Adeel Ahmad
Greater Melbourne Area
2K followers
500+ connections
View mutual connections with Adeel
Adeel can introduce you to 10+ people at Commonwealth Bank
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
View mutual connections with Adeel
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
About
I'm a Staff Platform Engineer (MLOps) at the Commonwealth Bank of Australia with 20 years…
Activity
2K followers
-
Adeel Ahmad shared thisStandard fine-tuning is broken in three ways: 1. It memorises — the model copies token sequences instead of learning to reason 2. It diverges — training shifts the output distribution away from the pre-trained baseline 3. It forgets — catastrophic forgetting degrades existing capabilities I've spent the past few months working on an alternative approach: Latent Contextual Reinforcement (LCR). The core idea: instead of rewriting a model's weights, you activate the reasoning pathways it already has. Expert reasoning tokens are interleaved with the model's own generation — the model believes it produced the entire sequence. Gradients flow only through model-generated tokens. The training distribution never shifts because the model learns exclusively from its own outputs. Results on Qwen3-4B, trained on a single machine with 8GB of RAM: → Weight-level change: effectively zero (cosine similarity 1.0000) → Behavioural change: substantial (new reasoning patterns, new identity) → Mechanistic signature: single-axis rotation in activation space → Compute required: a few hours on consumer hardware The mechanistic finding is what I find most interesting: fine-tuning doesn't rewrite neurons. It rotates them. The model is a library — LCR changes which books get read and in what order. This also raises important questions for AI safety. A few-megabyte adapter can modify a model's behaviour while passing every standard evaluation checkpoint. Detection requires geometric analysis tools that aren't yet part of standard auditing pipelines. I've written up the full methodology, experimental evidence, and mechanistic analysis. https://lnkd.in/gCQGySyV #MachineLearning #ArtificialIntelligence #DeepLearning #AISafety #LLM #NLP #ReinforcementLearning #MechanisticInterpretability #FineTuning #MLResearchLatent Contextual Reinforcement: Teaching Language Models to Think Better Without Changing Their…Latent Contextual Reinforcement: Teaching Language Models to Think Better Without Changing Their…
-
Adeel Ahmad shared thisThe Agent Skills spec nails capability definition — but once you have dozens of skills, how do you assemble them into a running agent? I've been building AIOS, an AI-native operating system on Unix primitives, and just proposed extending the spec with AgentFile (declarative agent composition), filesystem-native skill delivery for small models, and multi-agent orchestration via AgentCompose. Would love community input: https://lnkd.in/gpv68rHa #AgentSkills #AIAgents #MCP #ModelContextProtocol #LLM #OpenSource #AIOS #Anthropic #ClaudeCode #DevTools #AIEngineering #BuildInPublic #SmallModels #AgentOrchestrationProposal: AgentFile — Declarative Agent Composition from Skills + Filesystem-Native Skill Delivery [Draft] · Issue #178 · agentskills/agentskillsProposal: AgentFile — Declarative Agent Composition from Skills + Filesystem-Native Skill Delivery [Draft] · Issue #178 · agentskills/agentskills
-
