C (2:14)

Yeah, I mean, this is kind of.

B (2:15)

A crazy one where Perplexity offered allegedly.

C (2:20)

$34.5 billion to buy Google's Chrome browser, which is almost double.

B (2:24)

Perplexity's own valuation and way above what probably most people think Chrome is worth.

C (2:29)

And then while this is going on.

B (2:31)

There'S a federal judge is weighing in antitrust remedies after ruling Google illegally monopolized.

C (2:37)

Search perplexity seems to be signaling, like.

B (2:40)

Hey, if you force a sale, we're ready to buy it. And of course, I doubt Google's not interested in this. And they've said that selling Chrome would obviously hurt their business and even raise security risks.

C (2:52)

From their perspective, analysts don't think the.

B (2:55)

Sale is likely, but the bid itself shows how, I think, aggressive some of these AI companies have been, and they're very interested in getting a full hold in search.

C (3:03)

But to me, it feels like not.

B (3:05)

A real acquisition attempt and more like a PR move to stay in sort of this antitrust conversation. There's a lot going on. Let's kind of attach ourselves to that. We can make some headlines.

A (3:14)

Yeah, because I mean, the bid that they made 34.5 billion is almost double their own valuation. So unless they're going to go get a bunch of debt on that or something.

B (3:23)

Yeah. Can you get like a mortgage loan from a bank to buy Google Chrome?

A (3:27)

Yeah, exactly. I'll have to watch that meeting. So, yeah, and as you say, Google said it's not interested, but that's the point is to get it kind of into the news around antitrust. I mean, there's been a lot of noise around Google needing to break up. I think it's funny, we still talk about Google as Google even though the parent company is Alphabet. Everybody just seems to have forgotten that. And it's still Google and all the Google products.

B (3:50)

Yeah, I mean, it's kind of like Facebook with Meta and stuff like that.

C (3:53)

I mean, a lot of these conversations.

B (3:54)

Remind me of the early 2000s with all the antitrust stuff around Microsoft as well. And there's been sort of rumors of this coming for Alphabet for quite some time.

A (4:03)

Yeah. But also perplexity, I guess. Maybe another PR move perhaps. But they're talking about a slightly different rev share model with publishers.

C (4:14)

Yeah. So there they want to essentially introduce.

B (4:18)

This new version of their revenue share, and they plan to distribute the money when an AI assistant or search engine uses news articles to fulfill a task or answer a search question.

C (4:29)

So, I mean, it makes a lot of sense. Like if you are switching from offering blue links as search results to something.

B (4:36)

Where you're giving essentially a fully baked answer, can you monetize the answer in some way where you give credit back to where the source material came from.

C (4:45)

And I think this is something we touched on last time of how might.

B (4:47)

That impact the types of answers that you get?

C (4:50)

If you can essentially pay for the.

B (4:52)

Answer or sometimes be part of the answer because you're paying for it, are you necessarily surfacing the best answer and doing right by the users?

C (5:00)

Now, you can make the same argument for conventional search, but at least with conventional search search it's clear, well, somewhat.

B (5:06)

Clear which are the ads and which are the actual organic results.

C (5:10)

So I think it's something interesting to pay attention to. I am a little bit wary of.

B (5:15)

Ultimately how that could impact the value of what some of these models are able to provide users.

A (5:21)

Yeah, and I mean, I think to your point around sort of paying to play, it's going to be interesting. This is cured at publishers. So which publishers are they picking? Because they kind of let everyone in the door. This is just going to be clickbait article heaven, you know, where, oh, we're designing this article to be featured in an answer. And you know, interestingly, the model is moving to Comet plus, which is their browser offering. They're offering 80% of that revenue to the publishers, which is very interesting. I mean, that's a polite way of putting it, where they're going to like take literally just the revenue off a product and say, if you're featured, we're going to distribute. I mean, there must be some kind of ratios depending on how you come up in results. But yeah, it's a very interesting way to peg it. And 80% sounds like a lot to me, but I mean, that will come down probably.

B (6:15)

But yeah, yeah, definitely. It does seem like a lot.

C (6:17)

And I think it's one of those things where whenever there's kind of like.

B (6:20)

A new form factor with which to get your message out, like a new.

C (6:24)

Sort of marketing channel, what inevitably happens is the early adopters of that channel will be a lot of times very.

B (6:33)

Successful because they're kind of early to it. It's like if you're early to Facebook or something like that, or social media.

C (6:37)

Then it's like, oh, wow, it's this.

B (6:39)

Untapped channel where everybody else is stuck sending emails. Emails is really noisy. Let's exploit this new channel.

C (6:45)

The problem is once people show success.

B (6:48)

There, it's like everyone just drives in there and eventually the channel becomes this noisy channel that gets devalued and people pay less attention to.

C (6:56)

So I kind of think that we may be leaving the heyday of the.

B (7:01)

Chat LLM experience if this becomes too popular.

A (7:04)

Yeah, definitely. So, moving on Next one is intel, which has just sold 10% of itself to the US government.

C (7:12)

Yeah, I mean, I think this is.

B (7:13)

Pretty big news in the world of semiconductors. Essentially that 10% is worth $9 billion. It's one of the biggest government interventions in a private company since I think the 2008 auto bailout. The deal came directly out of negotiations between Trump and Intel CEO Lip Bu Tan and is framed as this way of strengthening America's chip making leadership. And recently I read the government AI Action plan for the US government and part of the strategy outlined in that document was making sure that the US stays ahead of the rest of the world when it comes to CPU GPU innovation.

