Ben Dias

"I really do believe in the benefit of long-term curiosity-driven research. But… addressing government priorities and helping companies to grow and importantly to scale are super important and we don't do enough on that."

36

15 Comments

This time of year I catch up on research papers I’ve been meaning to read — and since we’re developing a context-engineering benchmark at STAC Research, I’ve been digging into multi-turn AI. One paper that stood out comes from the NatWest AI Research team: Evaluating the sensitivity of LLMs to prior context. They systematically vary both the amount and type of prior context, then test multiple frontier and mid-sized models across these setups. Some of the things I learned: 1. Prior context influences outputs —Unrelated earlier conversation turns can hurt model performance, showing how sensitive LLMs are to what came before. 2. Interaction length matters — Longer multi-turn interactions accelerate performance drops, although most of the decline happens within the first few thousand tokens. 3. Bigger isn’t always better — Increasing model size doesn’t guarantee robustness across multi-turns; smaller models can outperform larger ones under complex context. 4. Task placement counts — Where you position the task description in the context sequence can significantly change performance, highlighting the importance of input structure. 5. Domain alignment is critical — Accuracy suffers when prior QA context comes from a different knowledge domain than the target query. What I like most is that this work highlights the limitations of single-turn benchmarks — they can be misleading if we assume they generalize to real-world, sustained interactions. Link to the paper in the comments.  cc: Greig Cowan

36

3 Comments

This was one of my first projects after joining GeoDS, and my first paper in the world of Geography/Geographic Data Science! We needed census data at small area geographies across the whole of the UK for another project (which will follow soonish...). Putting it together from the different national sources into a unified dataset, and validating that we were combining like-for-like variables, turned out to be pretty time consuming. So we've published it as an open dataset. If you have a project which requires census data across UK national boundaries, this will hopefully save you some time.

25

One of the aims of the Innovation Launchpad Network+ Researchers in Residence scheme is to allow university researchers to work with the Catapult Network to take their research and progress it closer to application in industry. The first session this afternoon at #techfellows2025 includes presentations from 4 of our RiR projects at the AMRC on their projects. First up are Eloïse Eimer and Ellis Taylor who are talking about the Repair of 3D printed components.

30

We were delighted to host the Careers in Data Science and AI within Academia and Government session with The Alan Turing Institute on Wednesday.   We were joined by the brilliant Emma Li, Data Science Careers Facilitator at the Turing, who delivered a fascinating talk introducing undergraduate students at the University of Essex to the breadth of roles available in Data Science and AI across government, research, defence and security.   Emma explored: What Data Science really is, and common myths about working in the field Why diversity is essential within Data Science and AI. The skills needed to thrive in these roles, both inside and outside defence and security Routes into public‑sector, government, and early‑career opportunities   We were also delighted to feature a Spotlight Session from Dr Yuchen Zhang (IADS), who shared his career pathway, his work at the Eloquence Ai project and offered valuable insights into navigating a career in data science.   A huge thank you to Emma Li, Yuchen Zhang, and The Alan Turing Institute Institute for an inspiring and informative event.   Students left with a clearer understanding of the opportunities ahead and how to take their next steps in this exciting field.   #TuringInstitute, #Eloquence AI #UniversityofEssex, #CSEE, #SMSAS, #EBS

12

I’ve been thinking a lot about how future generations won’t just use AI and advanced technology - they’ll grow up alongside it. The tools we’re building now will shape how they learn, create, and solve problems. Preparing for that future isn’t just about technical skills; it’s about teaching curiosity, ethics, and how to really think critically in a world where answers are instant. What we build today becomes the baseline tomorrow, the foundation that will spark creativity and innovation for more people. That’s sounds both exciting and a huge responsibility to me!

2

The biggest lie sold to Data Engineers in the last 5 years is that "Storage is cheap, so you don't need to model data anymore." Sure, storing the mess is cheap. But scanning it is bankrupting you. We stopped teaching Dimensional Modeling because it was "hard." We replaced it with raw dumps and "Schema-on-Read." Now, every simple query triggers a full table scan on a petabyte of JSON blobs. Here is the uncomfortable truth about your Cloud Bill: Lazy Architecture: You are burning $500 in Snowflake credits to answer a question that a Star Schema could answer for $0.05. Performance: Without Conformed Dimensions, your joins are exploding, and your dashboards are timing out. 💀 Johnny Winter 💀 isn't just teaching a "course." He is teaching Anti-Fragility. I can teach a Junior Python in a week. I can't teach them how to design a grain that keeps the CFO off my back. If you don't know the difference between a Slowly Changing Dimension and a Snapshot, you aren't a Senior Engineer. You are a liability that is going to burn our entire Q1 compute budget by Valentines day. Go learn this before your next architectural review. #DataModeling #Kimball #DataEngineering #CodeToCash #TechnicalDebt

