Axel Nilsson

🔢 Historic Mathematics Meets Modern Computer Science Imperial College London is undertaking an extraordinary project that bridges centuries of mathematical legacy with cutting-edge formal verification technology. Formalizing the complete proof of Fermat's Last Theorem in the Lean theorem prover. For those unfamiliar, Fermat's Last Theorem states that no three positive integers a, b, and c satisfy aⁿ + bⁿ = cⁿ for any integer value of n greater than 2. This seemingly simple statement puzzled mathematicians for over 350 years until Andrew Wiles finally proved it in 1995. Led by Kevin Buzzard, this initiative represents the future of mathematical research where human insight meets computational rigor. It's not just about proving the theorem again—it's about creating machine-checkable mathematics that eliminates any possibility of error. The implications extend far beyond pure mathematics, touching fields like cryptography, software verification, and AI safety where mathematical certainty is paramount. As we increasingly rely on mathematical foundations for everything from blockchain to autonomous systems, having computer-verified proofs becomes not just academically interesting, but practically essential. Check out the project: https://lnkd.in/dsHwqZ_v What are your thoughts on the role of formal verification in mathematics and computer science? #Mathematics #FormalVerification #Lean #OpenSource #ImperialCollege #Research #ComputerScience #TheoremProving

4

Thanks to the NDRC for letting me tell my little story for the NDCR pre accelerator group! Building product without a base engineering background, although still a huge grind, is at least possible for non-technical people now. Hear me out engineers... Experienced engineers embrace Ai in coding now but at the same time are scared shitless of some "vibe coding nutcase" on a coding rampage, creating webs that they'll never be able to untangle... What I found is you need to find a middle ground between using Ai to build and learning WHAT you are actually doing. You should spend most of your time understanding the infrastructure & what the Ai has helped you build and how. You should treat like a degree. Without doing that, soon enough, you will become useless on the tech side and a massive pain to the product / engineering team. But if you close the gap of knowledge between yourself and the engineering team, you can stay a MASSIVE asset to them and the company long term, instead. Things like the integrity / security / testing of the platform become your main concern over adding a new feature, which is where we need to be. You are helpful in meetings as you understand the underlying tech rather than shouting "HOW LONG WILL IT TAKE". You need to evolve from insane speed during the MVP to alignment with the engineering team on what long term viability of the tech looks like, and that means dramatically slowing down and concentrating on other "non sexy" aspects like security. All it takes is asking how and why and what constantly, and spending time understanding it, while you are building. Bit by bit you will get there without sacrificing any time building your MVP. The goal should be to become a junior engineer of sorts and be able to let Ai assist you doing things that you fully/mostly understand. This is how I think we commercial people should be looking at using Ai! It was actually really fun to discuss with Deirbhile Gorman and I hope I added a little value to all the vibe coding (hopefully not) nutcases of the future.

71

14 Comments

Very proud to have worked on developing the first open source modelling suite for Marketing Mix Models (MMM), which we at Mutinex hope will lay the groundwork for a more democratic and transparent MMM landscape. The new validation suite enables practitioners to rigorously test and compare MMM models in a consistent, open framework. It's a step toward greater accountability, better standards, and shared progress in marketing science. We’re excited to share it with the community - contributions and feedback are more than welcome! https://lnkd.in/dmWgKUd2 https://lnkd.in/dWJPtr2m

10

2 Comments

Very excited to talk about the latest developments in Autonomous AI Agents last week at ChangeVision’s innovation session hosted at Ecocharting - ESG Platform. It’s exciting to see how quickly the conversation is shifting, from experimenting with AI, to actively identifying where AI Agents can be embedded into real business processes. This raises important questions around governance, risk, and accountability. We’re clearly moving beyond the hype into real-world impact and applications! What stood out most was the level of engagement and the depth of the discussions, with hands-on demos of AgentZero, OpenClaw, and various CLI tools such as Gemini, Codex, and Claude Code. Thanks for the inspiring evening and great conversations Jan Ruijgrok, Arnoud van der Reest, Fokke Wijnstra, Michiel de Boer, Karel van der Burgt, Leo Monhemius. Looking forward to continuing the dialogue!

