Handling Urgent Tasks

The AFLOW model advances agentic workflow optimization by treating it as a search space it explores with Monte Carlo analysis. One outcome is small models outperforming GPT 4o at 4.5% of the cost. There is vast scope for optimization of agentic workflow, due to the unlmiited potential combinations. This is likely to yield more progress than in underlying model development. The researchers are sharing code as well as a detailed paper (link in comments). Some highlights of the paper: 🌟 Enhanced Workflow Adaptability Through Operators. Operators such as "Generate," "Review & Revise," and "Ensemble" act as reusable workflow building blocks in AFLOW, enabling the framework to adapt efficiently to diverse tasks. This modular approach improves search efficiency and ensures robust solutions across six benchmark datasets, underscoring the value of integrating predefined patterns into automated systems. 🔄 Cost-Effective Task Execution with Model Agnosticism. Workflows generated by AFLOW allow less powerful models to outperform larger ones in cost-effectiveness, particularly in high-complexity tasks like GSM8K and MBPP. This scalability in performance at reduced computational costs makes AFLOW a game-changer for deploying AI in budget-sensitive applications. 📊 Iterative Improvement with Monte Carlo Tree Search. AFLOW’s tree-based structure retains successful experiences and avoids redundant failures, facilitating iterative improvements. For example, in GSM8K, AFLOW autonomously crafted workflows that performed similarly to manually designed structures, showcasing its ability to innovate through optimized search strategies. 🔍 Model-Specific Workflow Tailoring. Different language models require tailored workflows for optimal performance. AFLOW demonstrated that workflows optimized for one model (e.g., GPT-4o-mini) might not transfer perfectly to another (e.g., DeepSeek-V2.5), emphasizing the need for context-specific adaptations in AI systems. Code sharing of research such as this is an incredible amplifier of progress. Expect to see others take this excellent work further.

I've tried 100s of time management techniques.  This is by far my favourite: I used to work 80 hrs/week and call it "productive." When really I was: - Attending pointless meetings - Fighting countless small fires - Being involved in every decision Now I work less than 70% the time and get 4x as much done. The Eisenhower Matrix helped me get there.  It teaches you to categorise tasks by importance and urgency. Here's how it works: 1. Do It Now (Urgent + Important) Examples: - Finalise pitch deck before investor meeting tomorrow. - Fix website crash during peak customer traffic. - Respond to press interview request before deadline. Best Practices: - Attack these tasks first each morning with full focus. - Set a strict deadline so urgency fuels execution. 2. Schedule It (Important + Not Urgent) Examples: - Plan quarterly strategy session with leadership team. - Map long-term hiring plan for next 18 months. - Build a personal brand content system for LinkedIn. Best Practices: - Protect time blocks in advance. Never leave them floating. - Tie them to measurable outcomes, not vague intentions. 3. Delegate It (Urgent + Not Important) Examples: - Handle inbound customer service queries this week. - Organise travel logistics for upcoming conference. - Update CRM with latest sales call notes. Best Practices: - Build playbooks so your team executes without confusion. - Delegate with deadlines to avoid wasting time. 4. Eliminate It (Not Urgent + Not Important) Examples: - Tweak logo colour palette again for fun. - Attend generic networking events with no ICP fit. - Review endless “best productivity tools” articles. Best Practices: - Audit weekly. Cut anything that doesn’t compound long-term. - Replace low-value busywork with rest, thinking, or selling. If you are always reacting to what feels urgent,   You'll never focus on what matters. Attend to the tasks in quadrant 1 efficiently,  Then spend 60-70% of your time in quadrant 2.    That's work that actually builds your business. Which quadrant are you spending too much time in right now?  Drop your thoughts in the comments. My newsletter, Step By Step, breaks down more frameworks like this. It's designed to help you build smarter without burning out. 200k+ builders use it to develop better systems. Join them here:  https://lnkd.in/eUTCQTWb ♻️ Repost this to help other founders manage their time.  And follow Chris Donnelly for more on building and running businesses. 

