"Bliss is was in that dawn to be alive.."

It has now been 10 years since I wrote my first angel cheque in an AI first startup ( David Plans BioBeats that later exited to Huma) and over 9 years since I invested as a fund in my first AI startup ( Tim Porter 's Gluru that exited to Dailpad) and almost a year since I picked up the UKBAA award for "Exit of the Year" for Peter Mountney and Dan Stoyanov led Odin Vision ( a 30X return exit to Olympus), but never have a felt more pleased and privileged to be involved in the UK's AI startup scene than last week at the Barclays Eagle Labs and Twin Path Ventures monthly AI Breakfast.

With an audience of 80+ primarily founders of AI startups, we talked and learned from a fantastic panel of UK LLM stalwarts comprising of Alistair Pullen co-founder of Cosine AI and Jamie Dborin (PhD) co-founder of TitanML (who gave a great rundown of both the power and limitation of Zero Shot RAG and why even with the expected arrival of Llama 3 and GPT 5 sometime soon, it is highly unlikely to do away with the need for AI experts to fine tune LLM's to improve their performance). The panel and audience were confident that in the majority of cases for LLM powered applications to be more than a curiosity and for them to actually carry out, with sometimes spectacular results, real world tasks at the same or better levels than human professionals, then Zero shot RAG will not be enough. In short for now data scientists and data engineers and not just "prompt engineers" will be needed.

(Check out this paper for the limits of Zero Shot RAG to generalise and perform on new to the LLM data-https://arxiv.org/pdf/2404.04125.pdf)

Both founders (which for full disclosure I invested an angel cheque in) have been working with LLM's well before the launch of Chat GPT 3.5 in November 2022, and we joked on stage that in the past 18 months these founders of young startups have become old-hand stalwarts of the present LLM powered revolution. Look out for announcements soon of new releases and partnerships from both Cosine and Tidal ML in the next few weeks.

We talked about the challenges of building Agenic AI and why most astute observers are forecasting this will be the next big thing in AI. The use of LLM Augmented Agents to solve anything more than short horizon trivial problems is proving a big engineering challenge despite the growth in the size of content windows. One example we discussed was Devin AI from Cognition AI which has proven to be a big improvement on previous agents attempts to replicate software developers but even then on the SWE-Bench benchmark, (a set of commonly used tests to try and assess the coding performance of generative AI tools), Devin at the time of launch was only getting 13.86% correct without assistance. So still a massive gap in capability for new technical solutions to aim at and why Agenic AI is still at the frontier of what is possible.

Despite these challenges, the panel and audience marvelled at the opportunity to build new undreamed of product features that LLM (not just LLM Augmented Agents) offer to this generation of tech startups and the positive impact it is now producing for their early adopting customers. We also talked about the eye-watering compute cost and engineering feat required to deploy foundation models at scale, especially if you need them to complete high frequency tasks, in the wild.

Proceeding this discussion was a presentation by my partner at Twin Path Ventures Katie Lockwood on the present LLM landscape facing startups looking to build applications on the back of LLM's. Katie's talk showed when evaluating and choosing which foundation model to choose from ( and there is increasingly large number of LLM candidates), AI start ups needed to focus on what they need the LLM to do and then engage in a far from straightforward trade-off, primarily between performance, speed and increasingly cost, to get to the best option for their application build and deployment. Based on the show of hands in the room, Open AI models are still the most popular with our London based AI startups, Claude is making headway, whilst Gemini is popular with those who need super large context windows and Groq for those who want speed. Whilst nearly all are experimenting (primarily on cost grounds) with open source alternatives. The rooms expectation is that open source models performance will keep on improving and more foundation models (beyond Alibaba's Qwen and Meta's Llama) from the likes of Open AI will be released with train data sets weights and nodes so AI startups can experiment with fine tuning.

In the networking over breakfast afterwards, AI founders shared their experiences on building applications with LLM's and most reported they were still amazed with their capabilities and yet often unbelievably and frustratingly found that even fine-tuned LLM's frequently make stupid inferences on sometimes simple tasks. The variety and inconsistencies of LLM's means for those startup building with them, they are not yet interchangeable commodities.