C (7:47)

So this seems to be somewhat in line with that.

B (7:49)

And we saw shares of intel jump based on this news. But I think it raises some big questions of like, can the government ownership really turn intel around when it comes to struggling to keep up with rivals like Nvidia, or does this make intel too big to fail in some fashion, like with taxpayers kind of now on the hook?

A (8:08)

Yeah, it is very interesting. As you say, it's like the biggest government intervention in the US company. I mean, this is a public company, so there's public shareholders as well who are like looking at this and going, well now one of our shareholders is the U.S. government. That's quite different to most companies does kind of have echoes of China, for example, where virtually all of their companies have state ownership in some respect. This just feels. Yeah. Could this be like tip of the iceberg where the government is starting to step in on certain. I mean, intel was struggling so it won't just sort of start jumping in on things. But yeah, I think too big to fail is a good way to think about it. Just given that these people that can fab chips effectively, it's very hard to do that. Only a couple of companies, two or maybe three, can do it properly. Obviously TSMC is the big one, Taiwanese. So they're kind of in a weird gray zone between China and the us. So yeah, this just feels. I mean, I can't say I support this exactly, but I can totally understand where it's coming from. And at the end of the day, Intel's had a really rocky ride recently. They're really struggling on the CEO front. And I've still got that book High Output Management by Andy Grove. And he was the one who took intel to its highs and then he died quite a while ago and they just don't seem to have hit their stride at all after that. I had many MacBook Pros with Intel processors that were absolutely shocking. So I was very happy when Apple moved onto its own silicon. So it's one to watch, certainly.

C (9:38)

I mean it's a good example. I mean, who knows what's going to.

B (9:41)

Ultimately happen to Intel.

C (9:42)

They were one of those companies where.

B (9:44)

If you went back 20 years ago, it's like hard to imagine intel being in the state that it is now.

C (9:50)

It's like there's all these companies that.

B (9:52)

I think we talk about today where it's impossible to think about the world without those companies.

C (9:58)

But these things happen.

B (9:59)

Not that Intel's necessarily going away, but.

C (10:01)

There is always the fear of a.

B (10:02)

Company becoming sort of like the next Xerox or something like that.

C (10:05)

And I think that some of the.

B (10:07)

Things that we're seeing now around this battle around search, even going back to the stuff we talked about with Perplexity.

C (10:13)

If all the high value searches, essentially.

B (10:15)

The equivalent of high value searches, go to ChatGPT, Gemini, Perplexity, these types of interfaces, then what does that ultimately do to a behemoth in the industry like a Google?

A (10:26)

Yeah. So moving on to our kind of final big headline, it's Meta, unfortunately, again. But we did try not to put Meta at the top of the show again. I think they're just trying to make news. So this is around the fact that they are now freezing AI hiring.

C (10:41)

If their strategy is to make news and press, it's a very expensive strategy. So we talked a bunch about Meta.

B (10:48)

AI hiring and AI talent battles that are out there in the last show and apparently Meta just hit the pause on AI hiring after months long spending spree where they pushed more than 50 researchers from OpenAI, Google, Anthropic and others. And sometimes there were offers of hundreds of millions of dollars.

C (11:07)

Zuckerberg's been personally involved in the recruiting.

B (11:10)

Even paying out 14 billion for scale AI to land as co founders as Meta's chief AI officer.

C (11:17)

So a lot of money, a lot of spend. So I'm not shocked that there could.

B (11:21)

Be some investor pushback from that.

C (11:23)

So Meta's frozen hiring and they're reorganizing.

B (11:25)

Their AI division into four groups with one focus on superintelligence.

C (11:29)

I think the other challenge is kind of like if you do a lot of acquisitions at once, it's one thing to be, hey, we acquired all these.

B (11:35)

Great teams and companies and software.

C (11:38)

It's a whole other challenge.

B (11:39)

How do you actually integrate that and make sense of it?

C (11:41)

And I think they could be in a similar state where just because you.

B (11:45)

Acquired all this talent doesn't mean that talent can work together behind one singular vision and actually deliver on something.

C (11:50)

So I think the jury's still out on terms of is Meta's lavish bets going to end up leading to a.

B (11:58)

New generation of innovation for the company and they'll be able to catch up in some of the AI races and offer something compelling.

A (12:05)

Yeah, it sort of feels like they made such a splash with the amount of money they were putting on the table, more of the package amounts they're putting on the table. And you can only imagine, sort of internally, there was probably just a lot of like, all right, how am I going to get on this AI team and triple my salary and all this kind of stuff? So there was probably quite a lot going on there. But excuse the pun, I feel I did make a bit of a Meta prediction last time, which was that this could be a vanity project. I can't see how. Just assembling kind of the super band of AI and hoping that the album is great. I just don't know how that works. I've never seen things work where people just throw money at a bunch of disparate people, put them in a room and say, go. I don't know. I can't think of many examples where that's worked. Or like, go, but go really fast. And like, we need you to go as fast as possible. That never tends to actually create great things. So let's see.

C (13:01)

And they spend a lot of money.

B (13:03)

Developing the Metaverse, which I don't.

C (13:06)

Was wildly successful. So I think if they end up.

B (13:09)

In a situation where they are spending wildly on AI initiatives that don't necessarily go anywhere, it could be problematic for the company in the long run.

