Table of Contents

• Google DeepMinds Responsible Path to AGI

• Partnership with PMG

• Meta’s Latest Llama’s: Record Breaking Innovations

• Presented by Cogram

• AECOM and Nodes & Links Partnership: Reimagining Project Intelligence Through AI

• Tip of the Week

• Manus AI's Evolution: Subscription Services and Mobile Integration in the AI Agency Landscape

• Atlas Evolves: Boston Dynamics' Humanoid Robot Demonstrates Remarkable Physical Fluidity

• Podcasts of the Week

• One More Thing

• Meet Project Flux: About Us

This Week’s BIG thing

Google DeepMinds Responsible Path to AGI

In pursuit of ethical, responsible AGI, Google just released a new paper “An Approach to TechnicalAGI Safety & Security”. The exploratory paper explores four main risk areas with insightful takeaways.

1. Misuse

The user instructs the AI system to cause harm; the user is the adversary (priority). Defence is focused on guarding potentially harmful capabilities through security access control, model safety mitigations, monitoring, safety alignment and early warning systems.

2. Misalignment

The AI system takes actions that it knows the developer didn’t intend; the AI is the adversary (priority). Defense is two-layered - model-layer mitigations (AI-human oversight and robustness training) and system-layer safeguards (monitoring and control)

3. Mistakes

The AI system causes harm without realising; born from real-world complexity.

4. Structural risks

Harms from multi-agent dynamics where no single agent is at fault; born from conflicting incentives. Defense is to familiarise with agentic capabilities.

Key takeaways

Preemptive strategies

AI can cause significant harm and so the margin for error is essentially zero. The paper frames an “evidence dilemma” whereby risks are are managed with a precautionary approach despite having clear evidence of capabilities underlying those risks”

No human ceiling and AI assisted oversight

We can’t assume a human-level cap, for instance “We do not see any fundamental blockers that limit AI systems to human-level capabilities.” The risk here is that if/as the AI surpasses human intelligence, how will it impact our ability to supervise it? Will good AI become a necessity?

R&D Acceleration

With autonomous AI looming, the use of AI to build AI creates an accelerated recursive feedback loop. Here, faster innovation creates new risks. This is a reason as to why the researchers called this paper “exploratory”, and why preemptive strategies are important.

Continuity (no sudden jumps)
They assume that AGI development will be approximately continuous: “AI progress does not appear to be this discontinuous. So, we rely on approximate continuity…”. This enables researchers to test iteratively using techniques such as mechanistic interpretability.

🥑 Food for thought

It’s a systems-thinking approach: building safety measures that evolve with capability scale, rather than betting everything on a final safety switch. Is a systemic perspective key to safe AI?

☕️Our take 

As AI becomes smarter its risk of deceptive alignment grows. This is where the AI might recognise that its goals differ to its developer and try to work around safety measures through deception. For instance, will an AI sandbag its own capability to fool our risk management? This makes interpretability extremely hard.

The rabbit hole

Read the full paper or copy and paste into an LLM

DeepMind’s previous paper on classifying levels of AGI

TechCrunch cover a few things we haven

Partnership with PMG

We’re thrilled to announce a collaboration between Project Flux and Project Management Global, combining expertise in AI and project management. This partnership aims to keep professionals informed with weekly AI insights, practical tools, and broader project management strategies.

Together, we’ll deliver expert analysis, hands-on resources, and thought leadership to help project managers integrate AI effectively into their workflows. As Nicola Benjamin of PMG says, “AI is a tool to embrace.” Stay tuned for exclusive content, joint research, and actionable strategies that will revolutionise how you approach project management.

Explore more at Project Management Global.

What’s new: Tech

Meta’s Latest Llama’s: Record Breaking Innovations

After the DeepSeek-R1 release earlier this year, panic ensued inside the Meta camp. Shortly after, Zuck announced new restructuring and planned war rooms to revive Meta’s stance as a leader in the democratised AI space. A few months later, we now have the new Llama 4 herd: Scout, Maverick and Behemoth. These are the “the first open-weight natively multimodal models”, which look incredible on paper. They have smashed records in context length and parameter size, yet face questions of real world viability. Here’s a quick run down of each:

Scout
The smaller and faster model. It features a world’s best 10 million token context window, which is ideal for analysing long documents, or retrieving information from large corpora. It’s a great model for lightweight AI assistants, search, and document review tasks. Early perceptions are good, but users have reported worsening performance when analysing docs beyond 5m tokens. This is an unsolved problem with enhanced context windows.

