Shivam Bansal

Interesting read from Anthropic on their effective way and challenges in building multi agent systems. https://lnkd.in/gkjjVvbM - Parallel sub‑agents super‑charge research: A single “orchestrator” Claude delegates to several specialised Sonnet 4 sub‑agents running at the same time. This parallel search approach boosts accuracy on broad, open‑ended questions by roughly 90 % vs a lone large model. - Performance is mostly about tokens (with tools and model tier close behind) In Anthropic’s BrowseComp tests, 80 % of the quality gains came from simply spending more tokens. Another 15 % was explained by how many tool calls were made and which Claude version was chosen—evidence that multi‑agent setups pay off when you can afford the token budget. IMO, This is good for getting high quality outputs, but not super viable due to the costs. - Prompt rules prevent coordination chaos Practical heuristics—match effort to query complexity, begin with broad searches, pick the specific tool best aligned to the task—stop agents from spawning unnecessary workers or duplicating work, turning an initially fragile prototype into a reliable production system. Also they have shared their prompts here: https://lnkd.in/gK4iJ_XD

22

1 Comment

Proud to have contributed to our collaborative project with Google Cloud Security on open-source software security! We've developed key tools like PROBER and a Vulnerability Prioritisation approach to strengthen the software supply chain. #SoftwareSecurity #SupplyChainSecurity #VulnerabilityManagement #CSIRO #Google

10

1 Comment

Excited to announce that Studitory is partnering with CSIRO's Data61, working closely with Xiwei (Sherry) Xu and the Applied AI systems team to develop a cutting-edge AI marking workflow to mimic real teacher marking patterns! This collaboration will enable us to fully automate the exam practice experience on Studitory, making learning smarter, more reliable and more personalized than ever. A huge thank you to Sherry and the Data61 team for joining us on this journey, excited to see how we can transform the learning space! #AI #EdTech #Innovation #Studitory #Data61 #Collaboration

30

1 Comment

What really stuck with me from this interview with Manos Raptopoulos: “𝐂𝐮𝐬𝐭𝐨𝐦𝐞𝐫𝐬 𝐰𝐚𝐧𝐭 𝐨𝐮𝐭𝐜𝐨𝐦𝐞𝐬, 𝐧𝐨𝐭 𝐞𝐱𝐩𝐞𝐫𝐢𝐦𝐞𝐧𝐭𝐬.” That’s exactly what I’m seeing across Australia and New Zealand. That there’s strong appetite for AI, but only where it’s 𝐠𝐫𝐨𝐮𝐧𝐝𝐞𝐝 𝐢𝐧 𝐫𝐞𝐚𝐥 𝐛𝐮𝐬𝐢𝐧𝐞𝐬𝐬 𝐯𝐚𝐥𝐮𝐞, trusted data and clear accountability. There’s a refreshing pragmatism here: AI adoption isn’t about speed for speed’s sake… it’s about 𝐜𝐥𝐚𝐫𝐢𝐭𝐲, 𝐭𝐫𝐮𝐬𝐭 𝐚𝐧𝐝 𝐨𝐮𝐭𝐜𝐨𝐦𝐞𝐬. What resonates most is the focus on:  • Business-case driven AI  • Enterprise-grade data foundations  • Precision and governance in the “last mile” This is how AI earns trust, and delivers value, at scale. Check it out: 🔗 https://lnkd.in/gApb39hw

11

AI hiring in Australia is not fairer. It is just “computer says no” at scale. Uni of Melbourne research on local AI hiring tools shows they punish messy, non linear CVs and can lock whole groups out of shortlists. SBS reports that a big chunk of Australian organisations already use AI somewhere in recruitment, and regulators are warning these systems can quietly reproduce discrimination while everyone hides behind the model. So if you are a grad or career switcher here, understand this: an auto rejection after an “AI assessment” does not always mean you were not good enough. It can just mean you did not look enough like yesterday’s idea of a “good hire”. If you are building AI for hiring in Australia and your product cannot show how it treats career gaps, non elite degrees and non standard paths, you are not fixing bias. You are industrialising it. If you have been knocked out at an AI screening stage here with zero explanation, put your story in the comments. Read the article here: https://lnkd.in/gtcZDbZB

