Artificial Intelligence

Announcing DreamDojo: our open-source, interactive world model that takes robot motor controls and generates the future in pixels. No engine, no meshes, no hand-authored dynamics. It's simulation 2.0. Time for robotics to take the bitter lesson pill. Real-world robot learning is bottlenecked by time, wear, safety, and resets. If we want Physical AI to move at pretraining speed, we need a simulator that adapts to pretraining scale. Our key insights: (1) human egocentric videos are a scalable source of first-person physics; (2) latent actions make them "robot-readable" across different hardware; (3) real-time inference unlocks live teleop, policy eval, and test-time planning *inside* a dream. We pre-train on 44K hours of human videos with zero robot-in-the-loop. Humans have already explored the combinatorics: we grasp, pour, fold, assemble, fail, retry—across cluttered scenes, shifting viewpoints, and hour-long task chains—at a scale no robot fleet could match. The missing piece: these videos have no action labels. So we introduce latent actions: a unified representation inferred directly from videos that captures "what changed between world states" without knowing the underlying hardware. This lets us train on any first-person video as if it came with motor commands attached. As a result, DreamDojo generalizes zero-shot to objects and environments never seen in any robot training set, because humans saw them first. Next, we post-train onto each robot to fit its specific hardware. Think of it as separating "how the world looks and behaves" from "how this particular robot actuates." The base model follows the general physical rules, then "snaps onto" the robot's unique mechanics. A world simulator is only useful if it runs fast enough to close the loop. We train a real-time version of DreamDojo that runs at 10 FPS, stable for over a minute of continuous rollout. This unlocks exciting possibilities: - Live teleoperation *inside* a dream. Connect a VR controller, stream actions into DreamDojo, and teleop a virtual robot in real time. We demo this on Unitree G1 with a PICO headset and one RTX 5090. - Policy evaluation. Roll out a policy checkpoint in the neural simulator instead of the real world. DreamDojo's predicted success rates strongly correlate with real-world results — accurate enough to rank checkpoints without burning a single motor. - Model-based planning. Sample multiple action proposals → simulate them all in parallel → pick the best future. Gains +17% real-world success out of the box on a fruit packing task. We open-source everything: weights, code, post-training dataset, eval set, and whitepaper with tons of details to reproduce. DreamDojo is based on NVIDIA Cosmos, which is open-weight too. 2026 is the year of World Models for physical AI. We want you to build with us. Happy scaling! - Website: https://lnkd.in/gmCXJs2V - Paper: https://lnkd.in/gnBHfWVw - Code & models: https://lnkd.in/gnd4AuiG 

Three AI recruiters look at the same 109 CVs. They agree only 14% of the time. That’s not the start of a joke. And that's not efficiency. That’s what I call 'Rank Roulette'. When I tested ChatGPT, Gemini and Grok against the same job spec and anonymised CV set, here’s what happened: • 14% overlap in shortlists → Four times out of five, the models disagreed. • ±2.5 places volatility → Yesterday’s #2 became today’s #5. • 55% of CVs never surfaced → Candidates vanished with no audit trail. • 96% recycled rationales → Fluent, but shallow logic. We’re told by vendors and in-house 'tinkerers' that LLMs can “shortlist in seconds”. The truth: they behave more like over-confident interns - smooth on the surface, but shockingly inconsistent. And the worst part? It’s not even random. In a follow-up piece, I explored why this happens: a technical quirk called batch non-determinism. In plain English: your candidate’s fate changes depending on what else the server was processing at that moment. Until volatility is tamed, hands-off AI screening with LLMs is more than risky. It’s completely unexplainable, indefensible and a governance nightmare. Go to the comments for 👉 Full research 👉 Follow-up on why AI recruiters play favourites

