Ahmed Khamassi

The pace of capital flowing into AI feels familiar. These two charts from Apollo Academy highlight just how quickly expectations (and valuations) have risen. History tells us that periods of exuberance are almost always followed by a correction. At Buoyant, we believe both sides of this story can be true: ✅ Yes, the current surge of investment is likely ahead of itself. ✅ And yes, the underlying transformation is as inevitable as the Internet’s rise 25 years ago. We’ve seen this same pattern in climate tech: early hype, painful correction, and then enduring progress as the world reorganizes around a fundamental shift. The companies that endure are the ones that solve real problems with staying power. Our responsibility as investors at Buoyant Ventures is not to chase the froth, but to stay grounded in fundamentals — backing entrepreneurs who are building durable businesses that will still matter long after the cycle cools. AI’s trajectory may be uneven, but like climate tech, its long-term impact is unstoppable. 🤛 https://lnkd.in/gga2igQJ https://lnkd.in/gfmRy-VG

30

3 Comments

AI or Behavioural Human-Centred AI? How can our AI investments achieve “higher trust, faster uptake and real ROI?” A neat article here arguing for #behavioralhumancentredai from De Cremer, Schweitzer, McGuire and Narayanan in Harvard Business Review. https://lnkd.in/g5cKgWq8 *.... “To align AI with how people actually think and work, leaders need an approach that applies behavioral science and change management principles. In our recent research, we offer exactly that, which we call Behavioral Human-Centered AI. The crux of it is that the success of AI adoption does not so much depend on the deployment of the most sophisticated and advanced technology, but rather on leadership decisions being fueled by behavioral insights about people’s flaws, biases, and habits...” Particularly like the ideas in here about working with diverse populations, measuring a broader range of outcomes, and designing for humans. Certainly, spaces we’re actively working on with our fantastic tech, customer and HR colleagues here at Commonwealth Bank. h/t @Nicola Nicol for the share. Engin Aygun Zoey Isherwood, PhD Rebecca Segrave Caitlin C. Tania Huynh #ai #tehnology #behavioralscience #behaviouralscience #humancentredai #nudge

207

12 Comments

𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗿𝗲𝗾𝘂𝗶𝗿𝗲𝗱 𝗳𝗼𝘂𝗿 𝗺𝗼𝗱𝗲𝗹𝘀 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝘀𝗶𝗺𝘂𝗹𝘁𝗮𝗻𝗲𝗼𝘂𝘀𝗹𝘆. By late 2024, most production systems needed two. This isn’t a story about scaling laws or bigger datasets. It’s about a mathematical insight that was hiding in the original RLHF equations all along—a closed-form solution that eliminated half the infrastructure overnight. In the latest essay for The AK Dispatch, I trace how preference optimisation displaced reinforcement learning as the default alignment method for language models. The piece covers: → The actual mathematics behind DPO, GRPO, and Constitutional AI—with plain-English explanations → Why DeepSeek-R1 and OpenAI’s o-series 𝘴𝘵𝘪𝘭𝘭 use explicit RL when everyone else abandoned it → A practitioner’s guide mapping 13 frontier models to their alignment methods and primary applications If you’re a CAIO, ML engineer, or technical leader evaluating alignment approaches for fine-tuning or vendor selection, this one’s for you. No hype. Just methods, maths, and decision frameworks. The “RL” in “RLHF” has become optional—except when it isn’t. Knowing when you still need it is what separates informed decisions from conventional wisdom. #AIStrategy #MachineLearning #EnterpriseAI #

1

One last late-night read... (or two) In his blog post (Oct 2025), Toby Ord (Senior Researcher at the Oxford Martin AI Governance Initiative) argues that #reinforcementlearning post-training enabled the transition to reasoning models and enabled productive inference scaling (longer chains of thought that actually help). Now that RL scaling itself appears to scale poorly and is likely to hit economic limits soon, the remaining “compute lever” is mostly #inferencescaling, which changes the economics and governance profile because costs shift to deployment and may reflect “more time spent thinking” rather than a durable increase in intelligence. In a more recent post, Ord claims that too few people are asking about how the costs of #AIagents are growing and the key question should be "How is the ‘hourly’ cost of LLM agents changing over time?" (considering the hourly costs for some models are now close to human costs). First blog link: https://lnkd.in/ehG2V3kj Second blog link: https://lnkd.in/eD5g73SC

