<?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 Rajeev Chhajer on Medium]]></title>
        <description><![CDATA[Stories by Rajeev Chhajer on Medium]]></description>
        <link>https://medium.com/@rajeev.chhajer?source=rss-c4c5ab7ae68a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*2vEnDQ6Ez0EuEAQy</url>
            <title>Stories by Rajeev Chhajer on Medium</title>
            <link>https://medium.com/@rajeev.chhajer?source=rss-c4c5ab7ae68a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 07:51:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rajeev.chhajer/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[The New ‘0 to 1’ Is ‘0 to 10’!]]></title>
            <link>https://medium.com/99p-labs/the-new-0-to-1-is-0-to-10-adff1211488d?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/adff1211488d</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 16:11:36 GMT</pubDate>
            <atom:updated>2026-11-19T16:11:36.221Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6wZA2ggnlJKjA0y9jFUqFw.png" /></figure><p>For years, the hardest part of starting something was the product itself. Building an MVP took real time, money and talent, and the moment you had one, you faced the next hard layer: getting customers to tell you the truth about it. Clearing both earned you the right to move on.</p><p>That sequence has broken. When building is nearly free, the product stops being the barrier, and everything that used to come after it gets pulled forward.</p><h3>The concept: every gate moved one step earlier</h3><p>The old path had three clear stages.</p><p>1. <strong>0 to 1:</strong> build a quick MVP and get early customer feedback. Both were hard.</p><p>2. <strong>1 to 10:</strong> with that validation in hand, test use, revenue or distribution in some small way.</p><p>3. <strong>10 to 100:</strong> if that worked, scale.</p><p>Now the MVP is close to frictionless, so the weight shifts. The new 0 to 1 has to carry what used to take two stages: a product, real feedback, and early evidence of use, revenue or distribution. The new 1 to 10 then asks for what used to wait until 10 to 100: the ability to build, sell or align, and distribute, much earlier than anyone used to expect.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*8VD0qKRvbfjA-o-GH3H4Cg.png" /></figure><p>old vs new stage gates · 0 to 1, 1 to 10, 10 to 100</p><p>Work that once sat in 1 to 10 and 10 to 100 is now pulled one stage forward.</p><p>The gates didn’t disappear. They moved forward, and they got heavier.</p><h3>For the entrepreneur</h3><p>A working demo used to be a signal. It showed you could build, and that alone set you apart. Today an investor or early customer may see ten versions of your demo in a month. The demo is the entry ticket, not the proof.</p><p>So the new 0 to 1 asks founders to show, in the earliest phase, what used to come later:</p><p>• <strong>Real feedback, not polite feedback.</strong> Customers who change their behavior, not just nod.</p><p>• <strong>A hint of use or revenue.</strong> A few people who come back, pay, or pull you into their workflow.</p><p>• <strong>A path to distribution.</strong> Some answer to how the next hundred users find you.</p><p>And the new 1 to 10 compresses what used to be scale-up work. Founders need to build, sell and distribute in parallel, early, often with a team of one or two plus agents. The founder who can only build is now competing with everyone.</p><h3>For the internal innovator</h3><p>Inside a large company, the same compression is happening. A small team can now prototype in days what once needed a funded project. That should speed innovation up. Often it doesn’t, because the gates inside the organization haven’t moved.</p><p>The internal version of 1 to 10 isn’t revenue or distribution. It’s alignment and adoption: a sponsor who will fund it, a business unit that will own it, and users who will change how they work. Internal innovators now need to show early signs of those in the 0 to 1 phase, just as founders must show early traction.</p><p>Here is where I think the real problem sits. <strong>Leaders tend to bet on the idea, not the person.</strong> And how they bet is shaped by incentives tied to their own position in the approval chain: who carries the risk, who gets credit, whose roadmap it touches. A strong idea with a misaligned sponsor stalls. A modest idea with the right person behind it often wins.</p><p>Startups learned long ago that the team matters as much as the idea. Many corporate processes still evaluate a slide deck, not the person who will carry it through the messy middle. When building is cheap, that gap becomes the bottleneck. It’s the same problem founders face, wearing a different badge.</p><h3>The cold start problem needs a new definition</h3><p>The classic cold start problem was about having nothing: no product, no users, no data, and no easy way to get the first of each. Building is now cheap, so that version is mostly solved. A new cold start has taken its place. A few ways to name it:</p><p>• <strong>Cold trust.</strong> Anyone can build a demo, so a demo earns no belief. The new cold start is getting the first people to believe you, not just to try what you made.</p><p>• <strong>Cold commitment.</strong> The hard first step isn’t a user. It’s the first person willing to commit something scarce: money, time, a roadmap slot, their reputation.</p><p>• <strong>Cold attention.</strong> Supply of new products and prototypes is exploding while attention stays fixed. The first hundred users, or the first executive’s hour, are harder to win than the first version is to build.</p><p>• <strong>Cold sponsorship.</strong> Inside companies, the cold start is finding the first leader who will bet on you, not just approve your idea.</p><p>If I had to pick one line: <strong>the new cold start problem is getting the first people to commit to something they could have built themselves, or have already seen ten times. I, personally, like it now being called the “Cold attention” problem.</strong></p><h3>What this asks of us</h3><p>If the new 0 to 1 is 0 to 10, our playbooks are a stage behind. Founders, investors and corporate leaders are still judging early work by whether it can be built. That question is mostly answered.</p><p>A few questions I’m still working through:</p><p>• What does good evidence look like at the new 0 to 1, before revenue or production?</p><p>• How would corporate innovation change if leaders were rewarded for betting on people, not just approving ideas?</p><p>• Which of the new cold starts is hardest to break: trust, commitment, attention or sponsorship?</p><p>I’d love to hear where you see the gates now.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=adff1211488d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/the-new-0-to-1-is-0-to-10-adff1211488d">The New ‘0 to 1’ Is ‘0 to 10’!</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Knowledge Is Made, Not Stored]]></title>
            <link>https://medium.com/99p-labs/knowledge-is-made-not-stored-f1280c5c5683?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/f1280c5c5683</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 12:33:14 GMT</pubDate>
            <atom:updated>2026-11-16T12:33:14.694Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Part 2 of an iterative series on context, knowledge, and how humans and agents get work done</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hcoaGdr1QO9gr-RPAiFMKQ.png" /></figure><h3>Where we left off</h3><p>Two weeks ago I published <a href="https://medium.com/99p-labs/nobody-knows-what-enough-context-means-2e543fba7e77">Nobody Knows What “Enough Context” Means</a>, version 0.1 of a belief system about context, knowledge, and how humans and agents share work. I promised to come back and mark what changed. Something changed faster than I expected.</p><p>Over a business dinner, Duane, Brian, Ryan and I started arguing about one word: <strong>knowledge</strong>. That argument turned into Ryan’s post, <a href="https://medium.com/99p-labs/why-many-ai-agents-fail-0ac558ffa7b7">Why Many AI Agents Fail</a>, and reading it made me redraw my loop and rewrite half my definitions after continuing conversations with Ryan. This is version 0.2.</p><h3>The argument that moved me</h3><p>Ryan’s point is simple and uncomfortable. Documents hold information. Knowledge gets made when someone applies that information to a goal, checks the result, and adjusts. You judge information by whether it is accurate. You judge knowledge by whether it gets the task done.</p><p>His kitchen analogy makes it concrete. A recipe card is accurate and relevant, but against a dairy-free ticket it is only information. The knowledge that gets the plate out is made on the spot, when the chef swaps the butter, tastes, and sends it.</p><p>That exposed a flaw in my first loop. I had Knowledge on the forward path, sitting next to Information and feeding Context before anyone acted. In Ryan’s terms, that is treating knowledge as something you retrieve. It is the exact confusion his post is about, and I had drawn it into my own diagram.</p><h3>The loop, redrawn</h3><p>The new loop moves Knowledge from the forward path to the return path, and adds a step my first version was missing: Check.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FhbUlwSMlYxI4V5U5NkKNQ.png" /></figure><p><strong>Where the work starts.</strong> Wisdom chooses the goal, and people hold it. The Problem box turns that goal into a definition: the decision the system supports, who uses it, and the test for a good answer. Ryan, following Dewey, puts this first, because a problem well put is half solved. The Problem feeds both Information and Context, since it decides what counts as relevant and what goes into the package.</p><p><strong>The forward path.</strong> Data becomes Information once someone selects it for the task. Information is assembled into Context, still highlighted, because it is where everything meets. The Agent <em>applies it</em> and produces an outcome. That wording is deliberate: agents apply knowledge, they don’t make it.</p><p><strong>Sufficient context?</strong> The dashed box stays under Context. My question is unchanged; only the name moved.</p><p><strong>The return path.</strong> Check is where the result gets tasted: evaluations, user corrections, expert review. People own this step, and my first loop skipped it entirely, letting outputs flow straight back as data. Now two things come back. Raw outputs still return to Data along the outer line. Checked lessons, with their reasons, return to feed the next round of Context.</p><p><strong>Why two paths.</strong> This is Ryan’s pork chop point. Raw output only adds to the pile. A lesson with its reason carries over to cases nobody planned for, the way the dairy-free swap moved from the chicken to the pork chop. Apply, check, adjust: that loop is how knowledge gets made.</p><h3>The vocabulary, revised</h3><p>Three definitions moved, and two new terms arrived.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t9tjEN7acV5ECWia29YCOw.png" /></figure><p><strong>Wisdom got its old meaning back.</strong> I had used it as a label for sufficient context. Ryan reserves it for choosing the goal, like the chef who pulls the chicken special even though the temperature log says it is safe. That is a better use of the word, and sufficiency is a clear enough name on its own.</p><p><strong>Knowledge became an act, not an asset.</strong> It now lives in the doing: apply, check, adjust. That is why it no longer appears in Context directly.</p><p><strong>Learnings are what we can actually keep.</strong> This is the one place I part slightly from Ryan’s framing. When the chef explains the dairy-free swap and someone writes it on the card with the reason, that note is valuable. But it is not knowledge yet. It is the residue of knowledge, waiting for the next person or agent to apply and check it again. I want to call that residue <em>learnings</em>, and I don’t want to call it knowledge, because the moment we do, we are back to storing knowledge in a database and believing we own it.</p><h3>What I believe today (v0.2)</h3><p>As promised, here is the list again with what changed marked.</p><p>1. <strong>Relevance is task-relative.</strong> <em>Held, and sharpened.</em> The task is not enough on its own; the problem definition is what decides relevance.</p><p>2. <strong>Context is assembled, not stored.</strong> <em>Held.</em> What we store is data and learnings. Neither is knowledge.</p><p>3. <strong>Sufficiency is the real target.