文字起こし 🔗
1
00:00:05,401 --> 00:00:08,930
Jennifer Reif: You are listening to the
Breaktime Tech Talks podcast, a bite-sized
2
00:00:08,951 --> 00:00:13,495
tech podcast for busy developers where
we'll briefly cover technical topics, news
3
00:00:13,495 --> 00:00:15,779
snippets, and more in short time blocks.
4
00:00:16,236 --> 00:00:19,943
I'm your host, Jennifer Reif, an
avid developer and problem solver
5
00:00:20,152 --> 00:00:24,116
with special interest in data,
learning, and all things technology.
6
00:00:25,039 --> 00:00:28,830
The last few weeks, I've had a little
bit more of an intense travel season
7
00:00:29,123 --> 00:00:32,787
than usual, and lots of presentation
prep that has been going on.
8
00:00:33,134 --> 00:00:37,709
I've had small pockets of time to catch
up on a few things now that I have a
9
00:00:37,709 --> 00:00:39,812
stretch of time where I'm remote again.
10
00:00:40,322 --> 00:00:44,773
Neo4j's virtual developer conference,
or NODES, was last week, so it will
11
00:00:44,774 --> 00:00:48,294
take some time for me to reset and pick
up where I left off on a few things.
12
00:00:48,669 --> 00:00:51,025
But I will fill you in on
what's going on right now.
13
00:00:51,835 --> 00:00:56,309
First off, just to wrap up
NODES 2026, that is Neo4j's
14
00:00:56,321 --> 00:00:59,910
online developer conference,
entirely free, entirely virtual.
15
00:01:00,164 --> 00:01:06,928
We opened with a keynote by Emil Eifrem,
Neo4j CEO and founder, and then Andrew Ng.
16
00:01:07,252 --> 00:01:10,267
There were a couple of takeaways that
I really enjoyed about this keynote.
17
00:01:10,313 --> 00:01:12,760
First of all, I do highly recommend
listening to this keynote.
18
00:01:12,793 --> 00:01:17,421
I found several tidbits that were just
really good to hear and reassuring
19
00:01:17,422 --> 00:01:18,762
and inspiring and interesting.
20
00:01:19,341 --> 00:01:23,399
The first is that computer science
fundamentals are still vital for
21
00:01:23,434 --> 00:01:26,583
understanding and guiding the
systems that we are building and
22
00:01:26,584 --> 00:01:28,093
the problems we are trying to solve.
23
00:01:28,365 --> 00:01:31,347
Because after all, the problems
we are solving are human problems.
24
00:01:31,719 --> 00:01:36,102
So having that human component
of that is going to help us build
25
00:01:36,157 --> 00:01:37,814
better solutions in the long term.
26
00:01:38,335 --> 00:01:41,427
The other little tidbit I picked
up from this is that skilled
27
00:01:41,442 --> 00:01:46,128
information or skilled judgments
are still held by humans.
28
00:01:46,181 --> 00:01:49,993
Andrew says it will be still held
by humans in the future as well.
29
00:01:50,136 --> 00:01:53,691
This is not something that's going away,
at least not for the foreseeable future.
30
00:01:53,984 --> 00:01:56,566
I really loved the
approach to this session.
31
00:01:56,567 --> 00:01:59,604
It was more of a fireside chat,
little bit back and forth.
32
00:02:00,303 --> 00:02:03,776
We got perspectives, and it was just
really interesting and inspiring to hear.
33
00:02:04,155 --> 00:02:08,562
There were a couple of sessions I wanted
to call out and a few little perspectives
34
00:02:08,563 --> 00:02:10,102
I picked up throughout the event.
35
00:02:10,464 --> 00:02:14,266
The first session I want to call out
is Evals as a Graph: Tracing Prompt
36
00:02:14,320 --> 00:02:16,788
to Judge to Outcome Across Releases.
37
00:02:17,089 --> 00:02:20,871
I really loved Rashi's storytelling
and how the team has lessons learned
38
00:02:20,882 --> 00:02:22,595
from their own healthcare solution.
39
00:02:22,877 --> 00:02:26,322
I found this session really
intriguing and found the speaker's
40
00:02:26,555 --> 00:02:29,045
storytelling perspective really
interesting and fun to listen to.
41
00:02:29,738 --> 00:02:34,287
The next session was Optimizing
Context Graph Agents with Deterministic
42
00:02:34,297 --> 00:02:36,091
Queries and Fast Decision Models.
