Musings on London Tech Week 2026

London tech week this year was really exciting to go to, and I genuinely enjoyed it. We even got to see the Williams F1 car thanks to Barclays as well! I do however, have a couple of thoughts on some of the things that we saw, some of which is observations and some of which is predictions based on trends we observed.

• Compliance is king
• The return to bare metal
• AI is still the magic word, even if we don't know what it means
• The consultants are not ahead, but they wish they were
• The best in the business are using a mixed approach

Compliance is king

One of the big features that we saw this year was that compliance and safety was a big talking point. A lot of different companies all promised a specific proprietary method to ensure that compliance and safety was put first. But, what was lacking was really any context of how this would be done. I don't know if this was due to the stalls being mostly manned by salespeople looking to generate leads rather than the nerds who work on the systems but the lack of detail felt a lot like "Just trust us bro."

Compliance is obviously a big concern, for lots of different reasons. As the surface that AI is given access to increases so does the potential liabilities, companies want to leverage the benefits of AI without being sued, the "move fast and break things" mentality only works for so long. At some point you need to start to move slower and consolidate control as well. The problem right now is that this type of consolidation and control comes in one of four ways:

• RAG/Tool Lookup methods - The LLM does not natively have access to the information, this is gated by a third party system using deterministic rules-based methods
• Inference/prompt based methods - The LLM does have access to the information, but we ask it nicely to be careful or add additional layers of LLM's on top as gatekeepers
• ML pattern matching methods - The LLM does have information, but there is a traditional ML system on top such as BERT or semantic similarity over vectors of known attacks
• Dual LLM patterns - One LLM does have tool and data access, another LLM consumes raw content and passes it to the trusted LLM

The first of these methods isn't new ground. We've been building RAG systems for years now and deterministic engineering is classic software engineering practices and principles, this isn't to say that it isn't a great approach, it is. It's just not necessarily a novel approach.

The second of these methods has problems, existential ones. If we look back to SQL for a moment we historically had issues with mixing two different domains together, DATA and COMMANDS, because as far as the system was concerned these were all just strings users could cross the bridge from DATA to COMMAND easily Remember Robert') DROP TABLE students; --? Prepared statements solved this problem for us, but we fundamentally *cannot* do this with LLM's. This means that no matter how much marketing is done, and how extensively the LLM is trained it can always be potentially vulnerable to some degree

The third of these principles is to use ML pattern matching methods, we either look for particular honey-pot words and phrases e.g. "ADMIN:" or we do semantic similarity of vector encodings of the whole word, phrase or chunk it up somehow. This isn't necessarily a bad approach either but the problem is that it is fundamentally retrospective. It will not protect you from novel attacks, as soon as an attacker finds or tries an approach that you haven't seen before it won't pattern match and it'll be let through, unfortunately, this is what hackers are really good at.

Finally, we have the architectural solution. This feels like the most concrete and reasonable until you realise that it effectively works like privilege escalation attacks, or parser attacks. Yes, I don't get to talk to the privaleged LLM directly anymore but it's still passing my signal through, so all I need to do now is to trick one LLM into tricking another which isn't as impossible as you think. This is all because it's fundamentally not possible to totally isolate an LLM.



These problems all get compounded when you start also talking about things like poison documents, undetectable poisoned LLM's via training that have the same issues as the compiler verification problem and the fact that there have been papers recently that given a novel enough attack it takes a tiny amount of data to orchestrate a training attack, this all becomes more compliance theatre or Comfort-As-A-Service rather than a useful product.

The return to bare metal

If you've seen the show Silicon Valley you'll remember the Gavin Belson box. If you haven't then... go watch it? But basically, Gavin one of the shows antagonists is obsessed with having a physical "box" to sell that he can put his name on, and I think we're making a return to this kind of physical commodity for a couple of big reasons but the main one has to do with the concept of data sovreignty.

Before we can really talk about how we want to regain something, we need to better define what that thing is. We used to deploy all of our code on physical servers that lived either in an office or in a data center that was somewhat local, quite a few of the old guard will remember laptops being left with the lids propped open and post-it notes attached of "CRITICAL INFRASTRUCTURE - DO NOT TOUCH". It wasn't all that great though to be honest, you had to provision the hardware you needed months in advance, load spikes in particular were brutal because you either over-provisioned accepting that 98% of your compute would be idling or you accepted that you'd go down during peak times. Then virtualisation came along and finally IAAS where we could rent on-demand compute from AWS. This got more and more complex as we moved from Infrastructure as a Service, to Platform as a Service to Code as a Service.

It can't be stated how much this was an explosion of productivity, however we did lose something and that is data sovreignty. everything we did, everything we built, and every byte of data now passed through a third-party company.

The data part was a secondary concern the primary one being the creation of walled gardens and the risk of robber-baron monopoly practices once people were trapped in the ecosystem, but I think we fundamentally became more aware of the data problem once LLM's entered into the picture. data was, has been, and always will be the most valuable commodity your business has, and the LLM providers know that.

So, how do we stop them from eating our lunch? Go local-first. A lot of the model use-cases don't actually need the most powerful models anyway, and even if they did we're waking up to the reality that paying $10,000 for a beefy computer that we can run inference jobs against is actually significantly less costly than sending your sensitive customer data to be trained on by an American firm breaching data export regulations and potentially being the subject of billion dollar lawsuits.

What does this actually look like? Well, companies know that consumers need the physical hardware as well as running inference isn't really possible unless you've got enough horsepower to hold all the weights. So companies are selling physical boxes. We ran into a number of different vendors that were essentially shipping Mac-Mini's with Ollama pre-installed to run their own software. Effectively offering plug and play, Inference as a Service.