</strong> <em>Held, renamed.</em> No longer called wisdom. The Check step is now the only place we can tell whether context was enough.</p><p>4. <strong>It’s a loop, not a pyramid.</strong> <em>Held, and now has two return paths:</em> raw outputs and checked learnings. But I also am now convinced that the pyramid is still important. That is what guided the loop creation.</p><p>5. <strong>The data problems never go away.</strong> <em>Held.</em> Learnings inherit them too: versioning, scope, confidence, and going stale.</p><p>6. <strong>A tool is a component, not a solution.</strong> <em>Unchanged.</em></p><p>7. <strong>The memory has to be ours.</strong> <em>Refined.</em> The learnings have to be ours. The knowledge was always ours, because it lives in people.</p><p>8. <strong>The problem comes first.</strong> <em>New.</em> If we can’t state the decision, the user, and the test, no context will be sufficient.</p><p>9. <strong>Agents apply knowledge; people check it.</strong> <em>New.</em> Until we see an agent reliably move from line cook to chef, the Check stays with people.</p><h3>Still open</h3><p>The name question from Part 1 got harder. If knowledge is made rather than stored, then “Knowledge Layer” names the one thing the system cannot hold. What it can hold, and assemble, and route back, is data, information, and learnings. I still don’t have a better name, and I’m more convinced we need one.</p><p>A few questions I’m carrying into the next round:</p><p>• Can a learning carry enough of its reason that an agent applies it correctly on a case nobody planned for, the way the pork chop ticket worked?</p><p>• How do we know when a learning has gone stale, before it quietly wins?</p><p>• Could an agent ever own part of the Check step, or does checking always need someone who holds the goal?</p><p>The shift in this version is small on paper and large in practice. Part 1 asked how much context is enough. Part 2 adds the harder question underneath it: enough for what problem, checked by whom. Thanks to Ryan, Duane, and Brian for the dinner that started it and Thanks to Ryan for sharing a lot of thoughts post-dinner. Keep arguing with me.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f1280c5c5683" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/knowledge-is-made-not-stored-f1280c5c5683">Knowledge Is Made, Not Stored</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Zero to One: Two Days Back at Carnegie Mellon University]]></title>
            <link>https://medium.com/99p-labs/beyond-zero-to-one-two-days-back-at-carnegie-mellon-university-f47f848f26af?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/f47f848f26af</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[business]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 19:19:23 GMT</pubDate>
            <atom:updated>2026-11-11T19:19:23.310Z</atom:updated>
            <content:encoded><![CDATA[<p>Back to Pittsburgh, with a different question</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NiFgBrnXQmGsbcFk2T8lpQ.png" /></figure><p>We spent November 9 and 10 at Carnegie Mellon’s School of Computer Science. It was the sixth stop in our <em>Sensing the Frontier</em> series. Carnegie Mellon is the partner we know closely, and that made this visit different. We were less focused on discovering what exists and more on deciding what to build together.</p><p>From our side, Kiura-san, from Global Honda R&amp;D, joined Duane, Brian, Ryan and me. The two days had four kinds of rooms: first conversations with faculty we had never met, an early working session on a sponsored project, report-outs from four capstone teams, and a leadership conversation with the Dean’s office about where the whole relationship goes next.</p><p>We carried one question into every room, and it came from something Ryan and I keep saying to each other. Five years ago, the hardest part of innovation was going from zero to one: getting a credible idea at all. AI has made that part easy. The hard part now is going from one to ten: choosing which of a hundred plausible ideas deserves real effort, building enough conviction to carry it forward and executing on it. So we asked, in different words each time: once ideas are cheap, what becomes valuable? Here is what we heard.</p><h3>New introductions</h3><h3>Ideas are cheap; conviction &amp; execution is not</h3><p>Niki Kittur’s analogy engine, now a CMU startup, breaks expert fixation by finding where a problem has already been solved in another field. His key point: an idea alone is now nearly worthless. Analogies matter because they point to places where something has already worked, and expert steering then turns promising ideas into real ones. The open question we want to explore together is how to filter and build conviction once ideas are abundant.</p><h3>Language is not the whole story</h3><p>Yonatan Bisk, who advocates for robot-learning faculty across CMU, believes that agentic robotics is about to hit a wall. Every handoff between language reasoning and physical control throws information away, and today’s models have no metacognition: they do not notice when a tool is quietly wrong. His most memorable example was a seasoned machinist palming five screws and feeding them one at a time. No manual describes that skill, and no amount of text will teach it. For a company that builds things with its hands, that landed hard. Tacit expertise has to be measured before it can be learned.</p><h3>Delegation is a design decision</h3><p>Aarti Singh and the NSF AI Institute for Societal Decision Making study when human–AI teams outperform humans or AI alone, and how trust changes over time. Their work with the Red Cross and emergency-management agencies maps directly to our product development: which decisions an engineer delegates, which ones they verify, and how you test that safely. Ryan raised the “extended mind” idea, that a tool can be part of how we think rather than something outside us. It is a useful lens for students who believe that using AI is cheating.</p><h3>Context is the new data problem</h3><p>Shreya Shankar joins CMU’s computer science faculty next year and is building database systems for AI. Her open-source query engine, Quail, lets you analyze agent traces at interactive speed. That is a pain point we live with every day in our own observability stack. Kiura-san raised organizational intelligence: most of what a team learns lives in failed experiments and meeting debates, not in the final report. Shreya reframed it as a data-cleaning problem nobody has solved: context that conflicts, goes stale, or is overruled two meetings later.</p><h3>The missing rung on the ladder</h3><p>Jonathan Aldrich directs CMU’s Master of Software Engineering and leads PRAISE, a preceptorship program built with Microsoft. AI now writes the entry-level code that used to train junior engineers, yet companies need people who can supervise AI, which takes senior judgment. PRAISE pairs junior engineers with senior mentors and a credit-bearing online certificate to close that gap faster. As Honda moves many mechanical engineers into software roles, this is close to home. Jonathan also argued that AI-era teams need <em>more</em> documentation, not less. AI will maintain it and humans will review it, because humans will no longer read all the code.</p><p>Over lunch, Elias Towe walked us through the Center for Quantum Computing and Information Technologies, where industry members vote on which projects get funded. Kiura-san will share Honda’s quantum challenge areas, from battery sensing to materials discovery.</p><h3>A new project, in its early days</h3><p>In September we started a one-year sponsored research project with Tim Dettmers, and this visit was our first long working session together. Tim’s agent-guided graph search combines three layers. A symbolic layer finds information precisely but misses a lot. An unstructured layer holds everything. A semi-structured layer in between lets an agent get an overview without burning tokens. With an agent steering, the combination gets both precision and coverage, and it already sets the state of the art on autonomous research benchmarks.</p><p>Our problem is messier. Building software for a software-defined vehicle means code on one side and, on the other, specifications, architecture diagrams, slide decks and spreadsheets, with little traceability between them. Intent leaks out at every step: from concept to requirements to code, and across spec versions, Slack threads and whiteboard notes. By the end of the afternoon we had a shared framing: <em>what is the minimal sufficient context an agent needs, at the right moment, to solve a given task?</em> Retrieval is how you get from data to information. Sufficiency is the harder question underneath.</p><p>The next step is a set of benchmark tasks of graded difficulty, starting from a toy example that is still complex enough to be honest. Duane offered a realistic final test. CMU’s multi-year robotics challenge teams generate years of code, papers and reports, and each new cohort of students mostly starts from scratch anyway. That is exactly our problem inside Honda. If a new team, working with an agent that understands what came before, does not have to start over, we will have learned something real. Kiura-san added a fast-and-slow idea: quick symbolic lookups for routine needs and slower, deeper reasoning when the task calls for it, the way people decide how hard to think.</p><p>It is early. That is the point of writing it down now.</p><h3>What the students showed us</h3><p>Four capstone teams shared their progress. Every one of them had found its way to the same theme from a different direction: capturing context so people can act on it.</p><p><strong>CSL: rethinking how industry and academia work together.</strong> After more than 35 interviews, this team named four structural problems: mismatched calendars, mismatched expectations, briefs that are too broad or too narrow, and success that nobody measures the same way. Their sharpest finding was about AI. Students who care about their problem use AI to go further; students who don’t use it to fill in the requirements. Matchmaking matters more than ever. We asked them for an opportunity sheet rather than a white paper: which bets would they place, and why?</p><p><strong>Heinz: road safety, with receipts.</strong> This team built an agentic assistant that pulls together scattered crash records, plans and public input for transportation planners in Columbus, Ohio. Every sentence carries a citation, and the system says plainly what the evidence does not cover. Running a small model on a laptop, it beat a frontier model on local specificity. Our challenge to them was the hardest one: how will you measure whether safety actually improves?</p><p><strong>Co-X: a context engine for people with too many meetings.</strong> This team is on its third or fourth prototype of a workspace that prepares you for each meeting, tracks topics across email and chat, and mirrors existing permissions so nothing leaks. They took seriously our one rule, <em>don’t add to the cognitive load</em>. Duane’s feedback named the gap every busy leader feels: across eight projects that keep changing, which ones are actually on track?</p><p><strong>R2C: from paper to code to product.</strong> This team took our open-source Research2Code backend and turned it into an AI-native research workspace. Before generating code, it surfaces the settings a paper leaves out and the shortcuts an agent would quietly take. It then builds and audits the code module by module. They also stretched it past code into simulations, research-gap finding and learning by doing. I asked them to go one step further, toward the loop between research and product: what does it take for prototype code to earn a product team’s trust?</p><p>Watching four teams independently arrive at “capture the why, not just the what” told us something about where the problem really sits.</p><h3>A conversation about the next decade</h3><p>We closed with Dean Martial Hebert, George Darakos and Chris Kissell. Kiura-san opened with a candid problem. Honda’s technology strategy was built carefully, but for a slower era. Roadmaps are hard to change, research and product drift apart, and we observe slowness in both to start on a new technology and to catch up once we do. He wants a way for our global university network to help keep that strategy alive, not just validated once a year.