<?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 Owen Brown on Medium]]></title>
        <description><![CDATA[Stories by Owen Brown on Medium]]></description>
        <link>https://medium.com/@owenbrown?source=rss-cf32699c9e88------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*EfbNWnc8V1q9F1YT7tByrQ.jpeg</url>
            <title>Stories by Owen Brown on Medium</title>
            <link>https://medium.com/@owenbrown?source=rss-cf32699c9e88------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 07:51:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@owenbrown/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[Prior art on cattle securitization and fraud protection]]></title>
            <link>https://medium.com/@owenbrown/prior-art-on-cattle-securitization-and-fraud-protection-f15e28cf84ec?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/f15e28cf84ec</guid>
            <category><![CDATA[fraud]]></category>
            <category><![CDATA[securitization]]></category>
            <category><![CDATA[crypto]]></category>
            <category><![CDATA[patent-protection]]></category>
            <category><![CDATA[livestock]]></category>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 18:30:47 GMT</pubDate>
            <atom:updated>2026-10-01T18:30:47.124Z</atom:updated>
            <content:encoded><![CDATA[<p>As beef prices increase, the securitization of cattle becomes increasingly important to meeting the world’s demand for beef. However, with complex financing comes fraud. There has been a slew of recent Ponzi schemes, where ranchers borrow against cattle that do not exist.</p><p>Technology can be used to verify the existence and health of cattle.</p><p>For example, an on-body animal monitors, such as Seismie’s ear tags, can continuously monitor the vital signals and location of cows and store that on a cryptographic ledger or traditional centralized database, so that investors can verify and analyze the health of a herd. This introduces challenges. Where there is opportunity for fraud, fraudsters will be creative.</p><p>For example, fraudsters may choose to put multiple tags on a single animal. Seismi is designing several methods to prevent this. One method is to track the unique seismographic data of an animal. We can use statistics and machine learning to know when two monitors are placed on the same animal. We can use other signals, such as distance. The problem is not unlike anomaly detection used for monitoring credit card transactions.</p><p>Data can be compressed and stored on cryptographic ledgers so that it cannot be tampered with later and so that it is accessible to investors.</p><p>Seismi’s monitors can cryptographically sign the data that is generated. Each device having a public key that is stored when the device is manufactured, and a private key that exist only on the device. This significantly raises the difficulty level of fraud. Coupled with tamper resistance, using technology similar to e-fuses eFuses and anti-downgrade fuses or the technology that Apple record water damage to hardware, Seismi’s devices will reliably record and transmit real data.</p><p>The public key associated with each device truly is public and can be stored in a cryptographic ledger, such that investors and auditors can verify the provenance of the data. In the event of false positives for tampering, the device will still function. In the event of tampering, a device will still function. It will just transmit that it has been tampered with.</p><p>Geographic data may prove more reliable than Seismi and temperature data. Data suggesting that two animals are unnaturally close to each other for even a short period of time may indicate double tagging of an animal.</p><p>In the event that fraud continues its current trend, Seismi intends to introduce audit features. Specifically, when an animal is tagged, the user who tagged it is cryptographically recorded so that they can be held accountable later, and an image of the animal is recorded. This record, as with the animals’ health data, can be stored on a cryptographic ledger. On-device signing provided by Android and iOS platforms, and C2PA technology for verifying provenance of images, can be further used to raise the cost of committing fraud.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f15e28cf84ec" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prior Art — Time locked bank account]]></title>
            <link>https://medium.com/@owenbrown/prior-art-time-locked-bank-account-72cfbe51492a?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/72cfbe51492a</guid>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Sat, 16 May 2026 13:52:05 GMT</pubDate>
            <atom:updated>2026-05-16T13:52:05.012Z</atom:updated>
