TRASCRIZIONE 🔗
WEBVTT
00:00:00.050 --> 00:00:02.649
Think about the breaker panel in your house.
00:00:02.770 --> 00:00:06.490
Put everything on one circuit and a single fault
00:00:06.490 --> 00:00:09.990
takes out the whole house. Then you are hunting
00:00:09.990 --> 00:00:13.349
with a flashlight. Label the breakers and you
00:00:13.349 --> 00:00:16.710
know it's the kitchen before you get up. My guest
00:00:16.710 --> 00:00:21.170
today builds AI for lawyers and his whole approach
00:00:21.170 --> 00:00:24.550
is about labeling the breakers. Everybody's heard
00:00:24.550 --> 00:00:28.730
about AI making up court cases. He says the more
00:00:28.730 --> 00:00:32.109
interesting hallucination is a real case that
00:00:32.109 --> 00:00:36.350
just doesn't apply. And he says even if a fully
00:00:36.350 --> 00:00:40.649
agentic system got you 20% better results, you'd
00:00:40.649 --> 00:00:43.929
never know how it breaks until it breaks. I'm
00:00:43.929 --> 00:00:46.429
Brian Teller from Teller's Tech, and this is
00:00:46.429 --> 00:00:47.495
Ship It Weekly.
00:01:04.395 --> 00:01:05.909
Welcome back to Ship It Weekly,
00:01:06.090 --> 00:01:09.150
where I filter the noise and focus on what matters
00:01:09.150 --> 00:01:11.709
when you are the one running infrastructure and
00:01:11.709 --> 00:01:15.069
owning reliability. Most weeks, it's a quick
00:01:15.069 --> 00:01:19.569
DevOps, SRE, platform, cloud, and security news
00:01:19.569 --> 00:01:23.069
recap. In between those, I do conversation episodes
00:01:23.069 --> 00:01:26.670
with people building and operating these systems.
00:01:27.209 --> 00:01:30.989
Ship It Weekly turns one this month. To celebrate,
00:01:31.250 --> 00:01:34.370
through mid-December, you are getting two conversations
00:01:34.370 --> 00:01:37.909
a week, Mondays and Wednesdays, with the news
00:01:37.909 --> 00:01:41.650
every Friday. Then we're back to our usual rhythm.
00:01:42.109 --> 00:01:45.629
Today, I'm joined by Devansh, founder and head
00:01:45.629 --> 00:01:49.769
of AI at Irys. He builds reasoning and verification
00:01:49.769 --> 00:01:54.489
systems for legal AI, which is a brutal place
00:01:54.489 --> 00:01:58.030
for hallucinations and confident wrong answers.
00:01:58.290 --> 00:02:02.170
We talk about why the trickiest legal hallucination
00:02:02.170 --> 00:02:07.030
is a real case used the wrong way. How his team
00:02:07.030 --> 00:02:11.189
splits the system into three tiers so they can
00:02:11.189 --> 00:02:14.509
find failures fast what it's like running AI
00:02:14.509 --> 00:02:17.689
in production when the model can change under
00:02:17.689 --> 00:02:21.629
you and why he calls reinforcement learning the
00:02:21.629 --> 00:02:26.858
most overhyped technique in AI all right let's jump in
00:02:31.574 --> 00:02:33.069
thank you for joining me thank you for
00:02:33.069 --> 00:02:36.319
having me Brian very excited to be here so tell
00:02:36.319 --> 00:02:39.699
me a little bit about Irys and and
Load full transcript
what what
00:02:39.699 --> 00:02:42.000
you're working on there and what problems you're
00:02:42.000 --> 00:02:44.620
running into with AI and what problems you're
00:02:44.620 --> 00:02:49.370
solving. Absolutely. So I think to summarize
00:02:49.370 --> 00:02:54.289
something very briefly, what we see in legal
00:02:54.289 --> 00:02:57.110
AI, and this is a large problem in AI in general,
00:02:57.169 --> 00:03:01.210
is that language models and the fundamental nature
00:03:01.210 --> 00:03:03.530
of AI research, et cetera, is getting really,
00:03:03.590 --> 00:03:05.789
really good at doing a lot of things somewhat
00:03:05.789 --> 00:03:08.750
well. You know, that's the whole phenomenon of
00:03:08.750 --> 00:03:12.280
AI slop. But when it comes to going from somewhat
00:03:12.280 --> 00:03:16.120
well to being exceptional and being reliable
00:03:16.120 --> 00:03:20.639
and like the last mile delivery concept, our
00:03:20.639 --> 00:03:24.340
systems are not able to keep up. And what we
00:03:24.340 --> 00:03:26.699
are essentially trying to do is we're re-engineering
00:03:26.699 --> 00:03:31.699
a lot of structure into the way AI is done with
00:03:31.699 --> 00:03:34.639
the whole understanding that if we can have structure.
00:03:35.189 --> 00:03:38.129
We can better predict ahead of time when a system
00:03:38.129 --> 00:03:40.250
would fail, when it would succeed, how to make
00:03:40.250 --> 00:03:42.169
the system work better. And when a user comes
00:03:42.169 --> 00:03:44.430
to us and says, hey, can you, this is not working
00:03:44.430 --> 00:03:46.530
out for us. Can you improve this? Or this works
00:03:46.530 --> 00:03:49.129
really well. Can we double down on this? The
00:03:49.129 --> 00:03:52.150
ability to have structure and engineering stuff
00:03:52.150 --> 00:03:54.810
allows us to work on that instead of just like,
00:03:54.889 --> 00:03:57.030
you know, hoping that a better system prompt
00:03:57.030 --> 00:03:59.469
or a better model will solve those problems because
00:03:59.469 --> 00:04:02.550
that's not really fair to the user or to ourselves
00:04:02.550 --> 00:04:05.349
if we're doing that. Yeah, that makes sense.
00:04:05.889 --> 00:04:09.229
So why does AI for professionals break down so
00:04:09.229 --> 00:04:11.930
often? And what does it take for a system to
00:04:11.930 --> 00:04:14.210
be trustworthy? I think that's a great question.