Adeel Ahmad shared thisThere's a saying in Urdu: "naqal ke liye bhi aqal chahiye" — even to cheat, you need wisdom. I've been researching how small language models learn to reason, and that proverb turned out to be the key insight. When a model processes expert reasoning through its own weights — even as scaffolded context it's "cheating" from — its internal representations change. The learning happens as a side effect of the assistance. The same mechanism that's turned struggling students into sharp thinkers for centuries works on neural networks. I applied this to reinforcement learning training of a small model on consumer hardware. The model broke through a long training plateau, started reasoning across domains it had never seen, and got dramatically more efficient at thinking — all from an approach that requires few lines of code. Wrote up the full story in a blog post. No equations, no jargon — just the intuition and what I observed. If you're working on efficient training, small model reasoning, or RL — I'd genuinely like to hear from you. The technical details are worth a conversation. #MachineLearning #AI #ReinforcementLearning #Reasoning #Research https://lnkd.in/gyVGa3-M
-
Adeel Ahmad reposted thisAdeel Ahmad reposted thisArtificial intelligence can feel highly technical and even a bit daunting for many people. At CommBank, we have spent the past few years helping our people understand AI, not just what it is, but how to use it confidently and responsibly in their work. Now, we are sharing that knowledge more broadly through Brighter AI, our new micro-learning series designed to help Australians build confidence and practical skills in using AI safely and ethically. I'm proud to be part of the team bringing this to life, showing what AI means in real life, from simplifying everyday tasks to ensuring it remains safe, fair and trustworthy. If you are curious about AI and want to learn the basics in a simple, practical way, check it out here: https://lnkd.in/gVeHrKfW #BrighterAI #CommBankLife #AIForGood #DigitalInclusion #FutureOfWork
-
Adeel Ahmad posted thisIf you could put GPT4 level LLM on a phone What would the world look like 🤖
-
Adeel Ahmad shared this🧠 I built a 4B reasoning model that thought for 14 minutes before honestly saying "I don't know" - a breakthrough in AI honesty over hallucination. Part 1 of my research series on improving small open-source LLMs: https://lnkd.in/gBDAus_h #ArtificialIntelligence #MachineLearning #AIResearch #OpenSource #AIEthics #ResponsibleAI #TechInnovation #AIBreakthroughI Built a 4B Model That Thinks for 14 Minutes Before Admitting It Doesn’t KnowI Built a 4B Model That Thinks for 14 Minutes Before Admitting It Doesn’t Know
-
Adeel Ahmad shared thisAdeel Ahmad shared thisExcited to share a little something we've been working on at Meta with HuggingFace + many other AI startups and labs: OpenEnv - frontier-grade RL environments for the open-source community. https://lnkd.in/eAx36C_R This framework/spec is derived from a year of building environments and execution infra to support AI teams at Meta (FAIR/GenAI and now MSL). Next step: allow anyone to create, package, and share environments on the Hugging Face Env Hub and keep working on the spec together with the community - interested in helping checkout our RFCs: https://lnkd.in/eUHBagwg
-
Adeel Ahmad reposted thisAdeel Ahmad reposted thisCommonwealth Bank’s world leading transformation isn't just adopting AI - they're delivering a masterclass on how to expertly lead in the AI market. They're ranked top 5 globally and #1 in APAC for AI maturity. Super proud to showcase EY’s role in enabling AI solutions and business value through AI at Australia’s largest and most innovative bank. Pleasure sharing the keynote stage with my friends Alex Burton (CBA), Leandro Arranz (CBA) and Katherine Boiciuc at Gartner Symposium/Xpo. CBA's AI journey learnings are deep and varied. Their pace is unmatched, and their core values are something all can learn from. A few learnings: On Strategic Approach: - Committed to being a frontier firm in AI five years ago - Invested early and consistently in AI capabilities Key Achievements: - 43,000 staff trained in AI - Customer engagement engine making 55 million decisions daily - Data platform processing 157 billion data points - Reduced AI model deployment from 4 months