Mario Thomas

Real interesting post trying to tease apart the increase in user dissatisfaction on X. In the world of LLMs it’s really important marketplaces and media sites have a qualitative and quantitative understanding of user sentiment. The reason being is that users may be able to decide how they interface with your inventory (through their own agents.) I’m sure a lot of people have LLMs handle all their social media already (including posts 😬) https://lnkd.in/gKP2UU_7

1

Yann LeCun's team just published something worth paying attention to. It's called #LeWorldModel. And if you've ever worked with AI systems that performed well in testing and then failed in the field, this paper explains part of why, and where the field is heading next. Here's the plain-language version: For AI to operate in the physical world - robots, autonomous systems, anything that moves and acts - it needs to build an internal model of how reality works. Not just recognize what it sees. Actually predict what happens next: if I move this object here, it goes there. If that door is closed, I have to open it first. Until now, building that kind of world model typically required large pre-trained models and complex training setups. LeWorldModel does it with roughly 15 million parameters, on a single GPU, trained from raw video, and it plans up to 48x faster than foundation-model alternatives. That's a signal that the architecture barrier is dropping fast, and the constraint has shifted. It's no longer a question of whether we can build these models. It's whether we can feed them the right reality. A world model is only as good as what it learns from. LeWorldModel trained on synthetic control tasks and hit real limits on the most complex environments, not because the math broke down, but because the training data didn't encode enough of how the real world actually behaves. The model that trains on richer, more varied, more physically grounded human experience data will be the one that generalizes. The architecture is no longer the ceiling. The data is. This is the thesis we're building on at Kaius. Paper linked in comments. #EmbodiedAI #WorldModels #Robotics #HumanData  

8

1 Comment

This week the UK Government is set to announce a major collaboration with OpenAI and Nvidia, along with associated investment. On Friday morning last week I was live on BBC news discussing why this investment in the UK’s AI infrastructure is important, how the government is seeking to ensure the UK remains competitive with AI on the global stage. Thanks to Neha Misri for the video and support! UST #AI #SovereignAI #AIInfrastructure

58

1 Comment

At EmergeGen we have been working on a unique Reference Architecture for Quantum Computing to lower the barrier of entry for teams working specifically on data extraction combined with AI, dynamic search and retrieval. Data Central and SuperOntology now available on Snowflake ❄️ or partner through Collibra If you would like to explore Quantum use cases please get in touch 📧🤙 #quantum #ai #ontology #QDay #snowflake #dataextraction Allan Francis Beechinor Christopher Harrison Dan Matthies, CFA CAIA

8

I've spent the last few months building a scenario model for the UK knowledge economy, drawing on the Autor/Dorn/Hanson China Shock study (the most rigorous empirical evidence we have on what happens when a workforce is structurally displaced), current enterprise AI adoption data, and a dynamic I think most people haven't clocked yet: AI investment is funded from payroll savings, not capital budgets. Which means economic weakness accelerates displacement rather than slowing it. The probability-weighted result: 78% likelihood of significant adverse outcomes where the central scenario projects 1.2–1.8 million knowledge workers displaced by 2030, and multiplier effects pushing the total beyond 4 million. Three-in-four odds against the UK knowledge economy as we know it. The full analysis is in today's piece. It covers why the UK is uniquely exposed (the top 10% of earners pay 60% of all income tax and they're overwhelmingly in the sectors with the highest AI exposure), why the political system won't cushion this in time, and what it actually looks like when unemployment hits 10%+. It ends with a question I think more people need to sit with: how much of your work is genuinely irreducible human judgment, and how much is pattern recognition that an agentic system will do for a fraction of your cost within 18 months? Link in comments 👇

32

7 Comments

Another breakthrough in #AI was published earlier today by Sapient Intelligence, and the linked headline by VentureBeat says it all. https://lnkd.in/giYJtEb9 TL;DR, the new #HierarchicalReasoningModel (#HRM) reasons more like a human by quickly thinking about the best way to solve a problem and then thinking more slowly to solve the problem. This new method can be trained on an order of magnitude fewer training examples and can solve multi-step puzzles that simply befuddle #CoT #LLM s, like #ChatGPT and #Claude, ultimately requiring far fewer resources.