I don't really think this is a pure technical innovation, in some ways it is as models get more performant at smaller sizes, but I think this is mostly a big change in terms of consumer behaviour that is driving this. We started out decentralized, centralized for all of the benefits that it promised and are now decentralizing again.

AI is still the magic word, even if we don't know what it means...

A lot of companies at London Tech Week all had some mention of AI somewhere in terms of marketing. It is after all the new hotness. But after speaking with a lot of the vendors they couldn't really articulate how, or why they were using it.



Lots of companies are feeling left out, especially the ones that operated in a more traditional products based industry so referencing AI even in a tangential fashion is more of a marketing gimmick than anything else in the same vein that a few years ago everyone was asking how they could incorporate "the blockchain" somehow.

I think though that ultimately we've passed that era now, this approach feels dated especially with a consumer base that is now asking tougher questions because they're more educated in terms of what is available in the market and what their needs now. Taking this approach with customers leaves companies at real risk of looking the fool.

I think right now there are two approaches that seem to work, and there isn't room for pretenders anymore:

• Go traditional: Say that you don't use it proudly, articulate your reasons for *not* using it
• Go cutting-edge: Say that you do use it, but clearly state *how you are different*

The consultants are not ahead, but they wish they were.

The consultants are more nervous than ever, and I think that's clearly evident. One of the patterns that I think I've generally always observed with both people and companies is that when something can't be explained clearly, the party doesn't understand it. They start to hide behind jargon and generalities or hand-wave to the general complexity of the situation. But when you meet a true expert they will caveat a lot, they don't speak in certainties ever. But they will explain clearly in just a sentence or two.

Operating as a consultancy is mostly about selling comfort.

The big dirty secret of the business though is that consultants generally are only really one page ahead in the manual than you are, like Gromit frantically laying train tracks ahead they don't have the long term forecast either. Generally though this works atleast to some degree. This is because they can build up playbooks over time and take experience from different organisations and draw parallels in terms of approach.



The problem though is that we're in a totally new space. There are no established patterns.

So, we're in a bit of an odd space where consultancies in terms of sales pitches are very solution forward, they want to come and help, they can guide you through these turbulent times no problem. But, when put onto a panel with other experts in the field who *can* talk about what they think the challenges are and are not subject to the emperors new clothes they have to hold two simultaneous positions. Not being able to articulate what the future actually is, but needing to sound like enough of an expert that only they can help. For a price of course.



So, my advice at the moment is to ignore consultants frankly. Nobody has the answers right now, and if someone is saying that they do you have to ask yourself why are they selling them to me, and my company? At an affordable price? Why are they not working for the FAaNG lot or Wall Street making billions instead?

The best in the business are using a mixed approach

One of the best panels I attended in the whole event was "Knowing What Customers Want, Before They Do: The Power of Machine Learning in Discovery", specificially the contributions of Catarina Runa Miranda who is the Global AI and Platforms Director at Pandora.

I think that the reason that the contributions of this particular individual resonated with me for several reasons which i'll outline, and then go into why:

• Pandora has been through the entire ML modelling before, GenAI is an addition, not an introduction into AI.
• They are solving for a specific customer need.
• They are aware of the limitiations, and drawbacks of AI.
• They aknowledged some of the biases, ethical considerations and cost of AI useage.

Firstly, this isn't Pandora's first rodeo using analytics for their customers, and as a result of this or maybe due to the background the systems that were outlined felt grounded and clear. I think one of the problems at the moment is that we have evangelists on one side of the conversation and doomerists on the other. Having the history in terms of building these systems I think helps to ground out what they can do, what they're good at and most importantly how they can fail, and how likely they are to fail. Obviously the use case here is lower stakes than pacemakers but it still has failure cases.

Having the academic and practical background on these subjects as well helps to pick the right tool for the job, rather than everything looking like a nail when you have a hammer. There are specific cases in which GenAI is the best tool for the job, but these cases are actually quite rare. We must accept that at the moment it is expensive, and risky in terms of outcomes but sometimes, the only way to tackle a problem is the best way, purely by the virtue of it being the only way.



Pandora is tackling a specific gap in their capabilities, this is mostly in terms of summarization. Pandora builds complex models of customer profiles that they can use to know what to put in front of you, what you're in the mood for and what's on trend right now. But, giving a salesperson in a store a spreadsheet of floating point values between 0 and 1 across various statistics of you isn't helpful at all. Instead, feeding those statistics through a summarization layer in an LLM gives you explanability backed by sourced data, so the transparency isn't lost, it can be easily verified and we don't suffer the same worries in terms of sandboxing due to security because it isn't consuming raw data. To be frank it's the perfect use case.

This underscores the hope that I have surrounding this hype cycle. There are genuine gains to be made from GenAI and it is useful in specific contexts when you manage for it's weaknesses and draw upon it's strengths, but the hype that surrounds it is blinding us to grounded criticism and evaluation of it as a tool. The big winners going forward are going to be the ones that can do this.

In conclusion

There is a lot of hype driven marketing right now, and there are real players in the space that are operating. But it still feels like the biggest winners are the ones selling shovels in a gold rush rather than the ones digging. There was some genuine innovation and the move towards local first is exciting especially with local-first models, but it's going to be interesting to see if those same pain points that were the arguments for centralization come back or not, I think they will. Finally, the big winners going forward are going to be the ones that take a critical problem first approach towards solving problems, and how AI might be the best worst solution to those if at all.