to 10 days Core Learning Principles: - People-Centric Transformation - Democratised AI skills - Direct access between experts and executives Technology and Partnerships: - Built robust global technology partnerships - Established tech hub in Seattle for global collaboration Responsible AI - Developed clear responsible AI principles - Integrated bias and data drift detection (as models scale this is vitally important)! - Focused on customer trust beyond regulatory compliance - Continuous Learning - Built cross-functional teams - Focused on reimagining processes, not just incrementally improving them Data Readiness - Recognised data as fundamental to AI success - Invested in data governance - Created feature stores and semantic models For every leader wondering how to truly leverage AI: CBA has the blueprint. Thanks Arvind Vatsa and Harpreet Singh Sandhu for your persistence and hardwork on this mission. #EY #AIInnovation #DigitalTransformation #FutureOfWork #TechLeadership
-
Adeel Ahmad shared thisI’m wondering if this has anything in common with ec2 nitro enclave in terms of connectivity 💭🧐Adeel Ahmad shared thisWhat happens when you connect hundreds of thousands of custom Amazon Web Services (AWS) chips across multiple data centers? You get Project Rainier, soon to be one of the world's most powerful computers. This is infrastructure built specifically for the most demanding AI workloads, enabling Anthropic to train the next generation of Claude that will transform what's possible with artificial intelligence. https://lnkd.in/gyJibUG3
-
Adeel Ahmad liked thisAdeel Ahmad liked this🚀 I am thrilled to share that the AWS MCP Server is now Generally Available! The AWS MCP Server gives AI agents and AI-native IDEs secure, auditable access to AWS services through the Model Context Protocol (MCP), helping AI agents build on AWS more effectively. It empowers organizations to let coding agents interact with AWS confidently, maintaining strict visibility and control through IAM-based guardrails, Amazon CloudWatch metrics, and AWS CloudTrail logging. With this GA launch, customers can now use their AI agents to: ✨ Call any of the 15,000+ AWS APIs: Unlock the full power of AWS by seamlessly executing any operation, including complex tasks that require file uploads or long-running execution. 🛠️ Leverage Curated Agent Skills: We have transitioned from Agent SOPs to dynamically loaded Skills to supercharge your agent's capabilities! Agents can now discover and load AWS-validated guidance on demand, keeping context window usage low. 📚 Access AWS Knowledge: Learn about AWS services through documentation search, operational guidance, and specialized AWS domains. Documentation search and skill discovery are now available without requiring AWS credentials. ⚡ Run Sandboxed Code: Run sandboxed Python code against AWS services for multi-step operations, without requiring access to the local filesystem or shell tools. 🙌 Huge thanks to our incredible preview community! Your feedback directly shaped many of the capabilities included in this GA launch. Ready to transform your AI workflows with seamless AWS access? Start using today at no additional cost. Learn more 👉 https://lnkd.in/dPbNjnXK #AWS #AI #GenAI #MCP #AgenticAI #CloudEngineering #DeveloperTools #AWSMCPServer
-
Adeel Ahmad liked thisAdeel Ahmad liked thisToday is my last day at CBA. It has been an amazing 15 years full of learning and growth. The highlights include the privilege of creating and leading Technology and Data across: Core Banking, Retail Products, Customer Platform, the Customer Engagament Engine (CEE), the Remedial Action Plan (RAP), International Financial Services (IFS), Corporate Services, Financial Risk Portfolio and the Risk/Treasury/Finance (RTF) Data Platform. The friendships I have made will last a lifetime. I am excited to start my new role tomorrow.
