<?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 Mark Kram on Medium]]></title>
        <description><![CDATA[Stories by Mark Kram on Medium]]></description>
        <link>https://medium.com/@jonnyfoka?source=rss-d79be51b90ce------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Mark Kram on Medium</title>
            <link>https://medium.com/@jonnyfoka?source=rss-d79be51b90ce------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 21:36:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jonnyfoka/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[Jev is a new AI model that skips the chat and just decides, fast and cheap]]></title>
            <link>https://medium.com/@jonnyfoka/jev-is-a-new-ai-model-that-skips-the-chat-and-just-decides-fast-and-cheap-80cfac6f3528?source=rss-d79be51b90ce------2</link>
            <guid isPermaLink="false">https://medium.com/p/80cfac6f3528</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[customer-360]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[classification-models]]></category>
            <dc:creator><![CDATA[Mark Kram]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 10:19:19 GMT</pubDate>
            <atom:updated>2026-09-23T10:19:19.825Z</atom:updated>
            <content:encoded><![CDATA[<p><em>TypeSafe AI’s new ‘System 1’ model returns calibrated probabilities instead of prose — tested against a standard entity-resolution benchmark at $0.0522 total.</em></p><p>I prompted a language model last week and the response wasn’t a sentence. It was 0.87. That’s how TypeSafe AI’s new model, Jev, answers a question: you describe a situation in plain text, ask something structured like “is this a match,” and it returns a probability instead of prose. TypeSafe calls Jev a “System 1” model, built for fast, instinctive decisions rather than the slow, deliberate reasoning chat models are tuned for. I wanted to know if that distinction was real or just branding, so I put it to a real test.</p><h3>Why this matters now</h3><p>Classification-shaped decisions, “is this spam?”, “is this a match?”, “which agent should handle this?”, have always fallen into an awkward gap. They’re too trivial to justify a full LLM call and too complex to build well with traditional machine learning on a normal budget and timeline. A custom classifier runs $10,000 to $1,000,000 to build, then eats another 25 to 75 percent of that figure every year in maintenance.</p><p>Take a concrete example: building a unified customer-360 profile. Retailers often juggle between 5 and 15 separate systems of record for the same customer, a website account, a loyalty program, a call-center CRM, a point-of-sale system, each assigning its own ID. Merging these into one profile starts with entity resolution: for every candidate pair of records, deciding whether they describe the same person, exactly the match or no-match classification this benchmark tests. Once matched, the merged record usually needs a second round of classification just to fit the target schema, is this a consumer or a business account, which product category does a purchase belong to, is an address residential or commercial. Real identity-resolution builds run $90,000 to $200,000 in engineering effort for the matching logic alone, on top of whatever platform a company is already paying for. That’s the actual shape of the gap this article opened with: dozens of small, easy-to-state classification decisions that are expensive to industrialize one at a time.</p><p>A general-purpose LLM API looks cheap per call until you notice that identical traffic can cost 43 times more or less depending purely on which model answers it. Meanwhile, 78 percent of companies say they aren’t ready to deploy AI agents specifically because their customer data doesn’t line up across systems. That’s the gap Jev is aimed at.</p><h3>What Jev actually does</h3><p>Mechanically, Jev skips the step that makes chat models slow. A normal LLM generates one token at a time, re-running the full model for each word, then whatever’s calling it has to parse that text back into a decision. Jev’s architecture uses what TypeSafe calls a parallel sampler: one forward pass produces a probability distribution directly over a small, declared-in-advance set of outcomes.</p><p>There are three outcome types. A yes/no check with a single confidence score is called Noul. A pick from up to 255 options, returned as a full probability distribution, is called Choice. A rating on a scale you define, with between 2 and 10 labeled levels, is called Score.</p><p>Take TypeSafe’s own example for the last one: a 3-level frustration scale, 0 for calm, 1 for frustrated, 2 for very frustrated. Jev doesn’t just pick one label, it returns a single number like 1.4, a probability-weighted blend across those levels, so 1.4 means it leaned toward “frustrated” but saw some real chance of “very frustrated” too.