8

1 Comment

Following up on my earlier post about building a Python tool for analysing company balance sheets from Yahoo Finance data — I’ve made some solid progress over the past week. My focus this round was on making the code more modular and maintainable: Replaced long if/elif chains with a dictionary-based structure, where user input now serves as keys and the corresponding values for input in other actions. Added a master script to run the extraction module — the project now works through imports rather than isolated scripts. Moved the internal mapping of balance sheet line items into an external JSON file, which now acts as a small, editable “library” for category assignments. Added a README to document structure, logic, and future steps. These changes make the program cleaner, easier to modify, and closer to something others could eventually extend. Next up, I’ll be looking into filtering redundant line items — Yahoo Finance sometimes returns duplicate data under slightly different names, so I’m exploring a systematic way to reconcile those. GitHub: https://lnkd.in/eCbgUhyx It’s been a great exercise in balancing code architecture and financial logic — small steps, but each one makes the system more coherent. Thanks to [Vansh Trivedi] for the earlier comment about modular structure — that’s exactly what inspired this refactor. #Python #Finance #DataAnalysis #FinancialModeling #GitHub #SoftwareDesign #FundamentalAnalysis

2

1 Comment

Great project from The Alan Turing Institute starting to demonstrate how digital twins can support complex real world problems. Also showing there are still worthwhile things in computing and data science that are not simply "AI".

7

3 Comments

DAY 3/100 - BLOGS ON CONVERSATIONAL AI Liverpool John Moores University Exploring Structured Output in LangChain for Conversational AI As conversational AI continues to evolve, one of the key challenges has been ensuring that the outputs generated by large language models (LLMs) are both structured and reliable. In my latest research, I've explored how LangChain is leading the way in solving this challenge through its powerful ability to define and manage structured outputs. Why Structured Output? - Structured output is critical when working with AI-driven applications, especially in industries where consistent data formats are necessary. - Whether it’s transforming raw text into structured data formats like JSON, Pydantic models, or TypedDict, LangChain provides the tools to streamline this process, making it easier for developers to integrate LLM outputs directly into their applications without complex parsing. - Key Insights from My Research: .with_structured_output() Method: LangChain enables easy mapping of schemas like Pydantic models to LLM outputs, ensuring consistent and predictable data responses. Tool Calling & JSON Mode: The framework allows for binding structured schemas to models as tools, ensuring the output follows a defined structure, with some LLMs supporting JSON Mode to produce a conforming JSON output. - Practical Applications: This approach helps in areas such as data extraction, knowledge graphs, and applications requiring standardized API responses, enhancing the scalability and reliability of AI integrations. - Read the Research Papers: - SLOT: Structuring the Output of Large Language Models - A model-agnostic approach that helps transform unstructured LLM outputs into precise structured formats. - SLOT Paper on Arxiv - https://lnkd.in/dfpZFkdd - Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning - A novel method to control LLM generation, ensuring outputs follow predefined structures. - Grammar-Constrained Paper on Arxiv - https://lnkd.in/dYFrNkie - Resources to Dive Deeper: - https://lnkd.in/dKuvTwS3 - https://lnkd.in/dxuUX-bU #ConversationalAI #LangChain #MachineLearning #AI #StructuredOutput #Research #DataScience #TechInnovation #NaturalLanguageProcessing

17

AI and big data are transforming science, but is that good or bad? Nine Dots Prize winner and neuroscientist Grace Huckins explores how AI is reshaping the way we do science, and how data is revolutionising neuroscience. 🧠📊 🎧 Listen to the Royal Institution podcast on Spotify 🔗 https://cup.org/4mBN01w

8

🚀 The new DATAMIND Mental Health Data Collection tool is here - designed to make mental health research faster, easier, and more intuitive. The platform brings mental health datasets from across the UK into one place, helping researchers: 🔍 Discover what datasets exist across the UK relevant to your topic. 📊 See what each dataset contains - from referrals and diagnoses to prescribing, inpatient care, safeguarding and more. ⚖️ Compare datasets by nation, age group, access route and data type. ⏱️ Save time on study planning. 👉 Explore the tool now: https://lnkd.in/eMgQk_gx