            <content:encoded><![CDATA[<p>AI is enabling social engineering and fraud at a scale unseen in modern times.</p><p>At the same time, companies and individuals can receive large value by enabling AI agents to make financial transactions directly.</p><p>For explicitly enabling AI agents to make small transactions, I think segregated spending accounts with limits are the most efficient method. In other word, give the AI agent a credit card.</p><p>With regards to fraud, I anticipate elderly and less well educated people will have their life savings wiped out. I think it is time to create a new bank account type: A time vault.</p><p>The concept behind a time vault is that certain spending activity is explicitly temporally gated. So, for example, if you have a million dollars in your savings account, it becomes not possible to wire that money out on a moments notice. The bank provides a mechanism where the person must initiate the transaction, then wait a pre-determined period of time. There could be detail policies, or simple tiers based on cumulative activity.</p><p>Additionally, for larger transactions, it makes sense to add a concept of co-signers. That is, an elderly parent’s account would be configured so that her trusted child needs to approve very large transactions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72cfbe51492a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prior Art: RDBS backed service to coordinate retail audit and compliance]]></title>
            <link>https://medium.com/@owenbrown/prior-art-rdbs-backed-service-to-coordinate-retail-audit-and-compliance-94612c7c3b2e?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/94612c7c3b2e</guid>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Fri, 16 Jan 2026 04:10:35 GMT</pubDate>
            <atom:updated>2026-01-16T04:10:35.941Z</atom:updated>
            <content:encoded><![CDATA[<p>Retailers, retail display manufacturers, brands that sell in retail, and third party labor providers struggle to coordinate actions when auditing retail fixture.</p><p>An internet-based platform could reduce frictions for all parties.</p><p>Third party labor providers perform audits using mobile phone and tablet web apps and apps. Audits include structured questions specific to each display type. A website and app-based platform allows the different parties — labor, manufacturer, retail, and brand — to ask questions and update audit structure.</p><p>Different parties can user the platform to review audit data, classify problems, ask each other questions, and ultimately order replacement parts to either the third party labor provider or directly to the retailer for installation, usually by third party labor or the retailer.</p><p>Machine learning, through tradition classification algorithms or through pre-trained algorithms and VLLMs, can be used to classify audit data.</p><p>The problem of reviewing audit data can be treated as a data labeling problem. Managers reviewing the audit data can delegate ordering of parts and issuing work orders, but must monitor their delegates. Traditional data labeling management practices can be used for this monitoring.</p><p>A simple platform, backed by a performant databases and composed of of traditional crud interfaces, can forms the foundation for improved management.</p><p>Layered on top of this platform, communication and issue management tools can further reduce frictions. Value created can be communicated using analytics — dashboards and exports of pivotable data.</p><p>Implementation is vanilla, not worth describing in further detail. This article existed as prior art to stop trolls.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=94612c7c3b2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prior art: cryptographic proof of real image generation for the purpose of preventing document…]]></title>
            <link>https://medium.com/@owenbrown/prior-art-cryptographic-proof-of-real-image-generation-for-the-purpose-of-preventing-document-fcb606d3f6ad?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/fcb606d3f6ad</guid>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Thu, 20 Nov 2025 18:34:10 GMT</pubDate>
            <atom:updated>2025-11-20T18:34:10.906Z</atom:updated>
            <content:encoded><![CDATA[<h3>Prior art: cryptographic proof of real image generation for the purpose of preventing document fraud</h3><p>The creation of fraudulent images for the purpose of fraud is growing rapidly. It is possible to prevent document fraud by using cryptography to verify that a specific device took an image and the image was not tampered.</p><p>Mobile phone manufacturers like Apple can cryptographically sign their images, using software or hardware, to prove that the image was generated directly by a camera.</p><p>This would be an effective deterrent to most forms of expense reimbursement fraud.</p><p>This is the expected workflow:</p><ul><li>A company issues a phone to an employee, or the employee brings their own phone</li><li>The employee installs expense reimbursement software (app) like Brex, Navan, etc., through the regular method (Apple App Store, Google Play Store, etc.)</li><li>The expense reimbursement app records the public key of the camera. Ideally the key is at the hardware level, so that it cannot be tampered with by software. For companies that issue devices, this key could be shared through the manufacture’s (Apple, etc.) device management program. But, this is not necessary to deter most fraud. Even a software-generated would block fraud by unsophisicated users.</li><li>The employee takes pictures of documents using their phone. The pictures of the documents are signed using the key. A cryptographic hash prevents tampering.</li><li>The expense reimbursement company checks that the uploaded image matches the one of the keys on file for that user, and that it has not been tampered.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fcb606d3f6ad" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prior art for using blue light to brighten clothing and remove stains.]]></title>