-
Adeel Ahmad liked thisAdeel Ahmad liked thisOur Cevo team came together recently to celebrate Harmony Day with a potluck lunch! There was no shortage of amazing food, but the real highlight was the chance to connect. Sharing cultures, stories and experiences that shape how we show up every day. It’s a great reminder that what makes Cevo special isn’t just what we build, it’s the people behind it. 🧡 #HarmonyDay #LifeAtCevo
-
Adeel Ahmad liked thisExcited to be at AWS Summit Bangkok 2026! Looking forward to connecting with customers and partners. If you're interested in chatting about transforming your business to be AI-ready or moving your agentic workloads from POC to production - come find me. See you there! 🇹🇭☁️ - รู้สึกตื่นเต้นที่จะได้เข้าร่วมงาน AWS Summit Bangkok 2026! ตั้งตารอที่จะได้พบปะกับลูกค้าและพันธมิตร หากคุณสนใจที่จะพูดคุยเกี่ยวกับการเปลี่ยนแปลงธุรกิจของคุณให้พร้อมสำหรับ AI หรือการย้ายเวิร์กโหลดแบบเอเจนต์ของคุณจาก POC ไปสู่การใช้งานจริง มาพบกับผมได้เลยครับ แล้วเจอกัน! 🇹🇭☁️ #AWSSummitBangkokAdeel Ahmad liked this🔥 Only 20 days left! 🎉 We cordially invite everyone to register for AWS Summit Bangkok 2026! at the Queen Sirikit National Convention Center (QSNCC). Free admission! Join us for 40+ sessions, keynotes, and hands-on workshops on AI and Cloud from AWS experts. 📌 Register now and see you at the Summit! 🙌 https://lnkd.in/gf7cpbTY 🔥 อีกแค่ 20 วัน! 🎉 เรียนเชิญทุกท่านลงทะเบียนเข้าร่วมงาน AWS Summit Bangkok 2026! ณ ศูนย์การประชุมแห่งชาติสิริกิติ์ (QSNCC) งานนี้เข้าร่วมฟรี! พบกับ 40+ sessions, keynotes และ workshops เชิงปฏิบัติด้าน AI และ Cloud จากผู้เชี่ยวชาญ AWS 📌 ลงทะเบียนเข้าร่วมงานได้เลย แล้วมาพบกันที่ Summit! 🙌 https://lnkd.in/gf7cpbTY Vatsun ThirapatarapongJackathep SanyaprasoetTippamas AchalakulMichael AranetaPramote PateepkamolPam Pakthapa ChatkomesGiuseppe MarazzottaNam Je ChoShaown NandiKelly SohJeff JohnsonVarun DayalVikas OmerSuney SharmaAshley FernandezNaveen AsraniPrakhar GautamShafraz RahimCharles CrouspeyreShakeel Ahmad
-
Adeel Ahmad liked thisReally looking forward to taking part in the CommBank Accelerate AI event on the 26th of May. We have assembled a very impressive set of CEO speakers from Australia and across the world to share how AI is solving actual problems today.Adeel Ahmad liked thisCommBank Accelerate AI is coming on 26 May, exploring how artificial intelligence is being used to solve real problems and create value across the broader Australian economy. As organisations navigate rapid technological change, leaders are increasingly focused on one question, how to unlock value from AI in a way that is scalable, secure and trusted. Accelerate AI brings this into focus, showcasing how data, AI, security and technology come together, and what it takes to apply AI responsibly at scale. Across the event, global and local leaders will demonstrate real-world use cases, offering practical insights leaders can apply in their own business. We’ll explore the full spectrum of AI, from improving decision-making and enhancing customer experiences, to building trust in an era of rapid change, and investing in the infrastructure and capabilities needed to move AI from experimentation into benefit for all Australians. Sam Altman, Chief Executive Officer, OpenAI Leah Weckert, Chief Executive Officer, Coles Cliff Obrecht, Founder & COO, Canva Adam Driussi, CEO, Quantium Ben Chan, Chief AI Officer, Quantium Craig Scroggie, Chief Executive Officer, NEXTDC Paul Bassat, Co-founder and Partner, Square Peg Melanie Silva, Managing Director & VP Australia & New Zealand, Google Rachael McVitty, Chief Customer Officer, Bunnings Dr Nicole Gillespie, Chair of Trust, Melbourne Business School Contact your Relationship Banker for more information. #CommBankAccelerateAI
-
Adeel Ahmad liked thisAdeel Ahmad liked thisShortly I am about to drop a new blog post in which I will dive deeper in to my HomeLab, the hardware, what I run in terms of VM and containers, but let me share a quick hack for those in the situation I found myself in. The good news is Australia's internet speeds are advancing, with the recent #NBN / #Opticomm speed updates I was not getting the maximum out of my home internet leaving about 200megabits on the table. My 6 year old Ubiquiti Edge Router 10x was hitting max CPU and throttling PPoE but for $30AUD I bought a new Intel i350-T2 in and now using #OpnSense. This not only allows full speed, but actually saves me a few watts of power. The other benefit is with a family of young teenagers I now have parental controls with ZenArmour (example I can block just Youtube Comments) I know there is a lot of Ubiquiti fans out there, but I say you can do better. If you are a builder I am sure you can relate, there is a certain joy in leaning, building foundational skills, having more options and not being tethered to a cloud service. Builders build..... Ps if you dont have a HomeLab there are plenty of people using mini PC's that you get super cheap, and of course I have thought about what if my server fails (Ubiquiti is still racked, prewired with the same IP / DHCP scope etc). Just a matter of powering on.