</p><p>Training uses a method TypeSafe calls Reinforcement Learning for Calibrated Decisions, and the difference from a normal chat model is specific. A chat model is trained to produce text a human rater approves of, not to make its stated confidence match how often it’s actually right. Jev is trained the other way: its probabilities are optimized directly against outcomes, so that when it says 70 percent, it should be right about 70 percent of the time across similar predictions.</p><p>That property is called calibration, and it’s measured with a metric called Expected Calibration Error: group predictions by confidence level, then check whether the actual accuracy in each group matches the confidence in that group. It’s a checkable claim, not just marketing copy, though TypeSafe’s own docs are careful to note that calibration holds across groups of predictions, not for any single answer.</p><p>TypeSafe borrows the “System 1” label from psychologist Daniel Kahneman’s split between fast, intuitive thinking and slow, deliberate reasoning. System 1 is the quick gut call. System 2 is the careful, step-by-step version. Applied to Jev, that maps to one forward pass with no visible reasoning trace. A chatbot’s chain-of-thought, by that same logic, is System 2 territory.</p><p>The label is real ML vocabulary, not invented for this launch. It borrows from a psychological theory that’s itself contested though: a widely cited critique argues the classic Type 1/Type 2 split is largely unfalsifiable and that most real thinking mixes features of both. Treat “System 1” as a description of behavior, one pass, no deliberation trace, rather than a claim about how the model thinks.</p><p>One more thing worth stating plainly: TypeSafe hasn’t said how large Jev actually is. Every source I found, official and independent, stops at “transformer-based”.</p><h3>Putting it to the test</h3><p>I tested this against the Amazon-Google product-matching benchmark, a standard entity-resolution dataset where the task is deciding whether two listings from different catalogs describe the same product. Across 2,570 evaluated pairs, Jev scored 95.95 percent accuracy, 98.76 percent precision, 93.07 percent recall, and an F1 of 95.83 percent. The full experiment, including a 720-call prompt sweep before the real run, cost $0.0522 total across 3,290 calls. That’s a nickel, not the fifty cents I’d guessed going in.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IXN5BICYw6EliZi0pgXXjg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5cgCyLxBxmMjoh2jZkKfuw.png" /><figcaption><em>The headline numbers and the error breakdown, straight from the run.</em></figcaption></figure><p>A couple of the actual pairs make the task concrete. One easy match: “sony turntable pslx350h” against “sony pslx350h belt drive turntable”, different wording, same product, Jev called it a match at 99 percent confidence. One harder case, a non-match: “sibelius 4 professional edition” against “sibelius compass music composition teaching software”, same manufacturer, genuinely different products, correctly flagged as non-match at 96 percent confidence. The prompt sweep that preceded the real run tested 24 variants and found a clear pattern: short, plain instructions listing the full record, title, manufacturer, price, beat elaborate instructions with explicit edge-case rules, 93 to 97 percent accuracy against 77 to 87 percent. Dropping manufacturer and price to save tokens hurt more than it helped, and spelling out caution rules hurt too: on one tuning pair, “brain builder” from “core learning” against a reworded listing from the same manufacturer, the winning prompt style caught the match at 92 percent confidence, while a more prescriptive, rule-heavy prompt missed it entirely, calling it a non-match at 47 percent confidence. The pattern held across the sweep: give the model the same fields a person doing this by hand would use, and tell it what to decide, not how to reason about it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/980/1*XLb8Bb4Yx1D_7cx6oRG2Jw.png" /><figcaption><em>Simple, plain instructions consistently beat elaborate ones.</em></figcaption></figure><p>Mean latency came in at 0.65 seconds per call, with a p99 of 0.81 seconds and zero timeouts across 2,569 calls. The errors skew conservative: 15 false matches against 89 missed ones, so Jev is more likely to say “different” when unsure than to invent a match. That’s worth knowing before you deploy it somewhere a missed match costs more than a flagged one, or the other way around.</p><p>TypeSafe’s own marketing claims up to roughly 193 times faster and 444 times cheaper than a comparable LLM, and describes that as the high end of realistic gains. What I actually measured against Claude Sonnet 5 and Haiku 4.5 on this exact workload was 25 to 51 times cheaper, a smaller multiple than the vendor’s number and still a different cost category entirely.