<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Jorge Ortiz on Medium]]></title>
        <description><![CDATA[Stories by Jorge Ortiz on Medium]]></description>
        <link>https://medium.com/@jortiz81?source=rss-a13135e021af------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*F8k51NFCY75yV9gEaFCCoA.jpeg</url>
            <title>Stories by Jorge Ortiz on Medium</title>
            <link>https://medium.com/@jortiz81?source=rss-a13135e021af------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 11:53:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jortiz81/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[CRAIG, Competency Awareness, and the Future of AI Governance]]></title>
            <link>https://medium.com/@jortiz81/craig-competency-awareness-and-the-future-of-ai-governance-9c0066bfdba0?source=rss-a13135e021af------2</link>
            <guid isPermaLink="false">https://medium.com/p/9c0066bfdba0</guid>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai-trust]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Jorge Ortiz]]></dc:creator>
            <pubDate>Sat, 13 Dec 2025 19:29:53 GMT</pubDate>
            <atom:updated>2025-12-13T19:45:45.092Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XLZgNv1jRTPByNNok8yonQ.png" /><figcaption>NeurIPS 2025: Competency Awareness and the Future of AI Governance</figcaption></figure><p>I lead the Rutgers technical research program in the NSF-funded <a href="https://moritzlaw.osu.edu/CRAIG">Center on Responsible AI and Governance (CRAIG)</a>, an Industry–University Cooperative Research Center focused on operational AI accountability.</p><p>At Rutgers, we work on a practical bottleneck for governance. We ask how to determine the operating conditions under which an AI system is reliable, and when it is not.</p><p>My lab works on AI for urban infrastructure, smart cities, and human-AI interaction, especially in health-facing settings, where failures carry real consequences. These systems are already deployed, often at scale, without any reliable way to determine when their outputs should be trusted.</p><p>Within CRAIG, we focus on competency-aware evaluation: measurable claims about what a system can do, the conditions under which those claims hold, and what it does when those conditions fail.</p><p>Governance defines what systems are supposed to do. Technical methods determine whether those claims can be tested, reproduced, or enforced. Without the latter, governance tends to reduce to documentation review.</p><h3>Why most AI audits fail before they start</h3><p>“Audit” gets used for two different things, and that ambiguity breaks efforts early.</p><ul><li><strong>Conformance audits</strong>, asking whether a system meets its stated specification.</li><li><strong>Harm audits</strong>, asking whether downstream impacts caused damage.</li></ul><p>Harm cannot be meaningfully assessed if there is no evidence that the system ever behaved according to its stated design.</p><p>Competency evidence is what separates an audit from a documentation exercise.</p><p>Done well, it gives auditors a map. They test the claimed capability region, probe the boundary where performance degrades, and verify out-of-scope behavior (abstain or hand off). Without that map, audits devolve into spot checks with unknown coverage.</p><h3>NYC is legislating algorithmic accountability without tools</h3><p>Cities are writing audit requirements that agencies do not yet have the capacity to execute.</p><p>In November 2025, the New York City Council introduced the <a href="https://www.cityandstateny.com/politics/2025/11/new-york-city-will-try-again-regulate-ai/409803/?utm_source=chatgpt.com">GUARD Act</a> (Guaranteeing Unbiased AI Regulation and Disclosure), which would create an Office of Algorithmic Accountability. The bill tries to make this operational. They will establish a city office that can assess the systems used by agencies, shape procurement, investigate claims, and publish its findings.</p><p>This arrives as <a href="https://time.com/7331036/mamdani-nyc-mayor-election-campaign-promises/">Mayor-elect Zohran Mamdani</a> prepares to take office in January, after campaigning on tighter limits around surveillance and automated decision systems. How his administration handles the Council’s framework is an open question.</p><p>The immediate problem is the absence of usable auditing tools.</p><p>There are not many. Most audits are manual, ad hoc, and fragile, relying on documentation reviews and interviews rather than reproducible evaluation. In financial services, regulators have decades of doctrine grounded in civil rights law and consumer protection statutes; the scaffolding exists, even if it is imperfect. In the public sector, for most algorithmic tools, that scaffolding does not exist yet.</p><p>Regulation is moving faster than the infrastructure needed to carry it out. More than one practitioner described being asked to “audit” a system without access to model internals, without a standardized evaluation protocol, and without a budget for tooling.</p><p><em>CRAIG can help fill that gap by building evaluation protocols and tooling that make conformance audits feasible and make harm audits interpretable.</em></p><p>Agentic systems make this worse. As AI systems take on planning, memory, tool use, and coordination across agents, the number of ways they can fail increases. A single LLM has failure modes you can often bound. A coordinated multi-agent system is a different object. If a multi-agent system cannot represent its own limits, it cannot be audited.</p><p>Berkeley’s <a href="https://sky.cs.berkeley.edu/project/mast/">MAST (Multi-Agent System Failure Taxonomy)</a> is a step toward making multi-agent failures nameable and therefore auditable.</p><p>As we move from single-model deployments to agentic systems, auditing becomes qualitatively harder. Without a taxonomy, you cannot even say what you are auditing.</p><h3>Why government partnerships stall in practice</h3><p>The binding constraints are procurement, budgets, and compliance, not model access or model quality.</p><p>Government agencies often cannot pay membership fees. Multiple practitioners said there is no budget line for it, and in some cases they are barred from paying it.</p><p>The practical workaround is contracting. Agencies can pay for audits, assessments, evaluation frameworks, and training and workforce development. Membership fees are not reimbursable; services are. For CRAIG, that points to a services model for government and a membership model for private industry.</p><p>Federal agencies are different. They have more flexible procurement authorities and may be able to join as members.</p><p>Data access is the other hard constraint.</p><p>Before any agency shares real data, we need infrastructure that meets HIPAA, FERPA, and various state and local data protection requirements. A high-security data environment is mandatory. Without it, serious partnerships do not happen.</p><p>It is unglamorous work, but without it the data never leaves the agency.</p><p><strong>GovAI Coalition (San José)</strong> is a practical convening point where public-sector practitioners share insights and experiences on what they can actually implement.</p><p><strong>Town+Gown:NYC</strong> is a contracting path through NYC DDC that reduces the paperwork overhead that usually stalls university–city work.</p><h3>Zeynep Tufekci on the wrong nightmares</h3><p>Zeynep Tufekci argued that we are focused on the wrong AI nightmares.</p><p>She argued we should worry about the collapse of proof — effort, authenticity, accuracy, sincerity, and even whether something is human-made.</p><p>These signals tell us whether something is real, meaningful, or human. AI is eroding each one. For example, synthetic job applications, deepfake video, chatbots using “I” without any lived experience, hallucinated citations, and art generated without human intention.</p><p>Her argument is that if society cannot authenticate anything, people will accept mass surveillance as the only alternative.</p><p>Her proposed solutions involve cryptographic methods, zero-knowledge proofs, hardware attestation, and modernized public key infrastructure.</p><p><em>Authentication provides external verification. Competency-aware evaluation provides internal boundaries for when a system’s outputs should be trusted.