59

1 Comment

Quantum has moved from lab experiment to business infrastructure. UK deep science meets US scale to deliver faster modelling, earlier decisions and real competitive advantage, today. Hear how UK/US collaboration is helping IBM capitalise on quantum advantage and see how UK-built systems are now operating inside a first of a kind Quantum AI data centre in Manhattan, integrating British tech with secure Digital Realty infrastructure. Watch the full video here: https://lnkd.in/eZSg7Fgb #GreaterTogether #Quantum #Science #Partnerships #UKInvestments

59

1 Comment

Super excited to announce my upcoming talk at Warwick R User Group: Can LLMs Actually Analyse Your Data? ✨🤖📊 📅 Monday 16th February 2026 at 5PM GMT "LLMs are increasingly being woven into data analysis workflows, and it's worth asking: where do they genuinely help, and where do they quietly get things wrong? This talk looks at practical uses of LLMs for working with data in R, including Posit's recently released querychat package, which lets you query data using natural language. We'll cover what works well, what to watch out for, and what's actually worth adopting right now."

45

2 Comments

🎬 Can data predict the Oscars again? For the second year running, I’ve written an article for The Telegraph using data and machine learning to predict this year’s Academy Award winners. 🏆 The Oscars are built on artistic judgment and industry politics, but patterns still emerge in the data: precursor awards, film runtimes, release timing, even the age of directors and actors can all shift the probabilities. Read the full piece here: https://lnkd.in/eTfCaKRa

30

New AI device brings early skin cancer diagnosis to remote communities The system provides vital support for patients unable to travel, enabling early diagnosis with just a photo and with no internet required. A Scottish research team has developed a pioneering AI-powered tool that could transform how skin cancer is diagnosed in some of the world’s most isolated regions. The device, created by PhD student Tess Watt at Heriot-Watt University, enables rapid diagnosis without needing internet access or direct contact with a dermatologist. Patients use a compact camera connected to a Raspberry Pi computer to photograph suspicious skin lesions. The system then compares the image against thousands of preloaded examples using advanced image recognition and delivers a diagnosis in real time. These results are then shared with local GP services, allowing treatment to begin without delay. The self-contained diagnostic system is among the first designed specifically for remote medical use. Watt said that home-based healthcare is vital, especially with growing delays in GP appointments. The device, currently 85 per cent accurate, is expected to improve further with access to more image datasets and machine learning enhancements. The team plans to trial the tool in real-world settings after securing NHS ethical approval. The initial rollout is aimed at rural Scottish communities, but the technology could benefit global populations with poor access to dermatological care. Heriot-Watt researchers also believe the device will aid patients who are infirm or housebound, making early diagnosis more accessible than ever.

8

4 Comments

🚀 Transform Your Career with the MSc Applied Data Science (Degree Apprenticeship) at the University of Buckingham. Are you looking to upskill in data science using your organisation’s Apprenticeship Levy funding? We’ve partnered with Vitalograph, Renata Franczak, and Owen Mcginty to develop future-ready data talent through our MSc Applied Data Science (Degree Apprenticeship). Our MSc programme is an integrated degree apprenticeship aligned with the Data Analytics Specialist pathway 📈 of the Level 7 Digital and Technology Solutions Specialist (Integrated Degree) (L7 DTSS) apprenticeship standard. A distinctive feature of our programme is the focus on inclusive leadership and innovation, designed to develop well-rounded, inspiring data leaders. 📅 Next cohort starts December 2025 👉 Learn more and apply: https://lnkd.in/eKXacBbS #DataScience #MachineLearning #Leadership #Apprenticeship #Degree #TalentDevelopment Buckinghamshire Council, Buckinghamshire Healthcare NHS Trust, Silverstone, Silverstone Technology Cluster, Milton Keynes City Council

7

Space friends: applications are now open for the 2026 Karman Fellowship! And they have some very helpful information sessions coming up next week -- see the original post. I truly loved being a Karman Fellow in 2024. My colleagues in my cohort were fascinating people: talented, driven, creative, passionate, and often with viewpoints and ideas radically different from my own. It was a pleasure to really dig in on some deep conversations about the present and the future of our aerospace community. Hot button topics were not off limits, but instead were engaged directly -- always with respect and courage, and usually with as much optimism as is possible to muster on some of the really hard issues. The superstars who run The Karman Project -- like Hannah Ashford, Öykü Gürsu, Adrian Flynn -- always have very tough challenge, as the range and caliber of Karman applicants is pretty amazing. But they are up to the task! And these webinars are a helpful way to learn more about the whole process. If we know each other in real life and you've got questions about what it's like being a fellow, let me know! I'm happy to jump on a call to discuss.

39

4 Comments