# Who Controls AI Progress? | Mark Esposito, PhD posted on the topic **By:** Mark Esposito, PhD **Published:** 2026-07-14T10:23:08.765Z **Source:** [LinkedIn](https://www.linkedin.com/posts/markesposito_a-new-piece-of-economic-analysis-from-our-share-7482741467405254656-wXt4) --- Who Controls AI Progress? This title was summarized by AI from the post below. Mark Esposito, PhD 6d Report this post A new piece of economic analysis from our work at micro1 with my Harvard-trained colleague LIU ZHANG. It asks a simple question with uncomfortable answers: who actually decides what counts as AI progress, and what happens when the signal runs out? Our argument: as models saturate existing benchmarks, the only real signal left sits in the hardest test items, and those depend on a thin, structurally scarce layer of expert human judgment. Whoever controls that layer holds quiet power over the story of AI progress. Read it here: https://lnkd.in/g7Mk6ndT 133 6 Comments Like Comment Mourad Mazouni 5d Report this comment Mark Esposito, PhD and @Liu Zhang, this raises a deeper evaluation question: Is the benchmark item still the right unit of measurement? In a sustained mathematical research program, the real test is whether a model can preserve the scientific object, assumptions, evidence, and unresolved state across linked decisions. We have just documented a frontier model failing precisely there while continuing to sound fully competent. Item-level scoring would barely see it. Is research-state integrity the missing layer in the current benchmark debate? Like Reply 1 Reaction Chenglin Li 4d Report this comment Hello Micro1 Support, I hope you are doing well. I am experiencing an issue with my Micro1 account. Yesterday, I was able to access my dashboard normally and view my active applications, interview history, and certified skills. However, today, after logging in with the same email address, I am immediately redirected to the “Create Your Profile” page instead of my existing dashboard. I have not created a new profile, as I am concerned that doing so may overwrite or disconnect my existing account. My account previously contained: • Multiple completed AI interviews. • Numerous active job applications under Hiring Manager Reviewing and other statuses. The following certified skills: • Native/near-native Chinese linguistic proficiency (written & verbal) • NLP/AI model output evaluation & feedback • Data annotation, labeling, and categorization for supervised learning • Technical documentation and cross-functional remote collaboration • Mandarin–English Translation & Interpretation • Localization & Cultural Adaptation • Editing/Proofreading & Linguistic Quality Assurance Could you please check whether this is a temporary system issue or an account/profile linking issue, and help restore access to my existing profile? No more previous content No more next content Like Reply 1 Reaction David Brin 5d Report this comment This is explored in my new book on Artificial Intelligence - AIlienMinds summary Optimists foretell a golden age of Al-managed abundance. Doomers cry: vast cyber-minds will crush old style humanity!... or make us irrelevant. Meanwhile, geniuses fostering the artificial intelligence boom clutch clichés rooted in our dismal past... or else in cheap sci-fi. Is there still time for perspective? - on 4 billion years of evolution? - or 60 centuries of feudal stagnation? - or how we handled prior tech revolutions? - or mistakes that keep getting repeated... - or ways this time may be different? From Al-driven unemployment to deceitful images, to hallucinating LLMs and tools for tyrants... to potential wondrous gifts by machines of loving grace... come evade the standard ruts. https://www.davidbrin.com/ailienminds.html Like Reply 1 Reaction micro1 5d Report this comment Great work, Mark Esposito, PhD & LIU ZHANG! Like Reply 2 Reactions 3 Reactions See more comments To view or add a comment, sign in More Relevant Posts Carlos Batista 3w Report this post There is nothing wrong with improving your prompts. But prompt optimisation has a ceiling. Eventually, the quality of AI output is limited less by how clearly you ask and more by how much of your real judgment, context, and decision logic AI can work with. Today’s article explores where that ceiling comes from. The Prompt Trap 👉 https://lnkd.in/duwV9GJG. 1 1 Comment Like Comment To view or add a comment, sign in Adnan Khan 2w Report this post A generative AI platform processing 400 court cases a day, with 90 to 95% of them failing. After SPR rebuilt the foundation, that dropped to 10 to 15% and went live in production. The takeaway: AI doesn't reach production on clever models. It gets there on solid engineering underneath. Read now: https://lnkd.in/gzegBPsx #GenerativeAI #AI #LegalTech 3 1 Comment Like Comment To view or add a comment, sign in Cerebratech Research and Consulting 15 followers 5d Report this post The delegation ladder: a framework for deciding what to hand to AI and what to keep The first time you let an AI choose your meeting time, you might not realize you’ve just made a delegation decision. This seemingly trivial choice mirrors a deeper pattern: as AI infiltrates workflows, we constantly decide what to hand off and what to keep, but often without a framework. Without intentionality, we risk either over-reliance […]Read more at Cerebratech https://lnkd.in/eHsHjKfy 1 Comment Like Comment To view