</p><h3>Where this actually fits</h3><p>None of the three use cases in the pitch, routing, guardrails, entity resolution, are new problems. Each already has an established place in production systems, filled today by a small classifier. vLLM’s own router uses a lightweight model called ModernBERT to pick between a fast path and a reasoning path, cutting latency and token spend by roughly half. OWASP’s guidance on securing LLM applications formalizes guardrails the same way: cheap checks handle most traffic, and only ambiguous cases escalate to a slower judge call.</p><p>Entity resolution at real scale works the same way. A Commonwealth Bank of Australia system processing up to 15.7 million records runs the same two-stage shape: filter down to plausible candidates, then score each one. That scoring step is exactly the match or no-match decision from the customer-360 example earlier, and historically it needed its own bespoke classifier, trained and maintained separately from whatever handled the B2C-versus-B2B schema check, or the routing decision, or the guardrail. Jev’s pitch is to be the one model behind all of those slots instead: the same API, reconfigured per question rather than retrained per task, so a team stops building and maintaining a separate small model for every one of these small decisions.</p><p>For a buyer, the bigger signal sits one level up. Gartner expects organizations to use small, task-specific models three times more than general-purpose LLMs by 2027, for exactly this reason: better fit and lower cost on domain-bound decisions. There’s a sovereignty angle worth taking seriously too. TypeSafe hasn’t open-sourced Jev’s weights and has no announced plan to; only its client tooling is public.</p><p>But the architecture clearly doesn’t need frontier scale, and that’s already proven outside TypeSafe entirely. Days after launch, an independent project called jeff shipped a real, self-hostable replacement built on a 400-million-parameter encoder, small enough for a single GPU or a laptop. It scores meaningfully lower than Jev on published benchmarks, around 64 percent against 90 percent on one intelligence test, so this isn’t parity. That’s proof the job doesn’t need a model too big to run yourself, not proof that jeff specifically is ready for production. Jev itself only launched on September 15, 2026, so there’s no long track record yet, no history of how it behaves under sustained real-world load over months. The practical move for now is a bounded pilot: run it on a limited slice of traffic, keep whatever system handles that decision today running in parallel as a fallback, and widen Jev’s role only as it proves itself, rather than replacing your only classifier with a brand-new one on day one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YhYF8YKjdJG97H4igsypUA.png" /><figcaption><em>jeff, built on a 400-million-parameter encoder, 220 stars within days of launch. Source: github.com/logan-markewich/jeff</em></figcaption></figure><h3>Beyond Jev itself</h3><p>Whether a genuinely small, self-hostable version of this eventually comes from TypeSafe or from the community racing to clone it, that’s the part worth watching. Jev itself might not be the model that matters in a year. The gap it’s pointing at, decisions too small for an LLM and too awkward for old-school ML, finally has a real third option.</p><h3>References</h3><p>[S1] TypeSafe AI. “Introducing System One Models &amp; Jev.” TypeSafe AI Blog, September 15, 2026. <a href="https://typesafe.ai/blog/introducing-system-one-models-and-jev">https://typesafe.ai/blog/introducing-system-one-models-and-jev</a>. Accessed September 21, 2026.</p><p>[S2] MarkTechPost. “TypeSafe AI Releases Jev: A System One Model That Returns Typed, Calibrated Decisions Instead of Text.” MarkTechPost, September 19, 2026. <a href="https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/">https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/</a>. Accessed September 21, 2026.</p><p>[S4] “From System 1 to System 2: A Survey of Reasoning Large Language Models.” arXiv:2502.17419. <a href="https://arxiv.org/abs/2502.17419">https://arxiv.org/abs/2502.17419</a>. Accessed September 21, 2026.</p><p>[S5] DataCamp. “Jev: TypeSafe’s System One Model That Never Hallucinates.” DataCamp Blog. <a href="https://www.datacamp.com/blog/system-one-models-jev">https://www.datacamp.com/blog/system-one-models-jev</a>. Accessed September 21, 2026.</p><p>[S6] MindStudio. “Jev Explained: TypeSafe AI’s Non-Autoregressive System-1 Model.” MindStudio Blog. <a href="https://www.mindstudio.ai/blog/jev-system-one-model-launch">https://www.mindstudio.ai/blog/jev-system-one-model-launch</a>. Accessed September 21, 2026.</p><p>[S10] TypeSafe AI. “Docs: Introduction.” TypeSafe AI Documentation. <a href="https://docs.typesafe.ai/introduction">https://docs.typesafe.ai/introduction</a>. Accessed September 21, 2026.</p><p>[S12] TypeSafe AI. “Docs: full reference.” TypeSafe AI Documentation. <a href="https://docs.typesafe.ai/llms-full.txt">https://docs.typesafe.ai/llms-full.txt</a>. Accessed September 21, 2026.