A (13:18)

Yeah, for sure. All right, so let's move on to our main topic. So something that Sean is quite the expert in. So we're just talking everything agentic. We're just kind of looking at where are we with agentic. And agentic workflows and so on is kind of what everyone is trying to anchor themselves to at the moment, is kind of my feeling. You know, whether you're anthropic or whether you're like a company that's actually been going for, say, five years and now you're, you know, you need to tag AI on your name. But it's not just AI. You're then saying, oh, but we're the agentic, blah, blah, blah. But the reality is kind of where we're trying to look at today. So maybe if we kind of look at this in three stages. Just like where is the disconnect between reality and what's going on? What is actually happening in agentic that's real. And then just looking at any of the main innovations in agentic workflows. But I'm going to hand over to Sean, you're the expert.

C (14:11)

Yeah. So I mean just I think looking.

B (14:14)

Broadly, there's a lot of, I think recent news around what is truly the ROI for this. Reports on, hey, most of these projects are going to be canceled in the next couple of years. It's unclear what the business value is. We don't really understand the risk controls.

C (14:30)

Like there's all these, both technical and.

B (14:32)

Also I think sort of business operational challenges around bringing these AI systems into production.

C (14:39)

And I get that. I think one of the challenges in.

B (14:43)

The industry, and I often talk about this with different businesses I work with.

C (14:46)

Is that there's the reality of what.

B (14:50)

You can do with agentic workflows or even models in general today and where you can get value as a company.

C (14:56)

And then there's sort of the hype.

B (14:57)

Cycle around this where people are vibe coding demos or you see these keynotes that look really impressive and things like that.

C (15:05)

And I think there's clearly a separation.

B (15:07)

Between what's happening in the zeitgeist of hype and what you can do in reality.

C (15:13)

And I think unfortunately a lot of.

B (15:14)

The hype is distracting for where there.

C (15:17)

Actually is value because I do think.

B (15:19)

That there's generally value.

C (15:21)

But most of the hype is kind.

B (15:23)

Of focused on aspirational things that I don't think are realistic right now.

A (15:27)

Yeah, like hype being oh, we can have a fully agentic employee is like a extreme example.

C (15:33)

I think that all that does is.

B (15:35)

Create fear in the market. And I also think it's unrealistic.

C (15:38)

There's so many challenges right now of.

B (15:40)

Actually being able to build something that is fully autonomous, that you could trust and run reliably.

C (15:46)

If you even think about like a.

B (15:47)

Multi agent system where you have a.

C (15:49)

Bunch of, let's say you basically have.

B (15:51)

A bunch of nodes in this, maybe in a graph or maybe it's a dynamic graph structure.

C (15:55)

Well, if each of those have a.

B (15:58)

Probability of success of say 90% success.

C (16:01)

And 10% failure, but each of that compounds, then you end up with essentially the probability calculation. You multiply the error rate. So by the end of that, if this thing's like 10 nodes deep, you might have only like a 30% chance of success. So that's not a very good success rate. So the Idea that that could replace.

B (16:21)

A person right now in most tasks is I think unrealistic.

C (16:24)

That being said, I think that there.

B (16:26)

Are a lot of sort of meat and potatoes unsexy use cases that are.

C (16:29)

Tremendously useful and a lot of it.

B (16:31)

Has more to do with these kind.

C (16:33)

Of human on the loop experiences where you use a model, maybe component of.

B (16:39)

It is sort of agentic in some fashion where there's some dynamic decision making.

C (16:43)

And it's able to create some sort of first pass type of document. Let's say you take support tickets and then you want to be able to give like a support engineer, arm them with a potential solution or at least.

B (16:56)

The materials that help them solve the problem faster.

C (16:58)

Well I think that is a realistic scenario. We run tests on those types of things where and had humans evaluate and score those, where we get quite good success rates. And if you can make those people significantly more efficient and kind of take away some of the not fun mundane tasks of just collecting all this data.

B (17:15)

Together, going inspecting certain logs and you can have an AI system basically go and do some of that grunt work.

C (17:21)

And kind of bring it together, that's very, very valuable. And I think from an ROI perspective if you have to pay somebody to.

B (17:28)

Kind of run around and stitch these.

C (17:29)

Things together today, that's a very expensive use of human talent where I think.

B (17:35)

The token cost is certainly justified.

C (17:37)

If you even think about the legal domain, there's a whole pocket of legal where you pay people to basically go into these rooms and find an obscure.

B (17:45)

Document needle in a haystack.

C (17:47)

Well LLMs are very, very good at.

B (17:48)

That kind of activity.

A (17:49)

Yeah, I guess maybe two of the core examples that keep coming up as success cases is the almost like it's either like IT support or employee onboarding. And where does agentic become helpful in that regard? Well with IT support you might in Slack say hey my VPN isn't working and it can go off and actually look up not only just the documentation for that VPN but it can actually, for example maybe in a browser look at the configuration of the VPN and sort of reason about why that might be a problem.

C (18:26)

Capital One's tech team isn't just talking.

B (18:28)

About multi agentic AI.

C (18:30)

They already deployed one. It's called chat concierge and a simplifier.

B (18:33)

In car shopping using self reflection and.

C (18:35)

Layered reasoning with live API checks, it.

B (18:38)

Doesn'T just help buyers find a car they love.

C (18:40)

It helps schedule a test drive, get.

B (18:41)

Pre approved for financing and estimate trade in value.

C (18:44)

Advanced intuitive and deployed.

B (18:46)

That's how they stack. That's technology at Capital One.

A (18:50)

The other example that again seems to be just getting reality check is good. It's coding again and a lot of it now is in the terminal cloud code, other such things. So it's interesting. We're already slightly moving away from in the IDE to just in the terminal. What's your take on those?