            <link>https://medium.com/@owenbrown/prior-art-for-using-blue-light-to-brighten-clothing-and-remove-stains-298fc4116aed?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/298fc4116aed</guid>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Fri, 12 Sep 2025 02:25:00 GMT</pubDate>
            <atom:updated>2025-09-12T02:25:00.762Z</atom:updated>
            <content:encoded><![CDATA[<p>Consider “Environmentally Friendly Photobleaching Method Using Visible Light for Removing Natural Stains from Clothing” by Tomohiro Sugahara, Hisanari Yoneda</p><p>This article is prior art that hopefully invalidates patents of painfully obvious technologies.</p><ol><li>A machine could be built that uses high intensity blue light to brighten clothing.</li><li>The machine may have a timer that turns it off.</li><li>The machine may have a dial that controls the intensity of the light.</li><li>The machine may use mechanism to avoid igniting the clothing. Mechanism include a transparent material that separate the lighting element from the clothing, heat sinks to disperse heat, liquid such as water to disperse heat.</li><li>The machine may be used in combination with chemicals such as hydrogen peroxide.</li><li>The machine may block other parts of the spectrum, so that only certain bands of light reach the clothing.</li><li>The machine may use LED lighting that emits light at specific frequencies.</li><li>The machine may use high frequency LED lighting to activate phlosphorescents that emit light at different temperatures.</li><li>The machine may use mirrors or magnifying glasses to focus the rays of the blue light over specific parts of the fabric that are stained.</li><li>The machine may be controlled by computer. The computer may actively direct the light at parts of the clothing that are active.</li><li>The machine may be operated by remote payment, credit card, pre-paid card, or coin. For example, this machine could exist at a laundry matt.</li><li>The machine could have people people drop off the their clothing, and then require some system to identify whose clothing is whose when picking up.</li><li>It could also be a home device that plugs in.</li></ol><p>0[\\\slP</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=298fc4116aed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prior Art 2025–06–23]]></title>
            <link>https://medium.com/@owenbrown/prior-art-2025-06-23-7c7b75edef30?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c7b75edef30</guid>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Mon, 23 Jun 2025 18:51:56 GMT</pubDate>
            <atom:updated>2025-06-23T18:51:56.534Z</atom:updated>
            <content:encoded><![CDATA[<p>I hate software patents, so I’d like to get these ideas out as public domain.</p><ol><li>System for storing markdown or plain text files in a single, ordered, file or folder system for joint access by AI, Large Language Models, or humans.</li><li>System for AI to propose changes to internal documents, publish documents to a web page, allow those documents to be reviewed by a human, then, if human approves through UI, make the proposed change the “production version” of the system or policy. For example, a an AI proposes changes, on its own or through prompting by a human, to a companies expense reimbursement policy. This results in a proposed change that a human reviews, clicks an “accept” button or accepts through voice command or messaging system message, and then the AI takes the action.</li><li>AI system that proposes payments to other parties or internally. Either on its own or through human prompting, the system creates a ledger, similar to a “git diff” with changes to the ledger. Human reviews the changes. If the human accepts the changes to the ledger, the AI then uses a payments platform (like stripe, brex, etc.) to programmaticly make the change.</li><li>Framework for organizing a companies internally documents for easy access by both humans and AI / large language models. Just as Django and other frameworks allow a human to create a project in an opinionated but flexible manner, a human user creates a new project. The AI may or may not ask the user a series of questions. Then, the AI creates a folder structure that sorts documents for easy access.</li><li>Context perserving data structuring. A system that fetches, hashes, caches, and summarizes internal and external documents into a structure that allows LLMS and other AI to read the documents, and only the subset of the document, in a way to conserve the use of tokens that are burned when rifling through unstructured or less intelligently structured systems.</li><li>System for proposing changes created by an LLM graphically. Specifically, a user and/or/with AI/LLM creates changes to a configuration, setting, a data table. The data the changes are then represented graphically, for a user to then accept, reject, or modify.</li><li>System 1–through-6 above delivered, delivered over messaging system or text-primary system. This application and user interface is obvious and should not be patentable.</li><li>System for recording a displaying context data usage across AI tasks.</li><li>System for programmatically measuring discoverability, understandability, and operability of computer and software systems by AI. For example, measuring the success of API calls generated by LLMs for a service.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c7b75edef30" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Efficiency of Internet Advertising — Continuous Price Discrimination]]></title>
