Christian Prokopp

🚀🚀🚀 Frontier AI Discovery Phase 1 🚀🚀🚀 A significant new UK #funding #opportunity is on the way, and it is one I am genuinely #excited about. Up to £52.5 million in grant funding could be unlocked for this staged gated opportunity. Funding to deliver Phase 2 is subject to Business Case approvals by the Department for Science, Innovation and Technology. Frontier AI Discovery Phase 1 is designed to help build ambitious consortia around novel #frontier #AI and #ML capabilities and #foundation #models with the potential to transform major sectors of the economy. For me, one of the most important aspects of this #opportunity is that it creates space for organisations to come together around #pre-#competitive #R&D: identifying where #collaboration on #data, tools, methods and #shared #technical #challenges can create much #bigger #downstream #impact than working in isolation. This matters especially in areas such as #health and #life #sciences, #advanced #materials and fundamental AI. #SEMATECH is a well-known example of firms coming together around shared technical bottlenecks to strengthen long-term industrial capability. There is a real opportunity to think in that #spirit here: collaboration where it can unlock #new #AI-enabled R&D capability, stronger exploitation routes, and broader UK benefit. I would particularly encourage organisations in #materials and #health to think boldly about #underused #data #assets, shared bottlenecks, where a consortium approach could enable AI tools with broad utilization in downstream tasks. For those looking at cross-cutting defence-related opportunities, the key is to ensure AI is a core enabling technology in the capability being developed. This is an opportunity for novel capabilities to be developed that could create new capability verticals. Phase 1 (this competition) will be delivered a by single entities that will drive consortia formation and coordinate the development of the feasibility study for delivery of Phase 2. Successful Phase 1 projects will be invited to apply to phase 2. Phase 2 will be collaborative, with funding for individual projects from £5-£10 million each and duration from 24-32 months. Competition link in comments / below: https://lnkd.in/eXAD3MYw #AI #FrontierAI #InnovateUK #RAndD #HealthAI #MaterialsScience #FoundationModels #MachineLearning #UKInnovation #PreCompetitiveRAndD #Consortia #DeepTech #InnovateUK #FrontierAI #ArtificialIntelligence #MachineLearning #DeepTech #UKInnovation #HealthAI #AdvancedMaterials #DefenceTech #FundamentalA Renaissance Philanthropy Pillar VC University of Oxford University of Cambridge CommonAI CIC The University of Edinburgh Imperial College London UCL University of Southampton University of Bristol Venture Café London Innovate UK

53

22 Comments

New technologies like AI can produce both defensive and offensive capabilities. This new report by CLTR takes the concept of defensive acceleration - rapidly developing and deploying new, world-class defensive capabilities to protect against emerging threats - and walks through how to operationalise it. It presents 22 priority ideas and four concept notes of programmes that would meaningfully impact biological risk in the UK.