6

2 Comments

A nice insight from John Martinis in the Financial Times. I agree that there is a critical balance to be struck between outsourcing technology overseas vs insourcing the best technologies from around the world, particularly in quantum. The UK is getting this balance right, investing in homegrown quantum industries, and I expect there'll be a new quantum funding allocation in the budget (as previously committed) that balances supporting the UK ecosystem and utilizing the best global technologies through collaboration across companies and countries. The UK could be a global leader in quantum computing with this approach - and I'm proud that Riverlane is already collaborating with national labs like Oak Ridge National Laboratory, and leading quantum hardware companies to accelerate their path to utility scale. John, a renowned quantum pioneer at Google, founded his own venture, Qolab, with which Riverlane is honoured to collaborate. It is through these strategic partnerships that the #quantum industry will truly advance and achieve its full potential. https://lnkd.in/eCUZ2Fyz

83

3 Comments

And it's signed! After a busy few weeks, I'm pleased to have led the AI strand of the 🇬🇧🇺🇸 UK-US Tech Prosperity Deal: https://lnkd.in/ebfCk_7i The Deal lays the foundation for even stronger bilateral collaboration on all things AI and already underpins a record wave of investment in AI and tech - with yesterday's announcements totalling more than £31 billion. I'm particularly pleased that the Tech Prosperity Deal will advance our collaboration on AI for science, in particular healthcare, bringing together more resources to amplify the use of AI in medical diagnosis and treatment, and look forward to working with new partners across the research community to implement it.

215

My latest blog 'LLMs, World Models, Reinforcement Learning and the Future of AI' aims to take a technically-grounded, neutral look at some of the key approaches that will drive the future of #AI, moving beyond the current debate between #LLM boosters and skeptics. #reinforcementlearning #worldmodel https://lnkd.in/eRmYHgj3

7

I am hoping that 2026 will be the 'decisive shift' where businesses finally move past the era of endless surveys and experimentation into the full-scale production of autonomous AI agents. This article definitely suggests so. #AI #aiagent https://lnkd.in/gJxkdWB6

13

6 Comments

I've been following the LLM field on synthetic personas for some time now. It's truly interesting, both as an efficiency play, but also to unlock deep, rapid insights for early product testing and prototyping. In this recent example, PyMC Labs and Colgate-Palmolive developed semantic similarity rating (SSR), achieving 90% of human test-retest reliability across 57 personal care product surveys (9,300 responses). The method maps LLM-generated text to Likert scales, enabling scalable concept testing without traditional panel biases or billion-dollar survey costs. Amazing. Source [arXiv:2510.08338]

33

6 Comments

We’ve mapped the AI and ML inferencing hardware landscape for capital markets, focusing on how different architectures perform within the constraints of trading venues and colocation environments. Drawing on years of STAC Research benchmark data, and our view of where the hardware industry is heading, our analysis compares CPUs, GPUs, FPGAs, SmartNICs, and ASICs across latency sensitivity, determinism, throughput, and operational deployability. The result is our view of where each hardware class fits for each category of trading application, e.g. tick-to-trade pipelines, feed normalisation, signal filtering, smart order routing, pre-trade risk controls, and real-time surveillance.

26

2 Comments

I am blown away, by the ability of LLMs when coupled with supporting software. We are now consistently deploying small LLMs to read, understand and reliably cite sources. The image below shows a local 7b model answering a question by compiling information from several documents to create a wholistic answer on a broad and unrefined question. The icons are links to the document and page the answer was found on so the validity of answer can quickly be checked for accuracy. Testing larger models, we see graceful increase in capabilities allowing more understanding applied to answers allowing complex reasoning over text. Larger models also scale in generation tasks. All of this without making a single API call. All documents safely on my machine. No cloud, no data sharing, no extra costs. #InsurTech #DataPrivacy #Compliance #LocalAI #CopliantAI

5

Yesterday we published the 2024 AI Sector Study, our annual analysis of the UK’s AI sector, providing estimates of the scale, scope and growth of firms developing AI products and services in the UK. Some key insights: - Companies: From 2023 to 2024, the number of AI firms jumped by 58% to around 5,800. - Employment: UK AI sector employment grew by 33% to around 86,000. - Revenue: Revenue saw a massive increase of 68%, reaching around £23.9 billion. - Gross Value Added (GVA): The sector's GVA more than doubled, increasing by 103% to £11.8 billion. - Investment: Investment in dedicated AI companies rebounded to £2.9 billion, surpassing the previous 2022 record. - Lots more in the report including breakdowns by sub sector, business size, region, and evidence on barriers to growth. Great to partner with Perspective Economics, Ipsos, and Glass.AI on this. And brilliant work from the DSIT team - Maeve Fitzmaurice, Luke Yates, and Mark Wingham https://lnkd.in/eM4_Jiev