</em></p><p>CRAIG should treat agentic failure characterization and secure data infrastructure as primary deliverables, not supporting work. Government engagement should default to contracted services rather than membership, and we should formalize practitioner input through an external advisory board.</p><h3>Where this leaves CRAIG</h3><p>Laws are emerging faster than technical frameworks, and audit requirements are being written before anyone has reliable tooling to execute them.</p><p><em>Competency-aware evaluation makes conformance auditable at all.</em> <em>Conformance evidence makes harm assessment interpretable.</em></p><p>Practitioners need this, even if they are not yet asking for it by name.</p><p><strong>If you are being asked to defend a conformance evaluation for a deployed system, email me at </strong><a href="/service/mailto:jorge.ortiz@rutgers.edu"><strong>jorge.ortiz@rutgers.edu</strong></a><strong>.</strong> The first step is to write down the system’s claimed specification, identify what evidence exists, and produce a protocol that a third party can rerun.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9c0066bfdba0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[NeurIPS 2025: What I Actually Learned in San Diego]]></title>
            <link>https://medium.com/@jortiz81/neurips-2025-what-i-actually-learned-in-san-diego-1576439edc50?source=rss-a13135e021af------2</link>
            <guid isPermaLink="false">https://medium.com/p/1576439edc50</guid>
            <category><![CDATA[machine-leanring]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[neurips-2025]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jorge Ortiz]]></dc:creator>
            <pubDate>Sat, 06 Dec 2025 15:38:14 GMT</pubDate>
            <atom:updated>2025-12-06T16:18:14.467Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*humRlaMv5vkvyRWXP4AN_g.png" /></figure><p><em>Jorge Ortiz | December 2025</em></p><p>I just got back from NeurIPS 2025 in San Diego. Two days of keynotes, poster sessions, and way too many hallway conversations. Here is what stuck with me, and a few things I did not expect.</p><h3>The scaling wall is real, and people are finally admitting it</h3><p>Everyone has been repeating “scale is all you need” for years. This year, that narrative started to crack. Yejin Choi’s keynote made the point clearly: brute-force scaling is hitting a wall.</p><p>The problem is not compute. If you have money, you can buy compute. The problem is data. The internet has been mined. There is not enough high-quality reasoning data left to sustain naive scaling, and synthetic data pipelines bring their own failure modes.</p><p>In the Prismatic Synthesis work, Choi’s team generated synthetic math data, then threw away 70 to 90 percent of it. Without aggressive filtering you get mode collapse. That is, synthetic data that all looks the same. <em>Their approach was to measure diversity in gradient space instead of just in text space. </em>The key finding was that diversity in gradient space tracks out-of-distribution generalization much better than simple text diversity metrics.</p><p>This aligns with what I have been thinking for <strong>competency-aware systems</strong>. If you train on homogeneous data, you build a system that is confidently wrong exactly where it has not seen enough variation. <em>For high-stakes settings, that is exactly what we cannot afford.</em></p><h3>The RL debate got heated</h3><p>I arrived at NeurIPS expecting reinforcement learning to be the default answer for “how to make models reason.” RLHF, RLAIF, and related recipes are everywhere. What I actually saw was a community that is very unsure whether RL works in the way people claim.</p><p>A few papers stood out.</p><p><strong>“Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?”</strong><br> The uncomfortable claim here is that RL is not creating new capabilities. The base model already encodes the right answers somewhere in its distribution. RL mostly reshapes the distribution so that those answers show up more often at the top. That is useful, but it is not the same as learning new reasoning skills.</p><p><strong>“Spurious Rewards” from UW</strong><br> This was even more unsettling. The authors gave models random rewards, or flat-out incorrect rewards, and still saw performance gains on Qwen. The same did not hold for Llama or OLMo, which makes it even stranger. Something about Qwen’s pre-training and inductive biases interacts with RL in a way we clearly do not understand.</p><p><strong>“Artificial Hive Mind” (Best Paper)</strong><br> This is the one that will stick with me. The authors showed that major LLMs give nearly identical outputs for open-ended queries. This is true within a model, across samples, and across different models. Ask a bunch of models to write a metaphor for time and you basically get two clusters: “time is a river” and “time is a thief.” That is the entire effective space of answers, for a concept that humans have explored for centuries.</p><p>Then there is Choi’s Pro-RL work. With a lot of effort, careful hyperparameter tuning, entropy control, and long training runs, they pushed a 1.5B model into the performance range of distilled 7B models. RL can move the needle, but only when you treat it as a serious engineering project, not a plug-and-play trick.</p><p>My takeaway: low-effort RL is mostly useless. High-effort RL can be powerful. The “RL does nothing” papers and the “RL is great” papers are both correct. They just live at different points on the effort axis.</p><h3>Rich Sutton on superintelligence from experience</h3><p>I only caught the second half of Rich Sutton’s talk on the Oak Architecture, but even that was enough to give me a lot to think about.</p><p>The idea is to design a fully domain-general agent that learns everything from experience, discovers its own abstractions, and avoids domain-specific inductive biases as much as possible. The Bitter Lesson taken seriously.</p><p>The part that stayed with me was the feature ranking mechanism. The agent:</p><ul><li>Generates new features from existing ones</li><li>Ranks features by some priority criterion</li><li>Creates sub-problems for highly ranked features</li><li>Learns options, meaning temporally extended actions, for each sub-problem</li></ul><p>That ranking step quietly assumes some meta-knowledge of what the agent does and does not know. You cannot assign learning priority without some notion of competency gaps. Sutton did not use that language, but the concept is there.</p><p>What he did not answer is how the agent discovers reward signals. He spoke about learning rewards from experience. Philosophically, that is a hard problem. You do not get “ought” from “is.” At some point, terminal values have to come from outside the system.</p><p>The other tension is between RL and causal inference. RL hides confounders behind the Markov assumption. If your state representation is rich enough, the past shields you from hidden causes. If it is not rich enough, you still proceed as if it were, and you fail silently. Causal models, in contrast, force you to declare what you think the structure is and what is or is not identifiable. That explicitness makes them look weaker, but it also makes them more honest.</p><p>The thing I keep circling back to is: why does the community tolerate opaque confounders in RL but treat them as disqualifying in causal inference? I suspect it is because RL typically claims “this policy is optimal given this state representation,” while causality claims “this is the actual effect.” The latter claim is easier to falsify and harder to walk back.</p><h3>Numerical non-determinism is worse than I thought</h3><p>This was the surprise of the conference for me. I walked into a talk expecting a routine GPU kernel optimization session. Instead, the authors showed something that essentially undermines a large chunk of current evaluation practice.</p><p>Change the batch size and the model output changes. Change the number of GPUs and the output changes. Change the GPU version and the output changes. All of this under greedy decoding, where you expect determinism.</p><p>The root cause is floating-point non-associativity. Different hardware and kernel implementations accumulate sums in different orders. Tiny rounding differences propagate through the network. For reasoning tasks where multiple tokens have nearly identical probabilities, these small differences can flip the argmax.