Your AI got more intelligent last quarter. Its benchmark scores improved. Error rates dropped. But what if less is sometimes more? Research from Anthropic suggests that the nature of the remaining errors have changed - more sophisticated models may also be less predictable. To come to this conclusion, the paper looks at AI failure through two lenses: bias and variance. Bias happens when a model is consistently wrong. Variance, in contrast, is inconsistent difference.  Whilst

Training for coordination - as well competence Most enterprise AI training programmes are designed to answer one question: how do we make individuals competent with AI? The Schelling Point Framework asks a different question: how do we create the conditions that means everyone adopts AI within this organisation? Both matter. But its the second is what determines whether adoption scales. In Part 1 we explored the experimental framework, explaining why focal points are key to w

Nineteen investors just gave a startup $47m - to collect the data an insurer's phone system could gather before 10am. Harper is an AI-native brokerage. It didn't exist before October 2024. Within months it wrote $6m in annualised premiums with 25 employees - because its entire model is built around one idea: every call, email and policy makes the next decision sharper. More operational data in, better decisions out. Continuously. But Harper's advantage isn't the AI. It's the

Welcome to all the AI news that matters this week - across biopharma, medtech, manufacturing and insurance. tl;dr: the price of position This week, the cost of having – or lacking – a clear position in the AI value chain became visible. In stock prices. In capital flows. And in and competitive strategy. In one case, the position was geopolitical. But despite common misconceptions, capital is not retreating from AI. It is moving directionally – toward specific chokepoints in t

Two exponentials walk into an advanced manufacturing plant. What happens next? We're about to find out. Because yesterday Nvidia shipped its Vera Rubin platform - promising up to 10x lower inference costs. And Anthropic bought Vercept, to make its platform even more capable of using computers autonomously. One announcement makes AI cheaper. The other makes it more able. They’re different forces. And as they collide, they'll multiply again on impact. The capability curve is st

An electrician uses AI to write quotes, saves 3 hours and sleeps soundly. A corporate strategist drafts reports, saves more and lies awake. Which may be a concern for employers planning to become AI-first, because how much we use a tool is often shaped by how it makes us feel. The electrician's identity lives in practical, physical outcomes. Admin is an overhead - scheduling, invoicing, chasing customers. Removing it feels like a gift today. And a gift to be used again tomorr

A research-grounded approach to training, coaching and culture change The Core Thesis There is no question that Enterprise AI is an information problem - people need to know how to use AI. But is that sufficient? Is it also a coordination problem, a cultural issue? Does it need to be a shared experience - common sense even? In our experience, most training programmes don't think this through, confining themselves to educating individuals about AI's capabilities. At Brightbeam

Its taken a full weekend to get our heads round this. Three wins at the Digital Business Ireland Awards last week. 🥇 Digital Impact of the Year: ALONE , one of our most rewarding clients; 🥇 Women in Digital of the Year: Melissa Proxenos ; and 🥇 Digital Trailblazer of the Year: Brian Hanly With our previous success, that adds up to a lot of glassware. We're now saving to buy a new shelf. Or perhaps we might just reinforce the current one? It's difficult to add anything els

Welcome to all the AI news that matters this week - across biopharma, medtech, life sciences, complex manufacturing and insurance. tl;dr: AI goes live The week's dominant signal across all four sectors was the same: AI moving from experimental to operational. In biopharma, Merck's partnership with Mayo Clinic   centres on a production data environment - where Merck will run AI models against de-identified clinical records at scale.   The deal signals that a competitive edge i

Agents, disruption and the question of readiness In December 2024, we published ' Cheaper than a Peanut ' - tracking the cost of compute from $20 trillion per GFLOP in 1945 to roughly a penny. Fourteen months later, that same penny can buy you far more. The peanut was a metaphor. But that metaphor is now too expensive. And as a result there are no longer waves of progress. They have united as converging exponentials: cost collapse, efficiency breakthroughs, benchmark annihil

The waves converged. What happened next? A companion to Cheaper than a Peanut (December 2024) In December 2024, Brightbeam published Cheaper than a Peanut , tracking the exponential decline in compute costs from (a theoretical) $20 trillion per GFLOP in 1945 to roughly a penny in 2023. We mapped the history of computing as a series of waves - mainframes, personal computing, the internet, cloud, mobile - each crashing ashore faster than the last. We argued that a second meta w

