So questions of
access equity is already an issue with machines in this category.
Machines in category 4 run around $400,000 each. (Of course, after
the dot.COM bubble implosion, Sunfire E10K's that startups paid
$100,000 ended up selling for pennies on the dollar. So after a
Neocloud company go out of business, maybe thse machines will be
affordable by individual developers. However, even then your partner
might not be enthusiastic about the heat and sound from one of these
data center servers being run in your office or living room --- not to
mention the electricity bill. :-)
Now, a H100 has 80 GB of High Bandwidth Memory (HBM) which has a
bandwidth of 2 TB/s. So a server with 8 H100's has a 640 GB of HBM
and an aggregate bandwidth of 16 TB/s. In contrast, a M5 Max with
128 GB and a 40-core GPU has a memory bandwidth of 614 GB/s. (A
normal M5 Macbook Pro has 153 GB/s memory bandwidth and maxes out at
32GB.)
Running the 671 billion parameter Deepseek R1 model at full precision
requires 1.5TB of VRAM --- so two of these 8x H100 servers. Of could
try to run them using quantization techniques, by collapsing each
parameter from 16 bits to 4 bits, or even 1 bit to reduce the size of
the ML rig required. However, you lose a lot of accuracy when you do
that, and when models are much more prone to hallucinate when you use
the more aggressive levels of quantization.
So it's one say to say, we should figure out how to try to run Sashiko
on a local LLM, using open-weight models. But it's going to be a lot
easier to propose such a thing than to actually do it.
What we *might* be able to try doing is to take an open-weight model
that can fit on a 128GB machine, and then fine-tuning it by feeding it
several years of LKML archives which we convniently have in public
inbox format. This might allow us to create a specialized model which
is optimized for the Linux kernel, and it would be interesting to see
how this compares with a general frontier model.
But even if we did that, it use would be limited to those people who
can afford one of these 128GB unified memory machines or can otherwise
get access to one of these machines.
Cheers,
- Ted
^ permalink raw reply [flat|nested] 14+ messages in thread* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-16 23:59 ` Theodore Tso
@ 2026-07-17 0:58 ` Mauro Carvalho Chehab
2026-07-17 2:27 ` Theodore Tso
0 siblings, 1 reply; 14+ messages in thread
From: Mauro Carvalho Chehab @ 2026-07-17 0:58 UTC (permalink / raw)
To: Theodore Tso; +Cc: Jonathan Corbet, Sasha Levin, ksummit
On Thu, 16 Jul 2026 19:59:02 -0400
"Theodore Tso" <tytso@mit.edu> wrote:
> On Thu, Jul 16, 2026 at 09:53:42PM -0500, Mauro Carvalho Chehab wrote:
> > > What
> > > will we do when the current round of corporate generosity ends and that
> > > tool goes away? Maybe I'm worrying too much, but this does seem, to me,
> > > like a possibility we should keep in mind.
> >
> > This is a serious concern. It sounds risky to rely on that, as there's
> > no free lunch. We need to rely on something that can be managed in
> > an affordable way, prioritizing models that can run on affortable GPUs
> > and are open source.
>
> This gets tricky. We can divide ML models into a couple of
> different classes:
>
> 1. Those that can fit on a mobile phone
> 2. Those that can fit on low-end developer machine (16GB-32GB
> unified memory, or 16GB of VRAM in your GPU)
> 3. Those that fit in High-end developer machines (128GB unified
> memory, such as could be found in a M5 Max Macbook Pro, a DGX
> Spark, or an AMD Strix Halo)
> 4. Those that fit into one or more Enterprise servers with 8 H100
> GPU's --- that is, frontier models.
>
> Machines in category 3 run about $4k (on the low-end, without a
> monitor) and go up from there. About six weeks ago, I invested in a
> M5 Max Macbook Pro with 128GB, and it set me back $6,214 USD
> (including tax). When Apple increased their prices due to the
> DRAMpocalypse, for the first time, I've seen a computer *appreciate*
> in value after being purchased --- by $1,400 USD. :-) So questions of
> access equity is already an issue with machines in this category.
I'd say the best would be to aim on (3) and (4). Not as powerful
as H100 GPUs, but it may end work.