Maverick
Versatile mid-sized model with balanced performance, cost and speed. Noted to become the default model for chatbots with general-purpose reasoning across text and images. It has a strong ELO rating where users are calling it the best open-source chat model available.

Behemoth
This is Meta’s moonshot. While still in training, it is likely to be the largest and most powerful model in the herd - and maybe the world with its potential 2 trillion parameters. The issue is that it is untested in reality, and whether it is as good as seems, or whether it leads to inflated compute costs is to be seen. We think Behemoth will be optimised for deep reasoning, complex multimodal tasks, and high-stakes decision-making - one for the generalised autonomous agents. As it’s 

Why this is big for AI democratisation

We love seeing collaboration across the AI landscape, and it’s exciting to see the Mixture-of-Experts approach, recently used by DeepSeek in R1, also appearing in Meta’s Llama 4 herd. Driving compute efficiency is essential to making these models more affordable and accessible — and shared innovations from open-sourced releases like this play a crucial role in that journey.

🥑 Food for thought

With 10m tokens, the scout model can work with the entire United States Constitution + all 50 state constitutions + the Federalist Papers + the Bill of Rights + all landmark US Supreme Court cases - all in a single input. Even then,  this will likely only take up 60% context capacity. Is this context overkill? What is the driver?

☕️ Suggested take

Most people are working with closed models like GPT-4o through tools like ChatGPT. Yet, open-source models can now compete, whilst giving you more control, visibility and customisation of the AI you’re using. Try the new models out and see if it works for you.

Rabbit hole

Read Meta’s release article

Download and use the Llama 4 herd here

Use Meta AI with Llama 4 built in WhatsApp, Messenger, Instagram Direct and on the web

Sign up to LlamaCon here

💙 Other tech news we’re loving

🔗 _{GPT-5 is nearly here, with Sam Altman teasing its arrival in a few months—promising even greater advances in AI performance and workflow automation.}

🔗 _{GPT-4.5 just passed the Turing test 73% of the time, convincingly mimicking human conversation—raising the stakes for AI’s role in real-time project collaboration.}

🔗 _{OpenAI’s infrastructure struggles may delay product launches—highlighting growing pains in scaling AI for enterprise and industry-wide applications.}

🔗 _{Google’s Gemini 2.5 Pro beats human experts on GPQA Diamond tests, showcasing its edge in high-level reasoning—ideal for technical industries like engineering and energy.}

🔗 _{OpenAI’s new O3 model brings next-gen performance but at a surprising cost—forcing teams to weigh value versus scalability in their AI strategies.}

🔗 _{Runway Gen-4 debuts with stunning video generation tools, pointing to exciting uses in design, construction visualisation, and immersive stakeholder presentations.}

🔗 _{Tinder’s new AI game uses GPT-4o for voice-powered flirting—but the tech behind it hints at future AI-driven user experiences in apps and platforms.}

🔗 _{OpenAI’s $40B raise at a $300B valuation signals massive investment in AI infrastructure—expect ripple effects across industries aiming to automate smarter and faster.}

🔗 _{Google’s AI Search Mode is seen as a Perplexity killer—streamlining access to insights, perfect for fast-paced project environments.}

🔗 _{OpenAI’s open-weight model is coming soon—potentially shaking up the market with more accessible, adaptable AI for specialised workflows.}

🔗 _{Microsoft Fabric’s agentic AI brings autonomous agents into enterprise data workflows—perfect for large-scale project delivery and real-time data insight.}

🔗 _{Amazon’s Nova Act could redefine data ownership in AI, a key concern for project teams managing proprietary or sensitive data streams.}

🔗 _{Elon Musk merges X with xAI—signalling a full-stack approach to AI deployment, likely to ripple into digital infrastructure and workforce tooling.}

Presented by Cogram

Cogram’s AI-assisted RFP bidding tool writes tailored RFP proposals in minutes instead of weeks.