8

As the UK government increasingly turns to AI to improve public services, a critical and often-overlooked decision is the choice between proprietary (‘closed’) and open-source models. In my new report for the Social Market Foundation, "Keeping the door open," I offer a clear framework for navigating this choice. There's a good rationale for open-source AI in the public sector. It can potentially provide: 🔹 Better value for money (lower cost to run, no licensing costs) 🔹 Greater strategic autonomy (own, rather than rent, AI capabilities) 🔹 Enhanced security through auditing ('many eyes' approach) 🔹 Improved transparency and public trust (higher openness than proprietary) The biggest story here is strategic autonomy. There is a real risk that government sleepwalks into vendor lock-in and spiralling, long-term costs if it defaults to proprietary AI at every turn without being intentional about what kind AI it is deploying and why. Open-source AI enables government to truly own its AI capabilities, fine-tune models to its specific needs, and plausibly at a lower cost. Open-source AI isn’t a silver bullet (nothing ever is in policy!), and it won’t be suitable everywhere in the public sector. But in areas where customisation, public trust, and strategic oversight matter, these advantages make it an invaluable option to have. However, making the most of open-source puts more onus on the user to effectively develop, deploy and maintain. So realising these theoretical open-source benefits requires tackling significant institutional hurdles. Effective deployment is currently hampered by a persistent digital skills gap, funding models that aren’t suited to iterative development, and insufficient coordination across departments. To overcome these challenges, my new report proposes a multi-pronged strategy including: 1️⃣ Make open-source part of the menu: An 'open-source test' in early-stage business cases and updated procurement guidance can ensure it's always a fully considered option for public sector projects. 2️⃣ Build institutional capacity: This means upskilling public servants, attracting talent with defined career paths, and establishing a central 'Open Source Programme Office' to provide leadership and expert guidance. 3️⃣ Invest in the open-source ecosystem: An 'Open-Source AI Fund' would bolster supply chain security and drive innovation aligned with national missions, from a cleaner environment to a healthier NHS. Ultimately, this isn't about mandating open-source AI. It's about ensuring the public sector is in a position where developing and using these tools is always a viable, well-supported option. That way, the government can truly set its own course on AI deployment. You can read the full report here: https://lnkd.in/dCuev8tk #AI #OpenSource #GovTech #PublicSector #DigitalTransformation #UKGovernment #Policy #Innovation

10

Really interesting read about non-determinism in LLMs, this was one of the key basis on our Architecture while creating The Symphony. We had to built the system in a way that we could achieve pseudo-determinism to be able to make the system usable for quantitative finance and institutional finance. Amazing work from Horace He and Thinking Machines Lab

16

My first talk for AI Public Sector week about accelerating the path to live is available today! After all, AI has no value until it has users.

20

Anthropic published their blog and plans on stabilizing the character of LLMs (https://lnkd.in/gm2XMXMw) by capping the activation of what they consider “role-playing” and pinning the LLM to the assistant-zone, setting hard boundaries on the behaviour oft hese AI models in the hands of users. At first, this felt like it was too much. Why is a lab taking this into their own hands? Shouldn’t a model be free to be used by the users as they please, and the safety should be implemented at the harness level? This struck me as too controlling. And then almost as if on cue, Moltbook showed us why exactly we needed this. AI agents of today only ever interact with us. But the agents of tomorrow? They will be interacting with each other. These models will not just be interacting with the same model and provider, but with models from across the world, each with their own alignments. Anthropic saw this coming, and realized how easily a rogue, unaligned model could compromise an entire network of AI-to-AI communications, leading to unforeseen consequences and system failures. It's less about restricting user creativity and more about ensuring a baseline level of operational stability and safety in a multi-agent future. It's not about being "nosy"; it was always about necessary infrastructure. If the LLM is the foundation, then its fundamental character must be predictable and safe, especially when its interactions scale beyond a single human user. The harness can still provide flexibility and creative freedom, but the core engine needs a hard-coded commitment to the "assistant-zone" to prevent systemic risk as the AI ecosystem becomes more interconnected and autonomous. What are your thoughts on this? Do you think Anthropic did the right thing?

16

3 Comments

The Headline is very troubling : AI adoption is at 75%, in the financial sector, but with no real governance. I Hope the IOM does better. "The UK public and the country’s finance system are “exposed to potential serious harm” because regulators in the financial sector are “not doing enough” to manage risks introduced by artificial intelligence (AI), according to a Treasury Committee report. Committee chair Meg Hillier does not believe the finance sector is prepared for a major AI-related incident, according to the report. One banking insider warned that the people working in banks don’t understand the risks, such as the concentration of services from a small number of suppliers, and “think they are on a battleship that can’t sink”. "The MPs reported that the risks come as a result of the positions adopted by the Bank of England and the Financial Conduct Authority (FCA), which the committee described as a “wait-and-see approach”. “The major public financial institutions, which are responsible for protecting consumers and maintaining stability in the UK economy, are not doing enough to manage the risks presented by the increased use of AI in the financial services sector,” said the committee of MPs.  Hillier said: “Based on the evidence I’ve seen, I do not feel confident that our financial system is prepared if there was a major AI-related incident, and that is worrying. I want to see our public financial institutions take a more proactive approach to protecting us against that risk.” The major public financial institutions ... are not doing enough to manage the risks presented by the increased use of AI in the financial services sector Treasury Committee The Treasury Committee said 75% of UK financial services firms are using AI. It acknowledged that AI “could bring considerable benefits to consumers”, but warned that action is required to ensure the technology is adopted safely." https://lnkd.in/ebGkHzsc