1

Your data scientists are brilliant. Your models are sophisticated. So why does it still take 6 months to get anything into production? The bottleneck isn't technical—it's organisational. I've just published a case study on epsilon cost scaling: how we went from deploying 1-2 models per year to dozens per quarter, by redesigning team structure rather than adding more talent. Platform teams. Embedded subject matter experts. Reusable patterns. Self-service deployment. The uncomfortable truth: building an organisation that scales AI is technological prowess and fit-for-purpose organisational design. 👇 https://buff.ly/shgWS3O #ProductionAI #DataScience #TechLeadership

7

Check out my post: Large base models have become commoditised. Where organisations truly differentiate is not in the base model itself but in how they handle knowledge: The PRE Framework describes the three complementary ways AI systems use knowledge : Prompted, Retrieved, and Embedded — to deliver accurate, governed, and scalable outcomes in enterprise settings. 

3

Yesterday I attended the Strachey Lecture at Oxford's Museum of Natural History, where Professor Ion Stoica shared insights from building Ray, vLLM, and Chatbot Arena. One point that stuck with me: he quoted David Patterson — "for better or worse, benchmarks shape a field." Ion built Chatbot Arena because existing LLM benchmarks were static, gameable, and didn't capture what actually matters to users. His solution: real humans, real questions, Elo ratings. To scale evaluation beyond pizza-fueled student labelling sessions, his team pioneered LLM-as-a-judge, using GPT-4 to evaluate responses, finding 85% agreement with human preferences. It reinforced why I'm working on TM-Bench - a benchmark for evaluating LLM capabilities in threat modeling. I use the same LLM-as-a-judge approach to assess threat model quality across multiple dimensions. If we want AI to genuinely improve security work, we need domain-specific evaluation that measures what matters. Benchmarks aren't glamorous infrastructure, but they shape what gets built. Thanks to Luc Plessier for the invite. #AI #Cybersecurity #ThreatModeling

25

Meeting quarterly budgets isn’t enough anymore—efficiency is the new multiple. History shows us that every leap forward was powered by more than technology—it was driven by operational reinvention: ·      The Printing Press democratized knowledge. ·      The Steam Engine scaled industry. ·      The Web connected markets globally. ·      Now, AI/GenAI is redefining efficiency and speed. The winners in this era won’t be the ones that just “meet budget.” They’ll be the ones who: ·      Scale operations to handle growth without leakage ·      Streamline processes to unlock hidden value ·      Embed digital tools & AI as enablers, not side projects Accordion’s recent research highlights that CFOs often fall short not in forecasts, but in operational readiness—harmonizing finance with operations, rationalizing systems, and staying exit-ready. That’s the real value driver. Quarterly results matter, but enduring enterprise value comes from operational scale, process discipline, and digital enablement. Efficiency is the new multiple.

7

Pleased to welcome RevOpsly into the Quantum Tiger ecosystem & into our Dark Matter Drive beta program. Their team is taking on one of the most complex & underserved challenges in modern business how revenue systems actually work beneath the surface & how GTM teams can operate with clarity rather than guesswork. Even in this early stage of their build, it is clear that RevOpsly understands the gaps that slow organisations down & the opportunity AI creates to close them. What excites us most is their ambition to design an operating system for Revenue Operations from first principles. It aligns directly with our mission to give emerging products the infrastructure they need to think, adapt & compute at the speed of real markets. With access to Dark Matter Drives, Revopsly can experiment freely, run heavy models without constraint & shape a platform that reflects the realities of how global GTM teams operate. This partnership is about supporting a team that is building carefully & deliberately for the long term. As they continue developing, Quantum Tiger will provide the foundation, compute depth & engineering support needed to accelerate their journey. We believe the future of revenue operations will be AI native, & RevOpsly has the vision & discipline to help lead that transition. We are glad to have them with us & look forward to what we will build together.