00:04:14.289 --> 00:04:17.430
And I think what it takes depends heavily on
00:04:17.430 --> 00:04:20.230
what your system is expected to do. Essentially,
00:04:20.250 --> 00:04:22.610
any software system you build is a contract between
00:04:22.610 --> 00:04:26.910
the system and the user, where you say the system
00:04:26.910 --> 00:04:28.930
has to do certain things, the user has to do
00:04:28.930 --> 00:04:30.870
certain things, they're expected to engage with
00:04:30.870 --> 00:04:33.360
the system in a certain way the struggle we're
00:04:33.360 --> 00:04:35.480
often facing is that these contracts aren't very
00:04:35.480 --> 00:04:37.560
well defined and they're always changing you
00:04:37.560 --> 00:04:41.100
know new AI tools and new AI workflows will come
00:04:41.100 --> 00:04:43.939
around and they'll completely change how you
00:04:43.939 --> 00:04:47.379
should be operating it or how to best use certain
00:04:47.379 --> 00:04:49.839
things and not use certain things so I think
00:04:49.839 --> 00:04:54.240
at its core of where things fail is often it's
00:04:54.240 --> 00:04:56.740
either users expect things that are not viable
00:04:56.740 --> 00:05:00.110
with the system or they're not educated and What
00:05:00.110 --> 00:05:02.050
we really need to be doing is figuring out better
00:05:02.050 --> 00:05:05.670
ways to integrate. We need better ways to identify
00:05:05.670 --> 00:05:08.730
what users are able to do well and what users
00:05:08.730 --> 00:05:11.569
are, what should be taken out of their hands
00:05:11.569 --> 00:05:14.189
and done through AI systems. Like right now it's
00:05:14.189 --> 00:05:16.290
very vibezy, it's kind of try everything,
00:05:16.470 --> 00:05:19.329
stick it, throw it at the wall. And this is where
00:05:19.329 --> 00:05:21.170
a lot of things are failing in practice because
00:05:21.170 --> 00:05:23.569
users don't, like users don't know what they
00:05:23.569 --> 00:05:26.290
have to verify, you know, where they should step
00:05:26.290 --> 00:05:29.779
in versus where the AI can be trusted without
00:05:29.779 --> 00:05:33.560
too much worry. Yeah. So in legal specifically,
00:05:33.920 --> 00:05:37.240
what is the actual failure mode that you're seeing?
00:05:37.480 --> 00:05:41.040
Is it wrong facts? Is it fake citations? We've
00:05:41.040 --> 00:05:43.379
heard some of that in the industry over the last
00:05:43.379 --> 00:05:45.779
couple of years. Or is it like missing nuance
00:05:45.779 --> 00:05:48.720
or misreading authority? What are the main issues
00:05:48.720 --> 00:05:51.800
that you see? I think the way we've broken it
00:05:51.800 --> 00:05:54.240
down, we often talk about three categories of
00:05:54.240 --> 00:05:57.970
hallucinations. One category of hallucination
00:05:57.970 --> 00:06:02.069
is you say what you're describing fake case citations.
00:06:02.189 --> 00:06:05.509
This AI is just saying like Devansh versus Brian,
00:06:05.790 --> 00:06:08.689
for the Supreme Court, when it doesn't exist.
00:06:08.990 --> 00:06:12.170
The second much more interesting hallucination
00:06:12.170 --> 00:06:16.209
is AI says it's a real hallucination. It's a
00:06:16.209 --> 00:06:18.930
real case and it is a real case, but it doesn't
00:06:18.930 --> 00:06:22.209
apply to the way it's doing because a lot of
00:06:22.209 --> 00:06:25.290
legal is interpretation it's not exactly like
00:06:25.290 --> 00:06:28.110
this case says this only and this will be applied
00:06:28.110 --> 00:06:30.209
in this context you say a lot of what lawyers
00:06:30.209 --> 00:06:33.110
do in litigation etc is essentially take what
00:06:33.110 --> 00:06:35.230
was already argued and say okay based on this
00:06:35.230 --> 00:06:37.569
we can apply it in this format or in this way
00:06:37.569 --> 00:06:42.019
so a lot of um AI tools where they're failing
00:06:42.019 --> 00:06:43.779
right now is they're misapplying case law they're
00:06:43.779 --> 00:06:46.199
they're creating bad legal strategy because of
00:06:46.199 --> 00:06:49.470
this and this is the part that a lot of basically
00:06:49.470 --> 00:06:51.589
nobody else is solving this was something that
00:06:51.589 --> 00:06:53.910
we were one of the first people to actually actively
00:06:53.910 --> 00:06:56.389
try to solve for and this is very much the next
00:06:56.389 --> 00:06:59.959
domain of problem solving where it would be formed
00:06:59.959 --> 00:07:02.920
is when you have very long contexts and very
00:07:02.920 --> 00:07:06.079
long amounts of information you have a lot of
00:07:06.079 --> 00:07:10.269
like conditional applications like this contract
00:07:10.269 --> 00:07:12.689
in this clause will be modified by another clause
00:07:12.689 --> 00:07:15.689
somewhere else, or in certain cases you could
00:07:15.689 --> 00:07:17.970
apply this kind of an argumentation to it, or
00:07:17.970 --> 00:07:20.290
certain law firms have certain ways of operating,
00:07:20.610 --> 00:07:22.790
certain lawyers have certain preferences for
00:07:22.790 --> 00:07:25.189
how they want to structure arguments, whether
00:07:25.189 --> 00:07:27.290
they want to rely on one kind of authority over
00:07:27.290 --> 00:07:30.589
another, etc. And that kind of judgment is not
00:07:30.589 --> 00:07:33.589
something that AI... gets naturally because again
00:07:33.589 --> 00:07:35.750
if you think about what AI is it's trained to
00:07:35.750 --> 00:07:37.829
give you the most probable output token at any
00:07:37.829 --> 00:07:40.149
given time which is a kind of reversion to the
00:07:40.149 --> 00:07:42.750
mean better models mean that you're reverting
00:07:42.750 --> 00:07:45.269
to a higher mean than before but you can never
00:07:45.269 --> 00:07:49.230
have this outlier-ish final touch application
00:07:49.230 --> 00:07:55.769
with just AI averages so on that lens we are um
00:07:55.769 --> 00:07:59.649
like you know How does an AI know when to apply
00:07:59.649 --> 00:08:02.389
what kind of conditionals and logic and how to
00:08:02.389 --> 00:08:05.189
chain different things together is one of the
00:08:05.189 --> 00:08:08.569
biggest breakthroughs that we had. And I think
00:08:08.569 --> 00:08:10.990
this is where the next generation of products
00:08:10.990 --> 00:08:14.990
will differentiate themselves. So in researching
00:08:14.990 --> 00:08:18.529
for this episode, I looked at Irys One. Can you
00:08:18.529 --> 00:08:20.850
walk me through the three-tier reasoning architecture?
00:08:21.509 --> 00:08:24.230
What are the tiers?
What do each of them mean?