This week MIT dropped a stat engineered to go viral: 95% of enterprise GenAI pilots are failing. Markets, predictably, had a minor existential crisis. Pundits whispered the B-word (“bubble”), traders rotated into defensive stocks, and your colleague forwarded you a link with “is AI overhyped???” in the subject line. Let’s be clear: the 95% failure rate isn’t a caution against AI. It’s a mirror held up to how deeply ossified enterprises are. Two truths can coexist: (1) The tech is very real. (2) Most companies are hilariously bad at deploying it. If you’re a startup, AI feels like a superpower. No legacy systems. No 17-step approval chains. No legal team asking whether ChatGPT has been “SOC2-audited.” You ship. You iterate. You win. If you’re an enterprise, your org chart looks like a game of Twister and your workflows were last updated when Friendswas still airing. You don’t need a better model - you need a cultural lobotomy. This isn’t an “AI bubble” popping. It’s the adoption lag every platform shift goes through. - Cloud in the 2010s: Endless proofs of concept before actual transformation. - Mobile in the 2000s: Enterprises thought an iPhone app was strategy. Spoiler: it wasn’t. - Internet in the 90s: Half of Fortune 500 CEOs declared “this is just a fad.” Some of those companies no longer exist. History rhymes. The lag isn’t a bug; it’s the default setting. Buried beneath the viral 95% headline are 3 lessons enterprises can actually use: ▪️ Back-office > front-office. The biggest ROI comes from back-office automation - finance ops, procurement, claims processing - yet over half of AI dollars go into sales and marketing. The treasure’s just buried in a different part of the org chart. ▪️Buy > build. Success rates hit ~67% when companies buy or partner with vendors. DIY attempts succeed a third as often. Unless it’s literally your full-time job to stay current on model architecture, you’ll fall behind. Your engineers don’t need to reinvent an LLM-powered wheel; they need to build where you’re actually differentiated. ▪️Integration > innovation. Pilots flop not because AI “doesn’t work,” but because enterprises don’t know how to weave it into workflows. The “learning gap” is the real killer. Spend as much energy on change management, process design, and user training as you do on the tool itself. Without redesigning processes, “AI adoption” is just a Peloton bought in January and used as a coat rack by March. You didn’t fail at fitness; you failed at follow-through. In five years, GenAI will be as invisible - and indispensable - as cloud is today. The difference between the winners and the laggards won’t be access to models, but the courage to rip up processes and rebuild them. The “95% failure” stat doesn’t mean AI is snake oil. It means enterprises are in Year 1 of a 10-year adoption curve. The market just confused growing pains for terminal illness.

AI is not failing because of bad ideas; it’s "failing" at enterprise scale because of two big gaps: 👉 Workforce Preparation 👉 Data Security for AI While I speak globally on both topics in depth, today I want to educate us on what it takes to secure data for AI—because 70–82% of AI projects pause or get cancelled at POC/MVP stage (source: #Gartner, #MIT). Why? One of the biggest reasons is a lack of readiness at the data layer. So let’s make it simple - there are 7 phases to securing data for AI—and each phase has direct business risk if ignored. 🔹 Phase 1: Data Sourcing Security - Validating the origin, ownership, and licensing rights of all ingested data. Why It Matters: You can’t build scalable AI with data you don’t own or can’t trace. 