Insights From Recent AI Research

MIT researchers paired 2,310 people into human-human and human-AI teams to create real ads in a collaborative workspace with some fascinating outcomes—tracking 183K messages, 2m copy edits, and over 5m ad impressions. The paper "Collaborating with AI Agents: Field Experiments on Teamwork, Productivity, and Performance" examined many facets of the dynamics of human-AI collaboration on what was most effective. Some of the valuable insights: 🤖 AI changes how teams talk and work together. Human-AI teams sent 45% more messages than human-only teams, with a focus on task execution—suggestions, instructions, and planning—while human teams sent more social and emotional messages. Despite this shift, both team types rated teamwork quality similarly, showing that collaboration can remain strong even when social interaction drops. 🧍➕🤖 One person plus AI can match or beat human teams. Individuals in human-AI teams produced 60% to 73% more ads than individuals in human-human teams, closing the productivity gap that usually favors groups. Despite having only one human per team, human-AI groups created just as many ads overall as two-human teams. 🧠 Human-AI success depends on psychological compatibility. When a conscientious person worked with a conscientious AI, message volume increased by 62%, signaling better engagement. But mismatches had negative effects—for example, extraverted humans working with conscientious AIs saw drops in text, image, and click quality across the board. 📊 AI lets people shift from doing to directing. Participants in human-AI teams made 60% fewer direct text edits compared to those in human-only teams. Instead of rewriting content themselves, they communicated what needed to be done—refocusing effort from manual changes to guiding and refining AI-generated output. 🔄 AI redistributes cognitive workload and changes who does what. With AI handling routine and complex text generation, humans shifted attention from editing to strategic input and idea generation. This redesigns roles within teams, suggesting new ways to organize work where humans steer, and AI constructs. Humans + AI is the future. This research provides more valuable foundations for understanding how to do this well.

Over the past few days I’ve been reading the Stockholm Resilience Centre’s new report AI for a Planet Under Pressure.  It’s a very comprehensive examination of how AI can support sustainability research, climate science and planetary resilience. It’s an important contribution at exactly the moment when society needs clarity on both the opportunities and the risks of advanced digital technologies. A few reflections from me: AI is already accelerating scientific discovery The report shows that AI is helping researchers uncover patterns in climate, biodiversity, freshwater and urban systems that were previously too complex or computationally demanding to analyse. From high-resolution climate downscaling to modelling Earth system tipping points, AI is opening new windows into how the planet is changing and where intervention is most urgent.  This is not just for relevant for scientific discovery. Climate science is becoming more predictive, more integrated, and more accessible One of the strongest messages is that AI is enabling a shift from isolated models to integrated cross-system understanding. For example, foundation models trained on vast geophysical datasets can support multiple climate-related tasks, from cyclone tracking to air quality forecasts, with far lower computational costs. This has real implications for researchers and public agencies that previously lacked the resources to run such models. But the benefits will depend on how responsibly we apply these tools The report is refreshingly balanced. It highlights very real concerns: the environmental footprint of compute, bias in data and models, uneven global representation in AI research, and the risk of over-reliance on systems that may still contain blind spots. Crucially, it argues that AI must support, not substitute, scientific judgement and local knowledge. What I find most compelling is the call for an “AI for sustainability science” agenda. This means moving beyond pilots and experiments and investing in the infrastructures, skills and governance frameworks that allow AI to strengthen climate research while remaining aligned with planetary boundaries and social equity. In other words: more capability, yes but also more responsibility, transparency and inclusion. For those of us working at the intersection of digitalisation, sustainability and resilience, this report is a timely reminder: AI’s contribution to climate action won’t be measured by novelty, but by whether it helps societies anticipate risks, steward ecosystems, and make better collective decisions under pressure. Well worth a read! https://lnkd.in/eMheXkkx #AI #Sustainability #ClimateScience #DigitalTransformation #Resilience #Research #TechForGood