</p><p>The implication is that 1 or 2 percent differences on benchmarks may simply be hardware noise. The authors proposed more deterministic kernels with fixed accumulation order, but those come with performance costs.</p><p>This is unglamorous work. It is infrastructure. My lab will not be competitive there, but we need to understand the limitations so we do not over-interpret small performance deltas.</p><h3>The governance conversations</h3><p>I had several confidential conversations about AI governance during the conference. I will not quote anyone or share specifics, but there were clear patterns.</p><p>The first is the enormous gap between principles and implementation. “Do not deploy biased AI” is easy to say. It is much harder to define, for a given system, what that means operationally and how to test for it.</p><p>A framing that stuck with me distinguishes two audit types:</p><ol><li>Conformance audits: Did the system do what it was designed to do?</li><li>Harm audits: Did the system cause real-world harm?</li></ol><p>Most of the AI ethics discussion focuses on harm audits and largely skips conformance. That is a problem. You cannot evaluate real-world harm if you do not know whether the system even behaved according to its own specification.</p><p>The second theme is infrastructure. High-security data platforms that satisfy healthcare, education, and social service regulations simultaneously are rare. Without that infrastructure, government agencies will not share deployment data, so we cannot study these systems at scale.</p><p>The third is legal and financial friction. Many government entities cannot legally pay membership fees to join research consortia. Budget categories do not exist for that. You end up needing to structure engagements as contracts for services with specific deliverables, even when what you want is a long-term research partnership.</p><p>Most academics underestimate that complexity.</p><h3>Papers to actually read</h3><p>Here is what I am telling my lab to read.</p><p><strong>On RL and reasoning</strong></p><ul><li>“Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?”</li><li>“Artificial Hive Mind” (Best Paper)</li><li>Pro-RL / Prolonged Reinforcement Learning</li><li>“Spurious Rewards” from UW</li></ul><p><strong>On data synthesis</strong></p><ul><li>“Prismatic Synthesis”</li><li>“Open Thought 3”</li></ul><p><strong>On spatial and embodied AI</strong></p><ul><li>“SAVI: Spatial Awareness via Seen and Hearing”</li><li>“Real-World Reinforcement Learning of Active Perception Behaviors”</li></ul><p><strong>On numerical stability</strong></p><ul><li>“Understanding and Mitigating Environmental Sources of Non-Determinism in LLM Inference”</li></ul><h3>What this means for my lab</h3><p>My lab works on AI for high-stakes human systems: healthcare, urban infrastructure, public services. Domains where confident errors have real consequences.</p><p>Everything I saw at NeurIPS pushes in the same direction. Competency awareness, meaning systems that know what they do not know, is the right framing. The community is dancing around it without naming it.</p><ul><li>Sutton’s feature ranking implicitly depends on it.</li><li>The Artificial Hive Mind result shows what happens when systems lack mechanisms to explore alternative reasoning paths.</li><li>Governance discussions keep running into the fact that deployed systems cannot signal when they are out of their depth.</li></ul><p>The connection I keep returning to is between competency awareness and causal modeling. Counterfactual reasoning lets you explore “what if” scenarios that sit outside the observed training distribution. Natural experiments can validate or invalidate causal hypotheses. This provides a structured way to describe where a model is competent and where it is not.</p><p>Concrete action items for my group:</p><ul><li>Draft a “Competency-Aware AI Manifesto” that defines the concept and outlines a research agenda.</li><li>Write the “Hidden Confounders of RL” position paper and push that discussion into the open.</li><li>Survey calibration and uncertainty benchmarks to identify where they fail for high-stakes settings.</li><li>Use our mobile robot platform to collect a competency dataset, with tasks that have clear success and failure criteria across varied conditions.</li></ul><p>No one else is going to name this problem for us. If we think competency-aware AI is central, we need to define it and demonstrate it.</p><h3>Bottom line</h3><p>NeurIPS 2025 felt like a turning point. The field is starting to recognize that scaling is not a free lunch, RL is not automatic reasoning, mode collapse is a real threat to diversity, and our infrastructure is less solid than we pretend.</p><p>For researchers without access to nine-figure compute budgets, the opportunity is in new architectures, new problem formulations, and new evaluation frameworks. Competency-aware AI sits exactly in that space.</p><p>We have plenty of work to do.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1576439edc50" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Next steps in reasoning]]></title>
            <link>https://medium.com/@jortiz81/next-steps-in-reasoning-4d4f5d94e1fa?source=rss-a13135e021af------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d4f5d94e1fa</guid>
            <category><![CDATA[sensing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[lrm]]></category>
            <dc:creator><![CDATA[Jorge Ortiz]]></dc:creator>
            <pubDate>Sun, 15 Jun 2025 14:40:23 GMT</pubDate>
            <atom:updated>2025-06-15T14:40:23.616Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Reasoning next steps…" src="https://cdn-images-1.medium.com/max/1024/1*183Qql9AwrNO8O80F3px7A.png" /></figure><p>Recently, I was interviewed by the Wall Street Journal about my thoughts the Large Reasoning Models for its feature titled <strong>“</strong><a href="https://www.wsj.com/tech/ai/artificial-superintelligence-overestimation-3f954065?st=rZyqSR"><strong>Why Super-intelligent AI Isn’t Taking Over Anytime Soon</strong>.</a><strong>”</strong></p><p>My takeaway: while today’s language models perform impressive wordplay, they underperform once each logical step is audited by code, as the problems get harder (or it tries to solve problems that are out of distribution). Like many in the community, I’ve spent the last few days dissecting how these models reason, where they break down, and how that insight steers my lab’s research.</p><h3>What is reasoning?</h3><p>Reasoning is the process by which a system constructs a coherent, logically valid sequence of steps (intermediate representations, decisions, or actions) that transforms an initial state into a goal state, constrained by a formal or learned rule set.</p><p>The sequence of steps can be either <strong>visible</strong> or <strong>opaque</strong>. Visible means every intermediate representation can be inspected; opaque means we see only the final goal state. Yet <strong>visible ≠ verifiable</strong> — seeing a chain of thought doesn’t prove the model relied on it. Large language models give us <strong>visibility without verifiability</strong>, whereas <a href="https://en.wikipedia.org/wiki/A*_search_algorithm">classical planners like <strong>A*</strong></a> provide both, because every intermediate state can be recomputed and checked. More specifically, A* logs each node expansion, cost update, and heuristic estimate. Because every entry is explicit, we can audit every step instead of just the endpoint. On a grid, it evaluates each candidate with <strong><em>f(n)=g(n)+h(n)</em></strong>, where <strong><em>g(n)</em></strong> is the exact cost from the start to <em>n</em> and <strong><em>h(n)</em></strong> is an admissible estimate of the cheapest cost from <em>n</em> to the goal. Expanding nodes in ascending <em>f-order</em> guarantees the first goal popped lies on an optimal path when <em>h</em> never over-estimates, yielding an explicit, logically valid — and verifiable — sequence from start to goal. Shojaee et al. stress that their step-by-step “verification” works only when each textual move can be mapped to a domain-specific simulator; outside such controlled settings, the trace remains visible but <strong>still not fully verifiable</strong>. This assumption narrows applicability to domains with deterministic simulators.