The polite expert problem In a recent paper, a team at Stanford described how they set out to prove whether agent teams could achieve 'strong synergy'. Would a team of different AI models outperform its best individual member?   They did not have human comparisons in mind. But they did reach for the tools organisational psychologists have used on human teams for decades. The same team-building exercises used in MBA programmes and corporate retreats. NASA Moon Survival. Lost a

tl;dr This week’s AI-based developments across   biopharma, medtech, life sciences, complex manufacturing and insurance   reveal an acceleration of digital intelligence that is structural rather than incremental.   AI is moving from advisory to autonomous, from augmenting human decision-making to executing workflows end-to-end. The enterprises that grasp this distinction are pulling ahead. Those treating AI as a bolt-on efficiency tool are discovering – through their share pr

Will AI really mean we end up doing less? What if we end up doing more? And what if task intensity keeps increasing? New research from Berkeley Haas - published in HBR - studied a 200-person tech company for eight months. They found AI intensifies work in three specific ways. 1. Task Expansion: Employees take on responsibilities outside their traditional roles as AI makes unfamiliar tasks feel accessible. Product managers write code, researchers handle engineering tasks and w

A little more than a month ago, we published   our Productivity Clock hypothesis . It holds that, for two years, markets have been pricing GenAI on narratives. And that, as a result, 2026 is going to feature a lot of actual measurement.   If AI is found to be sufficiently productive, there will have been no bubble. But if enterprise uptake - and therefore spend - is slower than currently forecast, investors will not be able to justify current prices with the updated future ca

So this is a big deal. Our very own Melissa Proxenos may be crowned CTO of the Year. But can we call ourselves surprised? Not so much. In the last 18 months, Melissa has been awarded: - Lead5050 Awards 2025: Mentor of the Year - CTO Craft 100 2025 - Computing Tech Women Celebration 50 2025 - Women in Tech APAC Awards 2024: Global Leadership - B&T Women Leading Tech Awards 2024: Engineering Plus, she's been a finalist 10 other times and nominated another six. Sometimes, talen

Why Some Things Take 10,000 Hours and Others Take 10 Seconds A Framework for Understanding Task Complexity Across Biological and Artificial Intelligence This paper is a companion to Critical Depth: When More Layers Unlock New Capabilities - which explores how neural network capabilities emerge suddenly at specific architectural depths, paralleling biological development. Here we ask: if certain tasks require minimum processing depth, what determines how long it takes to lear

Never liked geometry? Neither did geometry, particularly. But there's a new shape to get your head around. And it looks like a trident. Those of us who have marched off and started to put AI agents into our everyday workflows are increasingly reporting that we seem to be developing three key skills: 1. Curating the right work. When you can achieve so much, you have to ask: 'What's worth doing?' And the corollary: 'Which thousand ideas need to be given a firm no?' 2. Running t

How the environment changes along the way In our first part of this essay, we considered the difference between Amodei and Hassabis' conceptions of AGI. Now we can return to the question that kicked the entire enquiry off:   What does this mean for the environments organisations have to survive in? This feels quite important. The environment is, after all, a set of forces that decides who prospers. It includes customer behaviours, competitive tempo, regulatory regimes, supply

A Thought Paper on Parallels Between Deep Learning and Biological Intelligence The Phenomenon A recent paper from Princeton University won   Best Paper at NeurIPS 2025   — one of only four selected from over 20,000 submissions — by demonstrating something striking: when training neural networks for robotic control, performance doesn't improve gradually as you add layers. Instead, nothing much happens until you hit a specific "critical depth" — then capabilities suddenly emerg

Why there’s no real disagreement Our last essay ended with what we believe is a significant realisation: abundant digital intelligence doesn’t just accelerate progress, it changes what kinds of organisations can survive. AI creates selection pressure through new customer demands, competitive responses, internal behaviours and regulatory regimes. Which will mean organisational forms start to look exactly like what they are - animals evolved for a different ecosystem. Because

How 2026 produces divergent forms Our previous chapter argued that cheap intelligence reveals the organisational phenotype. The visible behaviour. The muscle memory. The habits that usually stay half-hidden behind the sheer effort of getting work done. This final chapter goes one step further. Because once you accept phenotype exists, something else must too - the process of selection. And when selection runs for long enough, divergence follows. In biology, speciation is wha

What's it like to work at Brightbeam? Here are the exact contents of an email sent by CEO Brian Hanly  earlier this week: Hi All, We had an important conversation in our management call this morning that I want to share with everyone. We're growing fast, and that's exciting. But growth means nothing if we burn out the people making it happen. So I want to be direct about something: no delivery timeline, no customer demand, is more important than your wellbeing. A simple share

How cheap intelligence reveals cultural truth Our previous chapter framed the problem of AI adoption as a set of choices. Because in choices is where agency lives. And where responsibility can be assigned. And where the potential of AI either converts into real gains, or leaks away. Even so, limits remain. Even when the choices are sensible. Or indeed optimal. And these limits sit in places less accessible to rapid change: inside strategy, inside tooling, inside programme pl