🔹 Phase 2: Data Infrastructure Security - Ensuring data warehouses, lakes, and pipelines that support your AI models are hardened and access-controlled. Why It Matters: Unsecured data environments are easy targets for bad actors making you exposed to data breaches, IP theft, and model poisoning. 🔹 Phase 3: Data In-Transit Security - Protecting data as it moves across internal or external systems, especially between cloud, APIs, and vendors. Why It Matters: Intercepted training data = compromised models. Think of it as shipping cash across town in an armored truck—or on a bicycle—your choice. 🔹 Phase 4: API Security for Foundational Models - Safeguarding the APIs you use to connect with LLMs and third-party GenAI platforms (OpenAI, Anthropic, etc.). Why It Matters: Unmonitored API calls can leak sensitive data into public models or expose internal IP. This isn’t just tech debt. It’s reputational and regulatory risk. 🔹 Phase 5: Foundational Model Protection - Defending your proprietary models and fine-tunes from external inference, theft, or malicious querying. Why It Matters: Prompt injection attacks are real. And your enterprise-trained model? It’s a business asset. You lock your office at night—do the same with your models. 🔹 Phase 6: Incident Response for AI Data Breaches - Having predefined protocols for breaches, hallucinations, or AI-generated harm—who’s notified, who investigates, how damage is mitigated. Why It Matters: AI-related incidents are happening. Legal needs response plans. Cyber needs escalation tiers. 🔹 Phase 7: CI/CD for Models (with Security Hooks) - Continuous integration and delivery pipelines for models, embedded with testing, governance, and version-control protocols. Why It Matter: Shipping models like software means risk comes faster—and so must detection. Governance must be baked into every deployment sprint. Want your AI strategy to succeed past MVP? Focus and lock down the data. #AI #DataSecurity #AILeadership #Cybersecurity #FutureOfWork #ResponsibleAI #SolRashidi #Data #Leadership

𝗢𝗻𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗠𝗢𝗦𝗧 𝗱𝗶𝘀𝗰𝘂𝘀𝘀��𝗱 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻: 𝗛𝗼𝘄 𝘁𝗼 𝗽𝗶𝗰𝗸 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝗟𝗟𝗠 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝘂𝘀𝗲 𝗰𝗮𝘀𝗲? The LLM landscape is booming and choosing the right LLM is now a business decision, not just a tech choice. One-size-fits-all? Forget it. Nearly all enterprises today rely on different models for different use cases and/or industry-specific fine-tuned models. There’s no universal “best” model — only the best fit for a given task. The latest LLM landscape (see below) shows how models stack up in capability (MMLU score), parameter size and accessibility — and the differences REALLY matter.  𝗟𝗲𝘁'𝘀 𝗯𝗿𝗲𝗮𝗸 𝗶𝘁 𝗱𝗼𝘄𝗻: ⬇️ 1️⃣ 𝗚𝗲𝗻𝗲𝗿𝗮𝗹𝗶𝘀𝘁 𝘃𝘀. 𝗦𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘀𝘁: - Need a broad, powerful AI? GPT-4, Claude Opus, Gemini 1.5 Pro — great for general reasoning and diverse applications.   - Need domain expertise? E.g. IBM Granite or Mistral models (Lightweight & Fast) can be an excellent choice — tailored for specific industries.  2️⃣ 𝗕𝗶𝗴 𝘃𝘀. 𝗦𝗹𝗶𝗺:  - Powerful, large models (GPT-4, Claude Opus, Gemini 1.5 Pro) = great reasoning, but expensive and slow. - Slim, efficient models (Mistral 7B, LLaMA 3, RWWK models) = faster, cheaper, easier to fine-tune. Perfect for on-device, edge AI, or latency-sensitive applications.  3️⃣ 𝗢𝗽𝗲𝗻 𝘃𝘀. 𝗖𝗹𝗼𝘀𝗲𝗱   - Need full control? Open-source models (LLaMA 3, Mistral, Llama) give you transparency and customization.   - Want cutting-edge performance? Closed models (GPT-4, Gemini, Claude) still lead in general intelligence.  𝗧𝗵𝗲 𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆? There is no "best" model — only the best one for your use case, but it's key to understand the differences to make an informed decision: - Running AI in production? Go slim, go fast. - Need state-of-the-art reasoning? Go big, go deep. - Building industry-specific AI? Go specialized and save some money with SLMs.  I love seeing how the AI and LLM stack is evolving, offering multiple directions depending on your specific use case. Source of the picture: informationisbeautiful.net