</p><p>[S16] ITRex Group. “Machine Learning (ML) Costs: Price Factors and Real-World Estimates.” ITRex Group Blog. <a href="https://itrexgroup.com/blog/machine-learning-costs-price-factors-and-estimates/">https://itrexgroup.com/blog/machine-learning-costs-price-factors-and-estimates/</a>. Accessed September 21, 2026.</p><p>[S17] CDP.com. “CDP Pricing: Models, Ranges, and Hidden Costs to Budget For.” <a href="https://cdp.com/articles/cdp-pricing/">https://cdp.com/articles/cdp-pricing/</a>. Accessed September 21, 2026.</p><p>[S18] CloudZero. “LLM API Pricing Comparison in 2026: Every Major Model Ranked by Cost.” CloudZero Blog, updated September 4, 2026. <a href="https://cloudzero.com/blog/llm-api-pricing-comparison/">https://cloudzero.com/blog/llm-api-pricing-comparison/</a>. Accessed September 21, 2026.</p><p>[S21] Mordor Intelligence. “Identity Resolution Software Market Size &amp; Share Analysis (2026–2031).” <a href="https://mordorintelligence.com/industry-reports/identity-resolution-software-market">https://mordorintelligence.com/industry-reports/identity-resolution-software-market</a>. Accessed September 21, 2026.</p><p>[S23] Gartner. “Gartner Predicts By 2027, Organizations Will Use Small, Task-Specific AI Models Three Times More Than General-Purpose LLMs.” Gartner Newsroom, April 9, 2025. <a href="https://www.gartner.com/en/newsroom/press-releases/2025-04-09-gartner-predicts-by-2027-organizations-will-use-small-task-specific-ai-models">https://www.gartner.com/en/newsroom/press-releases/2025-04-09-gartner-predicts-by-2027-organizations-will-use-small-task-specific-ai-models</a>. Accessed September 21, 2026.</p><p>[S29] Pavlovic, Maja (Queen Mary University London). “Understanding Model Calibration: A Gentle Introduction and Visual Exploration of Calibration and the Expected Calibration Error (ECE).” arXiv:2501.19047. <a href="https://www.alphaxiv.org/abs/2501.19047">https://www.alphaxiv.org/abs/2501.19047</a>. Accessed September 21, 2026.</p><p>[S31] vLLM Project. “vLLM Semantic Router.” vLLM Blog, September 11, 2025. <a href="https://blog.vllm.ai/2025/09/11/semantic-router.html">https://blog.vllm.ai/2025/09/11/semantic-router.html</a>. Accessed September 21, 2026.</p><p>[S32] Kannangara, Sandeepa, et al. (Commonwealth Bank of Australia). “A Robust and Efficient Pipeline for Enterprise-Level Large-Scale Entity Resolution (MERAI).” arXiv:2508.03767. <a href="https://www.alphaxiv.org/abs/2508.03767">https://www.alphaxiv.org/abs/2508.03767</a>. Accessed September 21, 2026.</p><p>[S33] OWASP. “LLM Prompt Injection Prevention Cheat Sheet.” OWASP Cheat Sheet Series. <a href="https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html">https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html</a>. Accessed September 21, 2026.</p><p>[S34] Perplexity (Sonar) research synthesis on routing, guardrail, and entity-resolution implementation patterns, compiled September 21, 2026.</p><p>[S35] TypeSafe AI. “Docs: System One, Confidence, and Confidence-Gated Routing.” TypeSafe AI Documentation. <a href="https://docs.typesafe.ai/concepts/system-one">https://docs.typesafe.ai/concepts/system-one</a>. Accessed September 21, 2026.</p><p>[S39] Markewich, Logan. “jeff: a self-hosted drop-in replacement for TypeSafe’s Jev.” GitHub repository. <a href="https://github.com/logan-markewich/jeff">https://github.com/logan-markewich/jeff</a>. Accessed September 21, 2026.</p><p>[S40] Dubey, Ashish. “TypeSafe AI Jev: Self-Hosting and Open Source Status.” TrueFoundry Blog, September 18, 2026. <a href="https://www.truefoundry.com/blog/typesafe-ai-jev">https://www.truefoundry.com/blog/typesafe-ai-jev</a>. Accessed September 21, 2026.</p><p>[S41] Melnikoff, A.I. and Bargh, J.A. “The Mythical Number Two.” Trends in Cognitive Sciences, 22(4), 280–293, 2018. <a href="https://pubmed.ncbi.nlm.nih.gov/29571664/">https://pubmed.ncbi.nlm.nih.gov/29571664/</a>. Accessed September 21, 2026.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=80cfac6f3528" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Football is a graph. The 2026 World Cup just made it impossible to miss]]></title>
            <link>https://medium.com/@jonnyfoka/football-is-a-graph-the-2026-world-cup-just-made-it-impossible-to-miss-1db00a401998?source=rss-d79be51b90ce------2</link>
            <guid isPermaLink="false">https://medium.com/p/1db00a401998</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[world-cup]]></category>
            <category><![CDATA[graph-machine-learning]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[football]]></category>
            <dc:creator><![CDATA[Mark Kram]]></dc:creator>
            <pubDate>Fri, 03 Jul 2026 12:19:22 GMT</pubDate>