</p><h3>Opacity in Practice</h3><p>The downside of opacity shows up in many high-profile systems. Models whose internal rollouts satisfy the definition of “reasoning” yet remain unverifiable include <a href="https://arxiv.org/pdf/2302.04798"><strong>MuZero</strong></a> (Schrittwieser et al., 2020), <a href="https://ai.meta.com/research/publications/v-jepa-2-self-supervised-video-models-enable-understanding-prediction-and-planning/"><strong>V-JEPA 2</strong></a> (Assran et al., 2024), and classic policy-gradient agents like <a href="https://arxiv.org/abs/1707.06347"><strong>PPO</strong></a> (Schulman et al., 2017).</p><p><strong>MuZero</strong> learns its own compact latent transition and value functions, then runs Monte-Carlo tree search entirely in latent space; observers see only the chosen action and final game score, not whether each simulated step obeyed real dynamics or consistent logic.</p><p><strong>V-JEPA 2</strong> encodes video frames into high-dimensional vectors, predicts several steps ahead in that embedding space, scores each hypothetical future with a learned value head, and selects the action path with the best score. The roll-forward chain exists only as sequences of 1 216-D tensors; no external constraint forces those tensors to correspond to physically plausible scenes, so the “reasoning” trajectory cannot be audited except by checking the final prediction against ground truth.</p><p>Even a standard policy-gradient agent such as <strong>PPO</strong> keeps an LSTM or transformer hidden state that aggregates past observations, planned sub-goals, and value estimates; the hidden state updates form an implicit thought process that guides the next action, but the vectors lack semantics and are therefore opaque. Post-hoc probing usually recovers only coarse correlations. The intermediate steps are not verifiable, even though they satisfy the reasoning definition by transforming the current state toward the goal under learned update rules.</p><h3>Why LLMs are different?</h3><p>LLMs are special because they <a href="https://arxiv.org/abs/2206.07682">have displayed emergent abilities</a> — capabilities that appear only once the model is sufficiently large — that were not explicitly designed for during training. Reasoning is one of them. When prompted with <em>“think step by step”</em> or similar cues, models emit consistent chains of intermediate states (“<a href="https://arxiv.org/abs/2201.11903">chain-of-thought</a>”) that let observers follow the trajectory from prompt to answer. Because these traces are visible text, researchers can <strong>measure, supervise, and distill</strong> the reasoning process into smaller models, spawning work on CoT verification, self-consistency, and CoT distillation [<a href="https://arxiv.org/html/2504.01738v2">https://arxiv.org/html/2504.01738v2</a>]. That visibility is the key distinction: unlike MuZero’s latent rollouts or PPO’s hidden states, LLM reasoning can be surfaced and improved.</p><h3>Media Hype vs. Academic Pushback</h3><p>Mainstream coverage treated emergent abilities as a mile marker on the road to AGI. <strong>Quanta</strong> likened the abrupt appearance of new skills <a href="https://www.quantamagazine.org/how-quickly-do-large-language-models-learn-unexpected-skills-20240213/">to a phase transition in scale</a>, implying that once the model is big enough, new capabilities ignite. <strong>WIRED</strong> <a href="https://www.wired.com/story/google-gemini-2-ai-assistant-release/">amplified the claim</a>, quoting Demis Hassabis on Gemini’s reasoning flashes as direct evidence that “true AGI” is now within striking distance. <strong>Politico</strong> <a href="https://www.politico.com/news/magazine/2023/05/08/manhattan-project-for-ai-safety-00095779">pushed the stakes further</a>, calling for a “Manhattan Project for AI safety” on the grounds that nations able to steer these emergent systems first could seize a decisive strategic edge.</p><p>Skeptics in academia, however, resisted the wave. A Stanford-led study, <a href="https://arxiv.org/abs/2304.15004">“Are Emergent Abilities of Large Language Models a Mirage?”,</a> showed that the sharp “jumps” vanish once you replace discontinuous evaluation metrics with smooth ones, concluding that emergence is often a statistical artifact rather than a real phase transition. Gary Marcus and Subbarao Kambhampati <a href="https://www.theguardian.com/commentisfree/2025/jun/10/billion-dollar-ai-puzzle-break-down">pointed to fresh puzzle-solving</a> failures as proof that scale alone doesn’t confer logic, calling AGI talk “overheated marketing.” Princeton’s Arvind Narayanan and Sayash Kapoor went further, branding the <a href="https://www.aisnakeoil.com/p/ai-scaling-myths">entire scaling narrative an “AI-scaling myth”</a> and arguing there is “virtually no chance” that bigger LLMs will reach general intelligence without new ideas.</p><h3>What Apple Found</h3><p>These opaque-yet-“reasoning” systems show why we need hard evidence on where LLMs/LRMs succeed or stall . The recent paper by <a href="https://machinelearning.apple.com/research/illusion-of-thinking">Shajaee et al. (Apple</a>) does a very good job of highlighting the properties and limitations of language models through a thorough empirical study. The limitations they found are as follows:</p><ol><li>LRMs do worse than LLMs at solving low-complexity problems due to “overthinking.” In fact, they tend to produce early tokens that solve the problem and proceed to use their token budgets to continue thinking, and this leads them astray.</li><li>LRMs outperform LLMs at medium-complexity problems. Here is where the thinking tokens seem to come in handy. The extra tokens used to generate intermediate steps help them eventually arrive at a solution.</li><li>LRMs and LLMs both fail at high-complexity problems, and the inflection point in performance for LRMs tends to happen around the same point in their reasoning steps. LRM performance on the Tower of Hanoi problem drops from 90% to 0% when the problem goes from 5 disks to 8 disks.</li></ol><p>There are some other noteworthy observations made in the paper. For example, they find that LRMs quit on solving hard problems and overthink easy ones. Why might this be? The thought is that the RL training, datasets used for training, and architectural limitations related to the attention mechanism and context window size limit the ability of the model to note how information is related to the goal, especially in long sequences.</p><h3>Where do we go from here?</h3><p>My position is pragmatic: Chain-of-thought alone plateaus after a few hops. What moves model capabilities forward is <strong>tool-augmented reasoning</strong> plus <strong>dense, task-shaped feedback</strong>. <a href="https://arxiv.org/abs/2210.03629"><strong>ReAct</strong></a> interleaves thoughts with API calls; <a href="https://arxiv.org/abs/2302.04761"><strong>Toolformer</strong></a> lets the model decide <strong>when</strong> to call external tools; <a href="https://arxiv.org/abs/2305.10601"><strong>Tree-of-Thought</strong></a> and <a href="https://arxiv.org/html/2505.16265v1"><strong>Think-RM</strong></a> add systematic search and long-horizon reward modelling.</p><p>Shajaee et al.’s error profile suggests three pain points and three training levers that already show promise.</p><p><strong>First, large reasoning models tend to overthink the easy cases. </strong>The fix is to insert a rule-checker between each thought and the next action. Whenever a candidate step fails a deterministic verifier — a symbolic planner, a type checker, even a unit-test harness — the model is forced to back up and try again. <a href="https://arxiv.org/abs/2406.02791">PSALM</a> follows exactly this pattern: the LLM proposes a move, the planner validates it, and any illegal moves are pruned away until only a provably correct plan survives.</p><p><strong>Second, the same models give up too early on hard problems.