00:08:24.310 --> 00:08:27.050
What do they do? Absolutely. And I think instead
00:08:27.050 --> 00:08:30.430
of... really talking about the technicals on
00:08:30.430 --> 00:08:33.169
uh what each of those tiers are I think first
00:08:33.169 --> 00:08:35.850
it's very important to just understand what kind
00:08:35.850 --> 00:08:39.230
of problems we have on our system when we're
00:08:39.230 --> 00:08:40.769
talking about stuff like how do you know when
00:08:40.769 --> 00:08:43.129
to apply which kind of logic these conditionals
00:08:43.129 --> 00:08:45.370
we're talking about contracts modify so the first
00:08:45.370 --> 00:08:48.049
tier would just be ingestion you need to be able
00:08:48.049 --> 00:08:52.169
to process your input and already here a lot of
00:08:52.169 --> 00:08:55.330
legal AI fails completely why would it fail Brian
00:08:55.330 --> 00:08:58.100
because if you think about input A lot of the
00:08:58.100 --> 00:09:00.840
times your input are not clean, one-shot texts
00:09:00.840 --> 00:09:03.899
with everything marked out. A lot of your users
00:09:03.899 --> 00:09:06.679
and viewers will probably have either dealt with
00:09:06.679 --> 00:09:09.460
it explicitly or would be figuring this out right
00:09:09.460 --> 00:09:12.299
now. Like if you had a PDF, you might have handwritten
00:09:12.299 --> 00:09:14.679
notes on it. You might have a PDF and an email
00:09:14.679 --> 00:09:16.919
chain where the email chain refers to some part
00:09:16.919 --> 00:09:19.860
of the PDF. But it's because those are two different
00:09:19.860 --> 00:09:21.980
documents. Those aren't going to be chunked in
00:09:21.980 --> 00:09:23.960
any meaningful way where it's going to happen.
00:09:24.480 --> 00:09:27.820
You will, like you just mentioned, Irys, one
00:09:27.820 --> 00:09:30.580
and the three tiers like you're referring to
00:09:30.580 --> 00:09:33.240
some document somewhere but that's not in the
00:09:33.240 --> 00:09:36.039
nature of this conversation so if you were just
00:09:36.039 --> 00:09:39.139
ingesting the way most people are currently doing
00:09:39.139 --> 00:09:41.200
it which is just document one document two document
00:09:41.200 --> 00:09:44.240
three you will miss these aspects or if some
00:09:44.240 --> 00:09:46.039
people are doing with chunks you will miss these
00:09:46.039 --> 00:09:48.340
aspects so the first thing that really sets us
00:09:48.340 --> 00:09:52.000
apart is the ingestion so ingestion is what allows
00:09:52.000 --> 00:09:55.200
us to map different concepts map different ideas
00:09:55.769 --> 00:09:58.289
now we do this through a knowledge graph that
00:09:58.289 --> 00:10:00.409
allows our AI to build kind of a context map
00:10:00.409 --> 00:10:02.690
with the point of it being that when our reasoning
00:10:02.690 --> 00:10:04.970
system is looking through something it's really
00:10:04.970 --> 00:10:09.370
looking at how those ideas are linked this clause
00:10:09.370 --> 00:10:12.129
where which claw if there was a clause a what
00:10:12.129 --> 00:10:14.750
other clauses is clause a related to and then
00:10:14.750 --> 00:10:16.649
what this allows us to do is we can have that
00:10:16.649 --> 00:10:20.029
thing be done at one ingestion time and two now
00:10:20.029 --> 00:10:22.649
we're rolling out a lot of work for background
00:10:22.649 --> 00:10:24.929
processes, checks, constant repairs, because
00:10:24.929 --> 00:10:27.250
what we're realizing is context is exploding
00:10:27.250 --> 00:10:29.230
so much. People are uploading so many documents
00:10:29.230 --> 00:10:32.269
that it's no longer viable for us to do one-time
00:10:32.269 --> 00:10:34.990
ingestion. It's much better for us to have constant
00:10:34.990 --> 00:10:38.610
batch of updates and improve the self-organizing
00:10:38.610 --> 00:10:43.480
graphs that way. Aspect two of this is we need
00:10:43.480 --> 00:10:46.600
to be able to then reason over this. Think of
00:10:47.190 --> 00:10:49.409
The ingestion step basically builds the world
00:10:49.409 --> 00:10:52.590
that the AI will navigate in. Now the AI needs
00:10:52.590 --> 00:10:56.090
to navigate this world better, take use of existing
00:10:56.090 --> 00:11:01.090
tools, get use of existing systems. That's where
00:11:01.090 --> 00:11:03.210
tier two comes in. That's where our research
00:11:03.210 --> 00:11:06.169
goes into latent space reasoning, hierarchical
00:11:06.169 --> 00:11:08.629
embeddings, et cetera. What all of these allow
00:11:08.629 --> 00:11:11.490
us to do is we are able to navigate the world
00:11:11.490 --> 00:11:14.389
better through these. We're analyzing the systems.
00:11:14.409 --> 00:11:18.889
We're like, okay, AI has to be like for this
00:11:18.889 --> 00:11:21.090
kind of a legal strategy we simulate like 50
00:11:21.090 --> 00:11:23.269
different outcomes and we're like this should
00:11:23.269 --> 00:11:25.429
work well this doesn't this might not work well
00:11:25.429 --> 00:11:27.549
this is how we might synthesize across aspects
00:11:27.549 --> 00:11:30.230
and then finally we get to the final tier which
00:11:30.230 --> 00:11:32.870
is the delivery and I'm breaking this down more
00:11:32.870 --> 00:11:36.149
from a user impact perspective as opposed to
00:11:36.149 --> 00:11:39.710
necessarily engineering layers or whatever because
00:11:39.710 --> 00:11:42.490
I think you can always get into arbitrarily deep
00:11:42.490 --> 00:11:44.769
engineering conversations but it's always better
00:11:44.769 --> 00:11:47.850
to I think ground that in a user perspective
00:11:47.850 --> 00:11:51.250
and how it reflects in users. And that way, when
00:11:51.250 --> 00:11:53.110
you're designing an improving system, you can
00:11:53.110 --> 00:11:56.009
always narrow it down. So the final tier is firm
00:11:56.009 --> 00:11:59.570
management, firm context, individual lawyer judgment
00:11:59.570 --> 00:12:01.929
and context, which is
as we're reasoning through
00:12:01.929 --> 00:12:03.590
things, as we mentioned earlier, certain law
00:12:03.590 --> 00:12:07.110
firms like to operate in certain ways or a certain
00:12:07.110 --> 00:12:09.169
lawyer likes to be more aggressive while others
00:12:09.169 --> 00:12:11.950
don't like to be as aggressive. If you have matter
00:12:11.950 --> 00:12:13.789
memory, firm memory on what has worked well,
00:12:13.809 --> 00:12:16.009
what hasn't worked well, this is where you want
00:12:16.009 --> 00:12:17.929
to start applying that judgment to really refine
00:12:17.929 --> 00:12:20.990
it and bring it closer to the end product that
00:12:20.990 --> 00:12:23.149
the user might need as opposed to just a generic
00:12:23.149 --> 00:12:28.029
AI a draft that anybody could spit out um so
00:12:28.029 --> 00:12:29.649
this is where the final tier comes in which is
00:12:29.649 --> 00:12:32.649
personalization and output uh orientation that
00:12:32.649 --> 00:12:35.370
way and essentially by breaking it down into
00:12:35.370 --> 00:12:38.929
these three tiers it's like we're able to systematically
00:12:38.929 --> 00:12:42.500
improve each aspect of it because you know if
00:12:42.500 --> 00:12:44.080
a user is saying this is missing information,
00:12:44.379 --> 00:12:46.360
we immediately know that we don't need to care
00:12:46.360 --> 00:12:48.200
about the personalization there. This is either
00:12:48.200 --> 00:12:51.340
in the ingestion layer, but data wasn't processed,
00:12:51.379 --> 00:12:53.820
or it's in the reasoning layer where we were
00:12:53.820 --> 00:12:56.500
navigating it. And then just by looking at that,
00:12:56.519 --> 00:12:58.480
we've cut down the search space, and then immediately
00:12:58.480 --> 00:13:00.460
we can just say, can we find this in the data
00:13:00.460 --> 00:13:02.620
anywhere? Oh, this already exists in the graph.