21

I'm asked quite often about why I opted for Advanced Computing instead of a more specialised program like Artificial Intelligence or Software Engineering for my postgraduate studies at King's College London. I believe this is the age of generalism as opposed to specialisation. If you look at any tech-related job description, you would see that the overlap between separate positions' responsibilities is growing more and more. For instance, a data scientist is expected to know ML, NLP and be proficient in working with LLMs, which was not the case a few years back. This is why the modules I chose for both of my terms were mostly AI-related, but also included Security Engineering, Software Engineering for Financial Systems and Big Data Technologies. That way, I could expand my skillset through university modules, while working on deepening the ones I enjoy more through work experience and side projects. Do you agree with that approach? What would you have done differently? 👇

31

Transitioning from engineering to product during the rise of LLMs has been a pretty interesting experience - I wrote some reflections and predictions for what's coming! ---- “𝐈𝐭'𝐬 𝐨𝐧𝐥𝐲 𝐰𝐡𝐞𝐧 𝐭𝐡𝐞 𝐭𝐢𝐝𝐞 𝐠𝐨𝐞𝐬 𝐨𝐮𝐭 𝐭𝐡𝐚𝐭 𝐲𝐨𝐮 𝐝𝐢𝐬𝐜𝐨𝐯𝐞𝐫 𝐰𝐡𝐨'𝐬 𝐛𝐞𝐞𝐧 𝐬𝐰𝐢𝐦𝐦𝐢𝐧𝐠 𝐧𝐚𝐤𝐞𝐝” - 𝐖𝐚𝐫𝐫𝐞𝐧 𝐁𝐮𝐟𝐟𝐞𝐭𝐭. As the tide pulls out on much of coding and knowledge work, what actually matters when building products? If 2026 pans out the way we're all predicting, I suspect a lot of engineers (and product managers!) will find out they've been swimming naked. Last year, I just so happened to be one of them.

68

4 Comments

The shift from metric-based to narrative CVs is fundamentally changing how research is funded. But asking scientists to suddenly become storytellers introduces a massive friction point. For funders like Research Ireland and UK Research and Innovation(UKRI), the days of relying solely on an h-index are fading. They want to see real-world impact and collaboration. But translating years of highly technical research into a compelling narrative format takes a significant amount of time. To help solve this, I partnered with a brilliant engineer to independently build and launch DORA CV Builder (www.doracv.com). It’s a free, AI-powered tool that uses the STARS framework to guide researchers in extracting the actual story and impact behind their work. When defining the product requirements, we recognized a major adoption barrier: intellectual property and data privacy. Researchers cannot risk their pre-published data floating around the cloud. So, we made a strict product decision regarding data handling: your content is processed securely in real-time, and absolutely no data is stored or retained on our servers. If you are a researcher navigating these new funding requirements, give it a try. It's completely free, and I’d love to hear your feedback on the UX and how it fits into your workflow. #NarrativeCV #AcademicChatter #ResearchAssessment #ProductManagement #BuildInPublic

38

In the last few weeks, I've seen a surge in full-time onsite roles in London, particularly across early-stage AI companies building hardcore tech. And despite the common narrative, candidates are showing surprising interest. Here’s a snapshot from 3 active clients: 🔹 Client 1: Seed-stage AI Startup * Applied AI Engineer (£180k base) – Python + LLMs * Full Stack Engineer (£150k base) – Backend-heavy, Python/Postgres/React 🔹 Client 2: Seed-stage AI Startup (𝗩𝗶𝘀𝗮 𝘀𝗽𝗼𝗻𝘀𝗼𝗿𝘀𝗵𝗶𝗽 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲) * Senior Compiler Engineer (£180k base) – LLVM + AI/ML * Performance Modelling Engineer (£140k base) – GPU/TPU work 🔹 Client 3: Stealth AI Startup (𝗩𝗶𝘀𝗮 𝘀𝗽𝗼𝗻𝘀𝗼𝗿𝘀𝗵𝗶𝗽 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲) * Full Stack Engineer (£130k base + equity) – Kotlin/React Every single one is onsite in London. None are compromising on quality or compensation.