            <atom:updated>2026-07-03T12:19:22.108Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>How a corner-of-the-screen wavy line, a corner kick and a Pinterest feed all speak the same mathematical language, and what that means for the rest of us.</strong></p><p>There’s a wavy line in the bottom-left corner of 2026 World Cup broadcasts that swings up when a team attacks and dips when the other team gets the ball back. It looks like a stock chart. It is not a stock chart. It’s a graph, in the same mathematical sense a molecule is a graph, and the room where it lives is quietly rewiring how the sport is played.</p><h3>What the wavy line is</h3><p>That line has a name. Fox Sports calls it “Match Momentum,” and Kathryn Xu at <strong>Defector </strong>watched it swing during the group-stage matches England-Ghana and England-Croatia, tried to figure out what it was, and wrote a piece essentially asking: what am I looking at?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*Mj9TloolYSXhXg_1IX8_Iw.jpeg" /><figcaption>source: <a href="https://defector.com/what-on-earth-is-match-momentum">https://defector.com/what-on-earth-is-match-momentum</a></figcaption></figure><p>The engine underneath is Opta’s Possession Value model: a probability that either team is about to score in the next ten seconds, updated every touch, weighted so only the last three or four minutes carry weight, and capped tight enough that a single scramble in the box doesn’t blow it up. When a player carries the ball into the danger zone, the number climbs. When they lose it, it dives.</p><p>The same shift is showing up all over the tournament. FIFA and Lenovo unveiled a “Football AI” bundle at CES this January: a knowledge assistant for all forty-eight competing teams, 3D player avatars for offside recreations, a stabilised Referee View. The world feed reaches 220 territories from a control room in Dallas. Something has shifted in the visual language of football.</p><h3>Football is (already) a graph</h3><p>Every one of those graphics is one word from the same mathematical alphabet: A<strong> graph</strong>.</p><p>Not the bar-chart kind. The network-science kind. A graph is a set of things and the relationships between them: nodes for the things, edges for the connections between them. Cities and roads are a graph. Power plants and power lines are a graph. Molecules are a graph. Your Instagram friends are a graph. As one open textbook puts it, edges are the channels through which content flows: cars along roads, power along wires, and on a football pitch, the ball along passes.</p><p>Football fits almost embarrassingly well. Players are the nodes. Passes and marking are the edges. Position, velocity, height ride on each node like tags.</p><p><strong>Two symmetries make the graph the right representation:</strong></p><p>First: what the model cares about is who is connected to whom, not what shirt number a player wears. Swap the numbers on the backs of two shirts and the graph is the same graph. The pattern of connections hasn’t moved.</p><p>Second: mirror the football pitch left-to-right, so that everything on the left half is now on the right and vice versa. A corner kick from the left corner flag becomes the mirror image of a corner from the right. Same shape, same tactical situation, just flipped. When a model respects this, it can learn from both mirror versions at once, which is why DeepMind’s TacticAI can find its footing on the roughly ten corners each team gets per match.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*hJ4PuO0wazCl_st_GZx_dA.jpeg" /><figcaption>source: Live Picture of MagentaTV analysis during broadcast</figcaption></figure><p>Now bring it back to that Match Momentum bar. Under the hood, Opta’s Possession Value chops the pitch into a grid of zones, treats each zone as a node, and lets the ball flow between them like traffic on a road network. Each zone gets a score for how likely a goal becomes from there. When the wavy line climbs, the attacking team just moved into a higher-value zone. When it dips, they just left one.</p><h3>The room where the graph is actually used</h3><p>Graph analytics is a specialised layer inside a broader data-science stack, and only the most forward clubs run it in earnest. What follows is a mix. Some run graph models directly. Others rely on broader data-science operations, with graph thinking as the most sophisticated end of a wider toolkit. All of them lean on clean structured data.</p><p>Cut to Anfield. Liverpool has run a Research Department since 2012: four scientists and two data engineers, first led by Cambridge physics PhD Ian Graham, later by his successor William Spearman. Their day job is building expected-possession-value models on live tracking data and hooking them to video, so an analyst can click any moment in a match and ask, “did we think this was dangerous, and did the model agree?”.</p><p>The model never talks directly to the manager. Graham has said as much: “Jürgen [Klopp] is very open and receptive to our area… he knows the analysis we do and how it feeds into the various reports and the weekly work”. Between the research desk and the touchline sit opposition analysts, post-match analysts and video editors.