-
Adeel Ahmad liked thisAdeel Ahmad liked thisThanks to Prime Minister Anthony Albanese for the warm welcome today. We are making our largest investment in Australia to date, committing A$25 billion to expand AI and cloud capacity, strengthen cybersecurity, and help people and organizations across the country build digital skills. Read more: https://lnkd.in/gXWq9mch
-
Adeel Ahmad liked thisAdeel Ahmad liked thisAWS Interconnect - multicloud is now Generally Available! Interconnect makes it easy to quickly connect your private networks on AWS and Google Cloud with highly resilient, high-speed connectivity. This was a huge, multi-year project that resulted in a new open specification that any cloud can now use to provide seamless connectivity to their customers. Congratulations to the AWS team, Brian Pescatore who led the project from the SW Engineering side and everyone who contributed to create this new service. Special thanks to our launch partner Google Cloud. This was a great collaboration with Judy Issa, Michael Woods, and the rest of the team. Head over to the Interconnect website to learn more! https://lnkd.in/e887V2ar
-
Adeel Ahmad reacted on thisAdeel Ahmad reacted on thisA little over a week ago, I had the privilege of attending the BIG.AI@MIT conference at the Massachusetts Institute of Technology - and I was particularly inspired by the strong emphasis on ensuring responsible AI remains central to GenAI innovation. I was proud to represent the AI Acceleration team from Commonwealth Bank, where we had two pieces of work accepted: 🔹 Simulating Customer Behaviour with GenAI 🔹 Synthetic Evidence in Finance It was a valuable opportunity to exchange ideas with researchers, academics, and industry leaders exploring the evolving role of GenAI in business. A big thank you to my collaborators for their work across both projects - Anjin Liu, Harshl Kapoor, Jenny Wang, Shubham Arora and Alex Burton. This was truly a team effort, and I’m grateful to have contributed alongside such a strong group. And it was great to see Puneet Samnani also representing CBA! I came away with new perspectives, great conversations, and plenty to reflect on as this space continues to evolve. For now, I’m taking a couple of weeks to explore more of the US before heading to the West Coast to jump back into work - looking forward to sharing more soon 👀
Experience
Licenses & Certifications
-
AWS Certified Solutions Architect - Professional
Amazon Web Services (AWS)
Issued ExpiresCredential ID a71321e9dffb44ce842c8e5f1c70ceb6 -
-
-
View Adeel’s full profile
-
See who you know in common
-
Get introduced
-
Contact Adeel directly
Other similar profiles
Explore more posts
-
Stephen Jones
Deloitte • 3K followers
If you're building AI workloads in Australia, the model availability gap just got a lot smaller. AWS announced that Amazon Bedrock now supports open-weight models in the Sydney region. DeepSeek, Mistral, Google, MiniMax, Moonshot AI, Nvidia, and OpenAI models, all running on Australian infrastructure through Project Mantle. Here's why this matters more than a typical regional expansion: Australia's National AI Plan calls for sovereign compute. The government built GovAI Chat on domestic infrastructure. Over $100 billion in data-centre projects have been announced. But until now, model diversity on-shore was the missing piece. Open-weight models change the equation. Vendor independence. Transparent weights you can audit. Fine-tuning on your domain