43
00:02:36,367 --> 00:02:39,833
Now, this one, there was a really
interesting tidbit I picked up, and
44
00:02:39,833 --> 00:02:44,736
that is that using JEV or another
type of judgment model for small
45
00:02:44,738 --> 00:02:50,110
criteria is a really great use case
for them, rather than asking them
46
00:02:50,110 --> 00:02:52,085
to make large, loaded judgments.
47
00:02:52,240 --> 00:02:57,252
Picking out very small, fine-tuned,
specific criteria and letting a decision
48
00:02:57,253 --> 00:03:01,487
model handle those things was a much more
valuable use of their time and expertise
49
00:03:01,795 --> 00:03:05,780
for the judgment models than for your
more general large language models.
50
00:03:06,613 --> 00:03:11,579
The last session It is From
Error to Fix: Building a GitHub
51
00:03:11,602 --> 00:03:13,189
Troubleshooting Graph with Neo4j.
52
00:03:13,491 --> 00:03:16,910
Now, this one, the piece that I picked
up here is they were talking a lot
53
00:03:16,910 --> 00:03:21,086
about using AI in a pipeline, and I
actually found this to be a common
54
00:03:21,086 --> 00:03:22,358
theme throughout NODES this year.
55
00:03:22,358 --> 00:03:27,325
There were several different projects
and use cases and developers that were
56
00:03:27,326 --> 00:03:34,432
talking about using graphs and using
AI in pipelines, in regular everyday
57
00:03:34,433 --> 00:03:39,876
business pipelines to pick up PRs,
to surface issues in their system, to
58
00:03:39,906 --> 00:03:42,405
bring in feedback to whatever it is.
59
00:03:42,804 --> 00:03:44,676
They were using a pipeline of some sort.
60
00:03:44,678 --> 00:03:51,059
So they're operating with AI as a
component in this much larger workflow.
61
00:03:51,166 --> 00:03:53,814
And I found this to be a much more
common theme than I've seen in the past
62
00:03:54,243 --> 00:03:55,625
throughout several of the sessions.
63
00:03:56,545 --> 00:04:00,915
Coming back from a bunch of speaking,
talking to a lot of developers in
64
Load full transcript
00:04:00,915 --> 00:04:06,066
real life, and then on the backside
of Nodes now, I'm spinning up several
65
00:04:06,066 --> 00:04:07,856
different personal exploration projects.
66
00:04:07,868 --> 00:04:11,298
I was toying with a few different things
and starting to pull on some threads.
67
00:04:11,298 --> 00:04:13,438
Well, I'm starting to dig a
little bit deeper in those.
68
00:04:13,759 --> 00:04:16,779
I have a bunch of new ideas
for exploration and potential
69
00:04:16,783 --> 00:04:20,483
session abstracts coming up in the
near future and into next year.
70
00:04:21,224 --> 00:04:25,675
The takeaway from many of my presentations
this year is that I focused on agents
71
00:04:25,677 --> 00:04:30,222
and MCP and other AI topics, but I ended
up always getting questions about the
72
00:04:30,223 --> 00:04:34,008
graph and the backend, because I would
always use a Neo4j graph somewhere
73
00:04:34,008 --> 00:04:38,228
in my stack, and I ended up getting
a lot of questions around the graph.
74
00:04:38,249 --> 00:04:39,100
What is this thing?
75
00:04:39,105 --> 00:04:39,743
How is it used?
76
00:04:39,744 --> 00:04:40,597
What does it look like?
77
00:04:40,751 --> 00:04:42,037
What value does it provide?
78
00:04:42,751 --> 00:04:45,625
And so I would spend some time going
into those things during my sessions,
79
00:04:45,912 --> 00:04:51,498
but I really didn't, realize that people
would be so intrigued with the backend.
80
00:04:51,781 --> 00:04:56,068
So I'm reworking some things and
building some new stuff that's
81
00:04:56,068 --> 00:05:01,105
focused around more of these
concepts of graphs and graphs in AI.
82
00:05:01,921 --> 00:05:05,569
I'm addressing some of these things with
a new project and an accompanying session
83
00:05:05,571 --> 00:05:09,335
that I hope to have up and running here
pretty soon, and it's also just fun to
84
00:05:09,335 --> 00:05:11,521
play with some basic graph things again.
85
00:05:11,664 --> 00:05:13,208
So I'm enjoying that as well.
86
00:05:13,641 --> 00:05:17,004
I also mentioned recently that I was
working on a benchmarking project.