</p><p>I described the white space we see. Capstones are fast and exploratory. Sponsored research is deep and specific. Between them there is no container for the continuous, contextual dialogue where both sides share what they are really working on, early enough to matter. Duane added the lesson from our long partnership with Ohio State: when it is easy to work together, you work together more.</p><p>Dean Hebert pushed us in a useful direction. Specific problems belong in sponsored projects, he said, but the collaborations that matter most are the ones nobody would have put on a roadmap. His example was a foundation model for sensing: not just vision and LiDAR, but heat, vibration, touch and radio frequency, the signals that actually fill factories and vehicles. It does not solve any one problem, and that is why it could transform many. Both sides, he said, tend to over-fit to faculty and topics we already know.</p><p>So we agreed to explore something new together: a jointly hosted workshop in 2027 that brings Honda’s technical leaders and CMU faculty into the same room. Its output would inform <em>both</em> Honda’s technology strategy and the school’s research direction, rather than being a Honda event held at CMU. Kiura-san shared how Japan is building new university–industry platforms, starting with Tohoku University, as one model to learn from.</p><p>None of this is decided. But it was the first of the many conversations I imagine focused on “where do we want to go together?” rather than “what can we do today?”</p><h3>What comes next</h3><p>We left Pittsburgh with a short, concrete list of action items.</p><p>If zero to one is now cheap, the scarce things are judgment, context and trust: knowing which idea to back, what the people before you already learned, and who to build it with. Every room this week was, in one way or another, about one of those three.</p><p>Soichiro Honda used to say that action without philosophy is a lethal weapon, and philosophy without action is worthless. Two days at CMU gave us plenty of philosophy. The next few months are about the action.</p><p>Thank you to Chris Kissell and Laura Alford for building two remarkable days, to Dean Hebert and George Darakos for their openness, and to every faculty member and student who shared their time and ideas. We will be back soon.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f47f848f26af" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/beyond-zero-to-one-two-days-back-at-carnegie-mellon-university-f47f848f26af">Beyond Zero to One: Two Days Back at Carnegie Mellon University</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nobody Knows What “Enough Context” Means]]></title>
            <link>https://medium.com/99p-labs/nobody-knows-what-enough-context-means-2e543fba7e77?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/2e543fba7e77</guid>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 15:51:22 GMT</pubDate>
            <atom:updated>2026-11-02T15:51:22.494Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ftYaiCOy1qNrUH2EJ60xuw.png" /></figure><p><em>Part 1 of an iterative series on context, knowledge, and how humans and agents get work done</em></p><h3>Why articulate this now</h3><p>AI agents are moving from novelty to coworkers in engineering work. They write code, review requirements, run tests, and summarize decisions. Almost everyone who works with them reaches the same conclusion: the quality of the work depends on the context the agent is given.</p><p>That realization has given us a phrase, the “knowledge layer”: a system that turns an organization’s scattered artifacts into memory that people and agents can use. It’s an appealing idea, and I believe in it. But it’s also an idea that hasn’t yet been articulated well, by us or, as far as I can tell, by the industry.</p><p>The market is moving faster than the thinking. New memory, context, and knowledge tools launch every week, and many of them are good. Each tends to frame the problem around what it already does, whether that’s storage, retrieval, or graph construction. When a big, loosely defined idea meets a well-packaged product, it’s easy to mistake one component for the whole.</p><p>This post is an attempt to step back and articulate the bigger idea: what problem we’re actually trying to solve, the vocabulary we need to discuss it, and the questions that remain open. It’s the first in a series, and it is deliberately unfinished. As we learn, I’ll come back and revise what I believe.</p><h3>Same slide, different problems</h3><p>We have a slide that describes a “Knowledge Layer”: a pyramid where operational knowledge refines into working memory and then into organizational memory, turning scattered engineering artifacts into durable, reusable memory for engineers and AI agents. I still believe in what that slide is reaching for. But when I listen to how people talk about it, I hear at least three different problems behind the same words.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eW9I-9ypwx9WO3nYPk2wbA.png" /></figure><p>the knowledge layer as first framed · three tiers over today’s scattered artifacts</p><p>In that framing, each tier refines the one below it, and value compounds toward organizational memory.</p><p><strong>The engineering lens.</strong> Here the knowledge layer is infrastructure: ingestion pipelines, storage, vector and graph databases, memory tools, access control. The questions are about where the data lives, how it’s indexed, and how fast it comes back.</p><p><strong>The methods lens.</strong> Here it’s a sequence of transformations: how raw artifacts become information, how information becomes knowledge, and how we evaluate whether any of it helped. The questions are about relevance, sufficiency, and measurable improvement in task completion.</p><p><strong>The organizational lens.</strong> Here it’s about capability that outlives any single project, tool, or person. The questions are about ownership, durability, and whether the organization ends up knowing more next year than it does today.</p><p>All three are right. Each is also partial. The trouble starts when one group solves its slice and assumes the whole thing is handled, or when two groups argue past each other because they’re describing different layers of the same problem. Before we can have useful disagreements, we need a shared vocabulary.</p><h3>A working vocabulary</h3><p>The old Data → Information → Knowledge → Wisdom progression is a useful starting point, with one change I think matters a lot: every step above data only makes sense relative to a task (something my colleague, Ryan Lingo stresses a lot). Relevance isn’t a property of a document. It’s a relationship between a document and something someone is trying to do.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F5lFOGJGSxOf_EmKh8tjng.png" /></figure><p>The data itself comes from three kinds of interaction. People produce artifacts on their own: documents, code, images, video, audio. People talk to each other through email, messages, and notes. And increasingly, people work with agents in sessions and discussions that are themselves a rich, mostly discarded, record of reasoning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ThEwgpM5Ts6eN-q666kZBw.png" /></figure><p>where the data comes from · three interaction types into collection and storage</p><p>The definition I most want us to argue about is <em>wisdom</em> or <em>sufficient context</em>. We can measure how much context we deliver. We can’t yet say how much is enough.</p><h3>It’s a loop, not a pyramid</h3><p>Our slide draws knowledge refining upward, with value compounding at the top. When I sketched it out on paper, I kept drawing something else: a cycle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4f5FyLeuQ4Ke60ZQW-Z7KA.png" /></figure><p>the context loop · five stages, one feedback path</p><p>The pyramid hides two things. First, every task produces data: code, decisions, test results, the transcript of an agent session. That output is raw material for the next task. Captured, it lets existing tasks get better or faster and makes new tasks possible. That loop is how engineering work evolves. Lose it, and the system stops learning.</p><p>Second, the problems we file under “data engineering” don’t stay at the bottom. Information and knowledge also need storage across different horizons (long-term, short-term, right now). They have a lifespan, they change over time, and they need lineage back to their sources. Every level inherits the same problems, which is why a storage tool, however good, only addresses one slice.</p><p>I’ve highlighted context because that’s where the loop is weakest today. We can assemble context. We can’t yet tell whether what we assembled was enough.</p><h3>Why “team context” is harder than it sounds</h3><p>Most context tools quietly assume one person, one agent, one task. Real engineering work doesn’t look like that. The relationship between people and tasks can be 1:1, 1:M, M:1, or M:N, all moving toward a shared goal. The moment you go beyond one person and one task, managing context stops being a storage problem and starts surfacing a set of harder ones.</p><p><strong>Access and security.</strong> Who is allowed to see what, and does that change when an agent acts on someone’s behalf? Context assembled for one engineer may include material another engineer, or an agent working for a supplier, should never see.</p><p><strong>Intent alignment.</strong> Two people can work on the same task with different goals in mind. An agent serving both needs to know whose intent it’s serving, and what to do when they conflict.</p><p><strong>Interpretation.</strong> The same artifact means different things to different readers. A test result reads as “done” to one engineer and “risky” to another. Context isn’t just what you hand over; it’s also how it will be read.</p><p><strong>Lifecycle.</strong> Context goes stale. Decisions get reversed, specs get revised, people leave. Knowing when a piece of knowledge has expired is as important as storing it in the first place.</p><p>I’m sure this list is incomplete. That’s part of the point of writing it down.</p><h3>Open questions, from both sides</h3><p>Here is where I want the positive friction. Every stage has a methods question and an engineering question, and neither can be answered well without the other. A storage design that ignores how relevance is judged will retrieve the wrong things quickly. A method that ignores lineage and access will work in a notebook and fail in an organization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9sc4UyfrimoAMb5d1-oBnw.png" /></figure><p>Two cross-cutting questions sit above the table. First, how do we evaluate any of this? “The tool returns results” is not an evaluation; “engineers finish this class of task faster with fewer escalations” might be. Second, how do we make sure the memory belongs to the organization rather than to whichever tool holds it this year?</p><h3>What I believe today (v0.1)</h3><p>These are working beliefs, not conclusions. I expect some of them to be wrong, and I’ll mark which ones changed when I revisit this post.</p><p>1. <strong>Relevance is task-relative.</strong> There is no such thing as relevant information in the abstract. Any system that ranks or filters without knowing the task is guessing.</p><p>2. <strong>Context is assembled, not stored.</strong> What we store is data and knowledge. Context is built fresh for each task, which makes assembly a first-class problem, not an afterthought of retrieval.</p><p>3. <strong>Sufficiency is the real target.</strong> Too little context and the agent improvises. Too much and it drowns. Measuring sufficiency is, to me, the most important unsolved question here.</p><p>4. <strong>It’s a loop, not a pyramid.</strong> Every task produces new data. If that data never flows back, the system stops learning on day one.</p><p>5. <strong>The data problems never go away.</strong> Storage, longevity, lineage, and access reappear at every level, from raw artifacts to judgment. Solving them once at the bottom isn’t enough.</p><p>6. <strong>A tool is a component, not a solution.</strong> Adopting one is a reasonable design decision. Declaring the problem solved because of it is not.</p><p>7. <strong>The memory has to be ours.</strong> Tools will change. The record of how our engineers think and decide should not leave with them.</p><h3>So, is “Knowledge Layer” the right name?</h3><p>We started this post with a slide and a phrase. Since then, the picture has shifted. Relevance turned out to be task-relative, context is assembled rather than stored, and the structure is a loop rather than a pyramid. The hardest question, how much context is enough, doesn’t appear in the original framing at all.