The AI industry has reached a strange crossroads where models are becoming more capable and more delusional at the same time, literally!! We’ve spent years optimizing for "correctness", but in doing so, we’ve accidentally built a generation of professional guessers. Current RL methods - including those used in the latest reasoning models - tend to treat every correct answer as equal. A model that reasons its way to a solution gets the same "pat on the back" as one that simply gets lucky. This creates a dangerous incentive: never admit doubt. If the goal is always to maximize reward, the model learns that a confident guess is better than a humble "I’m not sure". New research from MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), titled "Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty", addresses this head-on. By introducing Reinforcement Learning with Calibration Rewards (RLCR), researchers are proving that we can penalize overconfidence without sacrificing performance. It turns out that when you reward a model for being honest about its own uncertainty, it actually becomes more reliable across the board. In fields like medicine or finance, a model that claims 95% certainty while being right only half the time is a liability. True intelligence isn't just about the ability to process data; it’s about the self-awareness to know when the data isn't enough. The most exciting takeaway here is that reasoning about uncertainty isn't just a safety feature - it's a fundamental part of thinking. Moving away from binary rewards toward calibrated confidence is how we move from models that just "sound" smart to systems that we can actually trust with high-stakes decisions. Full length paper -> https://lnkd.in/gXm4K6EK #ArtificialIntelligence #MachineLearning #MIT #AIResearch #Reliability #LLMs #DataScience #ReinforcementLearning #ComputerScience #TechInnovation #FutureOfAI #NeuralNetworks #DeepLearning #ResponsibleAI #AIEthics #ModelCalibration #MITCSAIL #DecisionScience #AITrends #TechTrends2026

🚀 AI is no longer just solving problems — it’s creating original solutions. This recent breakthrough from #DeepMind really caught my attention. Their new system, #AlphaEvolve, combines the coding skills of #Gemini with an evolutionary approach — and the result is something we haven’t seen before. It’s not just remixing old code. It’s coming up with completely new algorithms, some of which outperform human-designed solutions that have been around for decades. One example? AlphaEvolve discovered a better way to handle matrix calculations — improving on the Strassen algorithm, which has been standard since 1968. To put things into perspective - that’s 56 years of human progress, beaten by a few weeks of experimentation from an AI. It also designed better algorithms for datacenter scheduling, chip design, and even helped improve the very AI models that power it. A full circle moment. What makes this exciting is the way it works: AlphaEvolve is given a problem and explores countless possibilities, keeping what works and evolving the rest — just like nature. It doesn’t rely on human-written code alone, and in many cases, it finds solutions no human has written before. This is important. For a long time, people assumed that LLMs are just pattern matchers — good at summarizing or rewriting what they’ve seen in training. This shows something different. With the right prompts, the right structure, and continuous feedback, we are now seeing sparks of true creativity in code. For those of us working on software, system design, or AI research — this opens up a new playbook. Instead of just asking AI to “generate” something for us, we can collaborate with it, feeding it ideas and letting it explore possibilities at scale. Like brainstorming with a very smart, very fast co-pilot. And as one of the researchers said — if AI can do this for tight algorithmic problems, imagine what it could do for larger, less-defined challenges in science, business, or policy. This is not about replacing humans. It’s about amplifying our capacity to explore. The bigger question though remains: "Are we ready to design systems that can explore, evolve, and experiment — at a level humans can barely keep up with?" I write about #artificialintelligence | #technology | #startups | #mentoring | #leadership | #financialindependence   PS: All views are personal Vignesh Kumar

For years, fine-tuning LLMs has required large amounts of data and human oversight. Small improvements can disrupt existing systems, requiring humans to go through and flag errors in order to fit the model to pre-existing workflows. This might work for smaller use cases, but it is clearly unsustainable at scale. However, recent research suggests that everything may be about to change. I have been particularly excited about two papers from Anthropic and Massachusetts Institute of Technology, which propose new methods that enable LLMs to reflect on their own outputs and refine performance without waiting for humans. Instead of passively waiting for correction, these models create an internal feedback loop, learning from their own reasoning in a way that could match, or even exceed, traditional supervised training in certain tasks. If these approaches mature, they could fundamentally reshape enterprise AI adoption. From chatbots that continually adjust their tone to better serve customers to research assistants that independently refine complex analyses, the potential applications are vast. In today’s AI Atlas, I explore how these breakthroughs work, where they could make the most immediate impact, and what limitations we still need to overcome.