87
00:05:17,248 --> 00:05:17,929
That went well.
88
00:05:17,942 --> 00:05:21,123
I did a virtual webinar
a few weeks ago on that.
89
00:05:22,023 --> 00:05:24,983
But have definitely just barely
scratched the surface on this.
90
00:05:25,003 --> 00:05:29,006
I just took an overview of the
initial beginner steps of this.
91
00:05:29,449 --> 00:05:33,383
Now, it has opened up some new threads
that I can start to pull on, and I am
92
00:05:33,384 --> 00:05:37,049
hoping to get to those when the schedule
gets a bit quieter here in a couple of
93
00:05:37,049 --> 00:05:38,927
weeks when we get closer to the holidays.
94
00:05:39,264 --> 00:05:44,211
Some things that I specifically want to
look at is how do we benchmark over time?
95
00:05:44,214 --> 00:05:47,897
'Cause obviously, I'm starting to build
that repertoire with this application,
96
00:05:47,899 --> 00:05:51,415
but I haven't really tested it over a long
period of time to see how it holds up.
97
00:05:51,821 --> 00:05:55,217
I'd also like to play around with
a few of the benchmarking tools.
98
00:05:55,365 --> 00:05:57,654
For instance, I've had several
questions about, how do I
99
00:05:57,660 --> 00:06:00,254
look at my AI system metrics?
100
00:06:00,507 --> 00:06:03,559
And there's tons of observability
tools that support this kind of
101
00:06:03,560 --> 00:06:06,012
thing now, but I actually haven't
played with any one of them.
102
00:06:06,013 --> 00:06:08,906
I've seen lots of product names and
vendor names and things that go along with
103
00:06:08,906 --> 00:06:12,289
that, so I would like to get a chance to
play around with some of that as well.
104
00:06:12,918 --> 00:06:17,724
Another tactic is in the
area of multi-agent projects.
105
00:06:18,018 --> 00:06:21,807
I have played around with a
multi-agent application here recently.
106
00:06:21,835 --> 00:06:25,912
I did a presentation on that, but I do
feel like it needs a little bit of work.
107
00:06:25,953 --> 00:06:28,968
I felt like there were a bunch of open
threads that I really want to pull on and
108
00:06:28,969 --> 00:06:30,625
narrow down the topic material a bit more.
109
00:06:30,625 --> 00:06:33,896
So I just felt like the things that I
tried to cover in that session, it was a
110
00:06:33,900 --> 00:06:39,885
little too broad, because there actually
were surprisingly a lot more varying
111
00:06:39,899 --> 00:06:43,158
routes that I could take, multiple
pathways that I didn't expect would be
112
00:06:43,159 --> 00:06:45,322
opened up when I broached this topic.
113
00:06:45,675 --> 00:06:48,327
So now I have a ton of different
things that I could potentially
114
00:06:48,327 --> 00:06:50,440
do in this avenue of multi-agents.
115
00:06:50,766 --> 00:06:52,163
My session just felt a little too broad.
116
00:06:52,164 --> 00:06:55,707
So I wanna take some of those pathways
a little bit, follow down those routes
117
00:06:55,707 --> 00:06:59,829
a little bit more, and see if I can
narrow my focus just a little bit and
118
00:06:59,830 --> 00:07:01,539
really surface, "Hey, this was good.
119
00:07:01,540 --> 00:07:05,831
This was not so good," or, "I found
this challenge or this success," or
120
00:07:05,831 --> 00:07:07,664
whatever it is in these various areas.
121
00:07:08,490 --> 00:07:10,419
I do have a couple of
upcoming events as well.
122
00:07:10,423 --> 00:07:11,931
I'm not completely done for the year.
123
00:07:12,179 --> 00:07:16,614
I have AI Dev New York coming here
at the very end of November into
124
00:07:16,614 --> 00:07:19,322
the beginning of December, so I'll
be there for a couple of days.
125
00:07:20,163 --> 00:07:22,638
If you're around, definitely,
come flag me down.
126
00:07:22,648 --> 00:07:23,530
I'm happy to chat.
127
00:07:23,796 --> 00:07:26,835
And then I have a virtual live
stream coming on December 3rd.
128
00:07:26,888 --> 00:07:29,726
I
might have a couple other things in
the works as well, but I just don't
129
00:07:29,738 --> 00:07:31,882
have details quite yet hashed out there.
130
00:07:32,486 --> 00:07:33,502
I will keep you posted, though.