</p><p>In light of that, I’m not sure “Knowledge Layer” still fits. “Knowledge” names one rung of the ladder. “Layer” suggests a static slab in an architecture diagram, something you install once. What we’re describing is a living loop: data becomes information and knowledge, gets assembled into context for a task, reaches a human or an agent, and the result flows back as new data.</p><p>So I’m looking for a new name, one that captures the data, information, knowledge, and wisdom progression as it plays out in agentic AI, where humans and agents share both the work and the context between them. Until something better emerges, “Knowledge Layer” stays as our working handle. If you have a candidate, I want to hear it.</p><h3>How this series will work</h3><p>This is version 0.1 of a belief system, written in public on purpose. Each future post will take one or two of the open questions above, share what we tried, and update the beliefs list, including the ones we got wrong. Some of that learning will come from experiments on our own team’s tools, where we are deliberately using ourselves as the first test bed.</p><p>To anyone working on this problem, whether you build the infrastructure or design the methods: the goal isn’t to agree with this framing. It’s to argue about the same thing. If you’re coming from the engineering lens, push on whether the methods questions are answerable. If you’re coming from the methods lens, push on whether the engineering can carry them. That friction is where the good design decisions will come from.</p><p>The tools will keep coming, and many will be useful. Our job is to know which part of the problem each one solves, and which parts are still waiting for us.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2e543fba7e77" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/nobody-knows-what-enough-context-means-2e543fba7e77">Nobody Knows What “Enough Context” Means</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Other ROI: Why AI Needs Two Definitions of Return]]></title>
            <link>https://medium.com/99p-labs/the-other-roi-why-ai-needs-two-definitions-of-return-0cd31eef224a?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/0cd31eef224a</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 22:03:45 GMT</pubDate>
            <atom:updated>2026-10-18T22:03:45.147Z</atom:updated>
            <content:encoded><![CDATA[<p>Every AI conversation with a leader eventually lands on the same question: <em>what’s the ROI?</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mPWnMsAFSwYRdgEp5HfDBQ.png" /></figure><p>It’s a fair question. It’s the question business executives and finance leaders are trained to ask, and for most of modern enterprise history it has been the right one. You invest capital, you expect a return, you measure the gap.</p><p>But here’s the uncomfortable truth that keeps coming up for me: for most enterprise AI work today, the honest answer is <em>we don’t know yet</em>. It is hard to calculate AI ROI. Not because teams are sloppy, and not because the technology lacks promise. We don’t know because we are trying to measure a new game with the scorecard from the old one.</p><p>I’ve come to believe that ROI isn’t the wrong question for AI. It’s just the wrong question asked <em>too early</em>. And it may need a second meaning before we’re ready to ask it in its traditional form. Here is my opinion piece about this topic.</p><h3>Why AI ROI is so hard to nail down and calculate right now</h3><p>Here’s the honest conditions we’re operating in. Calculating ROI assumes a reasonably stable system or technology: known costs, known capabilities, known outcomes. Almost none of that holds today with AI at every part of the software stack and in any workflow that AI is used in.</p><p><strong>The ground is constantly moving under our feet.</strong> We are living through an unprecedented speed of change. The model you evaluate this quarter is superseded next quarter. Architectures, tooling and best practices are reinvented every few months. The technology is not yet stable enough to implement an enterprise AI solution at scale in most cases with the confidence we’d expect from.</p><p><strong>People are still learning.</strong> Engineers, product owners, managers and executives are all catching up at the same time. Organizational fluency with AI lags far behind the technology itself.</p><p><strong>Token and overall cost economics are unpredictable.</strong> Nobody has a good handle on token spend or overall AI spend including professional services yet. Usage patterns shift as teams learn, agents loop, prompts grow, and pricing changes. Predictability, the foundation of any budget, isn’t there.</p><p><strong>Prototypes are easy; scale is hard.</strong> A proof of concept can come together in an afternoon. Turning it into a reliable, secure, integrated system that thousands of people depend on is a completely different problem. Completing prototypes is being used as a proxy for the solution working at scale and that is creating noise.</p><p><strong>Evaluations are hard.</strong> How do you know the system is good? How do you know it stayed good after the last model update? Evals are still an emerging discipline, and without them, “value” is mostly a feeling. Human judgement is still the key component in the workflow. Developers are still building the skill and leaders judging are still building the skill as well.</p><p><strong>The mindset shift is real.</strong> For decades, enterprise software has been deterministic: known input, known output. AI systems have probabilistic features. That shift unsettles how we design, test, trust and sign off on software.</p><p><strong>Nobody fully understands the risks.</strong> Hallucinations, data leakage, model drift, vendor dependency, regulatory exposure. We’re still mapping the territory.</p><p><strong>Governance is still old school.</strong> Most governance frameworks were built for deterministic systems with long release cycles. Applying them unchanged to AI either slows everything to a crawl or misses the risks that actually matter.</p><p><strong>User trust hasn’t caught up.</strong> There’s a confidence gap inside most organizations. Decision makers and developers often have high confidence in what these systems produce. The end users, the people expected to rely on AI in their daily work, are still struggling to judge the output, to trust it and to actually use it. A system nobody trusts is a system nobody adopts, and without adoption there is no return to measure.</p><p><strong>Our evaluation habits are being tested.</strong> In the previous era, project evaluations were rarely decided by data alone. A presenter could bring every test result and metric showing the system worked, and the decision often still came down to trust in the person presenting it. Today, many of us who evaluate projects are asking for a human in the loop before we trust an AI system. That instinct is understandable, and human oversight matters. But it’s worth noticing the pattern: in both cases, trust has been leading and evidence following. As leaders build a deeper understanding of how these systems behave, there’s an opportunity to shift toward evaluating on evidence, for people and for AI alike.</p><p><strong>The hype is drowning out judgment.</strong> Enterprise software vendors are racing to put AI into every product and every pitch. Their marketing is full of polished success stories that make it seem like everyone else has already cracked enterprise AI. The result is FOMO, and sometimes outright fear, in boardrooms: fear of falling behind, fear of being the company that missed the moment. That pressure doesn’t help anyone make rational decisions. Too often, we’re buying into the hype instead of asking whether a solution fits our problems, our data and our people.</p><p>Put all of this together and asking a team for a confident, <strong>quantitative ROI projection</strong> is is like asking explorers for an exact arrival date while they’re still drawing the map.</p><h3>The reframe: ROI has two phases</h3><p>Here’s what I’ve noticed. When a team brings AI into a real workflow, the most valuable thing that happens early on isn’t a cost saving. It’s that the team, and the organization around it, <strong><em>re-learns</em></strong> how that work should be done, and starts <strong><em>re-building</em></strong> the workflow around new capabilities.</p><p>That re-learning has a cost. Tokens, time, attention, failed experiments. And today, it’s mostly invisible on a spreadsheet.</p><p>So what if we kept the acronym but changed its meaning for the early phase? I think AI ROI comes in two phases:</p><p><strong>Phase 1 — ROI: Re-learning or</strong> <strong>Re-building on Investment.</strong> This is the phase before and during what I call the <em>AI Valley of Death</em>. The return here is organizational learning: what we discovered technically, what we discovered about the business, how the work itself changed, and which ideas we were able to discard quickly.</p><p><strong>Phase 2 — ROI: Return on Investment.</strong> This is the phase after the valley, once a solution has proven itself and is ready to scale. Here, traditional quantitative ROI comes back, and it should be measured rigorously.</p><p>The mistake many organizations make is demanding Phase 2 answers during Phase 1. That pressure doesn’t produce better projects. It produces inflated business cases, cautious teams, and a portfolio of safe pilots that never teach anyone anything. It leads to inaction and the cost of inaction can and most likely will be very expensive in the long run.</p><h3>The AI Valley of Death</h3><p>Here’s how I picture the journey an AI initiative goes through. Two curves matter: token spend and value realized.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RWNclsnzmudofHB6shybug.png" /></figure><p><strong>Exploration.</strong> Token use climbs fast. Teams experiment, prototype, try different models and approaches. Some value shows up in the form of prototypes and early insights, but spend clearly outpaces it. To a leader or CFO looking only at the cost line, this phase looks alarming.</p><p><strong>Decision: the AI Valley of Death.</strong> Spend drops as the team narrows down what actually works or get killed due to the sudden fear of high token bills. Value dips too, because prototypes don’t pay the bills and the real system isn’t built yet. This is the most dangerous part of the journey. It’s where projects get killed for the wrong reasons, or kept alive for the wrong reasons.</p><p><strong>Build &amp; Ship.</strong> For the projects that make it through, something different happens. Token use rises again but then stabilizes, now tied to real, repeatable workloads. Meanwhile, value compounds as the solution is adopted, integrated, and reused across the company. The two curves cross, and the gap keeps widening.</p><p>The key insight for leaders is this: <strong>the early spike in token spend isn’t a sign of waste. It’s the cost of learning </strong>(except keep a tab on people who are just using AI for cat memes and please don’t create token use leaderboards as well 😀)<strong> .</strong>Judging an AI project by its exploration-phase spend is like judging a student by their tuition bill.</p><h3>A different set of gates for Phase 1</h3><p>If the scorecard changes, the gates have to change too. Here’s the flow I’d like to propose for evaluating AI projects in the re-learning phase.</p><p><strong>1. Start with an elevator pitch.</strong> The team that wants to solve a problem writes a simple pitch: what problem is this, and what would a solution change including the work stlye? No long decks. No five-year ROI calculations.</p><p><strong>2. Lead with qualitative ROI.</strong> There should be no pressure to produce a quantitative ROI case. If the team <em>can</em> quantify the upside, great, but it isn’t required. What matters more is capturing the team’s intuition about the problem space. The people closest to the work usually sense where the leverage is long before they can prove it.</p><p>Alongside that, the team should answer two questions:</p><ul><li><strong>What will we learn?</strong> Technically (models, data, evals, integration) and from a business perspective (customers, processes, economics).</li><li><strong>How does the work change?</strong> Describe the before and after. Who does what today, and what would their day look like if this worked?