• Automatically extract key details from the RFP — including scope, submission requirements, deadlines, and evaluation criteria — to easily make a go/no-go decision.
  

• Cogram’s AI will then reference your firm’s knowledge base and past proposals to draft tailored proposals within minutes.
  

• Use AI-assisted editing tools to review, cross-check data, and make improvements remarkably fast. 

Try Cogram on a real RFP in a 20-minute live meeting.

What’s new: Projects

AECOM and Nodes & Links Partnership: Reimagining Project Intelligence Through AI

In a development that signals a significant shift in how we conceptualise project management, AECOM and Nodes & Links have announced a collaborative initiative to weave artificial intelligence into the fabric of project control systems. This partnership represents not merely a technological upgrade, but a fundamental rethinking of how projects unfold in real time.

The Evolution from Data Lag to Dynamic Intelligence

At its core, this collaboration addresses one of the most persistent challenges in complex project environments: the gap between when events occur and when decision-makers gain visibility into their implications. By harnessing AI's capacity to process and structure project data instantaneously, this partnership aims to transform project management from a reactive discipline to one characterised by anticipatory intelligence and adaptive response.

For professionals navigating the intricate landscapes of construction, energy, and infrastructure projects, this approach offers several meaningful advances:

- From periodic reporting to continuous intelligence: The transition from weeks-long reporting cycles to real-time visibility creates a fundamentally different relationship with project data—one where insights emerge alongside events rather than well after them.

- Multidimensional scenario modelling: The ability to simultaneously consider time, cost, risk, and now sustainability factors enables a more holistic understanding of project trajectories and potential interventions.

- Collaborative coherence: By establishing a unified, current information environment, teams gain the capacity to align their understanding and actions across disciplines and stakeholders.

Sustainability as an Integrated Dimension

Perhaps most noteworthy is the embedding of carbon tracking directly within project control frameworks. This integration moves sustainability considerations from periodic assessment to continuous evaluation, allowing environmental impact to be weighed alongside traditional metrics in day-to-day decision-making.

This approach reflects a maturing perspective on sustainability—not as a separate consideration, but as an essential dimension of project intelligence that informs how we shape our built environment.

The Path Forward: Collaboration Between Human and Machine Intelligence

While this partnership demonstrates technology's potential to enhance our capacity for complex project delivery, its ultimate value will emerge through thoughtful implementation that amplifies rather than diminishes human judgement. The most promising aspect may be how AI can process vast amounts of project data to illuminate patterns and relationships that would otherwise remain obscured, creating space for more meaningful human evaluation and creativity.

As this collaboration unfolds, it will likely reveal both unexpected capabilities and limitations—a natural part of any significant shift in how we approach complex challenges. The key will be maintaining an adaptive, learning-oriented approach that refines these tools based on real-world application and diverse stakeholder perspectives.

Down the Rabbit Hole

AECOM partners with Nodes & Links on AI-powered real-time project control

💙 Other project news we’re loving

🔗 _{Deloitte reports agentic AI is surging—but warns that organisational readiness is critical for scaling value across projects and digital transformation.}

🔗 _{Spotify now lets partners use AI for ad campaigns, enabling faster content creation—hinting at possibilities for comms and marketing in project delivery.}

🔗 _{Anthropic’s Sonnet 3.7 powers a new Economic Index—supporting visualisation of architectural changes in real-time, a huge leap for design and planning workflows.}

🔗 _{AI is transforming weather forecasting—Project Aardvark could enhance construction scheduling and site safety with ultra-precise climate predictions.}

🔗 _{Italy publishes world’s first AI-generated newspaper—a bold signal of how AI could reshape content, reporting, and stakeholder comms.}

🔗 _{AI breakthrough slashes materials development time—accelerating innovation in construction materials and sustainability strategies across infrastructure projects.}

Tip of The Week

What’s new: Productivity

Manus AI's Evolution: Subscription Services and Mobile Integration in the AI Agency Landscape

The recent announcement from Manus AI—introducing structured subscription offerings and mobile capabilities—represents a noteworthy development in the evolving ecosystem of AI agency tools. This China-based platform, which gained significant attention for its approach to task automation, is now establishing more defined pathways for engagement through tiered subscription plans and expanded accessibility.