00:13:02.700 --> 00:13:05.620
So it is ingestion there, but working fine. For
00:13:05.620 --> 00:13:08.879
some reason, it's not able to link to... it's
00:13:08.879 --> 00:13:10.919
not able to reason over it so is it a issue of
00:13:10.919 --> 00:13:13.679
the structure and how we've mapped it out or
00:13:13.679 --> 00:13:16.240
is it something wrong with our AI reasoning systems
00:13:16.240 --> 00:13:20.080
so by breaking these down like with very two
00:13:20.080 --> 00:13:22.279
three targeted queries we're already able to
00:13:22.279 --> 00:13:25.899
isolate where systems are failing and how we
00:13:25.899 --> 00:13:28.840
can more best improve them and that's probably
00:13:28.840 --> 00:13:31.379
one of the most important things for an agentic
00:13:31.379 --> 00:13:33.259
system to be building because you have a lot
00:13:33.259 --> 00:13:35.820
of moving parts so you want to isolate sources
00:13:35.820 --> 00:13:40.159
of error as quickly as possible So you had mentioned
00:13:40.159 --> 00:13:42.799
a knowledge graph when setting up your system
00:13:42.799 --> 00:13:46.120
with the tiering. What problem does that solve
00:13:46.120 --> 00:13:50.460
for you? So simply put, the knowledge graph exists
00:13:50.460 --> 00:13:54.159
to teach our AI layers where to look. Because
00:13:54.159 --> 00:13:57.519
if you think about what kinds of retrieval systems
00:13:57.519 --> 00:14:00.740
there are, you can build either something like
00:14:00.740 --> 00:14:04.200
what RAG systems do, which is chunking, where
00:14:04.200 --> 00:14:07.340
you're just trying to hope that all... relevant
00:14:07.340 --> 00:14:10.379
information is isolated within certain geographical
00:14:10.379 --> 00:14:14.019
proximities you know if your chunk size is 1024
00:14:14.019 --> 00:14:16.559
tokens you're just saying everything relevant
00:14:16.559 --> 00:14:19.240
for this chunk is contained within that geography
00:14:19.240 --> 00:14:23.019
or it's contained within similar semantic search
00:14:23.019 --> 00:14:26.240
geographies of 1024 tokens and if I do a top
00:14:26.240 --> 00:14:29.460
k of 5 or 20 I will be able to retrieve all of
00:14:29.460 --> 00:14:32.500
this my LLM can reason over that which is factually
00:14:32.500 --> 00:14:35.440
not true because a lot of the times if a user
00:14:35.440 --> 00:14:37.480
is asking hey where does this issue go wrong
00:14:37.480 --> 00:14:40.039
what does this issue mean how are you going to
00:14:40.039 --> 00:14:43.139
semantically chunk that and hope that every variable
00:14:43.139 --> 00:14:48.870
there is able to fix its problem um so on that
00:14:48.870 --> 00:14:52.610
lens there is no Like what the knowledge graph
00:14:52.610 --> 00:14:54.509
allows us to do is if we were to link different
00:14:54.509 --> 00:14:56.549
statutes together, if we were to link different
00:14:56.549 --> 00:14:59.690
entities together, it just tells our AI that
00:14:59.690 --> 00:15:03.029
when you're analyzing Brian's, like if somebody
00:15:03.029 --> 00:15:05.450
asks you a question about Brian, maybe Brian
00:15:05.450 --> 00:15:09.129
has some kind of a... with Devansh. Devansh might
00:15:09.129 --> 00:15:11.710
not show up on a semantic search as much as Brian
00:15:11.710 --> 00:15:15.009
does, but by linking those on the graph, it also
00:15:15.009 --> 00:15:18.409
knows to look at Devansh and then explore all
00:15:18.409 --> 00:15:20.529
the areas where Devansh and Brian have overlapping
00:15:20.529 --> 00:15:24.029
notes, overlapping judgments, and then take those
00:15:24.029 --> 00:15:26.269
aspects into account. So the knowledge graph
00:15:26.269 --> 00:15:29.830
essentially is the map on which our AI reasons
00:15:29.830 --> 00:15:32.679
and figures it out. It's not the direct. Like
00:15:32.679 --> 00:15:34.559
we're not directly doing graph RAG because graph
00:15:34.559 --> 00:15:36.620
RAG doesn't really work when you're trying to
00:15:36.620 --> 00:15:38.480
isolate specific chunks and you're trying to
00:15:38.480 --> 00:15:41.419
say, okay, quote, tell me
where this person said
00:15:41.419 --> 00:15:45.759
this or flag contradictions. But it's primitive
00:15:45.759 --> 00:15:48.879
for figuring out relationships, figuring out
00:15:48.879 --> 00:15:51.500
how things might be linked. Graphs are a very
00:15:51.500 --> 00:15:55.490
good first level. And then on top of that, we've
00:15:55.490 --> 00:15:58.210
built self-evolving structures so that when
00:15:58.210 --> 00:16:01.129
we extract insights and we start saying, okay,
00:16:01.169 --> 00:16:03.149
this is an assertion, this is a claim, or let
00:16:03.149 --> 00:16:05.750
me add this fact about Devansh. Like Devansh
00:16:05.750 --> 00:16:07.929
loves dogs, Brian loves dogs. Maybe that's something
00:16:07.929 --> 00:16:10.970
they bond over. Like when we have that relationship
00:16:10.970 --> 00:16:15.049
added, we should be able to propagate that insight
00:16:15.049 --> 00:16:17.210
across the whole graph. And how does that relationship
00:16:17.210 --> 00:16:20.230
change? Because there's an aspect of dogs. By
00:16:20.230 --> 00:16:22.389
using graphs, you can do stuff like message passing
00:16:22.389 --> 00:16:26.590
and do stuff like Bayesian math in order to propagate
00:16:26.590 --> 00:16:30.389
signals across the whole network and be able