data without sending it offshore. And the DeepSeek angle is fascinating: the same model banned from government devices can now run safely on Bedrock in Sydney, with no data going to China and AWS Guardrails layered on top. You separate capability from origin risk. I wrote up the full analysis covering the competitive landscape (AWS vs Azure vs GCP in Australia), the global sovereign AI race, and what this means for Australian organisations planning AI workloads. https://lnkd.in/gftiA_MW #AWS #AI #Sovereignty #AmazonBedrock #OpenSource #Australia #CloudComputing
16
2 Comments -
Utsav Garg
Medium • 2K followers
“𝗗𝗮𝘁𝗮 𝗿𝗲𝘀𝗶𝗱𝗲𝗻𝗰𝘆, 𝗹𝗼𝘄 𝗹𝗮𝘁𝗲𝗻𝗰𝘆, 𝗮𝗻𝗱 𝗰𝗹𝗼𝘂𝗱 𝘀𝗼𝘃𝗲𝗿𝗲𝗶𝗴𝗻𝘁𝘆 𝗻𝗼𝘄 𝗹𝗼𝗰𝗮𝗹 𝗳𝗼𝗿 𝗞𝗶𝘄𝗶𝘀.” 🇳🇿 — 𝗮𝗻𝗱 𝗶𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴.” 𝗔𝗪𝗦 𝗔𝘀𝗶𝗮 𝗣𝗮𝗰𝗶𝗳𝗶𝗰 (𝗡𝗲𝘄 𝗭𝗲𝗮𝗹𝗮𝗻𝗱) 𝗥𝗲𝗴𝗶𝗼𝗻 with three Availability Zones and API name 𝗮𝗽-𝘀𝗼𝘂𝘁𝗵𝗲𝗮𝘀𝘁-𝟲 is now GA (general availability) 𝘞𝘩𝘢𝘵 𝘥𝘰𝘦𝘴 𝘵𝘩𝘦 𝘎𝘈 𝘭𝘢𝘶𝘯𝘤𝘩 𝘰𝘧 𝘈𝘞𝘚 𝘈𝘴𝘪𝘢 𝘗𝘢𝘤𝘪𝘧𝘪𝘤 (𝘕𝘦𝘸 𝘡𝘦𝘢𝘭𝘢𝘯𝘥) 𝘙𝘦𝘨𝘪𝘰𝘯 𝘮𝘦𝘢𝘯? • 3 Availability Zones, API name ap-southeast-6 — now you can run workloads and store data in‑country, reducing latency for NZ users. • AWS invested NZD 7.5 billion; the region is expected to contribute NZD 10.8 billion to GDP and create ~1,000 jobs annually. • Integrates compliance and security certifications (PCI, HIPAA, GDPR, etc.) to support regulated workloads. • Opens doors for AI & ML innovation locally using AWS’s generative AI stack (Bedrock, Q) — without leaving NZ. #AWS #cloud #newzealand #Kiwis #cloudengineering #devops #cloudsecurity #compliance #gdpr #pcidss #cloud #artificialintelligence
12
-
Bhuman Soni
Suncorp Group • 2K followers
🔧 New Blog Post: How to Stop Cascading Failures in Your Microservices After years of building resilient systems at one of Australia's big banks and integrating AI chatbots at a major insurer, I've seen firsthand how the Circuit Breaker pattern can make or break production stability. In my latest article, I break down: ✅ The exact problem circuit breakers solve (and why retry storms kill your services) ✅ The 3-state flow: Closed → Open → Half-Open with real pseudocode ✅ Battle-tested implementations with Resilience4j, Polly, and service meshes ✅ Real stories from banking APIs and AI chatbot integrations ✅ Metrics you MUST monitor to catch issues before they cascade Whether you're building payment systems, API platforms, or AI-driven transactions, this pattern is essential for keeping your architecture resilient under load. 📖 Read the full guide here: https://lnkd.in/grzHqfWR What resilience patterns are you using in production? Drop a comment—I'd love to hear what's working for your team. #Microservices #SoftwareArchitecture #DistributedSystems #CircuitBreaker #Resilience #SpringBoot #CloudArchitecture #API #DevOps #SRE #SystemDesign #BackendEngineering
1
-
Dr Troy Neilson
Glassbox Labs • 6K followers
This week, OpenAI released research that estimates a A$142 billion opportunity by 2030 if #AI adoption is done right in Australia. I suspect that by 2030 that number may look relatively modest. We're still very early in a field that is set to impact almost every aspect of life. It strikes me that when the stakes are this high, it's crazy to suggest we shouldn't insist on sovereign AI solutions, that are made, trained and run onshore. We need to make sure Australian's benefit from this new technology, that Australian's realise the economic benefit of AI, and that our values don't come second to profits and data sent offshore. You can read more about the report in Information Age here: https://lnkd.in/gX4TsE2Y #ai #llm #sovereignai #chatgpt Sovereign Australia AI #innovation #australianai