> Machines in category 4 run around $400,000 each. (Of course, after
> the dot.COM bubble implosion, Sunfire E10K's that startups paid
> $100,000 ended up selling for pennies on the dollar. So after a
> Neocloud company go out of business, maybe thse machines will be
> affordable by individual developers. However, even then your partner
> might not be enthusiastic about the heat and sound from one of these
> data center servers being run in your office or living room --- not to
> mention the electricity bill. :-)
Those are really noisy. I don't think it is realistic to use it.
Maybe there is an alternative, but I've no idea about how it
actually works: having multiple machines sharing the same model.
I heard some people using DGX Spark and AMD Strix Halo on such
configurations, but I suspect that performance would seriously
drop.
> Now, a H100 has 80 GB of High Bandwidth Memory (HBM) which has a
> bandwidth of 2 TB/s. So a server with 8 H100's has a 640 GB of HBM
> and an aggregate bandwidth of 16 TB/s. In contrast, a M5 Max with
> 128 GB and a 40-core GPU has a memory bandwidth of 614 GB/s. (A
> normal M5 Macbook Pro has 153 GB/s memory bandwidth and maxes out at
> 32GB.)
>
> Running the 671 billion parameter Deepseek R1 model at full precision
> requires 1.5TB of VRAM --- so two of these 8x H100 servers. Of could
> try to run them using quantization techniques, by collapsing each
> parameter from 16 bits to 4 bits, or even 1 bit to reduce the size of
> the ML rig required. However, you lose a lot of accuracy when you do
> that, and when models are much more prone to hallucinate when you use
> the more aggressive levels of quantization.
There are kv-algorithms that provide a good compromise, like turbo-quant.
the precision is still good with 4 or 3 bits.
> So it's one say to say, we should figure out how to try to run Sashiko
> on a local LLM, using open-weight models. But it's going to be a lot
> easier to propose such a thing than to actually do it.
The main point is: do we really need 671B parameters? Those models
speak a lot of different languages, have medical databases, and a lot
of other random knowledge that are useless for kernel development.
I've been playing for a while with qwen 3.6 with 24KB context size,
36B parameters (3B activated), 4bits kv quantization and it does produce
some decent results. The main limitation is the context size: it is
probably not big enough to test big files (*)
(*) my GPU has 16GB and it is not dedicated to LLM - still, it does
present results on a reasonable time (a couple of minutes) and
with decent precision.
Sure, while we have free tokens for sashiko, we can afford using
bigger models, but at the same time we should invest some effort
to make it viable on smaller models as well.
> What we *might* be able to try doing is to take an open-weight model
> that can fit on a 128GB machine, and then fine-tuning it by feeding it
> several years of LKML archives which we convniently have in public
> inbox format. This might allow us to create a specialized model which
> is optimized for the Linux kernel, and it would be interesting to see
> how this compares with a general frontier model.
>
> But even if we did that, it use would be limited to those people who
> can afford one of these 128GB unified memory machines or can otherwise
> get access to one of these machines.
I think we should aim on training an open-wight model up to 36B parameters
with 3-4B parameters activated. Those will run easily on 64B unified
memory even with a big context and may still work on machines with
16GB or with 32GB VRAM GPU (with partial CPU offload - specially
with 16GB).
Thanks,
Mauro
^ permalink raw reply [flat|nested] 14+ messages in thread* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-17 0:58 ` Mauro Carvalho Chehab
@ 2026-07-17 2:27 ` Theodore Tso
0 siblings, 0 replies; 14+ messages in thread
From: Theodore Tso @ 2026-07-17 2:27 UTC (permalink / raw)
To: Mauro Carvalho Chehab; +Cc: Jonathan Corbet, Sasha Levin, ksummit
On Fri, Jul 17, 2026 at 02:58:12AM -0500, Mauro Carvalho Chehab wrote:
>
> The main point is: do we really need 671B parameters? Those models
> speak a lot of different languages, have medical databases, and a lot
> of other random knowledge that are useless for kernel development.
Fair; but I do think we need to fine-tune it with a lot LKML text so
that it has the knowledge that we actually need for kernel development
before I'd trust the smaller models.