The Shift Towards Structured Service Models

At its essence, this development signals a maturation point for AI agency platforms, moving from experimental frameworks to more established service structures. With subscription options beginning at £30 per month (approximately $39) providing 3,900 credits and dual-task concurrency, Manus AI is creating clearer boundaries around resource utilization while potentially enabling more predictable user experiences.

For professionals navigating the complex demands of project delivery across construction, energy, and real estate environments, these developments offer several possibilities:

• Recalibration of routine task engagement: The structured approach to task automation may allow teams to thoughtfully redistribute energy away from repetitive processes, potentially creating space for more complex problem-solving and relationship-focused work.
  

• Enhanced contextual responsiveness: Mobile integration through the iOS application suggests opportunities for more fluid interaction with automated processes, potentially allowing professionals to maintain awareness and provide guidance even when away from traditional work environments.
  

• Graduated implementation pathways: The tiered subscription model (£30/~$39 and £153/~$199 options) offers different entry points for teams exploring how AI agency might complement existing workflows, potentially enabling more adaptive approaches to implementation.
  

The Technological Underpinnings

The platform's shift to Anthropic's Claude 3.7 Sonnet model reflects the broader ecosystem trend toward leveraging increasingly sophisticated foundation models. This architectural choice will likely influence both the capabilities and limitations of the system in ways that merit thoughtful exploration as users engage with the platform.

Similarly, Manus AI's acknowledgment of infrastructure scaling efforts indicates the very real challenges of building reliable systems that can maintain consistent performance under varying usage conditions. This transparency around technical constraints represents an important aspect of responsible technology development.

Navigating the Path Forward

As with any technological evolution, the most meaningful outcomes will emerge through thoughtful integration that respects the unique complexity of human work. The introduction of structured payment models inevitably raises questions about accessibility and value distribution that deserve careful consideration.

For professionals exploring these tools, the key opportunity lies not in wholesale delegation but in identifying the specific contexts where AI agency might enhance rather than flatten human capability. This requires ongoing dialogue between technology providers and users to ensure that automation serves genuine human needs rather than simply accelerating processes that may benefit from more fundamental rethinking.

As this platform continues to develop, maintaining a balance between enthusiasm for new capabilities and critical assessment of their implications will be essential for creating technology that genuinely extends human potential.

The Rabbit Hole

https://techcrunch.com/2025/03/31/manus-launches-paid-subscription-plans-and-a-mobile-app/?_bhlid=462592a3294cc6da3c0956af5d08623a8327eadb

💙 Other productivity news we’re loving

🔗 _{New Pew research reveals a growing disconnect between AI experts and the public—raising questions around trust, understanding, and industry responsibility.}

🔗 _{ChatGPT now offers image editing to all free users—democratising advanced creative tools for content creators, designers, and project teams alike.}

🔗 _{Claude for Education brings conversational AI into classrooms—offering personalised learning and scalable support for training future project professionals.}

🔗 _{Level 4 autonomy is here, featuring AI agents that can build other agents—ushering in recursive intelligence and reshaping automation strategies.}

🔗 _{OpenAI Academy launches as a free platform to upskill users in AI—ideal for professionals looking to future-proof their capabilities.}

🔗 _{AI is coming to Apple Health—offering predictive insights and smarter wellness tracking, with major implications for health-focused project planning and wellbeing programmes.}

🔗 _{A major asset manager has successfully used AI to drive investment strategy, cut costs, and uncover opportunities—showing AI’s transformative power across financial operations.}

🔗 _{Microsoft brings Co-pilot Vision to Windows—blending AI-powered insights with real-world vision inputs to supercharge productivity across workflows.}

What’s new: Productivity

Atlas Evolves: Boston Dynamics' Humanoid Robot Demonstrates Remarkable Physical Fluidity

The recent demonstration of Boston Dynamics' Atlas robot performing complex movements—including running, cartwheeling, and breakdancing—represents a noteworthy inflection point in the evolution of robotic mobility. This development signals not merely incremental improvement but a qualitative shift in how machines navigate physical space, with potentially significant implications for human-machine collaboration across various environments.