00:16:30.389 --> 00:16:32.549
to get relatively good results. It's not super
00:16:32.549 --> 00:16:35.370
precise, but it's like there's a big difference
00:16:35.370 --> 00:16:38.990
between zero, doing the zero or letting an LLM
00:16:38.990 --> 00:16:41.129
implicitly do this for 10, 15 percent versus
00:16:41.129 --> 00:16:43.950
doing it 60 percent well. And that's really a
00:16:43.950 --> 00:16:46.700
big differentiator for us. How do you manage
00:16:46.700 --> 00:16:49.659
context when the corpus goes from like 10 docs
00:16:49.659 --> 00:16:53.639
to 10 gigabytes? Size isn't always that much
00:16:53.639 --> 00:16:57.370
of a problem. Because again, when you're laying
00:16:57.370 --> 00:16:59.370
this out this way, it just means your math will
00:16:59.370 --> 00:17:01.990
take a little bit longer to run and signals will
00:17:01.990 --> 00:17:04.369
take longer to process. But a lot of this is
00:17:04.369 --> 00:17:08.990
just math. So it's not as though it takes meaningfully
00:17:08.990 --> 00:17:12.250
longer amounts of time. We actually have very
00:17:12.250 --> 00:17:14.430
good long context reasoning and we're able to
00:17:14.430 --> 00:17:18.589
do this. The deeper issues with scale is just
00:17:18.589 --> 00:17:22.950
missing random nuances for which the best solution
00:17:22.950 --> 00:17:26.890
right now is just, you know, flag issues to users.
00:17:27.210 --> 00:17:29.950
So our system, we're working on rolling this
00:17:29.950 --> 00:17:32.690
out properly, but by giving users the ability
00:17:32.690 --> 00:17:35.869
to have absolute transparency into what we're
00:17:35.869 --> 00:17:38.069
looking at and what we're not looking at and
00:17:38.069 --> 00:17:40.329
being like, okay, these seem to be some open
00:17:40.329 --> 00:17:43.130
gaps. These seem to be some next level assumptions
00:17:43.130 --> 00:17:45.960
we're making we're basically able to isolate
00:17:45.960 --> 00:17:49.279
very quickly the second aspect to this is provenance
00:17:49.279 --> 00:17:51.660
with the knowledge graph with everything else
00:17:51.660 --> 00:17:54.380
when you have specific chunks and you're able
00:17:54.380 --> 00:17:57.099
to say this belongs here this belongs here uh
00:17:57.099 --> 00:17:59.619
the second aspect so the secondary benefit to
00:17:59.619 --> 00:18:01.859
this is when the AI is looking to analyze something
00:18:01.859 --> 00:18:07.160
it can look at where certain chunks tend to cluster
00:18:07.160 --> 00:18:09.799
if certain information tends to come in and out
00:18:09.799 --> 00:18:12.859
it's like a Wikipedia or like any index right
00:18:13.519 --> 00:18:16.140
it gets an idea of where it should look for what
00:18:16.140 --> 00:18:19.240
kind of information and that actually solves
00:18:19.240 --> 00:18:21.640
a lot more long context reasoning than you think
00:18:21.640 --> 00:18:25.559
because with very large corpus it's not like
00:18:25.559 --> 00:18:28.059
you need every it's not like there's a hundred
00:18:28.059 --> 00:18:29.859
chunks from a hundred different documents that
00:18:29.859 --> 00:18:33.880
matter it's still just my for most tasks it's
00:18:33.880 --> 00:18:35.900
still only like three four five documents that
00:18:35.900 --> 00:18:38.680
actually matter it's just the the difficulties
00:18:38.680 --> 00:18:41.890
in finding those three four five documents And
00:18:41.890 --> 00:18:44.369
being able to implement this structure helps
00:18:44.369 --> 00:18:47.950
there. Now, this approach kind of starts to break
00:18:47.950 --> 00:18:49.490
down when you're doing stuff like timelines,
00:18:49.750 --> 00:18:52.349
where you do really have to thread every document,
00:18:52.430 --> 00:18:55.809
every event chronologically, and especially because
00:18:55.809 --> 00:18:58.170
it's not like documents themselves are organized
00:18:58.170 --> 00:19:00.869
with chronologies and doing this. So this is
00:19:00.869 --> 00:19:03.910
really the next level of system that we're trying
00:19:03.910 --> 00:19:06.269
to build. We're trying to figure it out. That
00:19:06.269 --> 00:19:08.170
was one of the big inspirations for making this
00:19:08.170 --> 00:19:11.079
as self-evolving as possible if you add a new
00:19:11.079 --> 00:19:12.980
date instead of us having to manually insert
00:19:12.980 --> 00:19:16.400
that date and revise conclusions can we do this
00:19:16.400 --> 00:19:19.279
automatically and if it's causing any conflicts
00:19:19.279 --> 00:19:21.039
with the graph can we surface that to the user
00:19:21.039 --> 00:19:23.740
to be like hey can you this scene our graph system
00:19:23.740 --> 00:19:26.599
is kind of causing issues here can
you help us
00:19:26.599 --> 00:19:31.059
fix it uh but like honestly that is the biggest
00:19:31.059 --> 00:19:33.920
problem that anybody nobody has really sorted
00:19:33.920 --> 00:19:36.160
it out yet I've been talking to a lot of people
00:19:36.160 --> 00:19:39.170
not just in legal but in agentic coding systems
00:19:39.170 --> 00:19:44.069
where you have very quick deltas or in finance.
00:19:44.250 --> 00:19:47.410
And this is an open issue where everybody is
00:19:47.410 --> 00:19:50.589
kind of interested. And people have some theories,
00:19:50.650 --> 00:19:53.250
but we haven't super sorted that part out yet.