            <link>https://medium.com/@owenbrown/the-efficiency-of-internet-advertising-continuous-price-discrimination-f4fe0ddc8828?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/f4fe0ddc8828</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[internet]]></category>
            <category><![CDATA[advertising]]></category>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Sun, 12 Mar 2023 23:54:21 GMT</pubDate>
            <atom:updated>2023-03-12T23:57:57.304Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The Efficiency of Internet Advertising — Continuous Price Discrimination</strong></h3><p>Internet advertising is the most powerful form of price discrimination ever created. Before I explain why, let me summarize what price discrimination is, and why it benefits society.</p><h4>Price Discrimination</h4><p>Price discrimination is the practice of charging users different prices for essentially the same service. Three simple examples are:<br>1. Airlines charge different prices for an airline ticket on the same plane.<br>2. Grocery stores offer discounts to users who are willing to clip coupons.<br>3. Retailers put items on sale, temporarily reducing the price, to entice buyers who would not otherwise purchase.</p><h4>Price discrimination is efficient and benefits society</h4><p>Price discrimination allows producers to create value that would not otherwise be possible. For example, imagine a ferry route that costs $100 per day to operate. 4 passengers are willing to pay $10 per ticket to ride, and four other passengers are willing to pay $20 per ticket to ride.</p><p>If the ferry operator charges $10 per ticket, all passengers purchase a ticket, but the operator earns only $80, and cannot operate profitably. The operator will close the route.</p><p>If the ferry operator charges $20 per ticket, four passengers purchase a ticket and the operator again earns only $80 and will close the route.</p><p>However, if the ferry operator can find a way to charge, for example, $8 to the first group and $18 to the second group, all riders will purchase a ticket, the operator will collect $104, earn a $4 profit, and everyone is better off.</p><h4>Online advertising is continuous price discrimination</h4><p>With advertising, users pay for services through their attention. Attention is scarce and has value.</p><p>The brilliance of advertising is that the amount of value transferred to the advertiser is directly proportional to the user’s ability to pay. Wealthy users, who have a larger willingness and ability to purchase goods, will spend more on advertised products. The user’s attention is worth more, so they pay more.</p><p>Most price discrimination schemes are <strong>discrete</strong>, bucketing users into one of a small number of tiers. This pricing scheme creates kinks, where some users receive hardly any excess value, and the users who can pay the most for a product the most are always pay too little.</p><p>Internet advertising, by contrast, is <strong>continuous</strong>. A user with 1% more disposable income pays 1% more to use an advertising-based service. The continous nature of the system allows it to captures value from the long tail of ultra-wealthy users in a way that no other system can.</p><p>Can you think of any other pricing scheme that is able to charge a centa-millionaire 10x more for than a deca-millionaire, and charge a billionaire 10x more than a centa-millionaire, all while providing the exact same service? Yet through targeted advertising of $50,000 handbags and $30 million-dollar apartments, online advertising is able to do this.</p><h4>Google.com will never offer an ad-free version</h4><p>In response to the recent fire at Google, I’ve heard some people say, “Google only earns X dollars per user per year. Why not offer an option that charges X + Y dollars, with no ads?”</p><p>The flaw in this argument is that the attention of the users who can pay the higher price is worth much more than the amount they’d pay for the subscription. This is true no matter what price point you set.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f4fe0ddc8828" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prior art for ai-enabled digital assistant.]]></title>
            <link>https://medium.com/@owenbrown/prior-art-for-ai-enabled-digital-assistant-a22e3db7d79a?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/a22e3db7d79a</guid>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Thu, 09 Mar 2023 00:35:50 GMT</pubDate>
            <atom:updated>2023-03-09T00:35:50.227Z</atom:updated>
            <content:encoded><![CDATA[<p>I’d like to digital assistant based on a large language model is trained the following data sources:</p><ul><li>webpages I visit</li><li>emails sent to and from me</li><li>documents in my organizations knowledge base (i.e., Google Docs, Meta Workplace, Quip, etc.)</li><li>News article through news services such as Bloomberg, Apple News, Associated Press</li></ul><p>In addition, I would like an application that does the following:</p><ul><li>Understands my current website browsing behavior</li><li>Visits URLs (i.e., sends a GET request to the URL) of web pages, in advance of when I actually visit the web page, and stores the data locally, so that it can be retrieved quickly.</li><li>In some cases, stores a compressed version of the data locally. Perhaps stores the file in a Markdown format.</li><li>Presents all content to me in nice, compressed, easily readable list, similar to an email inbox.