13

1 Comment

AI accuracy and hallucinations are tricky topics. As adoption grows, the question is how we manage accuracy, especially in highly regulated areas. The balance between speed and relevance can be influenced by temperature settings, context, prompt design, and retrieval augmented generation. But is that enough for us to trust AI the way we trust humans? After all, people make mistakes too. The real answer lies in how we evaluate AI and the metrics we use to measure its performance. Without clear evaluation frameworks, it is hard to know whether improvements in accuracy are real or just perceived. Precision, recall, factuality checks, and human in the loop validation all have a role to play. Accuracy is not a one size fits all target. In some use cases, speed may outweigh precision, while in others, such as healthcare, finance or legal, even a small error could have major consequences. Trust in AI will not come from expecting it never to err, but from building systems that can measure, catch and correct those errors in ways that are transparent and reliable. How are you approaching accuracy and evaluation in your own use of AI? https://lnkd.in/eJTjbpjj

7

🚨 Released today, this data-driven update paints a detailed picture of how AI is evolving in Australia... Just released by CSIRO and the National AI Centre — the 2025 report on Australia’s Artificial Intelligence Ecosystem: Growth and Opportunities. Here’s what stood out: ➡️ AI jobs are surging: over 1,500 organisations sought AI-skilled workers in 2024, up from just 483 in 2015 ➡️ Research output doubled and patents quadrupled since 2015. But, there’s a lag in commercialisation with a 23:1 ratio of research publications to patents ➡️ Strong growth in small enterprise-led innovation, but the ecosystem remains fragile. 85% of AI companies employ fewer than 50 staff ➡️ Urban AI clusters are thriving (Melbourne, Sydney, Brisbane, Perth). Each with specialised capabilities from healthcare to mining. One of the highlights follows on from my previous posts about AI sovereignty. The report sees Australia’s dual identity as both AI taker and AI maker as more than a label — it’s a strategic stance: 🔧 AI taker: rapidly adopting global tools to optimise operations, deliver smarter services, and lift productivity 🌱 AI maker: building tailored AI solutions in high-potential sectors from precision agriculture to remote healthcare and mining automation 🎯 Strength lies in contextual application adapting AI to fit our unique societal, environmental, and industry conditions 🧬 Innovation corridors are emerging with research and industry co-locating to drive region-specific AI capability ❓It's a particularly positive spin on Australia's current sovereign AI position. But I think the big question remains, how do we turn this hybrid position into long-term impact and sovereign capability? Read the full report 👉 https://lnkd.in/gKKFBcYP #ArtificialIntelligence #AIEcosystem #TechPolicy #Innovation #ResponsibleAI #NAIC #CSIRO #AustraliaAI Disclaimer: GenAI helped me generate this content.

49

3 Comments

AI won't scale itself.   At Data & AI Edge 2025, over 200 of Australia’s top data and tech leaders gathered to answer one urgent question: How do we safely and strategically turn AI hype into enterprise-scale value at speed?   From Zero Copy architecture and executive alignment to GenAI governance and AI maturity models, the message was clear: AI transformation starts when data, talent, and business impact finally align.   Discover how leaders from AMP, Macquarie Technology Group, Salesforce, nbn® Australia, Alteryx, Tabcorp, and more are shifting from pilots to profit and reshaping AI's role in Australia’s future.   #AILeadership #EnterpriseAI #DataStrategy #GenAI #ADAPTInsights Jim Berry Dr. Michael G. Kollo Suzie Cardwell Damian Brennan Sonia Boije Richard Fleming Kavita Mistry Artak Amirbekyan, PhD Amy Shi-Nash, PhD Alan Jacobson Simon Kriss Dr. Peter Weill Prof. Marek Kowalkiewicz