00:19:55.109 --> 00:19:58.109
What's your strategy for dealing with it depends
00:19:58.109 --> 00:20:01.579
or ambiguity answers? It depends, because sometimes
00:20:01.579 --> 00:20:03.859
it depends is the best answer you can give. If
00:20:03.859 --> 00:20:05.759
the user is asking you what the best solution
00:20:05.759 --> 00:20:08.579
is, you don't want to be overly confident. Other
00:20:08.579 --> 00:20:10.380
times when they're asking for a legal draft,
00:20:10.619 --> 00:20:13.359
you don't want to say it depends, because lawyer
00:20:13.359 --> 00:20:16.180
drafts have to be very authoritative and stating
00:20:16.180 --> 00:20:18.819
wise. So first, it depends on what the end goal
00:20:18.819 --> 00:20:20.759
is. And we have broken this down extensively
00:20:20.759 --> 00:20:24.000
on like, what is the user asking for? And behind
00:20:24.000 --> 00:20:25.339
the scenes, this is one of the things we do,
00:20:25.400 --> 00:20:28.900
we create answer contracts. If it's a legal draft,
00:20:29.039 --> 00:20:31.000
if it's a memo, it has to have these properties.
00:20:31.140 --> 00:20:33.700
If it's something else, it has to have certain
00:20:33.700 --> 00:20:37.420
other properties. So that's a pretty good first
00:20:37.420 --> 00:20:40.640
order system to dealing with it depends. And
00:20:40.640 --> 00:20:43.380
I think that is something that our users have
00:20:43.380 --> 00:20:45.440
really appreciated. If you look at a lot of what
00:20:45.440 --> 00:20:47.500
Reddit says about us or other people say about
00:20:47.500 --> 00:20:52.339
us, or we were cited by some paper, this organization
00:20:52.339 --> 00:20:55.450
that was looking at tools for... legal judgment.
00:20:55.589 --> 00:20:57.809
One of the things people like about us is that
00:20:57.809 --> 00:21:00.210
we're able to help with brainstorming a much
00:21:00.210 --> 00:21:03.190
larger set of possible strategies, precisely
00:21:03.190 --> 00:21:06.910
because we have implicitly factored in uncertainties
00:21:06.910 --> 00:21:10.410
and counterfactual analysis and being like, we
00:21:10.410 --> 00:21:12.990
can go alternative theory, one, two, three, et
00:21:12.990 --> 00:21:15.529
cetera. Let me take like a DevOps angle here.
00:21:15.750 --> 00:21:18.549
What is the hardest part about operating an AI
00:21:18.549 --> 00:21:23.200
system like this in production? 100%. Being able
00:21:23.200 --> 00:21:27.099
to track everything, knowing control, understanding
00:21:27.099 --> 00:21:29.700
like, hey, if things are going wrong, where are
00:21:29.700 --> 00:21:31.339
they going wrong or where are they going right?
00:21:31.700 --> 00:21:36.160
One of the issues is we ourselves have such a
00:21:36.160 --> 00:21:39.000
limited understanding of what AI does well or
00:21:39.000 --> 00:21:43.160
not do well. And we as an AI engineers in general,
00:21:43.200 --> 00:21:46.690
like we are struggling very much to get they
00:21:46.690 --> 00:21:48.750
have benchmarks but benchmarks have completely
00:21:48.750 --> 00:21:51.750
gone out of touch with a lot of reality and then
00:21:51.750 --> 00:21:54.769
it gets further dependent on how exactly you
00:21:54.769 --> 00:21:56.690
set up your systems and which tools you're using
00:21:56.690 --> 00:21:59.910
which ones here how you're structuring them on
00:21:59.910 --> 00:22:04.849
that lens it becomes very difficult to um isolate
00:22:04.849 --> 00:22:07.430
error and that's why such a big part of our system
00:22:07.430 --> 00:22:10.230
is just on control and structure because even
00:22:10.230 --> 00:22:14.250
if you could get 20% better performance tomorrow
00:22:14.250 --> 00:22:17.930
by simply doing like a full fully agentic, fully
00:22:17.930 --> 00:22:20.690
recursive self-improving loop. The issue with
00:22:20.690 --> 00:22:22.970
those systems is you never know when they break
00:22:22.970 --> 00:22:25.170
and you never know how they break until they
00:22:25.170 --> 00:22:27.569
break. And with structure, what you're able to
00:22:27.569 --> 00:22:29.230
do is even when it's failing, you know where
00:22:29.230 --> 00:22:34.519
it's failing so you can isolate things. So I
00:22:35.230 --> 00:22:38.730
think definitely that is the area we're trying
00:22:38.730 --> 00:22:43.809
to most resolve is how can we ensure that every
00:22:43.809 --> 00:22:46.930
aspect of our system, when a user comes to us
00:22:46.930 --> 00:22:50.640
and says, why did it fail we know we can give
00:22:50.640 --> 00:22:53.619
them a clear answer on why and how we on what
00:22:53.619 --> 00:22:55.859
we're going to do about it because right now
00:22:55.859 --> 00:22:58.960
it's sometimes I can tell you the exact process
00:22:58.960 --> 00:23:02.079
on this it took this agentic path to come up
00:23:02.079 --> 00:23:04.279
to your conclusion but I can't tell you why it
00:23:04.279 --> 00:23:06.960
chose to take that path versus another path where
00:23:06.960 --> 00:23:09.779
in the uncertainty calculations it decided this
00:23:09.779 --> 00:23:12.799
and that's just like a general problem with AI
00:23:12.799 --> 00:23:15.440
interpretability so that's something we want
00:23:15.440 --> 00:23:18.990
to actively solve for So on the observability
00:23:18.990 --> 00:23:21.369
side, is there something specific that you do
00:23:21.369 -->
00:23:24.210
log and measure so you know when the system is
00:23:24.210 --> 00:23:25.344
getting worse or not?
00:23:25.843 --> 00:23:31.158
So it's an open discussion, obviously.
00:23:31.757 --> 00:23:35.220
Because, again, the user contracts themselves have been changing quite a bit
00:23:35.432 --> 00:23:40.924
on terms of what users are expecting our systems to do and not do.
00:23:41.771 --> 00:23:46.334
When we first started off, was just purely reason over these existing documents.
00:23:46.933 --> 00:23:49.029
Now it's gone from that to being
00:23:49.029 --> 00:23:51.210
like, let's do second or third order consequences,
00:23:51.410 --> 00:23:56.349
integrate real-time legal research. As these
00:23:56.609 --> 00:23:58.529
systems are getting better, we're using more
00:23:58.529 --> 00:24:01.509
of them. We're trying to figure out how the system
00:24:01.509 --> 00:24:06.549
is approaching the correct answers. I think number
00:24:06.549 --> 00:24:09.089
one right now, what we're tracking for, whether
00:24:09.089 --> 00:24:11.150
it works or not, is does it behave as expected?