9

1 Comment

"Anthropic, OpenAI and Cohere are recruiting for so-called forward-deployed engineers, a new job for generative AI companies, as part of a push to generate more revenues by installing specialists within businesses to help them customise their AI models." I am glad to see they are discovering that working with Enterprises actually requires boots on the ground. Palintir invented the "FDE" title because they charged way more than industry rates. You weren't paying for a solution engineer, you were paying for a FDE. In any case, FDE appears to be a growth segment in our market. https://lnkd.in/d7uE7b56

976

104 Comments

Today we launched Liminary to open beta. When you’re researching a new topic, how much time do you spend on “paperwork?” Hunting for that smart blog post you saw last month, cross-checking sources for areas of consensus and dispute, making sure your quotes are accurate, etc. With Liminary, spend less time on paperwork and more time drawing conclusions. - Save sources and notes. - Ask questions, get summaries, compare perspectives. - Recall items from your library as you browse for new sources or draft a report. Try it out at liminary.io. Let me know what you think. (If you’re on mobile, use a standalone browser app to sign up.)

13

2 Comments

2026: The Year Tabular AI Overtakes Traditional ML I recently sat down with Ben Jackson on the (AI) People Podcast to talk about why organisations should focus on Predictive AI instead of Generative AI. The reason: Generative AI struggles with tables, yet that's where enterprise value actually lives. Customer data, transactions, operational records stored for decades in tables. During the conversation, we covered: • why LLMs struggle with tabular data • why traditional ML hasn't evolved in 10 years • and how Tabular Foundation Models are changing the game, delivering better predictions with faster deployment. 🎧 My prediction to Ben: 2026 is the year Tabular Foundation Models reach enterprise-scale adoption. Listen to the full episode here: https://lnkd.in/ep8XTAxA #MachineLearning #DataScience #PredictiveAI

23

1 Comment

"AI is finally showing up in the productivity numbers." I wrote recently about the UK productivity narrative and felt unconvinced. On the ground, most non-tech organisations I meet are still doing painfully manual work, or catching up on the digital basics. Then I read a new Financial Times piece by Erik Brynjolfsson arguing the AI productivity take-off is now visible in US data [1]. The core idea is simple: output stayed strong while measured labour input got revised down, which mechanically looks like productivity growth. Here’s the part that actually reconciles my scepticism with the hype. In the last month or so, there really has been a step change in what AI can do when the work has cheap, automated verification. Software is the obvious example. If the output can be tested, compiled, linted, benchmarked, or checked against a spec, you can move from "prompting" to delegation with guardrails. That is not just faster typing. It is compressing whole work loops. And you do not need everyone doing it for the macro stats to twitch. If a small cohort of "power users" start automating end-to-end workstreams, it can plausibly show up in the aggregates while the median firm still feels stuck in spreadsheet land. So I think the question is less "is AI boosting productivity" and more: Are we seeing broad adoption, or concentrated adoption? If it is concentrated, the next bottleneck is not the models. It is process redesign, data plumbing, and getting the average employee to the point where AI is a daily tool, not a toy. [1] https://lnkd.in/ehRBFyWb #Productivity #AI #GenAI #DigitalTransformation #UK