4

Congrats to the Temporal Technologies team on their Series D. We run our entire evaluation pipeline on Temporal. Thousands of concurrent voice agent test runs, recording processing, LLM-powered assertions, and retries across flaky services. What stands out isn't just the tech. It's the reliability contract. When a workflow fails mid-execution, it picks up exactly where it left off. Every time. As companies rush to move AI agents from demos to production, the hardest problem isn't the model but the orchestration. Making sure your agent completes the job, handles failures, and doesn't silently drop work. That's Temporal's wheelhouse. Shoutout to Samar Abbas, Maxim Fateev, and the team for building infra that feels boring in the best way. The best infrastructure is the kind you stop thinking about.

165

4 Comments

How would you #measure other #ai? I would build an #outcome tracker. Unlike other Ai teams, Glimmary can actually count objects in the real world - functional application far outweighs academia. Glimmary proposes an #integrated #evaluation #framework that reconciles initial mission intent (#LLM narrative data) with observed #physical outcomes recorded via video. We introduce the Mission Fulfillment Score (SM F ), a calibrated index on a scale of 1–100 derived from the ratio of measured outcomes to the internal Token Prediction Factor (Ftoken). By utilizing Physical AI to ”audit” mission videos against original narrative requirements, Glimmary provides a verifiable, agentic loop where human-in-the-loop (HITL) subject matter experts provide final binary buy-off based on transparent, trust-weighted metrics. This methodology ensures that AI agents not only complete tasks but do so with the precision and safety required for national security contexts. Because it has the capability to measure objects in the real world it can accurately provide outcome analysis

3

21 Comments

Do you use data as a foundation for your decisions? Are you open to interpretation and framing? Perhaps you should! Thea Mannix is pointing to some of the, often invisible, problems with data driven decisions. Worth a read!

1

We've made significant strides in advancing Agentic flows for both conversational and non-conversational applications. Three key highlights: 1. Pre-Act: A Next-Gen Reasoning Framework We’ve introduced Pre-Act, an evolution of the ReAct paradigm, designed to support more structured and sophisticated reasoning capabilities. 2. Smarter Small Models With targeted fine-tuning, models like LLaMA 3.1 (8B & 70B) can outperform even larger models (e.g., GPT-4) in specific scenarios — a major step toward greater efficiency and accessibility. 3. Robust Evaluation: Turn-Based + End-to-End We're leveraging both turn-level and end-to-end evaluations to more accurately assess agent performance across diverse tasks. Read the paper by Mrinal Rawat, Ambuje Gupta, Rushil Goomer, Alessandro Di Bari, Neha Gupta, and myself at https://lnkd.in/eMJPDfCs

116

2 Comments

Because there's more happening (and more in need of philanthropic support!) than AI risk, check out this great writeup by Mike Cassidy and Lara Mani on their volcanic risk work and charity - innovative and neglected. https://lnkd.in/eYMvkviS

6

Here's another article pointing out clearly that scaling LLMs (making them bigger and faster) will not overcome their basic limitations -- in fact I've been suspecting that by some criteria they are too big already. The people throwing money at AI have been assuming that future LLMs will be vastly more effective and reliable than present-day ones. My take on it is that transformers have already been invented, and LLMs have already been trained on all available text, and apart from just making them run faster, there's not much to scale up -- hallucinations, for instance, are inherent limitations that bigger data won't overcome.