or add a comment, sign in Amanda Kavner 5d Report this post The delegation ladder: a framework for deciding what to hand to AI and what to keep The first time you let an AI choose your meeting time, you might not realize you’ve just made a delegation decision. This seemingly trivial choice mirrors a deeper pattern: as AI infiltrates workflows, we constantly decide what to hand off and what to keep, but often without a framework. Without intentionality, we risk either over-reliance […]Read more at Cerebratech https://lnkd.in/ePqyH_bA 1 1 Comment Like Comment To view or add a comment, sign in Boon AI 73 followers 1mo Report this post A line I keep coming back to: Truth is expensive to produce. Information is cheap to fly around. We've spent thirty years optimizing the second half of that sentence. We have endless information. We have almost no truth. AI is making the imbalance worse, not better — if we let it. Generative AI can produce infinite plausible information at near-zero cost. The signal-to-noise ratio of public discourse is collapsing. The companies that win this decade will be the ones that invert this — producing verified truth, at scale, for AI agents to act on. That's the bet Boon is making. 1 Like Comment To view or add a comment, sign in https://www.xpertsystems.ai/ 4,568 followers 1w Report this post An email in our synthetic enterprise references ticket MER-901. Here's the part that matters: hand someone only that ID — "MER-901" — and the exact ticket rebuilds on demand, byte-for-byte identical, without regenerating anything else in the simulation. We call it reconstruction. Any cited artifact can be rebuilt precisely, forever, from its identifier alone. It's the difference between synthetic data that looks coherent and synthetic data that is coherent — all the way down. https://lnkd.in/gxy9yu2m #SyntheticData #AI #DataInfrastructure Like Comment To view or add a comment, sign in Brendan Sechter 4d Report this post This is why you need to understand the product you are building if AI assistance is used. I quite literally just read the following snippet, and I get responses like this a lot. "The trap redesign I replaced my interval/refinement approach with yours, which is much better — it turns an undecidable semantic question into a syntactic scan:..." This is a case of simply understanding the problem. The more common case is the one where there is nominal shared understanding, but it drops from the AI's context window. In that case, you need a mechanism by which to remind the AI assistant. The crudest explanation is explaining the product. A vaguely less crude strategy is pointing out the right spec document to review. 1 Like Comment To view or add a comment, sign in Eese Kashi Product tech builder specializing in rapid, reliable delivery under ambiguity, with research interest in trustworthy AI interaction systems. 2w Edited Report this post AI - Semantic Interface Just as a graphical interface translates clicks into system events, a semantic interface is another way to translates those events. This also kills a lot of false debates. Instead of: "Can the AI Assistant do X?" The question becomes: "Should the Semantic Interface be allowed to request this transition?" That's a system question. Like Comment To view or add a comment, sign in Ana Fernández Senior SEO Consultant | Fundadora de My Rank Lab | +250M visitas mensuales para Reddit, Adobe y +40 empresas en LatAm y EE.UU. 4d Report this post 𝗜𝗳 𝗔𝗜 𝗻𝗲𝘃𝗲𝗿 𝗺𝗲𝗻𝘁𝗶𝗼𝗻𝘀 𝘆𝗼𝘂𝗿 𝗯𝗲𝘀𝘁 𝗳𝗲𝗮𝘁𝘂𝗿𝗲, 𝗰𝗵𝗲𝗰𝗸 𝗵𝗼𝘄 𝘆𝗼𝘂 𝘁𝗮𝗹𝗸 𝗮𝗯𝗼𝘂𝘁 𝗶𝘁 𝗼𝗻 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝘀𝗶𝘁𝗲. A company asked me why AI tools kept crediting a key capability to their competitors when their product had the same one. I audited their site. The feature was mentioned exactly once, inside a help center article. Their competitors described it on product pages, where AI systems actually pull answers from. AI can only recommend what it can find. I shared this case during a live conversation with David Bell about Previsible 2026 State of AI Discovery Report. David and his team analyzed 6.77 million LLM sessions across 166 properties over 19 months, and it is the most rigorous study on AI referral traffic available right now. If AI discovery matters to your business, start there. What's the full replay -> https://lnkd.in/dRb8KmHN …more 13 1 Comment Like Comment To view or add a comment, sign in Elias Salloum 2w Report this post There is a growing gap between the speed at which AI decisions need to be made and the quality of intelligence those decisions are grounded in. Organizations that close that gap will pull ahead. Those that don’t will spend more to fall further behind. Like Comment To view or add a comment, sign in 41,213 followers 2,285 Posts View Profile Follow More from this author AI Is Unlocking a New Era of Personalized Treatment Mark Esposito, PhD 2mo How AI Is Rewriting the Logic of Health Mark Esposito, PhD 2mo The Art of Breaking AI: Where Security Meets Imagination Mark Esposito, PhD 2mo Explore related topics The Economic Implications Of Automated Decision-Making Why You Need Human Judgment in AI Decision-Making Expert Judgment vs Data-Driven Climate Risk Models Importance of Benchmarks for AI Models Why Testing AI Systems Matters Explore content categories Career Productivity Finance Soft Skills & Emotional Intelligence Project Management Education Technology Leadership Ecommerce User Experience