In the face of an overwhelming volume of to-dos, turning to time management as a solution is a dead end. What do people who are really good at time management get? More work! Time management is important, but it's a productivity tool - not a solution to pressure. Instead, take aim at the three things that create volume pressure in the first place: tasks, decisions, and distractions. When you're faced with what feels like an overwhelming pile, consider the following: 1) What tasks have I taken on that are not linked to my major goals? Can they be deferred or deprioritized? 2) What decisions regularly create cognitive load for me? Are there any that can be replaced with policies or principles so I don't need to carefully weigh them each time? 3) How can I use structure to stop relying on will-power to reduce distractions? This can be as simple as a pomodoro timer, going on airplane mode for 30 mins, or physically isolating yourself in a conference room. If you pair time management with task, decision and distraction management you'll have a more sustainable approach over the long haul.

Velox - Execution Engine for Apache Spark, Presto? There are so many different compute engines today. Each engine is optimized for specific workloads, such as - SQL interactive analytics, stream processing, ML feature engineering. At Meta, this led to inefficiencies: different engines had different execution optimizations, inconsistent function behavior & duplicated engineering efforts. To standardize data processing across multiple workloads, Velox was built! It is an open-source C++ execution engine. A typical data engine consists of 5 components: - language frontend - intermediate representation - optimizer - execution engine - execution runtime The 'execution engine' is where the computations happen. Instead of different systems maintaining its own execution logic, Velox provides high-performance, reusable, and extensible components that integrate with existing engines. Core Features/Advantages: ✅ Efficiency: Implements advanced optimizations like SIMD, lazy evaluation, and adaptive query execution. ✅ Consistency: Ensures the same function behaviors across different data engines, reducing discrepancies for users. ✅ Engineering Efficiency: Reduces duplicate efforts by centralizing execution optimizations in one place Real-world applications: - Velox is already integrated into #Presto (Prestissimo) and Spark (Spruce) for SQL analytics, significantly improving performance. - It also powers stream processing (XStream), messaging (Scribe) & ML feature engineering/processing (TorchArrow, F3). Performance Gains? - The paper shows 6-7x speed improvements for SQL workloads over traditional Java-based Presto workers and reduced server usage by 3x, saving resources. - 3x fewer servers were needed to handle the same query workload. If you’re working on data infrastructure, you should take a look at Velox. These are also really interesting work and provides new direction towards modularity in data systems. Paper link in comments. #dataengineering #softwareengineering

Yet another reason estimates are ridiculous. One of the silliest things about time estimates is that the vast majority of time it takes for a team to finish something is spent waiting. For the average development team to create something of value, only 10-20% of the total start-to-finish completion time is spent actively working on the item. The majority of the time is spent waiting. 🔵 Waiting for Reviews 🔵 Waiting for team member hand-offs 🔵 Waiting on other teams or departments So much time is spent waiting… instead of asking, “How much time will it take WORKING to complete this?” You’d be better off asking, “How much time will it take WAITING to complete this?” This, of course, is impossible to answer since most teams have zero control (or even awareness) of waiting time. You’re far, far better off ditching time estimates entirely and focusing on reducing wait states instead. But how? 1] Use Flow Efficiency ↳ Few teams are even aware of the most critical flow metric: Flow Efficiency. ↳ Flow Efficiency tells you how much time is spent actively working on increments of value (features, assets, stories, etc.). ↳ Flow Efficiency (%) = Active Time / Total Time X 100 ↳ Any good workflow tool will calculate your Total Time (Cycle Time). 2] Determine Active Time ↳ To figure out Active Time, you need to track your wait states by adding a “Done” state to every existing stage in your workflow. ↳ For Example: Development -> Development Done -> Testing -> Testing Done -> Review -> Review Done -> Released ↳ The “Done” columns are your wait states.  ↳ Now, you can effectively determine Active Time for each item in your flow vs. Wait Time. 3] Improve Flow Efficiency ↳ Once you can visualize and track wait times, you can focus on fixing the worst offenders. ↳ Add team members, reduce work in progress, remove dependencies… there are many ways to minimize wait states. ↳ Any reduction made to any of your wait states will improve Flow Efficiency An average team will have a Flow Efficiency of 20%. Your team should achieve a Flow Efficiency of 40% or greater to be considered high-performing. Will this take some effort? Of course! But far less effort and total team time (and annoyance) than asking for estimates. Plus, the increase in productivity will far outweigh any loss in imagined predictability.