𝗧𝗵𝗲 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗦𝘁𝗮𝗶𝗿𝗰𝗮𝘀𝗲 represents the 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗲𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻 from passive AI models to fully autonomous systems. Each level builds upon the previous, creating a comprehensive framework for understanding how AI capabilities progress from basic to advanced: BASIC FOUNDATIONS: • 𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀: The foundation of modern AI systems, providing text generation capabilities • 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀 & 𝗩𝗲𝗰𝘁𝗼𝗿 𝗗𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀: Critical for semantic understanding and knowledge organization • 𝗣𝗿𝗼𝗺𝗽𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴: Optimization techniques to enhance model responses • 𝗔𝗣𝗜𝘀 & 𝗘𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗗𝗮𝘁𝗮 𝗔𝗰𝗰𝗲𝘀𝘀: Connecting AI to external knowledge sources and services INTERMEDIATE CAPABILITIES: • 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗠𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁: Handling complex conversations and maintaining user interaction history • 𝗠𝗲𝗺𝗼𝗿𝘆 & 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗠𝗲𝗰𝗵𝗮𝗻𝗶𝘀𝗺𝘀: Short and long-term memory systems enabling persistent knowledge • 𝗙𝘂𝗻𝗰𝘁𝗶𝗼𝗻 𝗖𝗮𝗹𝗹𝗶𝗻𝗴 & 𝗧𝗼𝗼𝗹 𝗨𝘀𝗲: Enabling AI to interface with external tools and perform actions • 𝗠𝘂𝗹𝘁𝗶-𝗦𝘁𝗲𝗽 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴: Breaking down complex tasks into manageable components • 𝗔𝗴𝗲𝗻𝘁-𝗢𝗿𝗶𝗲𝗻𝘁𝗲𝗱 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸𝘀: Specialized tools for orchestrating multiple AI components ADVANCED AUTONOMY: • 𝗠𝘂𝗹𝘁𝗶-𝗔𝗴𝗲𝗻𝘁 𝗖𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻: AI systems working together with specialized roles to solve complex problems • 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀: Structured processes allowing autonomous decision-making and action • 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗣𝗹𝗮𝗻𝗻𝗶𝗻𝗴 & 𝗗𝗲𝗰𝗶𝘀𝗶𝗼𝗻-𝗠𝗮𝗸𝗶𝗻𝗴: Independent goal-setting and strategy formulation • 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 & 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴: Optimization of behavior through feedback mechanisms • 𝗦𝗲𝗹𝗳-𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗔𝗜: Systems that improve based on experience and adapt to new situations • 𝗙𝘂𝗹𝗹𝘆 𝗔𝘂𝘁𝗼𝗻𝗼𝗺𝗼𝘂𝘀 𝗔𝗜: End-to-end execution of real-world tasks with minimal human intervention The Strategic Implications: • 𝗖𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝘃𝗲 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁𝗶𝗮𝘁𝗶𝗼𝗻: Organizations operating at higher levels gain exponential productivity advantages • 𝗦𝗸𝗶𝗹𝗹 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁: Engineers need to master each level before effectively implementing more advanced capabilities • 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗣𝗼𝘁𝗲𝗻𝘁𝗶𝗮𝗹: Higher levels enable entirely new use cases from autonomous research to complex workflow automation • 𝗥𝗲𝘀𝗼𝘂𝗿𝗰𝗲 𝗥𝗲𝗾𝘂𝗶𝗿𝗲𝗺𝗲𝗻𝘁𝘀: Advanced autonomy typically demands greater computational resources and engineering expertise The gap between organizations implementing advanced agent architectures versus those using basic LLM capabilities will define market leadership in the coming years. This progression isn't merely technical—it represents a fundamental shift in how AI delivers business value. Where does your approach to AI sit on this staircase?