131
00:07:34,386 --> 00:07:37,792
For a couple of the content
highlights this week, there's two.
132
00:07:37,823 --> 00:07:41,541
They are both Neo4j, but, this was
not something that I was actually
133
00:07:41,543 --> 00:07:43,555
aware of until right before NODES.
134
00:07:43,921 --> 00:07:47,952
So there's a new blog post
called Meet Relate by Neo4j, an
135
00:07:47,952 --> 00:07:49,684
embedded database for AI agents.
136
00:07:49,735 --> 00:07:54,265
This was written by Stephen Chen, Michael
Hunger, and Kate Smith, all Neo4j folks.
137
00:07:54,575 --> 00:07:58,271
So there is a new Neo4j embedded database.
138
00:07:58,355 --> 00:07:59,784
This is actually not the first.
139
00:07:59,797 --> 00:08:03,441
They used to have an embedded
database way back when.
140
00:08:03,514 --> 00:08:05,568
Actually, initially when
I joined, it was around.
141
00:08:05,814 --> 00:08:10,436
But this embedded database is designed
specifically for use by agents.
142
00:08:10,782 --> 00:08:13,780
It's alfo-also really helpful for
just getting started into graphs.
143
00:08:13,813 --> 00:08:16,975
I'm playing around with a couple of
projects there, so I hope to have
144
00:08:16,976 --> 00:08:18,188
some more updates on that soon.
145
00:08:18,209 --> 00:08:19,433
But definitely check it out.
146
00:08:20,184 --> 00:08:25,074
If you're playing around with agents and
you need some sort of memory layer or
147
00:08:25,076 --> 00:08:29,988
shared context or something, then this
might be a great place for you to start.
148
00:08:30,013 --> 00:08:35,285
It's a graph, again, primarily
for use by agents and hopefully
149
00:08:35,337 --> 00:08:36,443
us developers as well.
150
00:08:37,186 --> 00:08:40,637
Now, there's one more blog post that
actually follows up on this, and
151
00:08:40,637 --> 00:08:45,293
it's called Announcing the Relate
by Neo4j Hackathon 2026: Build AI
152
00:08:45,322 --> 00:08:47,377
Agents on an Embedded Graph Database.
153
00:08:47,665 --> 00:08:51,642
This is a subsequent blog post because
Neo4j opened up the same time that
154
00:08:51,643 --> 00:08:53,872
they released RelateDB for Neo4j.
155
00:08:53,874 --> 00:08:55,661
Oh, there was one thing I forgot to add.
156
00:08:55,883 --> 00:09:00,185
RelateDB is open source so that
means you can use it at will.
157
00:09:00,186 --> 00:09:03,110
There's no licensing and
issues going around with that.
158
00:09:03,444 --> 00:09:05,892
I think we're still waiting for
a few of the little pieces to get
159
00:09:05,893 --> 00:09:09,951
locked into place, but it should
make it really easy and nice to use.
160
00:09:10,641 --> 00:09:11,999
Now, going back to the hackathon.
161
00:09:12,313 --> 00:09:17,381
We have a hackathon that is on the
tail of this release where you can
162
00:09:17,383 --> 00:09:23,423
submit your projects using RelateDB
by Neo4j and, showcase your projects,
163
00:09:23,430 --> 00:09:24,674
the fun things that you do with it.
164
00:09:24,693 --> 00:09:28,710
I know I've seen tons of buzz
and lots of success with people
165
00:09:28,711 --> 00:09:29,654
submitting lots of things.
166
00:09:29,883 --> 00:09:32,263
If you wanna check that out, I
would highly recommend you publish
167
00:09:32,263 --> 00:09:33,811
your own projects with Relate.
168
00:09:34,057 --> 00:09:38,020
The hackathon ends on December
11th, and the top three winners
169
00:09:38,050 --> 00:09:40,439
will win a Mac Studio or Mac Minis.
170
00:09:40,670 --> 00:09:41,980
Really nice specs on these.
171
00:09:42,253 --> 00:09:43,543
Now, the fun doesn't stop.
172
00:09:43,802 --> 00:09:46,976
Things are buzzing, and I'm still
finding my rhythm after hopping
173
00:09:46,978 --> 00:09:48,265
around the country for a few weeks.
174
00:09:48,656 --> 00:09:51,973
As I start digging into projects
and code again, I'm looking forward
175
00:09:51,974 --> 00:09:53,801
to seeing how each thing develops.
176
00:09:54,173 --> 00:09:56,750
As always, thanks for
listening and happy coding.