✍️ Most teams spend millions on AI and still waste hours on busywork. 👋 Real gains start with workflow automation that actually works. Here’s how to make it happen: 1. Map the chaos   ↳ Don’t automate what you don’t understand.   ↳ Draw out every step.   ↳ Spot the manual handoffs and slowdowns.   ↳ Fix the process on paper.   ↳ Then automate. 2. Win fast, win small   ↳ No one will fund a year-long overhaul.   ↳ Grab one painful, repeatable task.   ↳ Automate it with Zapier or a custom GPT.   ↳ Prove results in weeks. 3. Keep people in the loop   ↳ Pure automation is a myth.   ↳ Build workflows where humans can step in, review, or approve.   ↳ Automation should make work easier—not eliminate good people. 4. Track real impact   ↳ Pick simple metrics:   ↳ Time saved.   ↳ Errors cut.   ↳ Output per person.   ↳ Show the numbers.   ↳ Get buy-in and more budget. 5. Let success snowball   ✅ Every win is a case study.   ✅ Document the pain and the payoff.   ✅ Share it.   ✅ Then find the next problem to automate. 👋 Workflow automation isn’t about replacing people or throwing money at software. It’s about discipline. 🎯 Find the pain.   🎯 Fix the steps.   🎯 Automate fast. That’s how you turn AI from hype into real money. What’s your biggest win - or toughest roadblock - in automating workflows? #WorkflowAutomation #AIProductivity #NoCode #AutomationStrategy #DigitalTransformation #FutureOfWork #AIWorkflows #ProcessImprovement

A 2-hour workflow just became 8 minutes. Here's what changed: 𝗧𝗵𝗲 𝗧𝗮𝘀𝗸: Find the best project management tool. For a 50-person team. Under $15K annual budget. Integrates with Slack and Google Workspace. Strong mobile app. 𝗢𝗹𝗱 𝗪𝗮𝘆: Step 1: Research (45 minutes) ↳ Google "best project management tools" ↳ Open 23 tabs ↳ Read 8 comparison articles ↳ Check G2 and Capterra reviews ↳ Visit 12 product websites Step 2: Filter (30 minutes) ↳ Build spreadsheet ↳ Check pricing for each ↳ Verify integrations manually ↳ Read feature lists ↳ Eliminate non-fits Step 3: Deep Dive (40 minutes) ↳ Watch demo videos ↳ Read user reviews ↳ Check mobile app ratings ↳ Look for deal-breakers ↳ Document findings Step 4: Comparison (15 minutes) ↳ Create comparison matrix ↳ List pros and cons ↳ Calculate total cost ↳ Rank options Total: 130 minutes 𝗡𝗲𝘄 𝗪𝗮𝘆: "Find project management tools for 50 people, under $15K annually, with Slack and Google Workspace integration, strong mobile app." ChatGPT shopping research: ↳ Asks clarifying questions (2 min) ↳ Searches across the internet (4 min) ↳ Delivers personalized buyer's guide (2 min) With pricing. With tradeoffs. With reviews. With recommendations. Total: 8 minutes 𝗪𝗵𝗮𝘁 𝗖𝗵𝗮𝗻𝗴𝗲𝗱: Not the research quality. The research speed. ChatGPT read everything you would have. Just 94% faster. 𝗧𝗵𝗲 𝗖𝗼𝘀𝘁 𝗕𝗿𝗲𝗮𝗸𝗱𝗼𝘄𝗻: Your time: $60/hour Old way: $130 in time New way: $8 in time Savings: $122 per decision 𝗦𝗰𝗮𝗹𝗲 𝗜𝘁: Your team makes 50 tool decisions per year. Old cost: $6,500 New cost: $400 That's $6,100 back. Per year. Just on research. 𝗪𝗵𝗮𝘁 𝗧𝗵𝗶𝘀 𝗠𝗲𝗮𝗻𝘀: You're not eliminating research. You're eliminating the boring parts. The tab-switching. The spreadsheet-building. The copy-pasting. What you keep: The judgment. The decision. The validation. 𝗧𝗵𝗲 𝗥𝗲𝗮𝗹 𝗤𝘂𝗲𝘀𝘁𝗶𝗼𝗻: What do you do with those 122 minutes? That's where competitive advantage lives. Not in faster research. In what you build with the time saved. What 2-hour workflow are you compressing? Found this helpful? Follow Arturo Ferreira