𝗪𝗲 𝗷𝘂𝘀𝘁 𝗰𝗿𝗼𝘀𝘀𝗲𝗱 𝘁𝗵𝗲 𝘁𝗵𝗿𝗲𝘀𝗵𝗼𝗹𝗱 𝘄𝗵𝗲𝗿𝗲 𝗔𝗜 𝘀𝘁𝗼𝗽𝘀 𝗯𝗲𝗶𝗻𝗴 “𝗲𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹” 𝗳𝗼𝗿 𝘀𝘂𝘀𝘁𝗮𝗶𝗻𝗮𝗯𝗶𝗹𝗶𝘁𝘆 I’m buzzing after seeing our latest research with #Reuters. After years implementing #IndustrialAI for sustainability, watching this shift happen in real-time feels significant. 𝟲𝟯% 𝗼𝗳 𝗼𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝘀 𝗵𝗮𝘃𝗲 𝗺𝗼𝘃𝗲𝗱 𝗯𝗲𝘆𝗼𝗻𝗱 𝗽𝗶𝗹𝗼𝘁𝘀. Implementation rates jumped from 13% to over 50% in a single year. Organizations deploying industrial AI are seeing: ⚪️ 𝟲𝟱% 𝗮𝗰𝗵𝗶𝗲𝘃𝗶𝗻𝗴 𝗲𝗻𝗲𝗿𝗴𝘆 𝘀𝗮𝘃𝗶𝗻𝗴𝘀 of 23% on average ⚪️ 𝟱𝟵% 𝗰𝘂𝘁𝘁𝗶𝗻𝗴 𝗖𝗢𝟮 𝗲𝗺𝗶𝘀𝘀𝗶𝗼𝗻𝘀 𝗯𝘆 𝟮𝟰% But there’s something even more fascinating underneath. 𝗪𝗲’𝗿𝗲 𝗘𝗻𝘁𝗲𝗿𝗶𝗻𝗴 𝗨𝗻𝗰𝗵𝗮𝗿𝘁𝗲𝗱 𝗧𝗲𝗿𝗿𝗶𝘁𝗼𝗿𝘆a I’m watching AI evolve from imitation learning—copying how humans solve problems—to exploration learning. 𝗔𝗜 𝗶𝘀 𝗻𝗼𝘄 𝘁𝗮𝗸𝗶𝗻𝗴 𝘂𝘀 𝗯𝗲𝘆𝗼𝗻𝗱 𝘄𝗵𝗮𝘁 𝗵𝘂𝗺𝗮𝗻 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗮𝗹𝗼𝗻𝗲 𝗰𝗼𝘂𝗹𝗱 𝗮𝗰𝗵𝗶𝗲𝘃𝗲. This isn’t incremental improvement. We’re talking radical innovation. AI can simulate entirely new designs that were previously impossible to conceive. When you’re juggling decarbonization, circularity, and societal changes simultaneously while navigating a “tsunami of regulations” - this capability becomes transformative. 𝗪𝗵𝗮𝘁 𝗞𝗲𝗲𝗽𝘀 𝗠𝗲 𝗨𝗽 𝗮𝘁 𝗡𝗶𝗴𝗵𝘁 (𝗜𝗻 𝗮 𝗚𝗼𝗼𝗱 𝗪𝗮𝘆) 𝟴𝟭% 𝗼𝗳 𝗼𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝘀 𝗯𝗲𝗹𝗶𝗲𝘃𝗲 𝗳𝘂𝘁𝘂𝗿𝗲 𝘀𝘂𝘀𝘁𝗮𝗶𝗻𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗶𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻 𝘄𝗶𝗹𝗹 𝗯𝗲 𝗔𝗜-𝗱𝗿𝗶𝘃𝗲𝗻. Not “AI will help.” But that AI will 𝗱𝗿𝗶𝘃𝗲 innovation. From what I’m seeing? They’re right. We’re using AI to capture regulations requiring hundreds of experts. We’re building multi-agent teams developing entirely new features. We’re optimizing complete process flows. The technical barriers are dissolving faster than expected. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗻𝗼𝘄 𝗶𝘀 𝗵𝗼𝘄 𝗾𝘂𝗶𝗰𝗸𝗹𝘆 𝘄𝗲 𝘀𝗰𝗮𝗹𝗲 𝘄𝗵𝗮𝘁’𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝘄𝗼𝗿𝗸𝗶𝗻𝗴. 𝗪𝗵𝘆 𝗧𝗵𝗶𝘀 𝗠𝗼𝗺𝗲𝗻𝘁 𝗙𝗲𝗲𝗹𝘀 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 For years: “Can AI really deliver on sustainability?” Now: “How fast can we deploy this across our operation?” That shift - from skepticism to urgency - tells me we’ve hit critical mass. The business cases are proven. 71% of leaders expect significant impact on the energy transition. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝘁𝗵𝗲 𝗰𝗼𝗻𝘃𝗲𝗿𝗴𝗲𝗻𝗰𝗲 𝗺𝗼𝗺𝗲𝗻𝘁 𝗜’𝘃𝗲 𝗯𝗲𝗲𝗻 𝘄𝗮𝗶𝘁𝗶𝗻𝗴 𝗳𝗼𝗿. What’s your take? Are you seeing this shift in your work?