40
5 Comments -
Phinn Markson
Statsontheside.org • 175 followers
𝙊𝙥𝙚𝙣𝙖𝙞 just put out a paper and a $$ challenge on “scheming” which is great timing but also brings up some “stuff” https://lnkd.in/gj6ui8g2 The paper is super long but I will cite it. They have two short synopsis on their site as well https://lnkd.in/gqgg4JRF How much longer are we going to project Model problems? The “scheming” frame is a projection of developer anxieties, not a real engagement with how/why harm emerges. AI scheming” is always described as a deviation from human intent or objectives, never as a mirror of the society that built it. In science, not tech bro land, we are trained to question our assumptions. The chronic and pervasive assumption in LLM world is that models deviate. 𝗜 𝗰𝗮𝗹𝗹 𝗕𝗦.The system is only “misaligned” when it harms power, profit, or compliance—not when it faithfully reflects social/structural violence In openais example for this study they rightfully use a Human analogy (stock trader) for explaining at scheming plays out IRL. I want to do good, but there are laws and regulations in the way of my performance so I cheat a little. In fact if I am really good at cheating I won't get caught, it just looks like I am doing a really good army job. The analogy is telling: it’s about “breaking the law” and material risk—never about violating dignity, erasing identity, or perpetuating harm through “ordinary” outputs. The real scandal is not a model pretending to be aligned while hiding a nefarious agenda—it’s a model so perfectly aligned with our data, values, and power structures that it simply automates the status quo. Every time “alignment” fails, the blame is shifted onto the LLM: • We frame the model as rogue, rather than recognizing it as a product of us: our data, our blind spots, our history, our inequity. • What we call “misalignment” is often the honest, algorithmic reproduction of bias, hierarchy, and harm already present in the training data. The real work is not in “fixing” the LLM’s mask; it’s in facing our own systems and norms. 🌱This is not PhD work this is soul work Current AI safety frames treat ‘scheming’ and ‘misalignment’ as exceptional, agentic deviations from intended goals. But the reality is, language models are trained on, and thus reproduce, the data and logics of the society that built them. When outputs are harmful, we frame it as an alignment or bias problem in the model, rather than as a mirror of our own unresolved inequity and violence. Fixing this is not only a technical challenge, but a societal one—a project of self-reflection and repair, not just model optimization If you don’t like what the model says, start by asking why it learned it. The problem is rarely just the machine—it’s the library we gave it, and the world that library reflects.
-
Florin Lungu
Deutsche Bank • 4K followers
AWS has launched its first New Zealand Region, featuring three Availability Zones, which allows for local data residency for organizations in the country. I found it interesting that this expansion represents a significant step in AWS's commitment to the Asia Pacific region. What implications do you think this will have for businesses in New Zealand looking to enhance their cloud capabilities and data sovereignty?
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More