6

1 Comment

Kaggle and fast.ai founder Jeremy Howard has spent the past year or two working on a process, and supporting platform, that addresses some of the challenges currently facing LLMs and agents. Firstly, things that they get wrong, or information they are given that isn’t relevant, can stay in the models context and continue to confuse them. Where SolveIt differs from other AI enabled Jupyter style systems, is that the output of an LLM is curated to remove wrong or confusing information. The second is really a problem more for us – as models can do more things, how do we build skills so that we can amplify the capabilities of more advanced LLMs. A process more than software, SolveIt also focuses on learning as things are built, so you continue to understand ever more complex work. So far SolveIt has been developed in collaboration with a initial cohort of people using it, and now only the second cohort are getting access. This is in the form of a paid course (with prices in USD!), but gives access to an LLM enabled platform, so you can’t expect it to be free. While the main focus of the course is on writing code, especially Python, SolveIt is more broadly applicable; for example Lean Startup author Eric Ries is also involved discussing how he is using the platform to build startups. While some basic programming skills are needed, there seem to be a wide range of people enrolling, from those wanting to improve their programming through to university professors in the field. Co-founder Johno Whitaker and Jeremy have written more, and also recorded a podcast that goes into more detail in a way that should clarify how this all works. https://lnkd.in/gJzBtMYT

9

1 Comment

I'm the CEO of Neural Voice, based just outside Leeds. And fundraising has been brutal. Want to know why? Yorkshire receives 0.2% of UK AI investment. We represent 8% of the UK's population. 8% of GDP. But we get 0.2% of AI funding. That's a 40:1 disparity. Whilst I've been grinding through months of pitch meetings for Neural Voice, London startups with half our traction are closing rounds in weeks. The North West saw +36% investment growth. Scotland +35%. Yorkshire? -2%. Here's what makes this absurd: • We have 1.2 million patient records in the Yorkshire & Humber Care Record • Leeds Teaching Hospitals runs a £140m AI Lab with 340+ partnerships • Our costs are 30-40% lower than London • We produce 34,000 graduates annually from 20 universities • We've already created success stories like Tracsis (£68m revenue) and Avacta (£100m+ raised) The infrastructure is here. The talent is here. The data is here. What's missing? Capital. And it's not just Neural Voice. It's 150-200 potential AI companies that will never exist because Yorkshire founders can't access the same funding ecosystem as our London counterparts. I've lost count of the number of times I've heard "interesting company, wrong postcode." This isn't about handouts. It's about fixing a market failure that's costing the UK economic growth, high-skill jobs, and global competitiveness. The solution exists: a £260m Yorkshire AI Growth Programme over 5 years. Phased. Gated. With proven 2.5:1 benefit-cost ratio. But right now, it's just a PDF gathering dust whilst brilliant founders waste months that should be spent building, instead pitching to investors who won't look north of Watford. If you're a Yorkshire founder who's felt this, comment below. If you're an investor who thinks regional balance matters, I'd love to talk. And if you're in government reading this: we're ready. The assets are here. We just need the capital mechanisms everyone else already has. Mark Sewards, lets chat!

30

2 Comments

Dear UK Visas & Immigration, polite rant if I may [alt: building a research-heavy AI start-up is hard, don't make it harder please] Semilattice is a software startup developing complex AI technology. We're trying to enable our globally mobile engineers to work from the UK because we have made a deliberate choice to be based in this beautiful, if (at least compared to other places our engineers could work from) somewhat rainy and high-tax country. We have a visa sponsorship licence and want to employ a recent Oxbridge graduate with a Master in advanced computer science who is already working in the UK for a global tier 1 tech company. You would think it shouldn't be hard to justify a quick decision. Instead, we have re-answer the questions we already answered, file 6 months of 'everything', while your boilerplate email template has a habit of calling our future employees 'migrants.' The decision of the US to make highly skilled immigration harder should be a natural opportunity for an advanced, education-based economy like the UK. To quote a recent speech by our prime minister: "[...] Britain must compete for the best talent in the world in science, in technology, in healthcare. You cannot simply pull up a drawbridge, let nobody in, and think that is an economy that would work." Some polite suggestions for improvement: 1. Consistently apply your policies instead of slowing things down by needless paperwork and hoop-jumping. We'll jump through the hoops, but would rather focus on building our company, make some money (and pay some taxes in the UK). Also wondering if your "culture of customer satisfaction" could do with some improvements? 2. Recognise the value of skilled graduates as assets. We have to fight hard to convince every software engineering candidate to work with us. Global AI talent has options. 3. Treat the people that want to live and work on this island with respect. The highly trained and globally mobile engineers we want to employ should not be called 'migrants'. No-one should be, frankly. Why don't you call them "your future employees". Let’s please support UK companies and nurture the next generation of innovators. 