15

The AI giants battled it out, and Claude won—even when it lost. New research from King's College London and Oxford put leading AI models through 32,000 rounds of game theory experiments. The results? We're not working with stochastic parrots but unique strategic minds. The study revealed distinct fingerprints for each LLM. - Google's Gemini: Ruthlessly adaptive, quick to betray attempts to cooperate - OpenAI's models: Persistently cooperative, even when being exploited - Anthropic's Claude: The sophisticated diplomat, most willing to restore cooperation after conflict What struck me most? When researchers analyzed the models' reasoning, they found genuine strategic thought. These models weighed time horizons, modeled opponents, and adapted their strategies. That's not pattern matching. It's decision-making. For business leaders selecting AI partners, LLMs aren't interchangeable. They take fundamentally different approaches to collaboration, trust, and conflict.   As for me, I'm betting on Claude from the team at Anthropic. The world's ruthless enough, tyvm. I'll take the AI that's playing for transformation. Full paper here: https://lnkd.in/eqHRqG8d

3

Excited to share this fun and wide-ranging conversation with Toby Babb on FinTech Focus TV by Harrington Starr. We covered some of the most exciting shifts in AI for market data management - from the rise of agentic workflows to how teams are using AI to share expertise at scale. We also looked ahead at what our team thinks could define the next generation of AI-driven market data management technology. Catch the full episode here: https://lnkd.in/eXzj3NHD Thanks again for the opportunity, Toby. #MarketData #MarketDataManagement #ai #AgenticAI #ArtificialIntelligence #TRGScreen #FinTechFocusTV

22

2 Comments

Interesting article on Palantir Technologies' use of Forward Deployed Engineers in their sales motion: https://lnkd.in/exXxXT4y ------- EXTRACT: Commercial model: proof before scale Palantir’s contracts start small. A first engagement might cover a short bootcamp and limited licenses. If value is proven, additional use cases, workflows, and data domains are layered in. Over time, the revenue mix tilts toward software subscription rather than services. Unlike consulting firms, services are a means to drive product adoption, not the primary revenue stream. Unlike most software vendors, Palantir is willing to fund its own engineering time upfront to land a meaningful customer.

1

At the All Things Open Conference, I had a fascinating (albeit short) conversation with Craig McLuckie about data provenance in the context of using unverified corpus for LLM training. Drawing parallels with the then and now of open source deployment, any developer worth his/her salt retains a repository of the open source licenses associated with the open source code he/she uses; but it took decades for founders, lawyers, and investors to include as an ordinary course business practice the vetting and validating proprietary open source code. As the next evolution of best practices for tech companies, it is imperative entrepreneurs and companies alike track data provenance of training corpus before getting hit with an IP infringement suit or an impossible diligence task...and if you don't know, now you know. More learnings at All Things Open Conference! #AllThingsOpen Arcee AI, Dan Estanqueiro

31

3 Comments

While brushing up on key data topics like Apache Iceberg deletion vectors (one per data file per snapshot in v3, per Lester Martin 🥑 insightful blog https://lnkd.in/dJjm6_CN), I’m comparing them to Databricks’ Delta Lake deletion vectors—which use Roaring Bitmaps for low-overhead row deletes and enable 10x faster MERGEs on Unity Catalog tables. This sparked questions: How do Iceberg’s semantics interact with vector/embedding layers in analytics/AI workloads? (Internal DVs vs. external indexes make a big difference.) Worth unpacking for lakehouse builders. Thoughts? #ApacheIceberg #Databricks #DeletionVectors #DeltaLake #DataEngineering

4

1 Comment

This was my favourite quote from the recent interview on the excellent Joe Walker Podcast with Francis Fukuyama on AGI and the Recommencement of History. Francis Fukuyama’s remark made me reflect that the way we frame “intelligence” shapes how we think about AI. If we treat narrow, technical reasoning as the ultimate human virtue, we risk sidelining the other types of intelligence (e.g. emotional, political, cultural) required for individuals, organisations and societies to genuinely thrive. Arguably, these types of intelligence become more important in a world where analytical intelligence is commodified by AI. I’d be interested in hearing your thoughts on the role different types of intelligence play in the development and governance of AI. Christina Wiremu-Brook Joel Backwell Paul Hubbard

21

2 Comments