> I've been playing for a while with qwen 3.6 with 24KB context size,
> 36B parameters (3B activated), 4bits kv quantization and it does produce
> some decent results. The main limitation is the context size: it is
> probably not big enough to test big files (*)
I've been playing with qwen3-next with 256k context context size, 80B
parameters (3B activated), with 8bit quantization, but I haven't been
willing to trust it with generating kernel code. I have experimenting
to see how it compares with Gemini 3.1 when creating a python script
to send e-mail[1] or creating a bash completion script[2] for my
fstests test appliance.
[1] https://github.com/tytso/xfstests-bld/commit/dfadb2014da446ecb967de51904ee531f7be8bd5
[2] https://github.com/tytso/xfstests-bld/commit/dfadb2014da446ecb967de51904ee531f7be8bd5
However, from Roman tells me, Sashiko is running muliple LLM passes
using a frontier model for each commit review. So what Sashiko does
is quite a bit more complicated than a series of prompts such as:
Create a python program which submits an e-mail message using the
Submission Port (port 587), It should enable encryption using
STARTTLS and it should obtain the username and password from a
config.ini file. Model the python program using the send-mail.py
in the sandbox directory. It should support the same command-line
options but instead of sending the e-mail using sendgrid, it should
send the e-mail using the Submission protocol.
Please add pydoc documentation to send-mail-smtp.py.
Please enhance the program you just created (send-mail-smtp.py) to
support specifying a path to a certificate file in the
configuration file in case the user doesn't want to use the system
provided top-level trusted certificates.
Please add support for a configuration file parameter which
specifies whether TLS should be mandatory, optional, or disabled.
Update the pydoc documentation as necessary.
Please integrate the functionality found in send-mail.py into
send-mail-smtp.py. Instead of getting the Sendgrid API key from an
environment variable, change it to obtain the Sendgrid API key from
the configuration file. If the Sendgrid API key is specified, use
sendgrid instead of the SMTP submission protocol. Update the pydoc
documentation strings.
.... which does work pretty well even on less capable models that I
can run locally.
I guess we could try running Sashiko using ollama-mlx on a Macbook
with 128GB, and see how it works, but my assumption is that the answer
is "not well" --- which is why I really want to look at fine-tuning
one of these smaller models first.
Cheers,
- Ted
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-16 18:36 ` Jonathan Corbet
2026-07-16 19:53 ` Mauro Carvalho Chehab
@ 2026-07-16 20:05 ` Bart Van Assche
2026-07-16 20:52 ` James Bottomley
2026-07-16 20:23 ` Liam R. Howlett
2026-07-16 21:23 ` Theodore Tso
3 siblings, 1 reply; 14+ messages in thread
From: Bart Van Assche @ 2026-07-16 20:05 UTC (permalink / raw)
To: Jonathan Corbet, Sasha Levin; +Cc: ksummit, Roman Gushchin
On 7/16/26 11:36 AM, Jonathan Corbet wrote:
> What will we do when the current round of corporate generosity ends
> and that tool goes away? Maybe I'm worrying too much, but this does
> seem, to me, like a possibility we should keep in mind.
My understanding is that Sashiko is open source and also that it
supports multiple LLMs. I do not know whether it supports open weight
LLMs. Roman, please correct me if I got anything wrong.
Bart.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-16 20:05 ` Bart Van Assche
@ 2026-07-16 20:52 ` James Bottomley
0 siblings, 0 replies; 14+ messages in thread
From: James Bottomley @ 2026-07-16 20:52 UTC (permalink / raw)
To: Bart Van Assche, Jonathan Corbet, Sasha Levin; +Cc: ksummit, Roman Gushchin
On Thu, 2026-07-16 at 13:05 -0700, Bart Van Assche wrote:
>
> On 7/16/26 11:36 AM, Jonathan Corbet wrote:
> > What will we do when the current round of corporate generosity ends
> > and that tool goes away? Maybe I'm worrying too much, but this
> > does seem, to me, like a possibility we should keep in mind.
>
> My understanding is that Sashiko is open source and also that it
> supports multiple LLMs. I do not know whether it supports open weight
> LLMs. Roman, please correct me if I got anything wrong.