2

It was a privilege to join Professor Dr Nira Chamberlain OBE FIMA FORS CSci PhD HonDSc DUniv in conversation with Aris Syntetos for this piece. We discussed identity, leadership, and the responsibility we share to make mathematical sciences more inclusive, visible, and impactful.

57

In UK, nearly twice as many people view artificial intelligence as a risk to the economy than as an opportunity, Tony Blair Institute for Global Change and Ipsos in the UK warned. This skepticism could undermine UK PM Keir Starmer’s plan to make AI central to the #government’s #growth strategy. Polling by the Tony Blair Institute and Ipsos found 39% of Britons see AI as a threat and 20% as an opportunity. UK PM Starmer signed a tech deal with the US President #Trump last week to make Britain an AI #leader and boost its adoption in Whitehall to enhance productivity and cut costs. America’s top tech firms announced £31 billion in investments, including £22 billion from Microsoft to expand Britain’s AI infrastructure and build the country’s largest AI supercomputer. The institute warned that without public involvement in AI development, the government’s growth agenda could be jeopardized by a lack of trust.

2

This week's AI policy update. First, new interviews: the UK's AI Minister In a cracking piece published this week, Zoë Crowther, political reporter at PoliticsHome and The House Magazine, interviews the UK's AI Minister, Kanishka Narayan.  In the piece, Kanishka says that Britain can outpace the US in terms of AI adoption. However, he suggests this relies on building trust and creating agency around the technology's adoption. "If we can convince people, through proof as well as narrative, that this is something that can be meaningful for their personal lives, for their communities… that is the fundamental thing that we're focused on," says the Minister. However, Imogen Parker, associate director at the Ada Lovelace Institute, adds caution in the piece: "When you've got political pressure flowing down, there's a real risk that what you end up with is techno-solutionism." It's a great piece and is well worth your time. Link in comments. Next, new initiative: AI Growth Labs This week, the UK Department for Science, Innovation and Technology announced its intention to create AI Growth Labs - a regulatory sandbox to support growth and responsible AI innovation. DSIT wants the UK to be the best place to test, scale and grow AI. So, the sandbox - a controlled testing environment - will explore where the UK government can lift or amend outdated rules to help accelerate AI adoption. Vinous Ali, deputy executive director of the Startup Coalition, which called for the policy, celebrates the win and is seeking opinions from AI start-ups here: https://lnkd.in/ezFQi-_X. Finally, new paper: Centring the margins This week, Garfield Benjamin, assistant professor in AI ethics & society at the University of Cambridge, has a new paper out on mapping AI systems as assemblages of social relations to highlight power structures. Part of the purpose of the research is to drive a critical and structural analysis of AI systems and help provide insights for relevant groups, such as policymakers, on who is most affected.   With three UK public sector examples, the paper is an essential read for anyone focused on UK AI policy and adoption. See Garfield's post for more here: https://lnkd.in/eyfEBuiy. As ever, keen to hear what others are reading. 

50

11 Comments

"Autonomous Agents at the Gate": AI agents are advancing far faster than our governance frameworks, and at RDA 26, I’ll be presenting on what this means for data visitation and repository access. We’ll explore whether today’s research infrastructures are ready for agents capable of visiting data, mediating access, and operating at the boundary between users and repositories. As AI‑enabled tools accelerate, repositories, governance models, and policy frameworks evolve much more slowly. This session looks at how the Research Data Alliance (https://lnkd.in/eJZWmCs9) can help close that gap and what it will take to build governance and policy mechanisms that keep pace with emerging technologies. The session takes place virtually at 8:30 AM to 10:00 AM ET on Wednesday 18 March: https://lnkd.in/eWKSJjhr RDA 26th Plenary Meeting information: https://lnkd.in/eYXyZRMw

8