In the audience it was great to take questions and hear insights from founders such as Bilal Hafeez from Macro Hive, Guillaume Bouchard from Checkstep, Kate Yanchenko from Lamoon and Jose Palazon from Sagittal and catch up with the likes of Stefan Ciesla-Grain from Ayora, Stephen Quick CFA from Fincrime Dynamics and Ilya Tolchenov from Delphos and Vlad Yanchenko from Machinet - all of which are certainly worth following since all are at different stages of building and releasing fantastic AI solutions to address challenges that are partly arising from the impact and power of recent LLM powered innovations and their uses.

The presentations ended with a simple but amazingly effective Demo by Prof Dan Franks co-founder of York based Causa of their causal ai API that can add causal prediction capabilities to any standard statistical model. Investors in a recent Sifted article announced that CAUSA were one of the top new AI startups to watch. After seeing the demonstration of their platform, my advice to investors would be to act now. Last week we also got the news from Yuxiang (Jimmy) Wu co-founder of Weco AI who did a demo of their platform at a previous AI Breakfast, that their AI Agent AIDE ( check out the results at https://www.weco.ai/blog/technical-report) is now better at generating highly functioning prediction models than the majourity of data scientists who enter Kaggle competitions (in a fraction of the time and at the fraction of the costs to boot).

( Full disclosure Twin Path Ventures invested in both).

As I said to have the honour and privilege to hear from and network with super smart people. The AI Breakfast has only ran in London for 3 consecutive months (shout out to Peter Hayes cofounder of Human Loop, Professor Emine Yilmaz from UCL, Camille Rougié founder of Plural AI and Tim Gordon of Best Practice AI who were previous speakers) but we hope it will and already has become a regular get-together for all those building AI startups in the UK.

There are many who insist that the present wave of AI is all hype, or that it is just too noisy, that it is better to hang back and see how it plays out, or that the deep pockets of big tech and the data/ client relationships/ brand reputation of existing incumbents will win out. That super talented founders of AI first startups can be ignored. That all the action is in SF, or that it is just too, too much. ( Please can we have our rules based B2B SaaS dominated Tech startup scene back is an often heard lament).

This is my view is just plain wrong.

AI ( and not just LLM's) is now here. Even if you ignore the hype and go for a more modest forecast of impact, the latest developments in AI will bring change in expected and unexpected ways. Hardly any operation or best practice method will remain the same, forcing all businesses, big or small, incumbent or brand new, to meet the challenge of re-imagining and rebuilding anew. Startups, especially startups in London and the UK, are more than up for the challenge. The UK might not have the deep pocketed USA VC's or Big Tech giants to fuel the emergence of tidal wave of AI first solutions on to the world ( although with the news that Microsoft is setting up, post digesting Inflection AI, in London, all the big AI research labs are now here) or the wall of state funding that bigger and better functioning economies in the Middle East, Far East and even France are deploying to maintain relevance in the new AI world. But we do have the technical and entrepreneurial talent, native and imported, and they do have the ambition and they are busy building.

If you don't believe me then check out some of the startups name checked in this blog. If you are a founder come to our next AI Breakfast event (https://www.eventbrite.co.uk/e/878896120957?aff=oddtdtcreator) on the 8th of May at Rise, created by Barclays ) to meet and share experiences with your fellow founders of AI first startups.

If you are a qualified individual investor looking for an opportunity to gain from this once in a lifetime wave of game changing innovations that AI first startups are bringing forth, then get in touch (our Twin Path Ventures S/EIS fund closes to new investors this month).

I began this blog by quoting ( with more than a hint of pretentiousness) Wordsworth. To end, I will quote another English poet in attempt to get across the thrill, the intellectual thrill of having the opportunity to talk to, learn from, when possible aid and support and even ( albeit with prudence and a hard headed economic motivation) invest in a few of the best of this generation of AI Startups that the UK is so lucky to produce and attract. The lines are from "Mad, Bad and Dangerous " to know Lord Byron. Somehow, when you are in London, the epicentre for AI startups in Europe, lucky to be surrounded by a crowd of ambitious, super smart and talented people, they seem appropriate ( if yet again super pretentious).

"To mingle with [with these people, in this) universe, and feel what I can never express but can not all conceal"

Come on, drop your world weary skepticism and get involved.