00:24:11.490 --> 00:24:14.569
Can we predict ahead of time what kind of...
00:24:15.789 --> 00:24:17.670
paths this would have followed because at least
00:24:17.670 --> 00:24:21.869
then we know that our system is stable and that's
00:24:21.869 --> 00:24:23.849
all that itself is a much bigger challenge that
00:24:23.849 --> 00:24:28.019
people realize because LLM update people don't
00:24:28.019 --> 00:24:30.000
know this but a lot labs are always changing
00:24:30.000 --> 00:24:32.839
their other LLMs behind the scenes and you're
00:24:32.839 --> 00:24:35.700
not it's not like those major releases like
00:24:35.700 --> 00:24:39.680
2.5 became 3. The behavior of 2.5 just changed
00:24:39.680 --> 00:24:41.279
because they were post-training something and
00:24:41.279 --> 00:24:44.319
they pushed that update or they quantized a model.
00:24:44.500 --> 00:24:47.380
You hear this a lot. They're changing something
00:24:47.380 --> 00:24:50.579
about the tokenizer. These come out of nowhere
00:24:50.579 --> 00:24:53.440
and that can completely change the behavior of
00:24:53.440 --> 00:24:57.839
your expected output. So on that lens, it becomes
00:24:57.839 --> 00:25:01.160
very difficult for us to control and even figure
00:25:01.160 --> 00:25:04.000
that out. But this is definitely the number one
00:25:04.000 --> 00:25:08.799
priority for us. What's the most overhyped technique
00:25:08.799 --> 00:25:12.279
in AI product land right now? Reinforcement learning,
00:25:12.500 --> 00:25:18.700
by far. Okay. It just doesn't work. Okay. RL
00:25:18.700 --> 00:25:24.079
was great and is best when you have a verifiable
00:25:24.079 --> 00:25:26.980
output. A truly verifiable output, not what we
00:25:26.980 --> 00:25:30.880
call RLVR. RLVR is mostly garbage. What reinforcement
00:25:30.880 --> 00:25:34.779
learning does is it, once you know this is exactly
00:25:34.779
--> 00:25:37.660
what I'm looking for, these are hard constraints
00:25:37.660 --> 00:25:41.970
for me RL is extremely good at figuring out the
00:25:41.970 --> 00:25:45.319
best way to get there very very quickly the problem
00:25:45.319 --> 00:25:48.299
with it is people are trying to use RL for open
00:25:48.299 --> 00:25:51.880
exploration problems make an LLM reason better
00:25:51.880 --> 00:25:54.579
make an LLM write better what does better mean
00:25:54.579 --> 00:25:57.880
better has no signals so we end up using proxies
00:25:57.880 --> 00:26:00.160
like okay if this code compiles in a certain
00:26:00.160 --> 00:26:03.410
amount of time it's better. If this sentence
00:26:03.410 --> 00:26:06.430
uses fewer emojis than this other sentence, it's
00:26:06.430 --> 00:26:08.769
better or worse. But you see how immediately
00:26:08.769 --> 00:26:12.569
that creates all kinds of issues. So people are
00:26:12.569 --> 00:26:14.869
investing heavily into RL because we don't know
00:26:14.869 --> 00:26:18.049
how else to do it. Or we have ideas, but there
00:26:18.049 --> 00:26:20.829
are very weird incentives for short-termism
00:26:20.829 --> 00:26:24.410
in AI research. And a lot of people who don't
00:26:24.410 --> 00:26:27.349
understand the details of reinforcement learning,
00:26:27.529 --> 00:26:29.750
I can break down the math of why reinforcement
00:26:29.750 --> 00:26:33.130
learning doesn't work. But people that haven't
00:26:33.130 --> 00:26:35.650
studied it as much or don't understand why big
00:26:35.650 --> 00:26:37.990
labs are investing in it are then taking the
00:26:37.990 --> 00:26:39.890
same approaches being done by big labs and trying
00:26:39.890 --> 00:26:42.809
to replicate it for startups. And you're hearing
00:26:42.809 --> 00:26:46.029
RL startups, RL environment startups. They're
00:26:46.029 --> 00:26:48.559
all going to sink. None of them are that good.
00:26:48.680 --> 00:26:51.460
And anybody that's here pitching reinforcement
00:26:51.460 --> 00:26:54.700
learning to you is either an idiot or a scammer.
00:26:56.339 --> 00:26:59.599
Interesting. Okay, so any other takeaways for
00:26:59.599 --> 00:27:03.019
our listeners around AI or prompt engineering
00:27:03.019 --> 00:27:07.640
or building AI systems? If I were to summarize
00:27:07.640 --> 00:27:11.440
everything in one go, all of the learnings we've
00:27:11.440 --> 00:27:14.380
had, both from the building Irys ourselves and
00:27:14.380 --> 00:27:16.500
also the open source community that we're engaged
00:27:16.500 --> 00:27:20.880
in and building it out of, I think the number
00:27:20.880 --> 00:27:25.670
one area AI is failing right now is that we don't
00:27:25.670 --> 00:27:30.160
have good measurements AI in general is failing
00:27:30.160 --> 00:27:33.680
with measurements and measurements are like where
00:27:33.680 --> 00:27:35.579
we're struggling is we don't understand even
00:27:35.579 --> 00:27:39.859
where how to extract meaningful signal
and too
00:27:39.859 --> 00:27:43.480
many like people trying to build AI are trying
00:27:43.480 --> 00:27:47.200
to build for high performance and high performance
00:27:47.200 --> 00:27:49.900
is what gets it it gets you the cool demos it
00:27:49.900 --> 00:27:52.480
gets you the good systems but often what that
00:27:52.480 --> 00:27:55.259
leads to is overtly complicated systems which
00:27:55.259 --> 00:27:57.119
you have to re-engineer from scratch because
00:27:57.119 --> 00:28:00.839
as people start finding everywhere it goes wrong
00:28:00.839 --> 00:28:04.279
you have to come back and question your assumptions
00:28:04.279 --> 00:28:06.079
and you realize hey these systems don't work
00:28:06.079 --> 00:28:09.880
or the system has certain issues so I think people
00:28:09.880 --> 00:28:13.019
should be engineering for um error isolation
00:28:13.019 --> 00:28:16.609
first and foremost be able to identify especially
00:28:16.609 --> 00:28:18.630
if they're using something like Claude Code or
00:28:18.630 --> 00:28:22.710
Codex or anything else to do their software engineering
00:28:22.710 --> 00:28:27.450
because these systems will just churn out heaps
00:28:27.450 --> 00:28:29.829
and heaps of code and then it's very hard to
00:28:29.829 --> 00:28:33.210
start figuring out where which systems work which
00:28:33.210 --> 00:28:35.849
system to work so optimize for maintenance not
00:28:35.849 --> 00:28:39.430
for generation because maintenance is where your
00:28:39.430 --> 00:28:42.589
system either lives or dies generation is like
00:28:42.589 --> 00:28:46.099
the It's what looks easy and it's easy to measure,
00:28:46.240 --> 00:28:50.450
but that's not really where your performance
00:28:50.450 --> 00:28:54.700
data can come from. Yeah, that makes sense. Where
00:28:54.700 --> 00:28:58.859
can people find your work, reach out to you,
00:28:58.920 --> 00:29:01.019
see what you're working on, see what Irys is
00:29:01.019 --> 00:29:05.859
working on? I think irys.ai, I-R-Y-S.A-I,
00:29:06.000 --> 00:29:10.539
easy. That link will contain a lot of research.