C (19:10)

Yeah, so I think one of the.

B (19:12)

Reasons why some of the stuff in engineering is kind of the tip of the spear when it comes to being able to leverage some of these. I guess like the edge of what these models are capable of right now.

C (19:24)

Is because you have ways of kind.

B (19:27)

Of validating the output.

C (19:28)

Because one of the biggest challenges with.

B (19:30)

A lot of these problems is it's very hard to validate whether what the model produced is actually correct or not.

C (19:36)

Like it might read totally fine, but.

B (19:39)

It'S hard to know for sure unless you're an expert of how do I actually validate this in a large scale system, how do I validate that or have some confidence at scale without having to have a human always in the loop.

C (19:51)

But with engineering type tasks, I have.

B (19:53)

Some checks and balances. Like if it's code generation, well, I.

C (19:56)

Can compile that code, I can run.

B (19:58)

The code in a sandbox environment, I can run unit tests, I can run integration tests.

C (20:03)

So I have a bunch of stuff that I can essentially have some level.

B (20:07)

Of confidence that it's executing correctly and.

C (20:10)

Doing the right thing. I think we're actually starting to see.

B (20:13)

Some similar things in sort of the.

C (20:14)

DevOps world, where if it's something like.

B (20:17)

Automatically diagnosing an issue in, I don't know, let's say cloud resources and there.

C (20:23)

Are ways of essentially trying things and then undoing it. So you need to be able to have something that is automated or at.

B (20:31)

Least you have a lot of confidence.

C (20:32)

Around that doesn't need as much hand holding. You need a way to kind of.

B (20:36)

Be able to roll back changes if you're allowing like an agent to take certain actions, like it has to be.

C (20:41)

Action and you have to be able to almost think of it like a.

B (20:43)

Commit log in a database where I.

C (20:45)

Can run some sort of transaction, but.

B (20:47)

I can undo that transaction. I have this kind of log that can go back and replay and so forth.

C (20:51)

And we've built a lot of technology around that over the years to help.

B (20:54)

Us solve some of these problems.

C (20:55)

And I think not all those things.

B (20:56)

Have been baked into some of the things that we're building and trying to productionize today.

C (21:01)

I think There's a big disconnect between being able to put together like a.

B (21:06)

Compelling demo internally at a company and then getting to a place where you can productionize that. And a lot of those challenges have to do with how do I eval and test these. We don't have great solutions for that today. It takes a lot of hand curation.

C (21:18)

To essentially do that. And how do I have some level.

B (21:21)

Of confidence of validating the output is correct.

C (21:23)

And then there's also this big challenge around sort of the data that describes the data. So when you're building any kind of.

B (21:32)

Gen AI experience for a business, the.

C (21:34)

Big hurdle is that these models are really powerful and they're very, very smart.

B (21:38)

Kind of about, like public information.

C (21:40)

But they're kind of dumb when it comes to your specific business, your specific task, your customer information. So you have to essentially go and collect that data and provide it to the model so that you can steer.

B (21:50)

It in the right direction.

C (21:51)

It has the right context, essentially during prompt time assembly to be able to.

B (21:56)

Come up with an intelligent response.

C (21:58)

And there's a number of challenges there. One is simply like, how do I.

B (22:01)

Go and retrieve the data from all these different locations?

C (22:03)

But on top of that, there's kind of this larger issue of even if I have the data, is the model necessarily going to understand how the data relates? Because when you're talking about human language, the way these models worked and are able to get this understanding in air quotes of human language is because they've been trained on essentially billions of documents where the relationships between words and sentences and so forth, the structure, they're able to form essentially a statistical model that helps them understand that by just seeing this multiple times. But those patterns don't exist in sort.

B (22:42)

Of the land of spreadsheets and databases.

C (22:44)

And where customer data is today, the model's not going to understand that some obscurely named column in your database represents.

B (22:52)

I don't know, like some sort of.

C (22:53)

Customer ID that has to be factored.

B (22:55)

In as a foreign key to another.

C (22:56)

Table without explicit instructions. So this metadata, the data about the data, is really, really important. And most companies don't have a good way of essentially conveying the metadata because.

B (23:07)

Most of it exists inside their heads.

A (23:09)

I mean, have you seen any real cases? Because this feels like from a fiction point of view, obvious. But any real cases of where companies are actually training their own foundational models on their data, which is different to say, let's just not putting out Avro here, but glean is, hey, we'll pull in all Your data from the organization across lots of different places and you can query that and get more contextually better results. But what we're talking about here is if you're going to set up an agentic workflow, well, which large language model are you using? And as you've just said, if you take one off the shelf, it's not going to have this super context or understanding of your business from all sorts of angles. And well, if it could have seen that spreadsheet before it answered the question, it would have probably given a much more accurate flow and output. So yeah, have you seen anything like that?

C (24:04)

I mean, some people do some level of fine tuning. There's not too many companies I've come.

B (24:08)

Across that are like from the ground up building an LLM, maybe a handful that you could count on like one.

C (24:14)

Hand, but most people aren't doing that some level of fine tuning. But even there there's challenges because with fine tuning, since you're adjusting not all models you can fine tune, there's not.

B (24:26)

Necessarily ways of doing that if there are these API based models that are hosted somewhere else.

C (24:30)

But then on top of that there's.

B (24:32)

A challenge of if you are making sort of these model weight adjustments and.

C (24:36)

Then new models come out, or I.