AI holds great potential for the semiconductor industry and will kick-start the next round of innovation for faster, cheaper and more energy-efficient computation – that was my message today at SPIE Advanced Lithography + Patterning. I discussed the potential and the challenges that AI holds for our industry.   The potential is clearly huge. AI is rapidly integrated into applications, and high-performance compute is expected to underpin growth towards $1 trillion of semiconductor sales by 2030. The challenges are around the computing needs of AI models and related energy consumption. The compute workload of training a leading AI model has increased 16x every 2 years in recent years – much faster than the increase in computing power delivered by Moore’s law, which is about 2x every 2 years. The energy needed to train a leading model has not grown so steeply but still rose 10x every 2 years. This computing need has been met by building supercomputers and massive data centers. If you extrapolate these trends, training a leading AI model would need the entire world-wide electricity supply in about 10 years. That’s clearly not realistic, so the trend has to break, by training algorithms becoming more efficient and by chips becoming more efficient. In other words, the needs of AI will stimulate immense innovation in chip design and manufacturing – and the potential value of AI to our society will put urgency and funding behind that drive. As a consequence, chip makers are pulling all levers to accelerate semiconductor scaling. This includes lithographic “2D” scaling: shrinking the dimensions of transistors to pack more into a square millimeter. It will also include “3D” integration, with innovations like backside power delivery, transistor designs like gate-all-around, as well as stacking chips in the package, where holistic lithography will play a critical role to deliver performance requirements. ASML will support these trends through a comprehensive, holistic lithography portfolio. Our 0.33 NA/0.55 NA EUV lithography systems allow chip makers to shrink dimensions at the lowest possible cost on their critical layers, while tightly matched and highly productive DUV systems will continue to reduce cost. More than ever, metrology and inspections tools – whose data is fed into lithography control solutions that keep the patterning process operating within tight specs to deliver the highest possible production yields – will be essential to deliver 2D scaling and 3D integration processes. 3D integration requires wafer-to-wafer bonding, and we have demonstrated the capability to map the stresses and distortions that bonding creates and to compensate for them, reducing overlay errors for post-bonding patterning by 10x or more.   It was a pleasure catching up with the industry’s lithography and patterning experts in San Jose. I’m excited to see our collective innovation power having a go at these challenges. Together, we will push technology forward.

The Irish Government has just announced plans to introduce the Regulation of Artificial Intelligence Bill in its Spring 2025 legislative programme, a pivotal piece of legislation aimed at giving full effect to the European Union’s Artificial Intelligence Act (EU Regulation 2024/1689). Even though the AI Act as a regulation has direct effect, this move is set to shape the national regulatory framework for AI governance in Ireland and establish national enforcement mechanisms in line with the EU’s approach. At the heart of the bill is the designation of Ireland’s National Competent Authorities: the entities that will be responsible for enforcing compliance with the AI Act. These authorities will oversee risk classification, conduct market surveillance, and impose penalties for violations. Given Ireland’s role as the EU base for major technology firms including Google, Anthropic, Meta, and TikTok, the effectiveness of its enforcement regime will be closely scrutinised across the EU and beyond. The Irish Government’s approach will be particularly significant due to the country’s track record in regulating the digital sector. Ireland’s Data Protection Commission (DPC) has wielded considerable influence over EU-wide enforcement of the GDPR, given the presence of multinational tech firms within the state. The DPC was designated as one of ireland’s nine fundamental rights authorities under the AI Act in November 2024. The bill will include provisions for penalties, though details remain unspecified. Under the EU AI Act, non-compliance can result in fines of up to €35 million or 7% of a company’s global annual turnover, whichever is higher. For Ireland, the challenge will be ensuring its enforcement framework has sufficient resources and expertise to oversee AI systems deployed within its jurisdiction. Tech industry leaders and legal experts will be closely monitoring how Ireland structures its national framework. The AI Act imposes strict obligations on high-risk AI applications, including those used in healthcare, banking, and recruitment. Companies will be required to maintain transparency, conduct impact assessments, and ensure that their AI systems do not lead to unlawful discrimination or harm. Ireland’s legislative initiative comes at a time of growing regulatory scrutiny over AI’s impact on society, innovation, and human rights. The AI Act represents the world’s most comprehensive attempt to regulate artificial intelligence, at a time other jurisdictions such as the USA are moving in the opposite regulatory direction. The Regulation of Artificial Intelligence Bill is still in its early stages, at the “Heads in Preparation” point. In the Irish legislative process, the Heads of a Bill serve as a blueprint for the eventual legislation. As Ireland moves toward full implementation of the AI Act, the government’s decisions on AI oversight will have significant implications for businesses, consumers, and the broader EU regulatory landscape.