</strong> Handing them a larger context window, as Shajaee et al. did, isn’t enough. Models leave that extra budget unused. The fix is two-fold. <strong>Token-budget scheduling</strong> must be paired with a <strong>use-it-or-lose-it reward</strong>: the agent earns credit for applying additional tokens to make measurable progress on tougher instances and is penalised for bailing out early. <a href="https://arxiv.org/abs/2412.18547">TALE</a> takes a step in this direction by allocating tokens in proportion to difficulty, but the next step is to train the model to recognise that extra budget as a resource it is expected to spend. With reward shaping that favours longer, constructive traces on hard tasks — and still trims verbosity on easy ones — we can keep the model engaged without encouraging rambling.</p><p><strong>Third, even when a chain of thought starts well, it often drifts off course late in a long proof.</strong> Dense, step-level rewards can keep the trajectory on track. Instead of scoring the output once at the end, a progressive reward model grades every intermediate step and nudges the search back toward validity whenever it strays. <a href="https://arxiv.org/abs/2310.10080">PRM</a> applies this idea and outperforms vanilla chain-of-thought on both GSM8K and MATH by rewarding each correct sub-derivation, not just the final answer.</p><p>Together, these techniques show a path beyond today’s limits: visible reasoning that stays faithful, efficient, and coherent all the way to the solution. This closes the loop between visible reasoning and verifiable action — the path I believe will push LLMs past their current limits.</p><h3>SnR Case Study: Tool-Augmented Reasoning</h3><p>My lab — The <strong>Sensing and Reasoning Lab</strong> (SnR) at Rutgers University — studies how to fuse rich, real-world sensing with advanced reasoning to literally “make sense” of the environment. For instance, our <a href="https://dl.acm.org/doi/abs/10.1145/3722570.3726890">SASS</a> edge stack ingests LiDAR, radar, and multi-camera video in real time, then <a href="https://ieeexplore.ieee.org/document/10893687">GeXSe</a> overlays each detection with a frame-level explanation.</p><p>All three fixes highlighted in the previous section fit neatly into our smart-space pipeline. The ReAct-style agent we’re building will: generate a thought, invoke SASS APIs or run A* in the digital twin if the verifier flags uncertainty, receive structured feedback, and continue reasoning with updated context and a fresh token budget.</p><p><strong>But seeing the world is not enough; we also need models that can reason over those multimodal streams.</strong> Chain-of-thought is powerful scaffolding, yet it stalls after a handful of hops. The remedy is explicit <strong>tool use</strong> plus reward structures that stretch the reasoning horizon. Work such as <strong>ReAct</strong>, which interleaves thoughts with API calls, and <strong>Toolformer</strong>, where the model teaches itself when to invoke outside tools, shows that language agents can ground their reasoning in verifiable actions. Frameworks like <strong>Tree-of-Thought</strong> and <strong>Think-RM</strong> push this further, adding systematic search and long-horizon reward models to keep an agent coherent over dozens of steps.</p><p><strong>We are operationalising that recipe in two testbeds.</strong> First, a <strong>triadic-dining</strong> table where the agent, a tabletop robot, and a diner negotiate utensil placement. Second, an <strong>in-cabin hand-off</strong> scenario where the agent explains lane-change decisions to a driver before transferring control. In both cases, the pipeline is the same: SASS delivers a fused scene graph, GeXSe tags it with visual rationales, and a ReAct-style agent reasons until it hits uncertainty. At that point, the model calls domain tools — sass.query_bbox() for a pedestrian pose, dtwin.astar() for a safe reroute, or edge.actuate(&quot;crosswalk&quot;, +7 s) to extend a signal, and folds the result back into its chain of thought. We train on a curriculum of multi-step tasks (“detect a near-miss, diagnose the cause, simulate a fix”) and score each episode on causal-inference accuracy <em>and</em> downstream control success.</p><p>The goal: <strong>turn visible reasoning into tool-verified action in live, sensor-rich environments.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d4f5d94e1fa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Reflections on the CMU LLM Agent (popup) workshop]]></title>
            <link>https://medium.com/@jortiz81/reflections-on-the-cmu-llm-agent-popup-workshop-f3b152de9f92?source=rss-a13135e021af------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3b152de9f92</guid>
            <dc:creator><![CDATA[Jorge Ortiz]]></dc:creator>
            <pubDate>Wed, 08 May 2024 03:25:04 GMT</pubDate>
            <atom:updated>2024-05-08T03:34:09.797Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/834/1*dhINTUHJV9bMT24xBJsxwg.png" /></figure><p>Last week, I had the opportunity to attend the <a href="https://cmu-agent-workshop.github.io/">CMU LLM Agent workshop</a>, an event remarkably organized by graduate students at Carnegie Mellon University in just 3–4 weeks, with support from <a href="https://phontron.com/">Professor Graham Neubig</a>. I want to express my gratitude to Dr. Neubig and his students for their incredible effort in putting together such a fabulous event on short notice, even securing a sponsor mere hours before it began. Their dedication and hard work provided a valuable service to the research community. The workshop brought together researchers to discuss the challenges and opportunities of developing LLM-based agents that can effectively use tools, interact with diverse environments, and engage in social interactions. Throughout the event, a set of fundamental issues emerged as key themes: LLM correctness, reasoning capabilities, evaluation methodologies, memory augmentation, training techniques, and architectural design choices. The talks and discussions primarily focused on these challenges that the field must grapple with to create LLMs that are robust, aligned, and truly intelligent.</p><p><strong>Motivation</strong></p><p>My research group’s active work on LLM-based systems [<a href="https://scarecrows-hri.github.io/assets/pdfs/paper_2.pdf">1</a>, <a href="https://human-llm-interaction.github.io/workshop/hri24/papers/hllmi24_paper_11.pdf">2</a>] over the past six months served as the primary motivation for attending this workshop. Around a year ago, we recognized the immense potential of LLMs and began deeply engaging with the rapidly growing literature in this space. However, we quickly realized that separating genuine scientific advances from sensationalism and unfounded claims is a major challenge. The signal-to-noise ratio in the field is currently quite low, making it difficult to distinguish between hype and principled progress. As we built our own LLM-based systems, we directly encountered many of the fundamental issues discussed at the workshop. Attending this event was an invaluable opportunity to connect with a community of researchers who are dedicated to rigorously tackling these challenges and establishing a solid foundation for LLM research and development. By fostering open discussion and a commitment to principled approaches, workshops like this are crucial for advancing the field in a responsible and scientifically grounded manner.</p><p>My students and I have been working on several projects that explore the use of LLMs in real-world applications. The first project, LibGen, is an AI software engineer for distributed systems whose main objective is to improve the quality and correctness of code with respect to functionality and the ability to reason about distributed system execution in a multimodal sensed environment. This module focuses on construction and compositionality, communication protocol correctness, and multimodal reasoning. The second project is an LLM-based interactive agent that combines numerical facts with nominal facts drawn from text to provide contextualized information retrieval [<a href="https://human-llm-interaction.github.io/workshop/hri24/papers/hllmi24_paper_11.pdf">2</a>]. More specifically, we examine the use of interactive agents that provide factual information about urban data in-situ and contextualize it within the cultural history and governing context of the space. In the third project, we focus on using LLMs for sensemaking in a distributed sensor deployment in an office [<a href="https://scarecrows-hri.github.io/assets/pdfs/paper_2.pdf">1</a>]. The goal is to track the spatiotemporal relationships between people and objects, particularly personal items, and to provide a natural language query interface for keeping track of these items. The fourth project involves an LLM-enhanced generation model that produces highly interpretable video output from non-video-based sensor data [<a href="https://arxiv.org/abs/2306.15857">3</a>]. This project combines diffusion models with LLMs for interpretable and explainable sensor modeling [<a href="https://scs.