B (24:38)

Want to try out a different model because another company comes out with the next big innovation models, I don't have an easy way to kind of transfer that knowledge.

C (24:46)

So it can become actually a sort.

B (24:48)

Of barrier to innovation.

C (24:50)

And most of the people that I.

B (24:51)

See doing some type of fine tuning have very specific reasons for doing it. Like they're quantizing a model to shrink it down and run it on an edge device and then they're fine tuning it to get the performance that they need for specific tasks.

C (25:03)

Or it might be really domain specific.

B (25:05)

But I would be cautious about doing.

C (25:08)

That from the get go. What most people are doing really is.

B (25:12)

Trying to figure out how do they contextualize the prompt. There's this whole area around context engineering now and also going gathering the data.

C (25:21)

And then how do you also encapsulate the metadata. So there's a lot of things going.

B (25:26)

On in the industry right now where.

C (25:28)

I think the new battleground for data.

B (25:30)

Is really about the metadata.

C (25:31)

Because the most valuable from an AI perspective for being able to take these.

B (25:36)

Models and make them task specific, probably.

C (25:38)

The most valuable data you can feed them, besides the raw data, is the.

B (25:41)

Data that describes the data.

C (25:42)

So that's like semantic layers, knowledge graphs, ontologies. All these technologies I think are having a huge Resurgence. But it takes a lot of work to actually get these things out of people's head and encode them so that.

B (25:54)

You can make the models reliable.

C (25:56)

So that's a big barrier. I think a lot of the challenges.

B (25:59)

Around building these systems successfully today has.

C (26:03)

Fundamentally to do with companies just have.

B (26:06)

Poor data posture to begin with. And it's all about the data. Like all these AI problems are fundamentally data problems.

C (26:11)

So we have these powerful models, but most companies aren't in a place from.

B (26:15)

A data perspective where they can even take advantage of them.

A (26:17)

Yeah, and that slightly runs against at the moment. If you look at any companies selling agentic workflows, they're all saying enterprise, that's where they're trying to get to. And yet here's you saying, but sure, but they don't have the data in a place that can make that a reality. And then okay, well could the enterprise do it as an internal project, for example? But it's that then classic thing of like, well, how much time are we going to sink into this prototype? Is it just to show everyone it's possible and then walk away from it? Or are you as an enterprise really invested in trying to automate away sort of drudgery of not necessarily coding because we're kind of doing that already, but like drudgery of other roles because I think the hey, we'll automate your drudgery is kind of the sell for most agentic to enterprise companies. But I just don't have a feeling that there's just a lot of uptake other than probably very low paid prototype type projects basically.

C (27:17)

I mean, I think the things that.

B (27:18)

I tend to see that actually hit.

C (27:21)

Production are a little bit more workflowy.

B (27:25)

In nature than these kind of truly agentic systems where you just have a bunch of nodes floating around the ether and you're going to figure out the graph dynamically.

C (27:33)

I think they're a lot more, hey, we want to take this particular input and be able to process it to some portion of automated processing to hand.

B (27:43)

Off to a person to then justify. So if you think about, I don't know, something like loan underwriting, well, loan.

C (27:49)

Could come in, that could be the input signal there. You know, like, okay, all inputs are.

B (27:53)

Going to be a loan application.

C (27:54)

So you already have constraints around what you need to interpret. It's not like anybody can just come and ask anything.

B (28:00)

So it's a much more like sort.

C (28:01)

Of closed world problem where I understand what the inputs are and I can.

B (28:04)

Break that down into kind of like a workflow that a person might execute themselves.

C (28:08)

And there'll probably be some form of maybe dynamic decision making as part of that. Like what which data resources and tools it might need to interact with. But it's a lot more fixed in terms of the graph structure of it than these like truly ejectic systems. And I think those are very reasonable.

B (28:25)

Places to start that do deliver value for business. And I see that, you know, a number of companies being able to be able to successfully put those into production.

C (28:32)

The other advantage of that, where you kind of make it, hey, we're going to solve this specific problem is if your data is a mess, you don't have to solve all your data. You don't have to boil the ocean from day one. You don't have to think like, oh my God, we have 150 petabytes of data. How are we going to define schemas and catalogs and have governance over all this? Well, you don't have to do all of that. You can pick a particular thing and be like, okay, well I know I need these data inputs. How do we make sure that data is a state where we can actually.

B (28:59)

Action it with an agent?

A (29:01)

And probably one area that we should touch on with again, why do things not hit production maybe as quickly as they could? And security is kind of a big one there. Simon Willison, who has a blog that seems like half of his posts just get up to the top of hacker news. I believe he's an exited technology person who spends most of his time writing now, which is awesome. And it's not behind a paywall, it's not behind one of those. He described this as the lethal trifecta. Well, he described this kind of before even maybe before agentic sort of really took off. But the lethal trifecta, three things as it would sound, access to your private data, exposure to untrusted content and the ability to externally communicate. And you know, this is just such a fertile place for all three of those, you know, access to private data. Well, companies basically only have private data, private to themselves, pretty much exposure to untrusted content. That's, you know, what prompts are you bringing in to drive the flow and then what could be maliciously added in the middle there and then the final one being then the ability to externally communicate. And we saw a case just two days ago, nx, which is a NPM package. And that followed quite a common pattern at the moment from the externally communicate bit, which is the developer is sort of asked, do you want to publish this into a Repo. And it does, but A, it's a public repo and B, it's got a whole bunch of information in it and that's all you need a window of like a minute because whoever has figured this one out is ready to listen for that exact repo name and grab the data and then off we go. So anyway, I've kind of gone on a little bit on that one, but yeah. What's your take on how we maybe get beyond some of these challenges?