AI is becoming a make-or-break factor for banks. But success will not depend on their ability to offer #AI, but on their competence in integrating it. Let’s take a look.   Banking is forecasted to feel the biggest impact from generative AI among sectors and industries as a percentage of their revenues with the additional value calculated between $200 bn and $340 bn annually (source: McKinsey). But why is the impact so powerful? One of the main reasons is because the abrupt surge of gen AI is exponentially increasing the speed with which #banking is being transformed. That is not to say that the transformation has started with or due to AI. On the contrary: during the past 10 to 15 years banking was already in the middle of transforming from a human-based, relationship-first industry to a more automated and technology-driven business following the #fintech revolution and the ascend of nimbler and more innovative competitors. But AI now does 2 things: —  It brings the transition to a new level, across 3 dimensions: speed, outcome and impact. —  It turbo-charges one of the biggest challenges in modern FS: the combination of AI and data that brings under the same roof two inherently opposing forces: mass and customization. In other words, AI seems to find a credible answer to achieving hyper-personalization. In a recent report Deloitte has provided realistic examples on how this is done across both cost efficiency and income growth: Cost efficiency: —  Workforce acceleration efficiencies across the board: 0–15% of total staff cost —  IT development and maintenance acceleration: 10–20% of IT staff cost —  Improved credit-risk assessment leading to 10-15% savings in impairment charges —  Improved FinCrime/fraud detection reducing litigation/redress charges and fraud losses Income growth: —  Next generation market analysis / predictive trading algorithms: 5–7% uplift on trading income —  Improved customer retention: 1–2% uplift on fees & commissions —  Improved customer acquisition through hyper-personalised marketing: 5-10% uplift from interest income and fees & commissions —  Tailored loan pricing based on credit risk assessment: 2–3% increase on net interest income Despite all the excitement around these estimated benefits, success will not be a walk in the park. It will depend on the banks’ ability to integrate AI in a seamless way into their day-to-day operations. Going forward AI will be re-writing much of the scenarios and use cases of the banking value chain. That doesn’t necessarily mean that they will all be different, but most will certainly be enhanced with impact spanning both across the back-end and the front-end. Given that resources are limited, one of the main challenges will be how to identify the ones to focus on. Factors such as #strategy, potential impact and a match with the existing skillset should be guiding the selection process.   Opinions: my own, Graphic source and use cases: Deloitte

Last week, I described four design patterns for AI agentic workflows that I believe will drive significant progress: Reflection, Tool use, Planning and Multi-agent collaboration. Instead of having an LLM generate its final output directly, an agentic workflow prompts the LLM multiple times, giving it opportunities to build step by step to higher-quality output. Here, I'd like to discuss Reflection. It's relatively quick to implement, and I've seen it lead to surprising performance gains. You may have had the experience of prompting ChatGPT/Claude/Gemini, receiving unsatisfactory output, delivering critical feedback to help the LLM improve its response, and then getting a better response. What if you automate the step of delivering critical feedback, so the model automatically criticizes its own output and improves its response? This is the crux of Reflection. Take the task of asking an LLM to write code. We can prompt it to generate the desired code directly to carry out some task X. Then, we can prompt it to reflect on its own output, perhaps as follows: Here’s code intended for task X: [previously generated code] Check the code carefully for correctness, style, and efficiency, and give constructive criticism for how to improve it. Sometimes this causes the LLM to spot problems and come up with constructive suggestions. Next, we can prompt the LLM with context including (i) the previously generated code and (ii) the constructive feedback, and ask it to use the feedback to rewrite the code. This can lead to a better response. Repeating the criticism/rewrite process might yield further improvements. This self-reflection process allows the LLM to spot gaps and improve its output on a variety of tasks including producing code, writing text, and answering questions. And we can go beyond self-reflection by giving the LLM tools that help evaluate its output; for example, running its code through a few unit tests to check whether it generates correct results on test cases or searching the web to double-check text output. Then it can reflect on any errors it found and come up with ideas for improvement. Further, we can implement Reflection using a multi-agent framework. I've found it convenient to create two agents, one prompted to generate good outputs and the other prompted to give constructive criticism of the first agent's output. The resulting discussion between the two agents leads to improved responses. Reflection is a relatively basic type of agentic workflow, but I've been delighted by how much it improved my applications’ results. If you’re interested in learning more about reflection, I recommend: - Self-Refine: Iterative Refinement with Self-Feedback, by Madaan et al. (2023) - Reflexion: Language Agents with Verbal Reinforcement Learning, by Shinn et al. (2023) - CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing, by Gou et al. (2024) [Original text: https://lnkd.in/g4bTuWtU ]