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=d2a1790f-4888-4f59-9c64-b0af00f8e2c7">4</a>]. In the fifth project, we are developing an LLM-based system for real-to-sim in buildings. The system examines building operational data, and the model’s objective is to parameterize a simulation to produce a digital twin. We have also been exploring S(mall)LMs, specifically examining new ways to reduce their size and fine-tune their performance for application-specific needs.</p><p>Throughout our work on these projects, we have encountered several fundamental problems and made key observations that we believe are critical for advancing LLM research and development:</p><p><strong>Problems and observations:</strong></p><ol><li>LLM-based coding agents get caught in loops and can’t self correct effectively.</li><li>LLMs are slow.</li><li>It is unclear what techniques or frameworks generalize across LLM sizes.</li><li>LLMs can be incorrigible at times and produce different outputs across runs.</li><li>Fine-tuning is expensive and memory is short. How do we manage memory effectively? Both contextual memory and parameter memory (i.e., the weights of the model).</li><li>How can LLMs make more effective use of tools and learn from using them effectively?</li><li>LLMs make very subtle mistakes because they’re optimized to produce answers we like. How can we effectively check their correctness?</li><li>How can LLMs learn about the world?</li><li>What is reasoning? We conjecture it is narrative plus logical consistency? Who decides?</li><li>More generally, how can we truly measure their performance?</li><li>How can we optimize their prompts systematically without humans having to produce the prompts?</li></ol><p>There were several lessons and takeaways from the workshop, but more importantly, many of the issues and approaches that have emerged and that the community is grappling with are ones that everyone is dealing with and through community I believe a set of principles and frameworks and research directions are emerging. Observations and solutions that generalize across GPT-base LLM architectures is emerging.</p><p><strong>Talks</strong></p><p>There were several talks and all had noteworthy takeaways. I summarize the specific ones I was able to attend and that were particularly relevant to the work we’re doing in my lab.</p><p><strong>Yu Su (OSU)</strong></p><p>Dr. Yu Su, a professor from Ohio State University, presented his work on enhancing large language models (LLMs). His research focuses on providing structured associative memory for LLMs, where associations are built on semantic and topical relationship mapping. This approach aims to create more sophisticated and context-aware language models that can better understand and respond to the intricacies of human communication. Additionally, Dr. Su introduced an updated component decomposition of an intelligent agent system, which provides a comprehensive framework for designing and developing advanced language agents.</p><p>Dr. Su discussed tool learning through <a href="https://arxiv.org/abs/2403.04746">simulated trial and error</a>. In this work, LLMs are placed in a simulated environment where they can learn from their successes and failures. The method involves multiple short-term and long-term memory components, allowing LLMs to assess, execute, and optimize API calls based on their outcomes. By iteratively refining their strategies and fetching past experiences into context memory, LLMs can develop enhanced problem-solving skills over time. Another work introduced by Dr. Su was the <strong>HippoRAG</strong> model. The model draws parallels between the neocortex and perception and reasoning, the parahippocampal region and working memory, and the hippocampus and associative memory. By mimicking these human memory processing structures, HippoRAG aims to create more robust and contextually aware systems that can process and retain information, adapt to changing environments, and make informed decisions based on past experiences and current contexts.</p><p>In addition, Dr. Su presented a conceptual framework for language agents, focusing on defining their core competencies and operating environments. These include grounding, reasoning, and tool use, highlighting the diverse skill set required for language agents to effectively navigate and interact with their surroundings. The environments in which these agents function range from human interaction to databases and APIs, emphasizing the need for adaptability and versatility [<a href="https://arxiv.org/pdf/2402.14672">link</a>]. Dr. Su drew an interesting analogy between the development of LLMs and biological intelligence, particularly focusing on the evolutionary impact of language on human cognitive development. He referred to Max Bennett’s book <a href="https://www.amazon.com/Brief-History-Intelligence-Humans-Breakthroughs/dp/0063286343"><strong>A Brief History of Intelligence</strong></a>, which suggests that the emergence of language has been a form of singularity, dramatically increasing the complexity and transmission of ideas across generations. Drawing from this analogy, Dr. Su highlights the potential of LLMs to revolutionize the way we process and interact with information.</p><figure><img alt="Singularity already happened." src="https://cdn-images-1.medium.com/max/1024/0*BZ7Q8qjyqwV7wt1e" /></figure><p><strong>Xinyun Chen (Google DeepMind)</strong></p><p>Dr. Xinyun Chen presented their research on self-correction mechanisms in large language models (LLMs), such as GPT-3.5 and GPT-4, and their potential to improve model performance without relying on external feedback. The first key concept discussed was the effectiveness of self-correction methods in LLMs without the need for an oracle or human feedback. Experiments measuring recall across different rounds of self-correction and various tasks, including GSM8K, CommonSenseQA, and HotpotQA, suggest that <strong>additional rounds of self-correction do not necessarily guarantee better results</strong>. The presentation also introduced strategies like multi-agent debate and self-consistency as potential approaches to refine model responses and enhance their accuracy. Another topic explored in the talk was the idea of using LLMs as optimizers to iteratively improve their predictions. In this approach, <strong>the LLM generates solutions, evaluates them using an objective function, and then refines the prompts and responses based on the scores received</strong>. This iterative cycle aims to continuously enhance the precision of the model’s outputs, allowing it to learn from its own generated data and improve its performance over time.</p><p>The presentation also highlighted the importance of self-debugging in LLMs and the various feedback formats that can be employed to facilitate this process. Different types of feedback, ranging from simple unit test feedback to more complex code explanations and traces, were discussed. These feedback methods vary in their level of detail and complexity, which can significantly impact the model’s ability to self-correct its errors effectively. <strong>Dr. Chen emphasized that the choice of feedback format plays a crucial role in determining the success of self-debugging mechanisms</strong>. Simple feedback formats, such as unit test results, provide a straightforward indication of whether the model’s output is correct or not. However, more detailed feedback, such as code explanations and traces, offer a deeper understanding of the model’s reasoning process and can help identify the root causes of errors more effectively. The presentation also provided an overview of the current state of self-correction techniques in LLMs and the ongoing efforts to improve their effectiveness.