Most people are wrong about where AI is right now. Three findings from the latest data flip the conventional wisdom. 1. The best AI products are the ones you barely use. Engagement depth with AI tools in North America fell 38% year over year, even as adoption stayed strong (Mixpanel, 2026). The reason: agent maturity. Tasks that used to take multiple prompts now finish in one run, or run in the background entirely. Every SaaS playbook said high engagement equals product strength. With agents, the opposite is true. 2. Vendor-led AI deployments succeed 2x more than internal builds. The viral stat said 95% of enterprise AI pilots fail. The buried finding from the MIT study told the real story. Vendor-led deployments: 67% success rate. Internal builds: 33% success rate. The “build it yourself” instinct that won the cloud era is the wrong playbook for AI. 3. The biggest AI risk inside companies is cultural, not technical. 54% of C-suite executives say adopting AI is tearing their company apart (Writer, 2026). 92% are cultivating an “AI elite” class of employees. 60% plan layoffs for non-adopters. Leaders preparing for technical migration are missing the harder problem. AI is splitting workforces in half, and the gap widens every quarter. The loudest AI conversations are about models and benchmarks. The quietest ones, about culture, partnerships, and what good usage actually looks like, are where the real advantage is being built. Which of these surprises you most?

🚀 Excited to share my latest Fortune column on truly groundbreaking academic work from my co-authors Professor Karim Lakhani and Fabrizio Dell'Acqua at Digital Data Design Institute at Harvard (D^3), where I serve as an executive fellow. This remarkable field experiment with 776 Procter & Gamble professionals fundamentally challenges what we thought we knew about teamwork. The research reveals the emergence of the "cybernetic teammate"—AI that doesn't just assist but actively participates in collaboration. Three breakthrough findings: 1. AI Can Replicate Team Benefits Individuals working with AI achieved nearly 40% performance gains—matching traditional two-person teams. AI is providing the same collaborative benefits we've long attributed to human teamwork. 2. Cross-Functional AI Teams Generate Breakthrough Innovation AI-augmented cross-functional teams were 3x more likely to produce top 10% solutions. This isn't marginal improvement—it's a multiplicative effect that neither human-only teams nor AI-enabled individuals could achieve alone. 3. AI Breaks Down Silos (For Real This Time) R&D specialists with AI proposed commercially viable solutions. Commercial professionals developed technically sound approaches. AI acted as a bridge, enabling each team member to think holistically across functions—achieving the "silo breaking" that leaders have struggled to accomplish through org chart reshuffles. Bonus finding: AI collaboration increased positive emotions by 64% in teams. This isn't cold, mechanical work—it's energizing and engaging. At Seven2, we're translating this research into practice with our portfolio companies, building these AI-augmented cross-functional teams to drive innovation and competitive advantage. This is the future of collaborative work—not AI replacing humans, but human-AI ensembles that combine the best of both worlds. Read the full analysis: https://lnkd.in/ef3f3pED #AI #Innovation #HBS #D3Institute #FutureOfWork #PrivateEquity #TeamDynamics