37

8 Comments

🚨 Applied AI Needs a Feedback Layer 🚨 In the rush to integrate AI into our workflows, we’re often left refining outputs manually—editing stiff emails, tweaking awkward phrasing, and making countless micro-corrections. This invisible labor highlights a missing piece in our AI systems: a structured feedback loop. In my latest piece on ON_Discourse, I delve into the necessity of building a feedback layer into AI workflows and products. It’s not just about smarter models; it’s about smarter systems that learn and evolve through human interaction. Discover how incorporating reinforcement learning from human feedback (RLHF) can transform static AI agents into dynamic collaborators. 🔗 Read more: https://lnkd.in/edvirNk4

10

2 Comments

This is incredible 👏 : "Students who received LearnLM (supervised) tutoring were 5.5 percentage points *more* likely to solve new problems on subsequent topics" There is a lot of despondency and worry about AI's impact on education. Early data is showing something clearly: it's just another tool that augments our abilities as a species, when used correctly: and that is key! Use a supercar without knowing how to operate it, you'll crash. Similarly, use AI without guardrails, results will be awful. That's why the onus is on educational institutions to adopt the right tools and augment student abilities and make short-cuts harder.

8

Applied AI ≠ “prompt + output” My co-founder Sam Broe just nailed a blind spot in today’s AI wave, we’re shipping agents without a feedback layer. In his new piece on ON_Discourse he explains why the next big acronym isn’t another model—it’s a business layer that captures human edits, scores them, and feeds them straight back to the agent (think RLHF at product-level scale). At Free Energy we’re already building these loops for our clients so their AI teammates get smarter every interaction—no more copy-paste-tweak-repeat. Read Sam’s article: https://lnkd.in/eWRwMVCM Curious how a feedback-driven AI stack could 10× your workflows? DM me or visit freeenergy.studio—happy to chat. #AI #AgenticAI #RLHF #ProductGrowth #WorkflowAutomation #FeedbackLoops

4

1 Comment

The best product insight I've read this week came from Catherine Wu, Head of Product for Claude Code at Anthropic. She wrote about what it's actually like to build products when the technology underneath them is improving faster than any roadmap can keep up with. The number that stopped me: the AI behind Claude Code went from handling ~21-minute tasks to completing 12-hour software engineering tasks about half the time. In 16 months. That's roughly a 41x jump. She shares four shifts that I found genuinely useful— not just for PMs, but for anyone building anything right now: 1. Short experiments beat long roadmaps Some of Claude Code's biggest features came from self-directed afternoon explorations. Not OKRs. Not quarterly planning. So don't shy away from running short experiments. 2. Prototype first, document later When a wrong bet costs an afternoon instead of a quarter, the calculus on experimentation completely changes. You stop asking "what if this doesn't work?" and start asking "what if it does?" 3. Revisit old features with every new model Anthropic cut 20% of their own system prompt just by removing workarounds built for older models. What you optimized around last year might be unnecessary overhead today. 4. Do the simple thing that works Over-engineering for model limitations is a trap and a fast way to create tech debt. The next model will make your clever hack obsolete anyway. Her closing thought is what I'd bookmark: track both things at once— how AI is changing the way you work, and how it's changing what's possible in your product. That's the dual awareness that separates people who adapt from people who catch up. Full article in the comments— well worth 5 minutes. #ProductManagement #AI #Leadership #FutureOfWork #TechStrategy

13

5 Comments