</li></ul><p><strong>3. Leaders make an intuitive judgment.</strong> With so much AI noise in every organization, leaders’ job at this gate is to filter. Is this an important problem? Does the team have real insight into it? Is the learning worth having even if the project doesn’t survive? This is judgment, not arithmetic, and leaders should own that. Leaders should really bet on people at this stage.</p><p>I’ve come to believe that early-stage bets come down to a simple equation:</p><p><strong>Bet = Theme × Team × Social Proof</strong></p><ul><li><strong>Theme:</strong> Is this a problem space that matters, and will it still matter in three years? A strong theme means that even if this specific solution fails, the learning compounds in a direction the organization needs to go.</li><li><strong>Team:</strong> Does the team proposing the project have genuine insight into the problem, the curiosity to learn fast, be able to execute and the honesty to stop when the evidence says stop?</li><li><strong>Social Proof:</strong> Are there real signals pulling for this? End users asking for it, peers or partners seeing the same opportunity, early prototypes getting used without being pushed. And does the team have execution credibility.</li></ul><p>The multiplication matters. A great theme with the wrong team goes nowhere. A great team chasing a weak theme spends its talent on the wrong problem. And without any social proof, even a strong theme and team are building in a vacuum.</p><p><strong>4. Budget for the valley.</strong> If the problem is important enough to pursue, leaders should plan for the shape of the journey described above. That means expecting token spend to rise sharply during exploration, and not panicking when it does. The leader’s role shifts from gatekeeper to guide: helping the team navigate the valley rather than pulling funding at the first spike.</p><h3>Teams should kill their own projects</h3><p>The leader’s mindset isn’t the only one that needs to change. Teams need a shift too.</p><p>In the old model, teams build and leaders decide when to stop. Projects drift for months because nobody wants to be the one who says <em>this isn’t working</em>. In the new model, that waiting is a luxury we can’t afford.</p><p>If a team sees that its initial intuition isn’t playing out, it shouldn’t wait for leadership to pull the plug. <strong>It should kill the project itself</strong>, document what it learned, and redirect its energy.</p><p>This is where I think the real Phase 1 return lives. <em>Killing something faster is the return on investment in re-learning and re-building.</em> Every fast, honest kill frees up budget and people, sharpens the organization’s intuition about where AI actually helps, and leaves behind knowledge that makes the next attempt better.</p><p>For this to work, leaders have to make it safe. A team that kills its own project early, with clear learnings, should be recognized, not penalized. If killing a project feels like failure, teams will keep zombie projects alive, and the organization will pay for it in tokens and in trust.</p><h3>Crossing over: when traditional ROI comes back</h3><p>I don’t think every project will make it across the valley. Most won’t, and that’s fine. But I believe two things will happen.</p><p>First, navigating the valley will itself become an organizational skill. The second time a team walks through it, they’ll recognize the terrain. They’ll know when token spend is a sign of healthy exploration and when it’s a sign of a dead end. They’ll build evals earlier and kill faster.</p><p>Second, a few projects will reach the other side. That’s where the curve changes. Token use stabilizes, while value compounds as the solution is adopted, integrated and scaled across the company.</p><p><strong>That is the moment to bring traditional ROI back.</strong> Once a solution has crossed the valley, it can and should be measured with the rigor Leaders &amp; CFOs expect: cost per outcome, productivity gains, revenue impact, scale economics. The quantitative case becomes real because the system is finally stable enough to measure.</p><p>In other words, traditional ROI isn’t abandoned. It’s earned.</p><h3>New rules for a new game</h3><p>Soichiro Honda(our company founder) famously said that success is 99 percent failure. I’ve always read that not as a tolerance for failure, but as a discipline of learning through it: try, fail, understand why, and try again better. The Honda Philosophy’s emphasis on <em>initiative</em> says something similar about people: those closest to the work should be trusted to act on what they see.</p><p>That spirit fits this moment remarkably well. The rules of the game have changed. We can’t play the new game with the old rules, and that applies to both sides of the table.</p><p><strong>For Leaders &amp; Executives:</strong></p><ul><li>Separate the two ROIs. Don’t demand Phase 2 numbers from Phase 1 work.</li><li>Judge early projects on problem importance, team insight, and expected learning.</li><li>Budget for the shape of the valley, and expect token spend to spike before it settles.</li><li>Reward fast, well-documented kills as much as successful launches.</li><li>Bring traditional ROI back with full rigor once a project has crossed over.</li></ul><p><strong>For teams:</strong></p><ul><li>Pitch the problem and the before-and-after of the work, not a fantasy spreadsheet.</li><li>Be explicit about what you expect to learn, technically and for the business.</li><li>Own the decision to stop. Don’t wait for someone else to kill your project.</li><li>Treat every kill as an asset: write down what you learned and share it.</li></ul><p>Return on investment will matter again, and when it does, it will matter enormously. But to get there, we first have to invest in something harder to measure: an organization’s ability to re-learn and re-build itself. That is the first ROI. The second one depends on it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0cd31eef224a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/the-other-roi-why-ai-needs-two-definitions-of-return-0cd31eef224a">The Other ROI: Why AI Needs Two Definitions of Return</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Seattle Signals: Two Days at the University of Washington]]></title>
            <link>https://medium.com/99p-labs/seattle-signals-two-days-at-the-university-of-washington-19914d717d4e?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/19914d717d4e</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Sun, 11 Oct 2026 13:00:50 GMT</pubDate>
            <atom:updated>2026-10-11T13:00:50.720Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*lsi3XoQaKgNejjYAoVY4kQ.png" /></figure><p>We spent October 6 and 7 at the University of Washington and left with at least two capstones ready to start and a longer list of things to rethink than we arrived with. This was the fifth stop in our <em>Sensing the Frontier</em> series, and our first real visit to UW.</p><p>From our team, Duane, Joan, Ryan and I made the trip to experience the campus and participate in various conversations.</p><p>Day one was in Human Centered Design &amp; Engineering (HCDE). We started over coffee with Vice Dean Jihui Yang and Chair Julie Kientz, then met Gary Hsieh, David McDonald, Emily Tseng and Sean Munson, with a capstone-focused lunch in between. Day two moved to the Paul G. Allen School, where Dan Grossman and Shwetak Patel gave us an overview, followed by Stephanie Wang and Baris Kasikci, Katharina Reinecke and Natasha Jaques. In between, the Information School hosted lightning talks from Amy Ko, Bill Howe, Chirag Shah, Alexis Hiniker and Xiyuan Shen, convened by Associate Dean Jevin West.</p><p>We carried one question into every room: as AI changes how research, engineering and learning get done, who should we be learning alongside? Here is what we heard.</p><h3>What we learned</h3><h3>AI raises the floor, not the ceiling</h3><p>Julie Kientz gave us two metaphors we keep repeating. AI is like “ask the audience” on <em>Who Wants to Be a Millionaire</em>: reliable on the easy questions, useless on the hard ones. And AI is a forklift: perfect for moving heavy boxes, pointless if you came to the gym to get stronger. Her students now ask two questions of every assignment: <em>can</em> I use AI here, and <em>should</em> I?</p><p>We face the same tension inside Honda. We want engineers to use AI far more without their engineering judgment quietly eroding. Gary Hsieh ran that exact problem through his causal pathway tool live in the room, walking from a long-term goal to barriers and strategies and asking for evidence at each step. His rules for avoiding “vending machine” AI: never offer just one option, and rarely answer without asking a question back.</p><h3>Translation is really transformation</h3><p>Gary’s Paper2Card turns research papers into design cards, tailored videos and Figma recommendations, so findings reach the people who build things. Ryan &amp; I saw an immediate parallel with our open-source Research2Code project, which turns papers into working code. AI’s most underrated capability may be changing the <em>form</em> of knowledge, not just summarizing it. Gary also measures impact by adoption, not by how good the artifact looks, a bar we should hold ourselves to.</p><h3>Organizational memory is about people, not repositories</h3><p>David McDonald’s early work at Bell Labs recommended <em>who to talk to</em>, not which document to read. Repositories go stale; people adapt. Today he uses LLMs to mediate editorial disputes on Wikipedia, a model for productive friction rather than the friction-avoiding behavior of most chatbots. As Duane put it, in the agentic era work documents itself, and the dialogue with an agent reveals where expertise is thin. That turns expertise location into capability growth, a direct input to our thinking on the knowledge and context layer.</p><h3>Simulated humans are too polite</h3><p>Synthetic personas, Joan’s current research focus, came up in almost every room. Natasha Jaques showed why naive LLM user simulators fail: they answer everything up front, stay consistent and never get frustrated. Her group fits personas to real human data with a coverage metric, and agents tuned on them satisfied real users more. Her approach of having one model build tasks at the edge of what another can solve gave us a fresh angle on evaluating agents against the unstated intent behind a spec.</p><h3>Perceived capability drives behavior</h3><p>Katharina Reinecke studies overreliance: people follow AI advice even when it is wrong, and showing them how often others verify outputs reduces it. Joan connected this to the Ridgeline: a truck bed that looks like it can carry gravel invites people to load it with gravel. Emily Tseng brought the same lens from red-teaming chatbots for psychological harm, and we compared notes on a robotaxi passenger boxed in by strangers with no override, a problem the AV industry has not solved.</p><h3>The in-cabin experience is a measurable signal</h3><p>Sean Munson’s personal self-experimentation work in health sparked a conversation about motion sickness, which may matter more in autonomous vehicles than perceived safety, because comfort is learnable and nausea is not. Xiyuan Shen showed that drivers select touchscreen targets at about half their parked efficiency, and that an interface adapting to cognitive load cut selection time and off-road glances. Alexis Hiniker reminded us why it matters: technology shapes rites of passage like a teenager learning to drive.</p><h3>Prototypes are easy; production is a mirage</h3><p>Stephanie Wang and Baris Kasikci are redesigning ML systems so that agents, not just human engineers, can safely optimize them. They named something we feel daily: coding agents make prototypes so easy that the distance to a reliable production system gets harder to explain. Amy Ko sees the same in end-user software: you can prompt your way to a first version, but later iterations depend on domain knowledge the model lacks.</p><h3>Two capstones to get things rolling</h3><p>The most concrete outcome of the visit is two capstones at UW.</p><p><strong>Industry Capstone: a sensing agent that knows when to slow down.</strong> Students will build an agent on a small embedded device, using camera and LiDAR and a local model. When it sees conditions the vehicle was not designed for, it can do only one thing: ask the vehicle to slow down.</p><p><strong>iSchool: stop telling AI what to do, tell it why.</strong> Based on Stephen Bungay’s <em>The Art of Action</em>, students will build a system that captures a project’s intent and links each task to it, so people and AI agents can adapt when plans change.</p><p>Both ask the question we heard all week: how do we keep human intent at the center when machines act faster than we can explain? We are also curious about the iSchool’s new AI clinic, an experiment in longer-running, smaller projects.</p><h3>What comes next</h3><p>We left Seattle with real momentum. We are writing up a summary of every conversation to reflect on internally. The capstones come first; deeper research will follow the trust they build.</p><p>Soichiro Honda said that 99% of what we do ends in failure, and that this is what makes the 1% possible. UW felt like a place that shares that spirit: a department that renamed and rebuilt itself, faculty who would rather tell us what to do than what not to do, and students who build the thing as well as design it.</p><p>Thank you to Liz Young, Danna Bowers and the whole leadership team at UW for orchestrating two remarkable days, and to every faculty member, student and staff member who gave us their time and their ideas. We cannot wait to get started.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19914d717d4e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/seattle-signals-two-days-at-the-university-of-washington-19914d717d4e">Seattle Signals: Two Days at the University of Washington</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Horizontal View: Building a Team LLM Wiki]]></title>