165

4 Comments

Our team at IBM Research is exploring new ways of how high-performance distributed storage systems like IBM Storage Scale can add value in the increasingly important AI inference use case. In our first blog on this topic, we experimentally demonstrate how simply offloading inference runtime state (KV Cache) to IBM Storage Scale can reduce time to first token by 8 to 12 times. Lots of more interesting things to come! 😁 https://lnkd.in/g5z6Gqjq Yue Zhu Radu Stoica Animesh Trivedi Jonathan Terner Frank Schmuck Jeremy Cohn Christof Schmitt Anthony Hsu Guy Margalit Vasily Tarasov Swaminathan Sundararaman Talia Gershon Vincent Hsu #ibmresearch #ibmstorage #ibmstoragescale #ai #llm #inference #kvcache

47

For the 10s of thousands of LinkedIn members here at the largest AI Summit in history in Delhi, please join our STEM Practice Presentation and Panel Discussion this afternoon: "AI Without the Cost: Rethinking Intelligence for a Constrained World" explores how to deliver impactful AI amid limits on capital, compute, energy, and talent. It highlights efficient models, frugal innovation, open ecosystems, and smarter deployment strategies that maximize value while minimizing resource intensity—enabling scalable, sustainable intelligence accessible to enterprises, governments, and emerging markets alike.

30

2 Comments

Incredibly proud of our team's work with UCL on this vital initiative. Using AI to tackle gender bias in scientific research isn't just about data—it's about creating a more equitable future for healthcare. Thrilled to see the impact our partnership with Women in Data® is making. #DataForGood

18

2 Comments

Sankar makes an excellent point about agents allowing foundational redesign of processes. How will this alter operations processes and supply chain configurations? #agentic #designthinking #Supplychain #Configuration Production and Operations Management Society (POMS)?

13

3 Comments

🚀🚀🚀UK AI Frontier start-up/SME Funding Opportunity🚀🚀🚀 Innovate UK’s AI #Champions: #Frontier AI Phase 1 opens on 17 March 2026 for UK-registered start-ups/SMEs developing genuinely #novel #AI and #ML technologies. This is aimed at teams pushing the frontier of #capability, where the innovation comes from #technical advance in #model or #system #architecture, #training #methodology, or core #control / #learning #algorithms, rather from routine application-layer assembly. The competition is looking for #ambitious #projects grounded in strong #scientific #evidence and #engineering #breakthroughs, with a clear route to defensibility, whether through protectable IP or other credible barriers to entry. We are looking for projects that start at #TRL 3 or below that can generate robust evidence that the core approach works in practice. The competition covers four themes: #AI-Enabled #Health and #Life #Sciences; #Advanced #Materials with AI; #Secure AI for #National #Security & #Defence; #Fundamental #AI. Within those, the priorities include areas such as #predictive #healthcare, #medicines #discovery, #materials #prediction and #physics-ML, #secure AI-enhanced #sensing and #decision support, and abstract #reasoning, #generalisation, #novel #architectures, #online #learning, #world #models, and #generalisable #explainable AI. This is Phase 1 of a three-phase, stage-gated pipeline intended to support the next wave of UK AI Champions from early technical validation through to scale-up. The 3 phase funding opportunity could unlock up to £2.7 million in grant funding + private investment over 3 years for successful projects. Funding for phase 2 and phase 3 is currently subject to business case approval. If you are building something that could deliver state-of-the-art benchmark performance or genuinely new-to-the-world capability, this is worth serious attention. If that sounds like your company, or one you know is doing real technical work in this space please re-share with your networks The competition web briefing is next week, link in the competition page: https://lnkd.in/e5UvVFmi #InnovateUK #FrontierAI #ArtificialIntelligence #MachineLearning #DeepTech #UKInnovation #HealthAI #AdvancedMaterials #DefenceTech #FundamentalA Renaissance Philanthropy Pillar VC University of Cambridge School of Informatics, The University of Edinburgh Imperial College London UCL University of Southampton University of Bristol University of York Venture Café London CommonAI CIC Innovate UK Business Connect

108

12 Comments