C (30:47)

Yeah, I mean it is tough.

B (30:48)

It's tough.

C (30:49)

I think that you're kind of going.

B (30:50)

Back to some of the things I said before where what I see companies being able to successfully do are these much more sort of closed world problems. Another one of the advantages there is that it's easier to understand the security model for it because you're basically constraining the number of options. It's all about how can I add as much determinism to this as possible? Because without that, where you have some chatbot type interface with an agent that has access to all kinds of stuff.

C (31:20)

It'S very hard to police that. How do you know what someone's going to input?

B (31:25)

It's basically an unbounded problem. How do I validate whether the outputs are correct?

C (31:30)

And I think another common challenge that.

B (31:32)

I see from a security perspective with companies that I talk to is that.

C (31:35)

Their platform teams, they want to be able to understand what each piece of.

B (31:40)

Software is accessing from a data perspective.

C (31:43)

And if every team is running with their own independent AI projects and they're building sort of bespoke tool integrations, maybe.

B (31:52)

Even through MCP, which they pull down.

C (31:55)

From GitHub and they're running it themselves or whatever, that's ridiculously difficult to control. You're sort of exponentially increasing the number.

B (32:03)

Of problems from where you could run into a data security issue.

C (32:06)

You're suddenly creating this huge footprint where.

B (32:09)

The platform teams, the governance teams don't really have insight into what's going on.

C (32:13)

And ideally the reason people invest in various.

B (32:17)

Or you have a platform team and you have, you've done your vendor procurements and you understand which data systems are.

C (32:23)

There, you have some sort of central control is to give you visibility and.

B (32:27)

Audit logs into what people are accessing. And I think a lot of companies are in such a hurry to rush into adopting AI that they're not always thinking about that. Building agentic AI apps isn't just about choosing the best. LLM agents need short term memory, long term recall and lightning fast retrieval. Without it you're left with clunky prototypes that never scale, you know. Redis, the world's fastest caching solution. It turns out fast data is the key to good context. And good context is essential for fast, accurate memory. It's what makes AI agents actually work with your data. Redis for AI. The right infrastructure, the right tools, the only way to scale. Learn more at Redis IO Genai.

D (33:10)

Building an app often feels like a balancing act. You want to ship features fast, chat, activity feeds, moderation, video, but building them from scratch is slow and complex. That's where Stream comes in. Stream provides developer friendly APIs that let you add real time communication without reinventing the wheel. Why Stream? First developer experience. Stream has clean open source SDKs and great docs and you can get a proof of concept running in hours. Second speed to build, experiment, prototype or launch features quickly. No credit card required to start. Stream also scales. Over 2000 global apps including Strava, Patreon, Nextdoor, Robinhood and Peloton rely on Stream to power in app communication for more than a billion users. Whether you're a startup or an enterprise, Stream handles the hard parts so you can focus on what makes your app unique. Get started today@getstream IO podcast.

A (34:16)

So yeah, I mean, what's maybe the TLDR here on where are we with agentic?

C (34:22)

I mean my take and my advice.

B (34:23)

Usually for companies is try to focus on these kind of closed world problems. If you want to be building stuff.

C (34:28)

There is a lot of value there. If you can take some human driven.

B (34:32)

Process today that takes people considerable time to like stitch stuff together, then it's really valuable to try to arm that person with the materials for them to be successful.

C (34:41)

If you look at things even from.

B (34:43)

Anomaly detection to root cause analysis, that's a great place to start.

C (34:47)

There's a lot of use cases even outside of engineering where some combination of.

B (34:53)

Events is triggering some alarm, where some human person has to go and investigate, like why are we having this alarm?

C (34:58)

And they have very little context other than an alarm went off. And then they have to do all.

B (35:02)

This investigative work, probably look at a.

C (35:04)

Bunch of different software systems, just understand why is that alarm going off, how.

B (35:09)

Big a problem this is.

C (35:10)

That's a great use case for having essentially, you know, call it an agent, call it an AI system that does a lot of that investigative work for you. So you arm the person when they come in to do this work with additional materials so they can cut down.

B (35:24)

The time to actually respond to that alarm and hopefully solution it.

A (35:28)

Awesome. Well, I Think we've covered kind of the key areas on sort of where we are with how our agentic workflows actually being used at the moment, I think is a bit of a confusion point because there's a lot of noise and that's usually what we try and dispel a little bit in this main section is take a topic where there's maybe a lot of noise and just kind of break down from a couple of angles what might actually be going on. And like what we've been seeing. I think it's always helpful, like when you're not living in one of the sort of tech epicenters, it really is, I think, such a disadvantage. You just don't get a bunch of conversations happening around you that you might sort of suddenly understand what might be going on actually. So hope that's been helpful for you as an audience. But as always, we move on to our, I think, as we keep saying, kind of our favorite part of the show where we have looked at hacker news. We just kind of keep a running tab of a couple of things that we might want to bring up. The first one I was going to bring up, so, yeah, I've got one kind of it's not exactly software. I think it will relate to most people listening, which was this group have got together a proposal to ban ghost jobs. So thank you to user Teaver for posting this. It got quite far up the list and yeah, this is basically just around someone who was struggling to find a job and was getting very frustrated with thinking that actually these jobs were what are called ghost jobs, that is jobs that are posted by companies who have absolutely no intention to fill that role. And I've got to say I've seen this firsthand. So I know that companies do this. They put out job ads where they might still take a first interview with someone, but they have very, very little actual impetus to fill that role. And a lot of it is often around like investor sentiment that you love to sort of an investor maybe looks up the company. Oh, wow, they've got 50 open positions or so. Yeah. And LinkedIn is quite frankly rife with these. So, yeah, the group got together and they wanted to put together kind of like a spec for like what a job ad should have in it. And you know, we've already moved to a place, I believe in certain states in the US where you have to state the salary range. So that's great. I think everyone's been appreciating that one.