The corporate generosity isn't the code, it's the free tokens required
to run it on the LLM. If that free supply dries up, we'll have the
code but won't be able to afford to run it.
Regards,
James
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-16 18:36 ` Jonathan Corbet
2026-07-16 19:53 ` Mauro Carvalho Chehab
2026-07-16 20:05 ` Bart Van Assche
@ 2026-07-16 20:23 ` Liam R. Howlett
2026-07-16 21:23 ` Theodore Tso
3 siblings, 0 replies; 14+ messages in thread
From: Liam R. Howlett @ 2026-07-16 20:23 UTC (permalink / raw)
To: Jonathan Corbet; +Cc: Sasha Levin, ksummit
On 26/07/16 12:36PM, Jonathan Corbet wrote:
> Sasha Levin <sashal@kernel.org> writes:
>
> > On Thu, Jul 16, 2026 at 09:09:27AM -0600, Jonathan Corbet wrote:
> >>The use of LLMs in the development process appears to be a clear theme for
> >>the upcoming summit. On top of what others have already suggested, I think
> >>we may want to consider these questions:
> >>
> >>- Do we want to continue naming specific LLMs in the Assisted-by tags, or
> >> put something more generic? I *think* that this thread:
> >>
> >> https://lore.kernel.org/all/20260701-work-coding-assistants-v1-1-a20a94d1d606@kernel.org/
> >>
> >> reached a consensus that "Assisted-by: LLM" was better than what we
> >> require now, but it might be good to ratify that in this setting.
> >
> > So originally I've added the full name of the tool and LLM because there was
> > interest in a later audit of the tools to determine how useful (or useless)
> > some of the tools are.
> >
> > If those folks aren't interested in doing so anymore, then sure - we can drop
> > it.
> >
> > But... I find it difficult to see the point of having the tag if we do that.
>
> Folks like Greg have, in the recent past, said that it is useful even
> without specific product-name information:
>
> https://lore.kernel.org/all/2026070227-payroll-eradicate-8f66@gregkh/
>
There are a few classes of LLM users, but as more and more people are
using LLM, the usefulness of this tag is converging on zero. We seem to
have three classes of LLM users.
We have people leaning hard on LLMs to push more code, but are known
developers. The uses of an LLM can be replaced by using an RFC tag for
things that _really_ should be looked at by others prior to acceptance.
That is, if they aren't using RFC already.
We have people using LLMs for writing tests or reviewing code. A tag
here is not useful as it seems a false sense of security for reviewers
and they may skip it. The LLMs are finding off-by-ones, but the larger
ideas are usually sound and the test cases are actually not all bad.
> >>- There are many first-time contributors coming in with LLM-generated
> >> patches. At times, I could swear that every one of them is focused on
> >> documentation typos, but the truth of the matter is that they are
> >> reaching into subsystems all over the kernel. We have some brand-new
> >> contributors making significant changes to dozens of subsystems. An
> >> experienced developer would be hard-put to truly understand what those
> >> changes are doing; a newcomer is unlikely to have that understanding,
> >> and is unlikely to be around to fix eventual problems.
> >>
> >> Our maintainers are not scaling to handle this new flood, and I fear we
> >> are going to see some unfortunate things merged. One LLM-driven newcomer
> >> recently nearly succeeded in establishing himself as the maintainer of
> >> lib/. How do we hold the line against this stuff while remaining open to
> >> new developers?
> >
> > Shouldn't it be a merits question rather than a tools question?
> >
> > If the commits are correct, does it matter if they were written with
> > an LLM? we can insist more on supplying tests and demonstrating
> > correctness, something we seem to be doing quite rarely right now.
People who don't know what they are doing, cannot write tests that cover
the code they are changing. They don't know what to ask the LLM to do.
Fundamentally, we've given some people a tool that stretches well
outside what they know, so requesting more testing will not result in
valid tests. You end up writing things like "can you please ask your
LLM to <task>", or more likely something they can copy/paste into it
because they're saying they did it alone.
Half the time you tell them why it won't/doesn't work and they come back
with arguments from an LLM as to why you are wrong.
Also, they'll respin the reviewed patches within minutes, and that can
cost you hours to validate if it/they are not acceptable patch(es).