</li><li>Click on items in the list should open the comment almost instantly, since the data stored locally.</li><li>As needed, machine learning should both surface the content I am interested in, and pre-fetch that content so that I don’t need to get it myself.</li><li>The AI model should consider the resources that are available to me, and follow links to download and train on data that might be of interest to me. For example, if I receive an email with a PDF or a link to a Google Doc or news article, the AI should train on the content of the article.</li><li>I should be able to ask the AI questions regarding the content that it is trained on. So, if I am sent the quarterly earnings report for a company, I should be able to ask the AI questions about the contents of the report.</li><li>Data should stored in a way that is private to me. Other people should not have access to this data. This could be achieved through several methods, including breaking the model up into seperate sub-compontents, some of which are accessible to me, and some of which are not.</li><li>Reduce storage and processing costs by storing and compress document that may interest multiple people by storing a single embedding that multiple people can access.</li><li>Balance maintaining my privacy while sharing training resources with other users.</li><li>Off significant customizability for my user interface.</li><li>Make the presentation of all data that interests my uniform. That is, convert all news articles into a single format. I can edit color, size of text, etc.</li><li>I should be able to scroll through articles that interest me by subject alone, the way I do with email.</li></ul><h4>Trust</h4><p>The machine should leverage a graph-based trust system. I should both explictly and implicitly trust humans and institutionts, who explicitly and implictly trust other humans and institutions. This creates a network, so that I can infer trust. That is, if I trust a person, and they trust a person, I have amount of trust for the person who is one link removed.</p><p>Explicit trust involves me stating that I “trust” a person through a user interface.</p><p>Implict trust might occur through sharing an article someone wrote.</p><p>The machine learning model should only train on trusted sources.</p><p>There should be a mechanism for flagging content as bad, untrusted, or dangerous.</p><p>The machine learning model should have a mechanism for un-learning or flagging information that is no longer trusted.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a22e3db7d79a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Wide Path or Narrow Path?]]></title>
            <link>https://medium.com/@owenbrown/wide-path-or-narrow-path-c6f499884fae?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6f499884fae</guid>
            <category><![CDATA[product]]></category>
            <category><![CDATA[technical-product-manager]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[project-management]]></category>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Thu, 02 Jan 2020 05:04:20 GMT</pubDate>
            <atom:updated>2020-01-30T05:14:00.864Z</atom:updated>
            <content:encoded><![CDATA[<h3>What military supply lines can teach us about technical product management</h3><h3>Tanks! Send the tanks now!</h3><p>A gruesome, weeks-long battle is being fought in the city of Lutz, Zubrowka. Your side holds the city, and you, as Chief Engineer, need to deliver reinforcements to the city. “Tanks! Send the tanks now!”, pleads the captain of the besieged troops.</p><p>Your army’s tanks, trucks, and infantry arrive at the port, where you are located. Three types of terrain separate the port from the city.</p><ul><li>First, a forest, two kilometers wide, filled with trees and brambles</li><li>Next, a mountain range, one kilometer wide</li><li>Finally, a river, eight meters across</li></ul><p>You quickly deduce that the two most important decisions you must make are what path to build to the city, and how wide a path to build. Your options are to build:</p><ul><li>a narrow path, allowing only infantry to pass</li><li>a wider path, allowing infantry and trucks to pass</li><li>the very wide path, allowing infantry, trucks, and tanks to pass</li></ul><p>You have chainsaws and machetes to cut through the forest. Your map shows several mountain passes wide and flat enough for tanks to pass through. You possess one inflatable bridge that can support the weight of a tank. There are no enemies on your side of the river.</p><h4>You weigh your options and decide to build a road wide enough for a tank</h4><p>Having decided the width of the road, choosing your route is straight forward. Using a map of the terrain, you decide on a path straight through the forest, from the port to the to nearest mountain pass wide enough and flat enough for a tank. The river 8 meters wide at all points, so you decide to cross just after the mountain pass. From there only 200 meters of grassland separate you from the city.</p><p>As you begin to clear a path through the forest wide enough for a tank, problems emerge immediately. You must cut down trees, and it takes longer than expected for your troops to learn to use chain saws effectively. Days pass.</p><p>Then, halfway through the forest, your troops discover a deep, wide ditch. You’re forced to select a different route. You end up abandoning most of the road you completed to this point.