            <link>https://medium.com/99p-labs/the-horizontal-view-building-a-team-llm-wiki-8afc7c6a1a8e?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/8afc7c6a1a8e</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 20:28:56 GMT</pubDate>
            <atom:updated>2026-10-04T20:28:56.202Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*61YU9VklGg9MHfn5VzZ7-A.png" /><figcaption>(Image from Ryan Lingo)</figcaption></figure><p>In July, a few of us sat down for a reflection meeting in Tochigi, Japan. Partway through, the conversation drifted to something ordinary: how one of the senior Honda leaders in the room keeps track of his work. He runs roughly fifteen projects at once. His notes were careful and well organized.</p><p>And yet the most valuable things he knew about those projects weren’t in the notes. They were in his head. Which decision in one project quietly constrained another. Which two teams were solving the same problem without knowing it. Where friction was building before anyone had named it.</p><p>We left with a reframe. This wasn’t a personal note-taking problem. It was a knowledge-layer problem with a potential huge upside for productivity gains, and in the AI era it might finally be a solvable one.</p><h3>We organize to find</h3><p>Most of how we manage information at work is built for retrieval. Folders by project. Notes by date. Status updates, meeting minutes, action-item trackers. The structure is vertical: drill into a project, scroll back through time, find the thing.</p><p>That works for the questions it was designed for. What’s the status? What did we decide last month? Who owes what by Friday?</p><p>But a different class of question has never had a home in that structure. What patterns repeat across our projects? Where are two efforts pulling against each other? What did a university capstone team learn last spring that our vehicle software team needs this week?</p><p>For those, we’ve relied on memory and systems thinking, usually the memory of whoever happened to be in enough rooms. That’s fragile. It doesn’t scale, it leaves when people change roles, and it’s weakest exactly when people are busiest.</p><h3>The horizontal view</h3><p>Here’s the hunch behind our recent work: the horizontal view of connections, correlations, tensions and friction across activities is now something a system can hold, not just a person.</p><p>Two things changed. First, an LLM can read everything. Every transcript, every set of notes, every project, without getting tired or forgetting what it read last Tuesday. That breadth is what lets it notice when the same constraint shows up in three unrelated conversations.</p><p>Second, it can meet you at whatever altitude you need. The same body of knowledge can be a one-line summary for an executive, a detailed trace of how a decision evolved for the engineer who inherits it, or an explain-it-like-I’m-five version for someone joining from another domain. Nobody has to choose the level of detail up front. The reader chooses it at the moment of asking.</p><p>Breadth across everything, depth on demand. That combination is what makes the horizontal view practical. Vertical organization doesn’t go away; it becomes the scaffolding the connections are built on.</p><h3>What we built</h3><p>Our starting point was Andrej Karpathy’s LLM Wiki idea. Instead of retrieving fragments from raw documents every time you ask a question, an LLM incrementally builds and maintains a persistent, structured wiki that gets richer as new material arrives. Karpathy framed it as a personal knowledge base. We wanted to see what happens when it’s built for a team.</p><p>The 99P Wiki takes our meeting transcripts and notes, from university capstones, internal syncs, partner collaborations and executive briefings, and keeps one living page per project. Each meeting updates its page: decisions, next steps, owners, with history and a link back to the source. A home page shows the week so far, its themes, and meeting volume by area. Anyone can ask a question in plain language and get an answer that cites the meeting it came from. Every Monday, a digest of the previous week lands in Slack.</p><p>Because every project lives in one place, it can also do what no single folder can: surface links, tensions and dependencies across projects.</p><p>The team went from idea to a live system in about a month. Today it covers roughly 120 meetings across about 25 project pages, and all of the AI runs inside on our internal platform.</p><h3>Asking instead of guessing</h3><p>One design choice mattered more than we expected. The system assumes as little as possible. When it isn’t sure about a name, a fact, or, very often, a transcription error, it doesn’t guess. It adds a question to a queue for a person to answer.</p><p>The answer corrects the record, and the system remembers it. The next time a colleague’s name gets mistranscribed the same way, it’s fixed automatically.</p><p>It’s a small feature with a large implication. A knowledge layer people rely on has to be honest about the edges of what it knows. Trust comes less from being right all the time than from being clear about when it might be wrong.</p><h3>A test bed for the Knowledge and Context layer</h3><p>We built the wiki for our own productivity, but that isn’t the main reason it matters to us. In the agentic systems we work on, we treat Knowledge and Context as its own layer: the part of the system that decides what an agent knows, how fresh that knowledge is, how much to trust it, and how to bring the right context to the right moment.</p><p>The wiki gives us a live test bed for that layer, with real data, real users and real mistakes. A few questions we want to pursue with it:</p><p>• <strong>Durability.</strong> When should a claim that keeps recurring harden into accepted knowledge, and how do you weigh evidence that isn’t equal?</p><p>• <strong>Evaluation.</strong> How do you measure the quality of a cross-project insight, not just whether retrieval found the right page?</p><p>• <strong>Proactivity.</strong> Can the system notice related past work during a conversation, in Slack for example, instead of waiting to be asked?</p><p>When we showed it to leadership, the first reaction was a request to use it for their own portfolio of projects. That’s the kind of feedback that tells you the problem is real.</p><h3>From scarce to shared</h3><p>For decades, the horizontal view of an organization has belonged to the people with the longest memories and the fullest calendars. That knowledge was valuable precisely because it was scarce.</p><p>What we’re seeing, early as it is, suggests it doesn’t have to be. Systems can now hold the connections, so people can spend their energy on judgment: deciding which tensions matter and what to do about them.</p><p>We’re about a month in. Next comes user feedback and iteration, and we’ll share what we learn, especially where it breaks. The part I’m very excited about is the “Philosophical Thinking” and “Technical Iterations” that will go into the <strong>Intention-&gt;Data-&gt;Information-&gt;Knowledge </strong>loop.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8afc7c6a1a8e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/the-horizontal-view-building-a-team-llm-wiki-8afc7c6a1a8e">The Horizontal View: Building a Team LLM Wiki</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rethinking the Industry–Academia Relationship — Part 3]]></title>
            <link>https://medium.com/99p-labs/rethinking-the-industry-academia-relationship-part-3-64fd3e061ae8?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/64fd3e061ae8</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Sun, 27 Sep 2026 17:13:47 GMT</pubDate>
            <atom:updated>2026-09-27T17:13:47.101Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What we learned by listening</em></p><p>Written by: Rajeev Chhajer and Ryan Lingo</p><p>In <a href="https://medium.com/99p-labs/rethinking-the-industry-academia-relationship-part-1-1ae90bbba7ec">Part 1</a> we mapped the gap between a fast-moving practice layer and the slow vehicles that connect industry and academia. In <a href="https://medium.com/99p-labs/rethinking-the-industry-academia-relationship-part-2-741635f0caa2">Part 2</a> we proposed three lenses for closing it (entrepreneurial thinking, open source building, and Liquid Research) and ended with a suggestion: the next step is a conversation.</p><p>We recently spent several days at a conference with the people who make university–industry partnerships work: research leaders, tech transfer and IP officers, corporate partnership leads, national labs, investors, and startup accelerators. One of us also joined a panel on how durable partnerships create value at every stage. Here is what we heard, and what we learned.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h9jK6fhcHIeMM-eMS8xJAw.jpeg" /><figcaption>Left to Right: James Williams (Teradata), Rajeev Chhajer (Honda), Chris Kissell (CMU) &amp; Cody Noghera (UCSD)</figcaption></figure><h3>1. The relationship is the return. The project is the receipt.</h3><p>We have always believed this relationship should not be transactional, and the room strongly agreed. The formal project is rarely the prize. Its most important output is usually the next conversation. What compounds is access: expertise the week you need it, faculty networks that open doors in days, and people who tell you when you are wrong. The strongest collaborations started as coffees years before they became projects.</p><p>Capstones, sponsored projects, and open source repositories are how a relationship gets exercised, not a substitute for it. Reimagining this relationship only happens where trust already exists.</p><h3>2. The horizons are compressing; our instruments haven’t.</h3><p>Most organizations file academic partnerships under Horizon 3, fundamental research years from market (see figure below). AI is breaking that assumption, compressing the path from research to application, sometimes from years to months. Keep academia in the long-horizon box and you miss the value available now: expertise to pressure-test decisions, early talent, and signal before it becomes consensus.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yQDv-NuuWIKcd2NCg4fT8A.png" /></figure><p>The tension is clear. Industry wants value within a quarter, while universities run on semesters and grant cycles. The fixes we heard (compress the vehicle, define honest quick wins, close what is not a fit) make sense, but they adapt existing instruments to a faster clock. They feel like band-aids. The bigger opportunity is new instruments built for compressed horizons, starting from a different question: not “what will this deliver in three years?” but “what did we learn this quarter that changed a decision?”