Yesterday I announced the start of my new Databricks Performance Tuning program, and the response has been incredible. For those who missed the update, I am dedicating the next three months entirely to this. My goal is to build a resource that moves you beyond just writing code and into engineering high-performance pipelines. This is an advanced specialized program. It assumes you already know PySpark and Databricks. While I have covered performance tuning in my Ultimate Data Engineering program before, this curriculum is completely re-designed to focus on - massive data volumes - real production scenarios - hands-on performance debugging to identify and fix real bottlenecks. We won’t be working with "toy datasets." We are going to deal with huge amounts of data and dive deep into actual debugging. You will learn how to investigate a slow-performing job, identify exactly where it is failing, and apply the right tuning techniques to fix it. Here is exactly what we will cover: - Spark Architecture and Execution Model (Driver, executors, task slots, Understanding Spark UI) - Performance Benchmarking and Bottleneck Identification - Optimized Storage and Data Layout Design (Partitioning strategy, file sizing, and layout decisions for high-performance I/O) - Optimizing Read Performance and Caching Strategies (Spark cache vs disk/IO cache and when each should be used) - Advanced Delta Lake Layout Optimizations - Data Skipping (Z-ORDER, Liquid Clustering, and metadata-driven pruning) - Small File Problem and Automated File Management (Auto Optimize and Optimized Writes) - Identifying and Eliminating Data Skew (Skew detection, skew joins, salting) - Shuffle Minimization and Network I/O Optimization (Reducing wide transformations and using broadcast joins effectively) - Identifying Spill Scenarios and Spill Mitigation Techniques (Memory pressure, disk spills, and tuning execution behavior) - Serialization and Memory Management for Performance (User defined functions, object overhead, and efficient data structures) - Adaptive Query Execution (AQE) (Dynamic coalescing, join strategy switching, and runtime plan optimization) - Choosing the Right Cluster Configuration for Different Workloads (Jobs vs all-purpose clusters and workload-driven sizing) - Selecting the Best Instance Types for Performance and Cost (Compute-optimized vs memory-optimized instances) - Choosing the right cluster size (Resource Estimation) - Databricks Photon Engine and Vectorized Execution - Predictive Optimization and Automated Background Maintenance I am putting everything into this because I want my students to walk away saying this is the best performance tuning course they have ever taken. I am expecting this to be a 7 weeks program. The batch is starting on coming Saturday. It might take around 3 months for me to fully deliver this. If you are ready to master these internals and optimize your production workloads, send me a DM to know more about the program.

Recently helped a client cut their AI development time by 40%. Here’s the exact process we followed to streamline their workflows. Step 1: Optimized model selection using a Pareto Frontier. We built a custom Pareto Frontier to balance accuracy and compute costs across multiple models. This allowed us to select models that were not only accurate but also computationally efficient, reducing training times by 25%. Step 2: Implemented data versioning with DVC. By introducing Data Version Control (DVC), we ensured consistent data pipelines and reproducibility. This eliminated data drift issues, enabling faster iteration and minimizing rollback times during model tuning. Step 3: Deployed a microservices architecture with Kubernetes. We containerized AI services and deployed them using Kubernetes, enabling auto-scaling and fault tolerance. This architecture allowed for parallel processing of tasks, significantly reducing the time spent on inference workloads. The result? A 40% reduction in development time, along with a 30% increase in overall model performance. Why does this matter? Because in AI, every second counts. Streamlining workflows isn’t just about speed—it’s about delivering superior results faster. If your AI projects are hitting bottlenecks, ask yourself: Are you leveraging the right tools and architectures to optimize both speed and performance?