</p><p>Dr. Chen talk also referred to her recent paper on <a href="https://arxiv.org/pdf/2311.17311">Universal Self-Consistency (USC)</a>; an extension of the standard self-consistency approach to support free-form generation tasks. <strong>USC leverages the LLM itself to select the most consistent answer among multiple candidate responses</strong>, notably boosting performance across diverse applications, including mathematical reasoning, code generation, long-context summarization, and open-ended question answering. It performs on par with standard self-consistency on tasks where answer extraction is feasible while offering more flexibility for open-ended generation by examining consistency based on criteria beyond exact match of the final answers. Ablation studies show that USC is robust to the ordering of candidate responses, and using more candidate responses generally improves results, although very long context can become challenging. Despite achieving strong results, there is still a gap compared to an oracle that always selects the best response, indicating that refining the USC framework to close this gap is an important direction for future work.</p><p>She also discussed her work on utilizing LLMs as optimizers. In this approach, called <a href="https://arxiv.org/pdf/2309.03409">Optimization by PROmpting (OPRO)</a>, the optimization task is described in natural language, and the LLM iteratively generates new solutions based on the problem description and previously evaluated solutions. The meta-prompt given to the LLM contains two key components: the optimization problem description, which includes the objective function and solution constraints, and the optimization trajectory, which consists of past solutions and their optimization scores sorted in ascending order. Including the optimization trajectory allows the LLM to identify patterns in high-scoring solutions and build upon them to generate potentially better ones.</p><p>Dr. Chen focused on the application of OPRO to<strong> prompt optimization</strong>, where the goal is to find a prompt that maximizes task accuracy. Evaluations on several LLMs and benchmarks like GSM8K and Big-Bench Hard showed that all tested LLMs were able to serve as optimizers, consistently improving the performance of generated prompts through iterative optimization until convergence. The best OPRO-optimized prompts matched the few-shot chain-of-thought prompting performance, outperforming zero-shot prompts by up to 8% on GSM8K and 50% on Big-Bench Hard tasks.</p><p><strong>Chris Paxton (Hello Robot)</strong></p><p>Dr. Paxton’s talk focused on developing low-cost, effective robots capable of assisting in home environments. His research addresses the challenge of open-vocabulary mobile manipulation, which involves enabling robots to handle and manipulate objects in diverse settings based on language instructions. This work aims to enhance robots’ utility in daily tasks through advanced planning and generalization skills, ensuring that they can perform effectively in varied, real-world scenarios.</p><p>The main takeaway from this talk is that LLMs have the potential to significantly improve robot performance in open-vocabulary mobile manipulation tasks. By leveraging the natural language understanding and generation capabilities of LLMs, robots can better interpret and execute complex instructions in a wide range of environments.</p><p>LLMs can enable robots to:</p><ol><li>Understand and follow high-level, abstract instructions provided in natural language.</li><li>Reason about the steps and actions required to complete a task based on the given instructions and the robot’s knowledge of its environment and capabilities.</li><li>Generalize to new objects, environments, and tasks by leveraging the vast knowledge encoded in the LLM.</li><li>Engage in dialogue with users to clarify ambiguous instructions, request additional information, or provide updates on task progress.</li><li>Adapt to changes in the environment or task by dynamically adjusting their behavior based on the LLM’s understanding of the situation.</li></ol><p>Further research is needed to address issues such as grounding language in the robot’s perceptual and action spaces, handling uncertainty and ambiguity, and developing appropriate control and monitoring mechanisms to prevent undesirable or unsafe behaviors.</p><p>[Part 2 — Takeways (coming soon)]</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3b152de9f92" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How AI and Computer Vision are Transforming Pitching Strategy in the MLB Playoffs]]></title>
            <link>https://medium.com/@jortiz81/how-ai-and-computer-vision-are-transforming-pitching-strategy-in-the-mlb-playoffs-115ee9d56b45?source=rss-a13135e021af------2</link>
            <guid isPermaLink="false">https://medium.com/p/115ee9d56b45</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[major-league-baseball]]></category>
            <category><![CDATA[sports]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Jorge Ortiz]]></dc:creator>
            <pubDate>Wed, 04 Oct 2023 14:46:11 GMT</pubDate>
            <atom:updated>2023-10-04T18:29:06.909Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/772/1*PSCejGL6AeL3-KoM5t2H5w.png" /></figure><p><strong>Real-Time Performance Analysis</strong></p><p>As the 2023 MLB playoffs get underway, teams are increasingly looking to cutting-edge artificial intelligence (AI) and computer vision (CV) to gain a competitive edge in their pitching strategies and decision-making processes. The convergence and synergy of advanced AI systems, CV capabilities, and large language models like GPT-4 are ushering in a new data-driven era in playoff baseball, expanding the analytical capabilities of teams unlike ever before. Real-time performance analysis is being dramatically enhanced by the ability of CV systems to analyze video feeds of games and provide immediate, dynamic insights into pitcher mechanics, batter stances, field conditions, and many other visual factors that impact play. The richness of visual data that can be gathered through real-time CV opens up a trove of new insights for teams to leverage, insights that human analysts would likely be unable to perceive on their own. When this wealth of visual data is combined with the ability of advanced AI to process statistics, historical data, and commentary in real time, teams now have an incredibly nuanced, enriched view of the game as it unfolds. They can detect subtle changes in pitcher mechanics that may indicate fatigue or injury risk or pick up on adjustments batters make at the plate. This level of real-time visual and statistical analysis enables entirely new possibilities for in-game strategy adjustments and immediate decision-making.</p><p><strong>Predictive Maintenance</strong></p><p>CV and AI are equally impactful for predictive maintenance and mitigating injury risk, especially for pitchers. The physical well-being of players is paramount in the high-stakes environment of the playoffs. By continuously monitoring pitchers through CV systems, teams can now detect early signs of overexertion, fatigue, changes in mechanics, or other indicators that may culminate in injury if not addressed proactively. Combining this CV data with the ability of AI to process and analyze historical workload and health data for pitchers enables teams to build robust predictive frameworks to forecast when a pitcher may be at heightened injury risk. This allows for timely interventions through substituting pitchers or adjusting rotations in order to mitigate injury scenarios before they happen and optimize player availability.</p><p><strong>Enhanced Matchup Analysis</strong></p><p>The combination of CV and AI also unlocks new frontiers in matchup analysis, providing teams with insights into the dynamics between specific pitchers and batters. A deeper historical analysis of batters&#39; performance against certain pitchers reveals trends and patterns in the matchups that human scouts and analysts would likely miss. By harnessing these richer, more nuanced insights into batter-pitcher matchups, teams can make more informed, strategic pitching decisions before and during playoff games to gain a possible edge.</p><p><strong>The Machine versus The Manager</strong></p><p>As teams rely more heavily on data-driven AI and CV insights for pitching strategy, an intriguing question arises — could artificial intelligence ever fully replace human managers in making gut-instinct pitching changes in high-leverage playoff situations?