B (37:34)

Yeah, that's the case in California.

A (37:35)

Right, right. Yeah, I think New York has it actually as well now like these kind of two hotspots for salaries. But yeah, they're also going, you know, like some of the spec was sort of like you must prove that what is the time frame that you were going to ensure that you had hired this person by and actually state that. So we need to hire this person within two weeks or two months. You need to state whether this could have been a backfilled position internally, et cetera. All very sensible stuff. The obvious thing is how on earth will this get implemented? Actually, who knows? I think it's great. It's not to in any way discourage people from taking a shot at these things how on earth this could get implemented. And I think it was interesting at the end of the article there was a sort of slightly glib ending which was. And it seems that when people get a job they completely forget all about this and don't care about it anymore. So yes, it's often a problem that people searching for a job and not having a ton of luck think about more. But I do think I've at least seen it. I can't say I've experienced it firsthand on my side, but I have definitely seen basically ghost job ads go out, things that just had zero intention of being filled. So it's not a fiction is all I can say.

B (38:45)

Yeah, I'm sure it happens.

C (38:47)

I think if there was something where.

B (38:49)

It did become some sort of policy.

C (38:52)

A number of companies would follow it.

B (38:53)

Just like the stating the salary, for example. I think it would probably be difficult.

C (38:59)

To police that widely.

B (39:01)

But a lot of companies are going to comply anyway if it's sort of part of baked in the rules. So maybe that helps in some fashion.

C (39:07)

I think if you have a bunch.

B (39:08)

Of those sort of requirements in place, then it at least maybe makes companies think a little bit more deeply about putting the effort essentially into posting a job when they have no real intention of hiring.

A (39:19)

Yeah. What did you turn up, John?

C (39:21)

Yeah, so I found this pretty cool blog post. I love these blog posts now where.

B (39:27)

You can make them really interactive.

C (39:29)

So this post that goes over big.

B (39:31)

O notation posted by Sam Hu is the name on Hacker News and his.

C (39:36)

Website samhoo.dev but essentially the explainer explains.

B (39:40)

Big O notation, which is if you're not familiar, it's been a long time.

C (39:44)

Since you studied computer science is how.

B (39:46)

We measure how algorithms scale as input grows.

C (39:49)

And it essentially looks at the article breaks down with really clear examples constant time or big O of one log.

B (39:58)

Time linear or quadratic.

C (40:00)

And it uses these JavaScript snippets as.

B (40:02)

Examples with these interactive visuals to show you why, for example, looping through an array is linear. Bubble sorts quadratic. Binary search is logarithmic.

C (40:11)

I've learned all those things in school.

B (40:13)

Now at this point, many, many years ago.

C (40:15)

But it's a really cool demonstration.

B (40:16)

Even if you are familiar with these.

C (40:18)

Concepts and if you're just learning them.

B (40:19)

I think it's a great way to kind of reinforce what you're learning.

C (40:22)

It also calls out some coding pitfalls as well.

B (40:25)

So there's some practical advice in there. Like if you use IndexOf inside a loop, how does that impact performance and what's happening sort of under the hood?

C (40:33)

So you think about not just the code that you're writing, but what are.

B (40:36)

Maybe some of the data structures that underpin some of the libraries that you're using or some of the functions that you're using.

A (40:42)

Yeah, that's awesome. And as you say, this is becoming a bit more, I don't want to say popular, but seeing a few more of these sort of very interactive blog posts, I'm missing the name, but there's someone who puts out maybe once every quarter something, you know, how bicycles work or how watches work. And it's just insane, the detail and like what you can do sliding things around. So, yeah, I love this where it actually is, you know, coding related and allows people to kind of maybe see some of the gotchas with stuff that, as you say, you might have learned back in high school or maybe university. But it's probably a bit of a foggy topic for you now, so that's very cool. Yeah. My second one, this is one that was a bit personal to me because so admittedly you didn't get loads of upvotes, but it crossed my radar by user vocram. It was about software bug renders Wahoo GPS units unusable. Now I'm up in Scotland and I've got like a little head unit on my bike and I went out and came back and found that the GPS had been pinging Iceland and it had been pinging some other islands that were definitely not Scotland. And I was like, okay, maybe it's just, you know, Wahoo having a bad day. Then I noticed that all the right data was pegged to 2006 and I was like, okay, well that's like super weird. And yeah, basically Wahoo pushed out a Update for their Gen1 units. And I think it's actually just that the bug itself is Kind of interesting because it's actually to do with how GPS data is transmitted and it's that it sounds like they'd upgraded their code base to 10 bit GPS encoding and these units just couldn't handle that. So there's a whole bunch to unpack there. Like how did this get through? No one realized that none of these very widely sold units couldn't handle this data format. They did push out a fix within a few days, but yeah, basically rendered them kind of useless for like a week. I didn't also realize that timestamps appear in GPS like transmission data and that apparently also then that was why it was pinging places like Iceland and wherever because it was getting super confused as to what was going on. So yeah, that was a bit of a, bit of a mess. And just on a similar topic, my Chromecast had the same issue earlier this year. They pushed an update and it just bricked the Chromecast. Now that was really interesting because unusually I went and googled it or whatever I did, the hacker news did probably before I did anything. And actually that was the right thing to do because if I had decided to, you know, turn off, turn on, that would have caused another week of fixing because they managed to roll out a fix for people that had not turned it off first and then they said, I'm really sorry, but if you've turned it off and turned it back on, we will try to fix it for you. And they did, I think eventually. But yeah, just this idea that we've got all these things we might have bought. I mean both certainly the Wahoo unit is like 10 years old, but still pushing updates, but the ability to brick them from an update now is pretty high. Yeah, just watch out for that.