</p><h3>3. Start with a relationship, not a project</h3><p>The most useful conversations were about changing structure, not just speed: build the relationship and the agreement first, then let projects emerge. Versions included master agreements signed before topics are chosen, tiered agreements with a no-IP fast lane, and small trial engagements that earn larger ones. These mostly fix contracting, but they are the foundation any new instrument needs: the ability to start in days without renegotiating each time.</p><h3>4. Every experiment needs a decision behind it</h3><p>We heard a simple rule: only ask a research question if someone will make a decision with the answer. Success is not a working demo or a finished paper. It is learning enough to say “keep going” or “stop,” and a clear stop is a good result. That is the measure we want for our small bets. Today’s instruments are judged by what they produce. A new one would start from the decision it needs to inform.</p><h3>5. Problem definition happens where the work happens</h3><p>The most common failure in lab technology is an impressive demo that no one needs. The advice was consistent: look for real pull, not a big addressable market, and go ask before you build, ideally before you have anything to show. Today, discovery is an add-on to projects. We think problem framing should be joint work done at the source, with both partners going to look together. For Liquid Research, a good problem posting would be the output of that visit.</p><h3>6. Candor and continuity keep a relationship alive</h3><p>Trusted partners say no early and exit gracefully. Yet many partnerships depend on one or two champions and fade when they leave. On the panel we shared our view: silence damages trust more than “no.” We decouple commitment from allocation, treat relationships as a team practice, and publish field notes so memory becomes institutional. Most fixes depend on individual discipline. A new instrument would build candor and continuity into its design.</p><h3>7. AI makes the network behind the expert more valuable</h3><p>The professor is the door; the network behind the professor is the thing. Our best signal has come from several uncoordinated groups independently pointing at the same problem, which gave us conviction quarters early. AI can find collaborators in hours, but it cannot supply judgment or foresight, or catch its own confident mistakes. As discovery gets cheap, value moves to trust and judgment. Yet today’s instruments fund deliverables, not ongoing access to that judgment.</p><h3>8. The matching layer exists, but in fragments</h3><p>We heard working versions of Liquid Research without the name: studios pairing shelved IP with entrepreneurs, companies bringing problems to universities in search of hungry teams, and startups de-risking technology faster than internal R&amp;D. Students, especially undergraduates, were called the most underused capacity, and building a venture breaks the need-experience-to-get-experience loop. Capstones work well but struggle to scale, and industry’s real constraint is time, not money. Each version serves one institution’s pipeline. The opportunity is connecting them into something shared.</p><h3>9. Incentives and openness are the hardest layer</h3><p>Two issues came up that no instrument can work around. The first is incentives. Faculty are rewarded mainly for papers and grants, not for working with industry. That is starting to change: some universities now count commercialization and industry mentoring toward promotion. Faculty are also more likely to try something new once a peer has done it first. The second is openness. Universities want to publish and share. Companies often need to keep work private. There is no good model yet for doing both on the same project. Until these are addressed, adjusting existing instruments will only go so far. This is where real redesign has to start.</p><h3>What it adds up to</h3><p>Much of what we believe was confirmed. Most of what works today adapts existing instruments, which makes it a foundation rather than the reimagining. The opportunity is new instruments, designed from first principles where things are hardest: scale, problem definition, incentives, and openness.</p><h3>The questions we are leaving with</h3><p>• What would an instrument designed for compressed horizons look like?</p><p>• How do you scale industry-embedded learning without losing what makes it work?</p><p>• How do you measure a relationship’s value, not just its projects?</p><p>• What IP framework lets open and proprietary work share a table?</p><h3>The conversation continues</h3><p>We went to listen and came back with sharper convictions and better questions. One more thing stays with us: AI is changing learning itself, including how students learn, how faculty teach, and how professionals keep up. Most instruments assume learning happens in courses and degrees, before work begins. Rethinking learning may be what lets us reimagine the instruments built on top of it.</p><p>The question now is not how to adapt what we have, but what to build together. If you are working on that, or think we have it wrong, we want to hear from you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=64fd3e061ae8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/rethinking-the-industry-academia-relationship-part-3-64fd3e061ae8">Rethinking the Industry–Academia Relationship — Part 3</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building New Relationships: Field Notes from University of Massachusetts, Amherst]]></title>
            <link>https://medium.com/99p-labs/building-new-relationships-field-notes-from-university-of-massachusetts-amherst-9d724fb45c6b?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/9d724fb45c6b</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Sun, 20 Sep 2026 23:30:49 GMT</pubDate>
            <atom:updated>2026-09-20T23:30:49.165Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A September day at UMass Amherst spent talking about capstones, open source opportunities, and what a software team’s AI memory should look like.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X4yGOqOwzD1zoA-O-XtKCw.png" /></figure><p>We have been more recently coming to the campus of U Mass, Amherst. Our relationship with Smith College goes back more than half a decade, and with U Mass in the neighborhood, we decided to explore opportunities there. Last spring we ran our first capstone at UMass Amherst with a master’s program there. This trip was about going deeper: a morning with the College of Engineering, an afternoon with the College of Information and Computer Sciences, and a research conversation with software engineering faculty to close the day. Three very different rooms, and yet by evening the conversations had braided themselves into one theme — how you design for ambiguity, whether the learner is a student, a sponsor, or a coding agent.</p><p><strong>Three mental models of collaboration</strong></p><p>In any given year our small team runs about twenty-five capstones across our university partners. That volume has given us a taxonomy we shared here, and it landed well enough that it is worth writing down.</p><p>There are two mental models that reliably produce bad experiences for an industry sponsor. The first is the consulting model: solve our problem, we don’t much care how, tell us at the end that it is solved. Students become cheap labor and everyone leaves disappointed. The second is what we call the genie-in-a-bottle model: take a problem your full-time engineers have not cracked, hand it to a student team for a semester, and hope for magic.</p><p>Our model is closer to an inverse consulting project. There will be a result either a prototype, a demo — but the end result is not what we are buying or hoping for. What we care about is the journey: how a team decomposes a problem, whether it is even the right problem, what trade-offs they discover, what the space looks like once someone has walked through it. The pace of change in software and AI is now so fast that year-scale research plans routinely get overtaken by a single model release. Short student innovation cycles, run in parallel sometimes several teams on the same problem, arriving at genuinely different answers and this gives us an interactive map of a space that is otherwise too vast to search. We call this sensing, and it is the reason these visits exist.</p><p><strong>Engineering the space for ambiguity</strong></p><p>The morning conversation with the College of Engineering (Jim Lagrant &amp; supported by Emily Nutwell) was a masterclass in the machinery behind authentic experience. Their industry-sponsored capstones run a full academic year, with faculty spending the spring and summer before working with sponsors to shape a problem statement that is achievable, honest, and marketed well enough that students say “that’s exciting, and I think I can do it.” The curriculum is the design process itself: scoping, converting wants and needs into engineering specifications, generating multiple candidate solutions, down-selecting, prototyping, and testing against both the analysis and the customer.</p><p>What struck us most was the shared diagnosis of what students find hardest. They have been trained for years on problems that have an answer, and a capstone is often the first time they meet a problem that does not. The room to play we offer is much larger than most students’ idea of room to play, and when they first get uncluttered, anxiety spikes. The craft is giving students enough room to think creatively without so little scaffolding that they are simply lost. One faculty member has removed rubrics from his classes entirely this year, arguing that in the AI era rubrics have curdled into something students optimize against rather than learn from. The replacement expectation is telling of how AI is re-shaping learning and also the acknowledgement that we know you don’t know everything, so show us the conversations you had with people who know more than you.</p><p>We also compared notes on decomposition. Their emerging model accepts problems complex enough that multiple teams, each staying in its home department and curriculum, work on different facets and must make their pieces fit, much like industry. We shared our own experiment in the same spirit: a multi-year platform that stays on campus as a persistent asset, with one cohort of teams building the mechanical and sensing foundations so that later cohorts can take on compute architecture and software. Both approaches point at the same insight: the interesting problems are bigger than one team and longer than one semester, and the academic calendar is a constraint to be designed around. This discussion definitely got us very excited about the potential to build from 1 capstone to a couple more on campus.</p><p><strong>Open source as curriculum</strong></p><p>The afternoon with the College of Information and Computer Sciences (Casey Maloney &amp; Emily Nutwell) turned to access. The experiential programs that exist today: summer research cohorts, data science for the common good, an open source apprenticeship pairing students with alumni mentors are outstanding and massively oversubscribed, with acceptance rates in the single digits is what we heard. Students who work jobs outside their coursework often cannot participate at all. The conclusion our hosts have reached is one we share: the way to make these experiences equitable is to bake them into the curriculum rather than bolt them on beside it.</p><p>Open source kept surfacing as the most promising vehicle. It sidesteps the intellectual property friction that slows industry projects, it produces public portfolios at a moment when students need them, and it matches where the industry itself is heading; more and more startups now open source first and build the business second. We talked concretely about building on and co-designing an open source contribution curriculum together, possibly piloted as a small group independent study in the spring, with mentorship layered so that faculty teach while industry and alumni mentors guide the teams. We are putting an open source project of our own into the world in the coming weeks, and we offered it and our support as raw material. We are happy to do experiments and try new methods out; the curriculum does not need to be perfect before the experiment starts.</p><p><strong>A research thread: what should a team remember?