</p><p>Other data-first clubs say the same thing in blunter form. Barcelona’s Sergio Llana, head of sport analytics: “Without data analysis and club algorithms, we would now be at a disadvantage”. Toulouse’s president Damien Comolli, one of the most on-record data-first decision-makers in top-flight football: “We are almost exclusively data driven in our processes, whether that is recruiting a player, a coach, analyzing opponents, or looking at our own team and players”. And in Germany, the DFB-Akademie, the DFL, Sportec Solutions and the University of Tübingen jointly run a graph-neural-network programme that won the MIT Sloan research award. It taught a model to detect overlapping runs at AUC 0.93. Applied finding: Germany made 23% fewer overlapping runs than their team average in the Euro 2020/2021 opener against France.</p><h3>From set pieces to open play</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*pTpmBxD4COPCXSYkfSyr-g.jpeg" /><figcaption>source: <a href="https://deepmind.google/blog/tacticai-ai-assistant-for-football-tactics/">https://deepmind.google/blog/tacticai-ai-assistant-for-football-tactics/</a></figcaption></figure><p>DeepMind and Liverpool published the flagship result in <strong>Nature Communications</strong> in 2024. TacticAI turned each corner kick into a fully-connected graph of eleven-versus-eleven players and taught a graph neural network to reason across it. It predicted the corner receiver 78.2% of the time, matching Liverpool’s own experts, who preferred its generated setups over the real ones in 90% of blind comparisons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/938/1*oYi_0Ke4FKymCt4eW8oUww.png" /><figcaption>source: <a href="https://deepmind.google/blog/tacticai-ai-assistant-for-football-tactics/">https://deepmind.google/blog/tacticai-ai-assistant-for-football-tactics/</a></figcaption></figure><p>Then on June 11 this year, at Google for Brazil, the frontier moved. Palmeiras became the first club to extend TacticAI to open play, “up to 8 seconds ahead,” through a drag-and-drop interface that lets a coach reposition players and watch the simulated defence rearrange itself. Palmeiras coach Kevin Middleton’s honest caveat on X: “it needs 22 tracked players and a data science department to run it”.</p><p>There’s also a black-box problem hanging over all of this: graph neural networks are notoriously hard to explain to non-technical staff, so the analyst-and-video layer between model and coach never fully goes away.</p><h3>Where the wonder has to keep its own promise</h3><p>Peer review is a spectrum, not a switch. TacticAI on corners is <strong>Nature Communications</strong>: peer-reviewed, honest numbers, private Liverpool data. The Palmeiras “up to 8 seconds” claim is a company announcement. No published accuracy at 4, 6 or 8 seconds.</p><p>The honest quantitative ceiling looks less shiny than the headline. The best openly-reproducible graph neural network on real tracking data, released by U.S. Soccer as an arXiv preprint at the end of 2024, hits ROC-AUC of 0.78 for men’s football and 0.83 for women’s, with practical improvements from tweaking player runs of just 1.4 to 3.8 percentage points. That is the state of the science outside the marketing deck. It’s still a serious step forward. It’s not magic.</p><p>And still, Premier League tracking data isn’t seven minutes late any more. Michael D’Auria of Genius Sports confirms his system now delivers position data in under a second in every top-flight ground. And FIFA’s decision to ban its Football AI Pro assistant from live in-play use is a FIFA policy for fair competition, not an IFAB Law of the Game. Nothing in the Laws stops a club from running its own graph model live on its own bench.</p><h3>The same (graph-) pattern is everywhere</h3><p>Once you see this pattern, you start to see it in every serious AI story of the last five years.</p><p>DeepMind’s GraphCast is a graph neural network on a multi-scale mesh of the atmosphere. It runs a 10-day global forecast in under a minute on one TPU, and it beats Europe’s flagship ECMWF model on more than 90% of the variables it was benchmarked on, and on 99.7% of them inside the troposphere.</p><p>Pinterest’s PinSage runs on a graph of three billion pins and eighteen billion links between them; engagement rose 30% in A/B tests. Uber Eats builds two bipartite graphs (users to dishes, users to restaurants), embeds them with a variant of GraphSAGE, and ships that as the single most influential feature in its ranker. Amazon Science reports about 16% better demand forecasting from a spatio-temporal multi-graph network over the prior baseline. And in payments and banking, graphs at scale flag fraud rings that a spreadsheet couldn’t see, with vendors like PayPal running billions of edges per day through them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8YlZTEj8MxSvOTXjtGeT8g.png" /></figure><p>Same abstract move. Wildly different domains. Once your brain does the graph swap once, it doesn’t stop.