13

1 Comment

Not all model outputs are created equal. Exploring a great tool for quantifying uncertainty in LLMs. As we integrate Large Language Models (LLMs) into more critical applications, understanding when to trust their output is just as important as the output itself. This is where Uncertainty Quantification (UQ) becomes essential. Open-source library called uQLM (Uncertainty Quantification for Language Models) that tackles this exact problem. It provides a comprehensive toolkit to measure an LLM's confidence in its own generations, going beyond a simple softmax probability. The repo includes implementations for several state-of-the-art methods: > Semantic entropy for detecting uncertainty in free-form generations > Pragmatic entropy and Pragmatic calibration > Ensemble-based methods This is crucial for building reliable AI systems in high-stakes fields like healthcare, finance, or legal, where a model's miscalibrated confidence can lead to significant errors. By quantifying doubt, we can build safeguards, trigger human review, and ultimately create more trustworthy and responsible AI. Major kudos to the AI team at CVS Health for open-sourcing this valuable contribution to the ML community. It's a concrete step towards more robust and transparent AI. Check out the repository here: https://lnkd.in/gHNwXEjb #AI #MachineLearning #LLM #UncertaintyQuantification #ResponsibleAI #ExplainableAI #XAI #MLOps #ArtificialIntelligence #OpenSource #HealthTech

10

The Australian Responsible AI Index, sponsored by the National Artificial Intelligence Centre (NAIC), provides an in-depth analysis of how Australian organisations are adopting and implementing Responsible AI (RAI) practices. Now in its fourth iteration, the 2025 Index has been updated to align with Australia's Voluntary AI Safety Standard (VAISS), tracking the maturity of AI systems across the 10 guardrails that underpin safe and responsible AI deployment. The index categorises organisations into four maturity levels – Emerging, Developing, Implementing, and Leading – based on their adoption of key RAI practices such as fairness, accountability, transparency, explainability, and safety. Also available is the Responsible AI Self-Assessment Tool, helping Australian organisations evaluate their maturity in responsible AI practices, producing: ✅ A personalised report to understand where you are in your Responsible AI journey ✅ Your RAI maturity score, benchmarked against peers ✅ Tailored guidance on your next steps Check it out! 👇 https://lnkd.in/gp78HQ4E

20

RAG was phase one.The next competitive layer may be observational memory. New evidence suggests memory-augmented agents can reduce costs by up to 10x and outperform classic RAG on long-horizon tasks. This is not a marginal improvement. It’s an architectural shift. Enterprise AI will not be defined by larger models, but by how we design memory, continuity, and evaluation at system level. Once agents retain structured memory, evaluation shifts from accuracy per prompt to stability and alignment over time. The question is no longer “Which model?” It is “What cognitive architecture are we building?” #AgenticAI #AISystems #AIArchitecture #EnterpriseAI #Innovation https://lnkd.in/ebSGNp5g

57

5 Comments

Last month, CHeBA’s health data scientist Rory Chen delivered an engaging workshop with R-Ladies Global on using #Shiny to build interactive apps and dashboards. The session highlighted the potential of interactive data visualisation to make complex research accessible and impactful - transforming datasets into user-friendly insights that can be explored by clicking, scrolling, and swiping. Rory is also the lead creator of the Data Repository Explorer (DataRepExp), an open-source Shiny app published in the Journal of Open Source Software. Designed to simplify access to research data, it enables high-level comparison across multiple studies - addressing major challenges in data access and harmonisation. Workshops like Rory’s demonstrate what’s possible when researchers are equipped with the right tools. With targeted infrastructure, we can scale this capacity and strengthen the bridge between research and real-world impact. 🔗 Watch now: https://lnkd.in/gV8GiAVa