I've honestly thought about a rule where someone can no longer have
patches accepted again until the next release as they have exceeded a
threshold of time wastage. (A nack merit?)
>
> It's definitely a merit question. But we're not always all that good at
> determining whether a commit is correct, and we depend a lot on the
> contributor understanding their work and being around if something goes
> wrong with it. That is part of "merit" too. How confident are we of
> that merit when a brand-new developer makes significant changes to a
> dozen or more unrelated subsystems?
The last class of users..
We have people writing patch sets across many subsystems and/or first
time people rewriting entire sections of the kernel to 'fix' things.
Some of these people are clearly passing off LLM work as their own,
without any LLM tag. I am at a loss on how to deal with these. Either
they are encouraged to try over and over, or they blow up at the
suggesting this is LLM and/or a NACK. Neither method is good and nether
reduces the flow of bad code.
In any case, the tag is useless in this case because it's not going to
be used regardless of what we do.
So on Greg's note that having an LLM tag is a signal, I question the
SNR now and if the signal is the same indicator across all patches.
>
> >>- Our process is becoming increasingly dependent on proprietary tools. We
> >> have done that before and, in 2005, it went pretty badly for us - and
> >> could have been worse. How do we prepare for the inevitable rugpull? I
> >> raised this last year, and it was largely brushed off, but I still think
> >> it's something we should be concerned about.
> >
> > Are we dependent on them, or do we just find them very useful? If
> > Claude/Codex/etc goes away next month, will it stall any of our processes?
> >
> > We have AI reviews, we have many AI tools that help both authors and
> > maintainers, but I don't think that any of them play an integral part of our
> > process.
>
> The related discussions have featured a number of maintainers talking
> about how much time Sashiko has saved them. I believe them. How long
> will it take until nobody does that level of patch review anymore? What
> will we do when the current round of corporate generosity ends and that
> tool goes away? Maybe I'm worrying too much, but this does seem, to me,
> like a possibility we should keep in mind.
>
Agreed. I think we are already dependent enough on them to have a
visible impact on throughput. That's going to increase as models become
better. When the products shift from market capture to monetization,
we're going to have a dip in productivity and quality, at best.
Thanks,
Liam
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [MAINTAINERS SUMMIT] Other LLM-related topics - tags, newcomers, etc
2026-07-16 18:36 ` Jonathan Corbet
` (2 preceding siblings ...)
2026-07-16 20:23 ` Liam R. Howlett
@ 2026-07-16 21:23 ` Theodore Tso
3 siblings, 0 replies; 14+ messages in thread
From: Theodore Tso @ 2026-07-16 21:23 UTC (permalink / raw)
To: Jonathan Corbet; +Cc: Sasha Levin, ksummit
On Thu, Jul 16, 2026 at 12:36:53PM -0500, Jonathan Corbet wrote:
> The related discussions have featured a number of maintainers talking
> about how much time Sashiko has saved them. I believe them. How long
> will it take until nobody does that level of patch review anymore? What
> will we do when the current round of corporate generosity ends and that
> tool goes away? Maybe I'm worrying too much, but this does seem, to me,
> like a possibility we should keep in mind.
Sashiko finds issues that human reviewers wouldn't necessarily notice.
Before Sashiko, at least for ext4, I'd find those sorts of issues by
running 24 VM hours worth of fstests, and rely on that testing to find
problems that I might miss when doing the initial patch review. Could
I have found it by spending 3 times as much time looking at each
commit, and thinking deeply? Sure. But I don't have the time for
that, so I've *already* relied on Google contributing roughly two
dollars of VM time for each 24 hour regression test run. Sashiko
might be a bit more expensive on a per-patch basis, but just as I
never viewed fstests as a replacement for human-level review, I don't
think Sashiko is complete substitute human review.
We still need experienced human reviewers to ask whether the fix is
being done in the right place, or whether the right solution is to
*remove* code as opposed to *adding* code, etc. This is critical if
we want the code to be maintainable in the long-term. I agree that if
reviewers stop doing this, we would be in a world of hurt --- but that
can happen if people were to start relying on kunit tests passing as a
substitute for code review.
- Ted
^ permalink raw reply [flat|nested] 14+ messages in thread