</p><p>When your troops finally arrive at the mountain pass, they discover that the ground will crumble under the weight of the tank. You’ve already built the road through the forest, so it’s too late to select another mountain pass. You’re forced to reinforce the pass. Days are lost.</p><p>Finally, you have some luck. The inflatable bridge deploys flawlessly. As your tanks roll over the bridge, you radio the besieged captain, telling him that you’re a mere 200 meters away. He sounds relieved, mentioning that he was about to request paratroopers.</p><p>KABOOM. It’s at that moment that you learn that the 200 meters of grassland between the river and the city is a mine field. You’ll need to move your bridge.</p><p>Your troops report that there is a clear path to the city a mere 100 meters south. However, your inflatable bridge isn’t designed to be moved. As you scramble to decide what to do next, the city is captured.</p><h4>What went wrong</h4><p>Stripped of your command and rank, peeling potatoes in the bunker, you reflect on how things went so badly.</p><p>You realize that your mistake was starting with a road wide enough for a tank.</p><p>Instead, you should have started by building a narrow path for infantry, all the way to the city.</p><p>In half a day, your troops, using chainsaws or even simple machetes, could have cut a path through the forest’s brambles wide enough for infantry to pass through.</p><p>As your troops passed through the forest, they would have discovered that the original route was untenable. The narrow path, up to that point, would still have been abandoned, but only a half day’s work would have been lost. More importantly, when the path for the tanks was eventually built, the correct route would have been chosen.</p><p>Once your troops were traveling through the mountain pass, they would have been able to test whether the pass could support tanks. Having not yet built the road for a tank, you could have then selected a different mountain pass.</p><p>Instead of using an inflatable bridge, your troops would have spent only a few hours to set up a zip line to move infantry and light equipment across the river. When your troops encountered the minefield, moving the zip line would have been trivial. Most importantly, you’d have known the correct place to deploy the inflatable bridge for the tanks.</p><p>Once your troops arrived in the city, there may also have been advantages that you could not have anticipated. You might have learned that snipers and infantry armed with rocket launchers are nearly as effective as tanks. At a minimum, the arrival of additional troops would have raised morale.</p><p>Your decision to build a wide road from the start lost the battle.</p><h4>How this applies to software</h4><p>Customers will often ask for the first version of a product to have many features. In the world of data, customers ask for every type of data that they think they might at some point need.</p><p>Building these features before creating a working product that delivers value to the customer is wrong. Contrary to expectation, building all the features from the start <em>delays the delivery of those features</em>. The reason is that there will be unanticipated changes to the design of the product, and each additional feature will increase the amount of re-work or even lock in a bad design.</p><p>Instead, build the narrowest, most simple product possible, that fully delivers value to the customer and validates the overall design. When you do this, unanticipated problems will occur, and changes to the design will be made. However, you’ll be able to easily navigate around these obstacles because there will be considerably less code to re-write.</p><p>Design decisions should create a path for the final product to be delivered. A simple rule of thumb is to minimize writing, debugging, and deploying code that you know will be discarded. Usually, simply leaving out features does not result in code being discarded.</p><h4>Communicate the approach to the team</h4><p>Building a minimal, complete product first is the fastest, most direct path to building a fully featured product. However, the end user is often distrustful that their needs are understood and will be met. Fearful, they will argue to build a full-featured product in stages.</p><p>To avoid this destructive behavior, educate all team members, including engineers and the stakeholders, on the reason for building a minimal product.</p><p>Make it clear to stakeholders that you understand their concerns.</p><p>Communicate in writing what will be delivered in your initial release. Communicate explicitly the features that will not be included, and why. Communicate design decisions with which stakeholders disagree. Communicate the appropriate method for appealing your decisions.</p><p>Finally, communicate how product decisions are made. Make sure people don’t confuse the right to complain, give advice, and openly debate with the right to make decisions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6f499884fae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Methods for decreasing latency on conversational platforms (Alexa and Google Assistant)]]></title>
            <link>https://medium.com/@owenbrown/methods-for-decreasing-latency-on-conversational-platforms-alexa-and-google-assistant-abdc755bd04c?source=rss-cf32699c9e88------2</link>
            <guid isPermaLink="false">https://medium.com/p/abdc755bd04c</guid>
            <category><![CDATA[voices]]></category>
            <category><![CDATA[alexa]]></category>