</p><p>Certainly, the advanced analytical capabilities provided by AI, CV, and LLMs are granting teams unprecedented strategic advantages. Optimizing pitching rotations, predicting injury risks, and gaining tiny marginal edges through matchup data are all hugely impactful.</p><p>However, baseball has always been a game defined as much by emotion, intuition, and “feel” for the pace of play as it has by data and probabilities. In the biggest moments of important postseason games, when managers must make split-second pitching change decisions, there are often intangible factors at play that no AI system could ever fully account for or understand.</p><p>Momentum swings, crowd energy, the look in a pitcher’s eye, or the vibe between a manager and a struggling player are all <strong><em>profoundly human factors</em></strong> that come into play. While large language models and other AI technologies are getting us closer to capturing more of the nuance and context that drives human decision-making, a manager’s embodied experience with the game still fundamentally provides a perspective that is difficult to replicate with vision and language algorithms alone.</p><p>Humans have five senses, not just two. Although deep neural networks can surpass human capabilities on specialized tasks, it is our ability to fluidly combine multiple senses into a subjective experience that AI cannot currently match.</p><p>So while AI will continue to complement and enhance human decision-making in baseball, it seems unlikely it will completely replace that age-old gut instinct managers rely on during the most pressure-filled playoff moments anytime soon. Together, pitching strategy will be defined by the symbiotic relationship between data-driven AI insights and the immeasurable human element.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=115ee9d56b45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Proctoring RemoteExams in the age of Corona]]></title>
            <link>https://medium.com/high-order-bit-blog/proctoring-remoteexams-in-the-age-of-corona-30b36e38860b?source=rss-a13135e021af------2</link>
            <guid isPermaLink="false">https://medium.com/p/30b36e38860b</guid>
            <dc:creator><![CDATA[Jorge Ortiz]]></dc:creator>
            <pubDate>Mon, 20 Apr 2020 18:56:31 GMT</pubDate>
            <atom:updated>2020-04-20T18:56:31.533Z</atom:updated>
            <content:encoded><![CDATA[<p>Two weeks ago, I had to proctor a midterm exam for my introduction to c/c++ class, programming methodology I. This is a core class in the ECE department at Rutgers University, so it’s a large class of about 160 sophomores. The scale, made it a bit challenging to design a method for remotely monitoring students while they take a midterm that was designed to be given in class.</p><p>In the proctor design, I decided to make full use of the staff. We have three graduate teaching assistant, two graders, and myself. That’s six staff members total. We make extensive use of <a href="https://en.wikipedia.org/wiki/Sakai_(software)">sakai</a>, google <a href="https://www.google.com/drive/">drive</a>, <a href="http://slack.com/">slack</a> and <a href="https://www.webex.com/">webex</a>. Every student at Rutgers has a personal webex account, so we want to use of this as well.</p><h4>Objectives:</h4><ol><li>Minimize the probability of cheating.</li><li>Minimize the number of students that can be monitored by each staff member.</li><li>Allow students to ask questions in real time.</li><li>Share the same set of announcements.</li><li>Enable a private communication channel between staff member during the exam.</li></ol><h3>Scale &amp; Real-time Communication</h3><p>To scale the proctoring, we simply divided the number of students by the number of proctors and randomly selected students to assign to a unique proctor. Each proctor was required to set up their own <a href="https://help.webex.com/ld-mzc23k-CiscoWebexTraining/Webex-Training#Get-Started">webex training</a> session and provide the information to each of their assigned students via email. For sharing the information with the staff, we also required that each proctor make the same information available via our class google-drive directory. That way, if a student lost the information, they could ask anyone of us for the information as we’d be able to provide it. Moreoever, the results of the assignment were also stored as a google sheet in the shared drive, for the same reason. If a student misplaced the information they could send an email to anyone in the staff.</p><p>Webex training allows students to asks questions in a separate side channel, make creating an audio link between the proctor and the student where the other participants on the call cannot hear it. This allows us to deal with questions in real time.</p><p>In addition, often during exams, there are issues (bugs) found by students in the exam, questions that need clarification that all students should know about, and general announcements about the amount of time remaining and how to turn it in. I create a shared, google presentation and then communicated with the staff via separate slack channel, with the link to the doc. They were instructed to share the doc in their training session, via the screen-sharing option. This took the place of writing announcements and time-remaining on a board.</p><p>The exam was shared via a google-doc link to the PDF in restricted viewing access writes enabled. That is, it cannot be downloaded or modified. The link was taken down after the exam was finished.</p><h3>Cheating</h3><p>To minimize the probability of cheating, we decided to make it mandatory that each student enable their camera while taking the exam. We provide guidelines for how it should look and include a statement in exam that acknowledges that they will be recorded. Recording them serves two purposes: 1) if they know they’re being recorded, they’re less likely to attempt to cheat and 2) if we grade them and suspect them of cheating, we can always go back to the video to look for suspicious behavior. All students are required to attend only the session they are assigned, in order to prevent them from making the excuse that they were in another session (in case they decided to cheat offline).</p><h3>Submissions</h3><p>When the students completed their exam, they were required to upload it to their personal “drop box” folder. They could either scan it or take a picture of their write-up with the answers on it. Before they could leave, the proctor had to check the submission to make sure it is readable or make them re-submit.</p><h3>Grading</h3><p>We created a shared google sheet with the class rosters and score for each question. I provided a grading rubric and we each graded the same set of students which we proctored. Every member of the staff marked up a PDF of each students’ drop box submission and returned it via drop box as well. The google sheet was exported and analyzed by a python script that tallied the grades and determined letter-grade thresholds. These images were saved and shared via a class lecture (as well as shared with the students after the exams were handed back).</p><h3>Grading Policy</h3><p>The re-grade policy is “all or nothing grading”. That is, if they want an exam question regraded, we must regrade all questions. That way we can make sure we didn’t make any other grading errors.</p><h3>Take away</h3><p>That’s it. It was a bit of a mess but all things considered and considering the size of the class, it went very smoothly. One thing to note, before the exam, each proctor held a practice session, where students in their assignment were invited to join. This would reveal any issues with the technology before the exam took place. During these sessions, many students became familiar with the UI and connection issues. Some discovered they had issues with their laptop webcam and decided to use their phones to log into webex using the mobile app.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30b36e38860b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/high-order-bit-blog/proctoring-remoteexams-in-the-age-of-corona-30b36e38860b">Proctoring RemoteExams in the age of Corona</a> was originally published in <a href="https://medium.com/high-order-bit-blog">High-Order-bit blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>