00:29:11.160 --> 00:29:13.440
If you're interested in the open source AI research
00:29:13.440 --> 00:29:17.200
work we do, the chocolatemilkcult.ai also,
00:29:17.339 --> 00:29:20.339
that's where we publish a lot of stuff, publish
00:29:20.339 --> 00:29:22.539
on Substack. If they just Google Devansh AI,
00:29:22.819 --> 00:29:25.880
they'll probably find out all my links, etc.
00:29:26.799 --> 00:29:29.539
But irys.ai if you're looking for one place
00:29:29.539 --> 00:29:32.279
to start. Cool. Any closing thoughts you want
00:29:32.279 --> 00:29:34.819
to leave our audience with? I think the one closing
00:29:34.819 --> 00:29:39.319
thought that I keep trying to emphasize to everybody
00:29:39.319 --> 00:29:42.329
is none of us really know where AI is headed
00:29:42.329 --> 00:29:45.690
and people can confidently say one thing or another
00:29:45.690 --> 00:29:50.150
but truly
there are no true experts so I think
00:29:50.150 --> 00:29:52.670
too many people right now are kind of being passive
00:29:52.670 --> 00:29:55.589
with AI which is like oh let me wait for somebody
00:29:55.589 --> 00:29:59.410
to tell me how to best build the system best
00:29:59.410 --> 00:30:02.690
do this or let's do that I think right now it's
00:30:02.690 --> 00:30:04.769
better to take a more proactive approach and
00:30:04.769 --> 00:30:08.440
get in the mud test things out yourself build
00:30:08.440 --> 00:30:12.220
things out yourself because you never know where
00:30:12.220 --> 00:30:14.220
the next big signal comes from you never know
00:30:14.220 --> 00:30:19.640
where it come comes out of and ultimately I can
00:30:19.640 --> 00:30:21.420
sit here and tell you some things but those are
00:30:21.420 --> 00:30:25.460
very biased by my own experiences and how I've
00:30:25.460 --> 00:30:27.500
built things it might not fully apply to you
00:30:27.500 --> 00:30:30.140
to build what you know everybody's talking about
00:30:30.140 --> 00:30:32.480
taste and judgment these days the only way to
00:30:32.480 --> 00:30:35.119
build taste and judgment is for you to fail repeatedly
00:30:35.119 --> 00:30:38.609
and then have that meta process analysis of where
00:30:38.609 --> 00:30:41.950
systems fail and where they succeed. So that
00:30:41.950 --> 00:30:45.230
would just be a generalist, I think, mentality
00:30:45.230 --> 00:30:47.910
to take about all of this. Oh, it makes sense.
00:30:48.670 --> 00:30:51.490
Really appreciate it, Devansh. I will leave links
00:30:51.490 --> 00:30:53.670
for all of the websites that you mentioned down
00:30:53.670 --> 00:30:55.269
in the show notes. And thank you so much for
00:30:55.269 --> 00:30:57.430
coming on. Really appreciate it. Thank you, Brian.
00:30:57.650 --> 00:31:01.250
That was my conversation with Devansh from Irys.
00:31:01.589 --> 00:31:04.769
The biggest takeaway for me is that you can't
00:31:04.769 --> 00:31:08.420
fix what you can't find Devansh said a fully
00:31:08.420 --> 00:31:13.000
agentic system might perform 20% better but you
00:31:13.000 --> 00:31:16.920
never know how it breaks until it breaks the
00:31:16.920 --> 00:31:20.539
way I see it his three tiers are labels on the
00:31:20.539 --> 00:31:23.880
breaker panel if a lawyer says something's missing
00:31:23.880 --> 00:31:28.420
his team knows it's ingestion or reasoning and
00:31:28.420 --> 00:31:30.880
a couple of targeted queries tell them which
00:31:30.880 --> 00:31:35.309
second the dangerous failure isn't always the
00:31:35.309 --> 00:31:39.990
loud one. A made-up case falls apart once somebody
00:31:39.990 --> 00:31:43.529
looks it up. A real case applied the wrong way
00:31:43.529 --> 00:31:47.369
looks fine. It's the same thing in infrastructure.
00:31:47.829 --> 00:31:51.690
A config that's valid but wrong hurts more than
00:31:51.690 --> 00:31:55.650
one that won't parse. Third, the model can change
00:31:55.650 --> 00:32:00.069
under you. Devansh said labs tweak models behind
00:32:00.069 --> 00:32:04.640
the scenes. And your output shifts with no new
00:32:04.640 --> 00:32:08.440
version number so the first thing his team tracks
00:32:08.440 --> 00:32:12.980
is whether the system behaves as expected to
00:32:12.980 --> 00:32:16.700
me that's drift detection and we know how to
00:32:16.700 --> 00:32:20.059
watch for drift and the line that I taped to
00:32:20.059 --> 00:32:23.880
my monitor optimize for maintenance not generation
00:32:23.880 --> 00:32:28.700
coding agents will churn out heaps of code but
00:32:28.700 --> 00:32:33.170
maintenance is where a system lives or dies You
00:32:33.170 --> 00:32:37.769
can find Devansh at irys.ai plus his open source
00:32:37.769 --> 00:32:42.069
research on Substack. Or just search Devansh
00:32:42.069 --> 00:32:45.509
AI through the links in the show notes. Follow
00:32:45.509 --> 00:32:48.630
or subscribe to Ship It Weekly wherever you listen
00:32:48.630 --> 00:32:52.829
and find previous episodes at shipitweekly.fm.
00:32:53.289 --> 00:32:56.230
Thanks for listening and I'll see you later this
00:32:56.230 --> 00:32:56.490
week.