C (43:37)

The last one I had here, which is something I feel like I talked.

B (43:41)

About at some point, but it's kind.

C (43:42)

Of interesting is so Florida State University.

B (43:44)

Did this study where they showed that chat based AI tools are essentially not only shaping online writing, but they're influencing everyday spoken English. So they looked at 22 million words from unscripted speech and they found a sharp increase in AI associated buzzwords like Delve. There was a period where ChatGPT overused Delve all the time. I bet em dash in writing is.

C (44:07)

Less about speech, but it's being used.

B (44:08)

All over the place. And essentially their findings suggest AI driving this measurable language change.

C (44:14)

And I think the curious thing about this is what is sort of the.

B (44:18)

Widespread impact of this. We train these models basically on human.

C (44:23)

Language and my thought process is eventually not that too. Distant future, there'll be more AI generated.

B (44:30)

Written content online than there is human generated by significant margin. And then we're training the models on that.

C (44:36)

What's that kind of do? And then on top of that, we're now approaching, apparently we're training our own brains in the way that we think. The way we're sort of influencing the way that we speak is coming from these models. So there's all this weird sort of Mobius strip bias that's coming into play here. I don't know what ends up being the large scale impact of this, but it does worry me a little bit.

A (44:57)

Yeah, this is interesting, the fact that buzzwords like delve and intricate and surpass were in there. I think this is interesting because Singapore, where I usually learn very proper English and they say delve a lot. It's not a word that I would use very often. So it's almost like it's being trained on these very proper texts that bias.

C (45:17)

Came from because in Nigeria, people who speak English there in business context, they.

B (45:23)

Use delve more than Westerners tend to use delve.

C (45:25)

And when OpenAI did their initial human reinforcement learning, they had hired a bunch.

B (45:30)

Of English speakers from Nigeria. So it ended up biasing the model during the training.

C (45:35)

They've corrected it because of people basically complained about it. But there's other biases that end up happening.

B (45:41)

You see this all the time.

C (45:42)

And that's why I think there's certain signals when you look at writing.

B (45:46)

A lot of times if you play around with these models a lot, you can tell, okay, this was written by AI or they overused AI because there's certain things that kind of just feel like it came from a model.

A (45:56)

The word I'd love to know in this context is elevate. I'm absolutely sick to death of seeing elevate in every marketing promotional material out there. But this was kind of predated AI and my wife says, this is AI And I'm like, I think elevate was just getting a bit popular, but I think AI has kind of pervaded it. So I get very fed up when I see elevate.

C (46:16)

I mean, we already do that, right? Where even outside of AI, if someone.

B (46:20)

You know, a particular company is successful in marketing or and uses certain terms.

C (46:25)

And language, other companies copy it. It's kind of like naming of children. You can look at these spikes in name popularity and then they get oversaturated.

B (46:35)

And people stop naming their kids John for a generation.

C (46:38)

And then it comes back again because what's old is new. I think it's similar with the other terms as well.

A (46:44)

That just gave me a funny thought where probably in the future we'll go, oh yeah, there's tons of those names. That was the GPT5 era, right where that was the most common name that was asked for when when people said what should I name my kid? And that name appeared with that model and then so on and so forth. Or maybe you were named by Llama and again that means you've got a slightly different name. Looking ahead, predictions as we try to just give a very completely non serious prediction for the month ahead. What you've been thinking about, Sean?

C (47:13)

Yeah, so I talked a little bit about this. I think that the new battle for.

B (47:17)

Data is going to be over metadata.

C (47:18)

Earlier also of kind of like context.

B (47:21)

Serving for agents and AI and I.

C (47:23)

Think that market's kind of just beginning. So this is not necessarily metadata, but.

B (47:28)

Databricks just acquired a company called Tecton. Tecton's well known for being like a low latency feature store, but they're also now kind of have like a gen AI offering around like context store.

C (47:39)

My prediction is essentially that Snowflake is.

B (47:41)

Going to follow suit. They're going to have some sort of similar acquisition or product offering in the not too distant future.

A (47:47)

Gotcha. Mine is meta. Will ramp up their AI hiring. No, I'm kidding. Yeah, I think mine is. Well again this is for me just super non serious. There'll be some other device that's totally bricked by a software update. So you know, we've just had the Wahoo's Chromecast not that long ago, so who knows, maybe like PS4s will get bricked. I've got a PS4, I've got two of them actually. So it would be very fitting if those got bricked for a week or something with an update. So let's go with that. So yeah, thank you for tuning in. Apologies for how I sound. I have been enthusiastic about this episode so don't read too much in the voice. Thank you Sean for really taking over on this one. Appreciate it.

C (48:30)

Yeah, thank you everyone for tuning in.

B (48:32)

And we'll see you next month.

A (48:33)

See you next time.