</p><h3>Two doors, one room</h3><p>This is the corner of applied ML I work in: graph data science and graph machine learning, where the same lens that just walked you through a World Cup broadcast becomes a technique you can actually deploy. Depending on why you’re still reading, one of these two doors is for you.</p><p><strong>If you want to try it yourself</strong>. Thirty minutes (and a little help of AI): `pip install torch-geometric networkx statsbombpy`, pull one open StatsBomb match, and draw its passing graph in fifteen lines of Python. A weekend: Neo4j’s free Graph Data Science Fundamentals course, plus two lectures from Stanford’s CS224W on YouTube (taught by Jure Leskovec, yes, the PinSage co-author). Deep dive: the full CS224W lecture series, Will Hamilton’s <strong>Graph Representation Learning</strong> book and Petar Veličković’s DeepMind lectures (yes, the TacticAI co-author). Already dangerous? Reproduce “DEFCON defensive-value GNN” on StatsBomb 360 open data, where no public code exists yet.</p><p><strong>If you’re wondering whether this applies to your business.</strong> The honest answer is <strong>probably yes</strong>, but the shape depends on your data. If your business runs on connections (customers who influence customers, suppliers who feed suppliers, transactions that link accounts, devices that talk to devices, documents that reference documents) you already have a graph. The interesting question isn’t whether you have one. It’s what a well-chosen graph model changes about a decision you make often and expensively. If you’d like to think out loud with someone about whether graph data science makes sense for your team, I’m happy to grab a virtual coffee: <a href="https://www.linkedin.com/in/mark-kram-503b2bb0/"><strong>My LinkedIn</strong></a></p><p>Whatever pulled you into this article, the ping of a Match Momentum bar, a corner-kick prediction, a Pinterest feed you can’t stop scrolling, you were watching a graph. The best time to learn to draw one yourself, or to find out whether one fits your business, is the same weekend the world is watching them win football matches.</p><h3>References</h3><p>[S1] Sanchez-Lengeling, Reif, Pearce, Wiltschko. “A Gentle Introduction to Graph Neural Networks.” *Distill*, 2 September 2021. <a href="https://distill.pub/2021/gnn-intro/">https://distill.pub/2021/gnn-intro/</a></p><p>[S2] Omar Lizardo, Isaac Jilbert. “Social Networks: An Introduction, §3.2 The Building Blocks of Graphs.” UCLA open textbook. <a href="https://bookdown.org/omarlizardo/_main/3-2-the-building-blocks-of-graphs-edges-and-nodes.html">https://bookdown.org/omarlizardo/_main/3-2-the-building-blocks-of-graphs-edges-and-nodes.html</a></p><p>[S3] Zhe Wang, Petar Veličković. “TacticAI: an AI assistant for football tactics.” Google DeepMind blog, 19 March 2024 (updated March 2026). <a href="https://deepmind.google/blog/tacticai-ai-assistant-for-football-tactics/">https://deepmind.google/blog/tacticai-ai-assistant-for-football-tactics/</a></p><p>[S6] FIFA. “FIFA and Lenovo unveil multiple AI-powered innovations ahead of FIFA World Cup 2026.” Media release, 7 January 2026. <a href="https://inside.fifa.com/media-releases/lenovo-tech-world-ai-powered-innovations-world-cup-2026">https://inside.fifa.com/media-releases/lenovo-tech-world-ai-powered-innovations-world-cup-2026</a></p><p>[S7] FIFA. “FIFA World Cup 2026 broadcast partnerships set new global benchmark.” Media release, 11 June 2026. <a href="https://inside.fifa.com/tournament-organisation/commercial/fifa-tv/media-releases/world-cup-2026-broadcast-partnerships-global-benchmark-record-reach-innovation">https://inside.fifa.com/tournament-organisation/commercial/fifa-tv/media-releases/world-cup-2026-broadcast-partnerships-global-benchmark-record-reach-innovation</a></p><p>[S8] Kathryn Xu. “What On Earth Is ‘Match Momentum’?” *Defector*, 25 June 2026. <a href="https://defector.com/what-on-earth-is-match-momentum">https://defector.com/what-on-earth-is-match-momentum</a></p><p>[S10] Jonny Whitmore. “What is Match Momentum?” *Opta Analyst*, November 2021 (updated May 2025). <a href="https://theanalyst.com/articles/what-is-match-momentum">https://theanalyst.com/articles/what-is-match-momentum</a></p><p>[S14] Google DeepMind. “TacticAI open-play extension with Palmeiras.” X post, 11 June 2026. <a href="https://x.com/GoogleDeepMind/status/2065093482088169719">https://x.com/GoogleDeepMind/status/2065093482088169719</a></p><p>[S15] Training Ground Guru. “Ian Graham: How Liverpool integrate data, analysis and coaching.” <a href="http://archive.trainingground.guru/articles/ian-graham-how-liverpool-integrate-data,-analysis-and-coaching">http://archive.trainingground.guru/articles/ian-graham-how-liverpool-integrate-data,-analysis-and-coaching</a></p><p>[S16] Liverpool FC. “Behind the Badge: Ian Graham.” <a href="https://www.liverpoolfc.com/news/behind-the-badge/398645-ian-graham-liverpool-fc-behind-the-badge">https://www.liverpoolfc.com/news/behind-the-badge/398645-ian-graham-liverpool-fc-behind-the-badge</a></p><p>[S19] CBS Sports. “Toulouse’s