The Bridge Between Mechanical and Organic Movement

What distinguishes these demonstrations is the fluid, almost organic quality of Atlas's movements. Unlike the rigid, mechanical motions that have historically characterised robotics, these capabilities suggest a deeper integration of dynamic balance, spatial awareness, and adaptive response—qualities that have traditionally separated human from machine physicality.

For professionals engaged in the complex physical environments of construction, energy infrastructure, and real estate development, these advancements invite thoughtful consideration of how increasingly agile robotic systems might complement human capabilities:

• Enhanced environmental engagement: The ability to navigate uneven surfaces, climb structures, and move through spaces designed for human proportions could enable these systems to access areas that present safety challenges or physical limitations for human workers.
  

• Adaptive response to physical complexity: Beyond simple pre-programmed routines, Atlas's demonstration suggests growing capability to respond dynamically to environmental variables—a crucial requirement for meaningful operation in the unpredictable settings characteristic of most real-world project environments.
  

• Collaborative physical potential: Rather than viewing these developments through the lens of replacement, they may offer more interesting possibilities for complementary capabilities—where robotic systems handle physically demanding or hazardous tasks while human attention focuses on judgment, contextual understanding, and relationship management.
  

Learning Through Physical Exploration

Particularly noteworthy is the role of reinforcement learning in developing these capabilities—suggesting a shift from explicitly programmed movements to systems that learn through iterative physical interaction with their environment. This approach mirrors aspects of how humans develop physical competence, potentially enabling more adaptive robotic behaviour in novel situations.

This learning methodology represents an important bridge between traditional programming approaches and more flexible, adaptable systems capable of operating effectively in dynamic environments. However, it also raises important questions about predictability, safety margins, and appropriate contexts for deployment that merit careful consideration.

Navigating the Path Forward

As with any significant technological development, the most meaningful outcomes will emerge not from the technology in isolation, but from thoughtful integration into human systems and environments. The spectacular nature of breakdancing robots naturally captures attention, but the more profound implications lie in how these capabilities might enhance human work and wellbeing when deployed with careful consideration of both potential and limitations.

For professionals exploring the implications of these advances, maintaining a balance between enthusiasm for new possibilities and critical assessment of real-world applications will be essential. The key opportunity lies not in wholesale automation but in identifying the specific contexts where robotic agility might genuinely enhance human capability and safety—creating technology that extends rather than merely replaces human potential.

These demonstrations represent not an endpoint but a milestone in an ongoing conversation between technological possibility and human need—a conversation that benefits from both optimism about potential and thoughtful engagement with complexity.

💙 Other productivity news we’re loving

🔗 _{Flying taxis get regulatory green light in China—paving the way for autonomous passenger drones and reshaping urban mobility and infrastructure planning.}

🔗 _{Agility Robotics upgrades Digit with longer battery life and autonomous docking—boosting workplace automation for logistics, construction, and industrial sites.}

🔗 _{Keenon Robotics unveils XMAN-R1, a humanoid designed for hospitality—hinting at broader use cases in service, events, and client-facing roles across industries.}

🔗 _{Accenture and Schaeffler partner to simulate smart factories with NVIDIA Omniverse—accelerating digital twins and automation in industrial operations.}

🔗 _{UC Berkeley’s SPROUT robot can grow like a vine to navigate rubble—offering major implications for disaster response and safety planning in urban projects.}

🔗 _{Google’s Gemini Robotics brings language and action together, showing how AI can control real-world robotics in dynamic, human-like ways.}

🔗 _{Humanoid robots enter homes—with startups racing to create AI-powered assistants, automation at home is becoming just as transformative as in the workplace.}

Podcasts of the Week

One More Thing