6

Corporate Australia is sleepwalking into an AI irrelevance crisis. Only 65% of Australian organisations plan to increase AI investment this year — nearly 20 points below the global average. We're the world's 13th-largest economy but account for less than 2% of global AI investment. Meanwhile, insurance quoting apps went live in ChatGPT last month. AI discovery is up 527% YoY. Agentic commerce is projected to hit 70% adoption by 2028. The world is not waiting for us. I keep hearing "we need to get governance right first." Yes. But governance done slowly is governance done badly. The organisations failing at agentic AI aren't the ones who moved fast. They're the ones who treated governance as a gate, not a capability built in parallel. The ones getting it right are doing three things simultaneously: → Named accountability for every AI agent's decisions. Not a team. A person. → Proportional observability investment alongside every capability build. → Interoperability standards from day one, not after fifty agents exist in isolation. Four questions every Australian board should be asking: 1. What is our agentic AI programme — not pilot, programme? 2. Who is named accountable for each agent's decisions? 3. What's our build-to-observability spend ratio? 4. Are our products discoverable by AI agents? Australia has the talent. The regulatory maturity. AUD 150 billion in data centre infrastructure being built. What we lack is urgency. The biggest risk isn't moving too fast. It's moving too slowly and waking up in 2028 wondering how we fell so far behind. #AgenticAI #DataAndAI #CorporateAustralia #DigitalTransformation #Governance #APAC

71

52 Comments

Interesting news for open science and genomics! UNSW Sydney, Pawsey, and AMD have introduced Slorado — an open-source, real-time nanopore DNA basecaller. Slorado accelerates whole-genome analysis and expands open, high-performance genomics for researchers worldwide. The Setonix supercomputer at the Pawsey Supercomputing Research Centre, powered by AMD Instinct GPUs, served as the primary testbed for Slorado’s performance validation. See the article for more details! #AMD #Genomics #OpenScience #HPC #ROCm #AMDBrandAmbassador

4

🎤 I'm speaking at Enterprise AI Sydney next week and would love to see you there! 📅 Tuesday 4th March | 🕙 10:10am AI Governance That Doesn't Kill Velocity: A Practical Framework One of the biggest tensions in AI product development right now is how to move fast without breaking trust — especially in regulated industries. I'll be sharing the frameworks we've built at MYOB to keep governance and velocity working together, not against each other. If you're in Sydney and working through similar challenges, come say hi 👋 🔗 Enterprise AI Sydney - https://lnkd.in/gutdyf5k

35

4 Comments

Scaling Agents via Continual Pre-training Alibaba’s Tongyi Lab introduced an Agentic Continual Pre-training (Agentic CPT) - a novel training paradigm that builds robust agentic foundation models before fine-tuning. Their model, AgentFounder-30B sets new benchmarks across 10+ evaluations outperforms existing open-source deep research agents. Key Features: - Agentic CPT Pipeline → Bridges pre-training and post-training for stronger alignment. - First-order & High-order Action Synthesis (FAS & HAS) → Scalable generation of planning and reasoning data. - Two-Stage Training Strategy → Short + long context learning (32K → 128K). - State-of-the-Art Results → 39.9% (BrowseComp-en), 43.3% (BrowseComp-zh), 72.8% (GAIA), 31.5% (HLE). - Scalable & Efficient → Demonstrates robust scaling laws with both model size and data volume. Github Repo Link : https://lnkd.in/gjQ67pi2 Read full paper here : https://lnkd.in/gBVXuTF2 👉 Join Telegram group for handpicked resources, learning materials and updates! Link : https://t.me/aibulletin56

77

14 Comments

New blog post: Using deep learning in PyTorch on rainfall data to predict Australian area remoteness Who would have thought I'd be back with another post so soon! Not me, that's for sure. Anyway, in this episode, I continue my exploration of rainfall data for each ABS SA2 in Australia by computing a set of time-series features and then using them to train deep neural networks in PyTorch to classify whether an unseen rainfall time series is from a major city, inner/outer regional Australia, or remote/very remote Australia. What's more, I switch between R and Python within the one document, highlighting the power of multilingual integration in modern markdown! I found that time-series features which capture temporal dynamics in rainfall can meaningfully classify the remoteness of ABS SA2s, which suggests that there are differences in temporal patterns of rainfall unique to the five different remoteness classifications. This is interesting, because it highlights that looking beyond simple/intuitive hypotheses such as 'cities probably get more rain than the outback' can reveal potentially novel insights about the differential patterns. This could lead to further inquiry about whether these dynamics shift over time (i.e., due to climate change) or other important scientific questions. In any case, if nothing else, deep learning is awesome and I had fun! Please check out the post via the link below if you are interested. https://lnkd.in/giQ-UHKR #statistics #deeplearning #machinelearning #climate #timeseries

8