Damien Comolli explains American-owned club’s almost exclusively data-driven transfer approach.” <a href="https://www.cbssports.com/soccer/news/toulouses-damien-comolli-explains-american-owned-clubs-almost-exclusively-data-driven-transfer-approach/">https://www.cbssports.com/soccer/news/toulouses-damien-comolli-explains-american-owned-clubs-almost-exclusively-data-driven-transfer-approach/</a></p><p>[S20] Google. “Google for Brazil 2026.” Company blog. <a href="https://blog.google/company-news/inside-google/around-the-globe/google-latin-america/google-for-brazil-2026/">https://blog.google/company-news/inside-google/around-the-globe/google-latin-america/google-for-brazil-2026/</a></p><p>[S21] Gabriel Anzer, Pascal Bauer, Ulf Brefeld, Dennis Fassmeyer. “Detection of tactical patterns using semi-supervised graph neural networks.” MIT Sloan Sports Analytics Conference, 2022. <a href="https://www.researchgate.net/profile/Pascal-Bauer-6/publication/359079429_Detection_of_tactical_patterns_using_semi-supervised_graph_neural_networks/">https://www.researchgate.net/profile/Pascal-Bauer-6/publication/359079429_Detection_of_tactical_patterns_using_semi-supervised_graph_neural_networks/</a></p><p>[S23] Barça Innovation Hub. “Sergio Llana: Without data analysis and club algorithms, we would now be at a disadvantage.” <a href="https://barcainnovationhub.fcbarcelona.com/blog/without-data-analysis-and-club-algorithms-we-would-now-be-at-a-disadvantage/">https://barcainnovationhub.fcbarcelona.com/blog/without-data-analysis-and-club-algorithms-we-would-now-be-at-a-disadvantage/</a></p><p>[S25] kloppy. Documentation (PySport). <a href="https://kloppy.pysport.org/">https://kloppy.pysport.org/</a></p><p>[S27] The Conversation. “Can AI improve football teams’ success from corner kicks? Liverpool and others are betting it can.” <a href="https://theconversation.com/can-ai-improve-football-teams-success-from-corner-kicks-liverpool-and-others-are-betting-it-can-225894">https://theconversation.com/can-ai-improve-football-teams-success-from-corner-kicks-liverpool-and-others-are-betting-it-can-225894</a></p><p>[S28] Ricardo Bekkers, Neeraj Sahasrabudhe. “A Graph Neural Network deep-dive into successful counterattacks.” U.S. Soccer / Unravel Sports, arXiv:2411.17450, 2024. <a href="https://arxiv.org/html/2411.17450v2">https://arxiv.org/html/2411.17450v2</a></p><p>[S30] Genius Sports / Training Ground Guru. “The Future of Tracking Data: Michael D’Auria interview.” <a href="https://www.geniussports.com/content-hub/future-of-tracking-data/">https://www.geniussports.com/content-hub/future-of-tracking-data/</a></p><p>[S31] FIFA. “Football AI Pro.” Official product page. <a href="https://inside.fifa.com/innovation/innovating-the-game/football-ai-pro">https://inside.fifa.com/innovation/innovating-the-game/football-ai-pro</a></p><p>[S33] Ruining He. “PinSage: A new graph convolutional neural network for web-scale recommender systems.” Pinterest Engineering, KDD 2018. <a href="https://medium.com/pinterest-engineering/pinsage-a-new-graph-convolutional-neural-network-for-web-scale-recommender-systems-88795a107f48">https://medium.com/pinterest-engineering/pinsage-a-new-graph-convolutional-neural-network-for-web-scale-recommender-systems-88795a107f48</a></p><p>[S34] Remi Lam et al. “GraphCast: AI model for faster and more accurate global weather forecasting.” Google DeepMind, November 2023 (*Science*, 2023). <a href="https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/">https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/</a></p><p>[S35] Uber Engineering. “Food discovery with Uber Eats: Using graph learning to power recommendations.” <a href="https://www.uber.com/blog/uber-eats-graph-learning/">https://www.uber.com/blog/uber-eats-graph-learning/</a></p><p>[S37] Amazon Science. “Spatio-temporal multi-graph networks for demand forecasting in online marketplaces.” <a href="https://www.amazon.science/publications/spatio-temporal-multi-graph-networks-for-demand-forecasting-in-online-marketplaces">https://www.amazon.science/publications/spatio-temporal-multi-graph-networks-for-demand-forecasting-in-online-marketplaces</a></p><p>[S38] StatsBomb Open Data (with Metrica, SkillCorner, Wyscout, kloppy). <a href="https://github.com/statsbomb/open-data">https://github.com/statsbomb/open-data</a></p><p>[S39] Graph ML learning resources composite: PyTorch Geometric, NetworkX, DGL, Neo4j GraphAcademy, Stanford CS224W (Jure Leskovec), Will Hamilton’s *Graph Representation Learning*, Petar Veličković’s DeepMind lectures, Friends of Tracking, Soccermatics.</p><p>[S40] Composite: TigerGraph banking fraud + NVIDIA developer blog on PayPal graph fraud detection.</p><p>[S41] Kim et al. “DEFCON: Defensive Contribution Graph Neural Network.” arXiv:2512.10355, December 2025. <a href="https://arxiv.org/abs/2512.10355">https://arxiv.org/abs/2512.10355</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1db00a401998" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>