</strong></p><p>The day ended in a more research depth topic. It was a working conversation with software engineering faculty and researcher (Prof. Juan Zhai) about a problem we think is about to matter enormously. Developers and coding agents now build software together, and every one of those interactions produces rich context that today simply evaporates when the session ends. The single-player version of the problem is carrying context across your own sessions and that is being worked on across the industry. The multiplayer version is wide open: software is a team game, and there is no good answer yet for how the contexts of several people, each working with their own agents, combine into a shared, trustworthy picture of a project’s state.</p><p>The framing we kept circling was justified memory: a team memory that does not just store claims about the system but ties them to evidence, tracks when claims conflict, and records how conflicts were resolved. Get that right and the project state stops being a lagging indicator but something you learn at standup, or when code is finally committed and becomes a leading one. A project manager could see friction forming before it becomes a blocker. Two engineers whose in-progress work is drifting toward a collision could negotiate before the conflict exists rather than resolving it after. And in settings where teams span organizational walls, the question inverts from “who has access to what” to “what is the sufficient information each person needs to complete their task.” We left with homework on both sides and every intention of continuing the thread.</p><p><strong>Sense, share, build</strong></p><p>We have written before that university visits are how we sense, share, and build. In U Mass, Amherst, the sensing was less about any single technique and more about convergence: educators designing capstones, program builders designing curricula, and researchers designing agent memory systems are all wrestling with the same question: how do you structure an environment so that learning survives contact with ambiguity? We shared the taxonomy and scar tissue from six years of running student innovation at scale. And the build column has real candidates: a second UMass capstone cycle in the spring, participation in the open source apprenticeship’s next round, co-design of an open source curriculum, and a possible research collaboration on team memory for human-agent software development. Which of these turn into work is the conversation of the coming months. That the conversation exists is why we keep making the trip.</p><p>A huge thanks to Emily Nutwell, Jim Lagrant, Casey Maloney &amp; Juan Zhai for taking time to meet with this.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9d724fb45c6b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/building-new-relationships-field-notes-from-university-of-massachusetts-amherst-9d724fb45c6b">Building New Relationships: Field Notes from University of Massachusetts, Amherst</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Hard Parts Stay Hard]]></title>
            <link>https://medium.com/99p-labs/the-hard-parts-stay-hard-0bc9d12b37c1?source=rss-c4c5ab7ae68a------2</link>
            <guid isPermaLink="false">https://medium.com/p/0bc9d12b37c1</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[liberal-arts-college]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Rajeev Chhajer]]></dc:creator>
            <pubDate>Sun, 20 Sep 2026 23:30:36 GMT</pubDate>
            <atom:updated>2026-09-20T23:30:36.381Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Field notes from two days at Smith College</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aQ8KGMq-MPaPv4jLSyH5EQ.png" /></figure><p>Most of our field notes in this series come from research universities: from labs where the frontier is being built. This visit was different. Smith College is a liberal arts college of about 2,500 undergraduates in Northampton, Massachusetts, and what we went to sense was not a new architecture or a new benchmark. It was something we think matters just as much: how a community that prizes slow, careful, humanistic thinking is metabolizing a technology that rewards speed and what a long-term industry partnership should look like when it is built on that ground.</p><p>We have been coming to Smith for a while. Our team has sponsored the engineering Design Clinic for roughly eight years and the Statistical &amp; Data Sciences (SDS) Capstone for five or six. This trip we spent two days of back-to-back conversations with SDS and engineering faculty, the Conway Innovation &amp; Entrepreneurship Center, the Design Thinking Initiative, the Lazarus Center for Career Development, the Provost’s office, and a packed seminar room of students. It was about taking the next bold step. The title of the talk Ryan gave on day one doubled as the thesis of the whole visit: use AI boldly and question it relentlessly. Smith, it turns out, is a campus unusually well built for the second half of that sentence.</p><p><strong>A campus that refuses the easy answer</strong></p><p>Every campus we visit is metabolizing AI differently, and Smith’s way is distinctive. The college’s institutional stance on AI is, deliberately, not to take a stance. Some disciplines are embracing these tools, others are deeply skeptical, and the college holds that disagreement in the open, treating it as part of a liberal arts education rather than a problem to be solved. There are moments, one academic leader reminded us, when slowness and friction are the better mode, and a campus should be able to say so.</p><p>For an industry partner, this is not an obstacle. It is exactly the environment you want if you believe, as we do, that the scarce skill of the next decade is not using AI but judging it. Several faculty described what students call “AI guilt”: a feeling, even among willing users, that they are getting away with something. The answer is not to argue that guilt away, but to convert it, through experience, into discernment. You cannot question a technology relentlessly if you refuse to touch it; you cannot use it boldly if you have never learned to doubt it.</p><p><strong>What a summer case study taught us</strong></p><p>The most concrete evidence came from a case study we ran with Smith this past summer. A design problem that a previous student team had worked through over a full academic year with traditional methods was taken on again but this time by a smaller team, over a single summer, deliberately augmented by AI tools and mentored along the way.</p><p>The headline result: the summer team came close, not all the way, but close enough to where the year-long team had landed. The finer print is more interesting. The easy parts of the project got dramatically faster. The hard parts stayed hard, and still deserved the time. And because the students had been challenged from day one to question every output, they caught an error in the AI’s work that a less skeptical team would have shipped and reworked it properly. When we caught up with them during this visit, they were more open to these tools than they had been in June, and still not fully convinced. We consider that the correct end state for a summer.</p><p>The faculty conversation around this case study was some of the best pedagogical thinking we have heard on any campus. There is a real difference, one professor noted, between AI helping students get their work done and AI helping them learn and the benefit seems to skew toward those who already have a foundation to build on. Critical thinking requires a bedrock of skills underneath it; a student who has never truly written code cannot yet tell the difference between asking an AI how to do something and asking it to do the work. None of this is an argument against AI in education. It is an argument for sequencing and for capstones, where students are at their most prepared, as the right place to go all in.</p><p><strong>The meta-skill is taste</strong></p><p>At a lunch roundtable, faculty asked us the question their students ask them: what will employers actually expect? Our honest answer is that the interview question of this era is becoming “how will you use AI to do this job?” and that the durable skill underneath it is shifting. When solving a well-specified problem gets cheap, the premium moves to deciding which problem is worth solving: mapping a messy situation, identifying stakeholders, extracting what they actually need rather than what they first ask for, and only then reaching for the technology. Some in industry call the pattern forward-deployed engineering. A liberal arts college should recognize it as its home turf.</p><p>The most productive disagreement of the lunch was about motivation. If a machine can solve problems, one professor pressed, and taste in problems can only be earned by solving them, do we risk raising a generation that never builds the taste? Someone offered the car and the body: when machines took over locomotion, exercise did not disappear but it did become a choice, and we lived through the consequences for those who chose not to. Our working hypothesis is the three-team version: a team that uses AI as an answer machine, a team that refuses it, and a team that uses it to get the work done and to learn at the same time. We believe the third team wins in both the first order and second order but we said plainly that this needs study, not assertion. The silver lining the table converged on is one a liberal arts college can love: in a world where you no longer have to learn to get the answer, learning becomes something you choose because you want it. Experiential learning does not get less important in that world. It becomes the whole game.</p><p><strong>Building the connective tissue</strong></p><p>The structural problem with even a great capstone relationship is that it resets. We arrive each fall, engagement peaks, the project ends, and momentum returns to zero until the next cohort. Much of this visit was spent designing, with partners across the college, the connective tissue that would let themes carry across academic years instead of expiring with them.</p><p>The pieces under discussion are deliberately small seeds rather than a grand apparatus: a standing summer research program, like the one piloted this year, to bridge one academic year to the next; post-baccalaureate research associates likely housed in Smith’s centers rather than any single department who carry a theme forward and connect capstones across disciplines; part-time externships during the academic year, so that a student’s campus job can be future-facing experiential work rather than a shift in the dining hall; and a framework that would let Smith faculty spend sabbaticals engaged with our research teams. Underneath all of it is a lesson we keep re-learning with our university partners: when you lower the barrier to starting something together, far more things get started.</p><p>Two research platforms will anchor the technical side. This year’s Design Clinic teams begin the first phase of an autonomous golf cart not as a traditional autonomy project, but as a testbed for a question we genuinely do not know the answer to: can an agentic system be made hardware-agnostic, so that the intelligence that operates a golf cart can operate a boat, or a drone? The second platform is an AI-augmented software development workflow that students from any discipline not just computer science can use as an instrument for their own research. Both are intended to be open, durable, and owned as much by the campus as by us.</p><p><strong>What we’re taking home</strong></p><p>A short list, in the spirit of the series. First, next visit we want student roundtables, not just talks: ten students around a table taught us more in an hour than any packed lecture hall. Second, an idea from the lunch worth developing: a course, or a module, on reading AI news critically, separating what the technology is doing from what headlines claim it is doing is itself a data-literacy problem, and Smith is built to teach it. Third, a reminder we needed: the students who resisted these tools the longest, once they engaged, were the ones who caught the hallucination. The questioning half of our seminar title is not the obstacle to the bold half. It is what makes boldness safe.</p><p>We go to the spot. We’ll keep going.</p><p>Huge thank you to Katie Zapert, Patti Thornton, Mike Kinsinger, Ben Baumer and all the faculty and students who participated in deep 2-way conversations with us.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0bc9d12b37c1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/99p-labs/the-hard-parts-stay-hard-0bc9d12b37c1">The Hard Parts Stay Hard</a> was originally published in <a href="https://medium.com/99p-labs">99P Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>