            <category><![CDATA[google-home]]></category>
            <category><![CDATA[actions-on-google]]></category>
            <category><![CDATA[echo]]></category>
            <dc:creator><![CDATA[Owen Brown]]></dc:creator>
            <pubDate>Thu, 24 Jan 2019 17:50:19 GMT</pubDate>
            <atom:updated>2019-02-06T19:37:31.464Z</atom:updated>
            <content:encoded><![CDATA[<h3>Methods for decreasing latency on conversational platforms (Alexa and Google Assistant), IOT, and mobile apps</h3><p>This article describes several methods for reducing latency experienced by users on conversational platforms.</p><p>Many of these methods could also be implemented on apps (iOS or Android), or on conversational IOT hardware.</p><p>My intention to two fold — to inspire continued evolution of these platforms and the ecosystem, and to serve as prior art should any of these methods ever be patented.</p><h3>Business need</h3><p>Voice conversational platforms such as Actions on Google and Amazon Alexa Skills allow third parties to create voice applications.</p><p>The platform offer bad experiences because user experience considerable latency.</p><p>This paper describes a method for reducing this latency.</p><h3>Current system</h3><p>Currently, each time a user sends send a response the agent, the response travels through the users hardware device, through software on the device, to the platform, and from the platform, and usually from the platform to a server that is managed by the third-party developer.</p><p>For example, when a user speaks to Amazon Alexa, the hardware device, which is usually an Amazon Echo, transmits information to the Alexa Platform, which processes the signal, then transmits information, either via an HTTPS post or via direct-invocation of an AWS Lambda function, to the third-party developer.</p><h3>Method for decreasing latency via caching</h3><p>The total latency to the user can be decreased from and average of 4 seconds, to less than one second, using the following method.</p><p>When a user initially invokes the application, and on each turn of the conversation, the third-party developer identifies most likely queries that a user will give in the next turn of the conversation. Queries includes button presses and other inputs, in addition to spoken or typed queries. This list of probable queries, and the and the responses that should be spoken to the user for those queres, are loaded into the the local memory of the hardware. Note that responses include UI elements such as pictures, menus, buttons, and carousels.</p><p>When the user speaks their response, the response is processed locally, on the phone, checking to see whether the user’s response matches any of the probable responses that are loaded in the memory of the hardware.</p><p>If the user query is matched, then the response to the users query that is already in the hardware’s memory is immediately spoken to the user. This is possible, and results in a vastly improved user experience, because the users query does not need to be transmitted to the platform or to the third party prior to speaking the response to the user.</p><p>While the response is being read to the user, the query is transmitted to the platform, and the next set of probable queries and responses are loaded into the hardware devices memory.</p><h3>Method for decreasing latency via processing of partially spoken queries</h3><p>A further method for decreasing latency is to begin processing the response to a users query prior to the user finishing speaking or typing their query.</p><p>To do this, key words and phrases are transmitted, from the third party developer, to the hardware’s memory. This list of keywords and phrases map to a much larger number of responses then can be loaded into the hardware’s memory. This true because a response may contain audio or a pictures that take up considerable memory. Keywords, by contrast, are small.</p><p>When these keywords or phrases are spoken, a request is immediately placed to the third party developer for the response associated with this keyword or phrase. The final, complete response of the user may or may not be mapped to the this response. The processing of the complete query may occur locally on the hardware, or on the platform.</p><p>However, if the complete query is mapped to this response, then the response can be returned to the user nearly immediately, if the intent recognition was done on the hardware, or without waiting for a response for the third party, if the intent recognition was completed on the platform.</p><h3>Method for decreasing latency via predictive processing using machine learning</h3><p>Machine learning could be implement, by the platform, to predict which queries a user is most likely give, in any situation.</p><p>The burden of predicting the user’s response could be shifted from the third-party developer to the platform. Learnings from different third-party developers could then be shared, to create one efficient model.</p><h3>Method for decreasing latency via memoization</h3><p>Latency can be further reduced by allowing developers to specify that certain queries will always map to certain intents.</p><p>Platforms currently transmit information about the user, the query, and the context, which is a list of parameters stored at the platform level.</p><p>The third-party developer returns a response to the platform. This response is a function of these three things, (the user, the query, the context), plus other information.</p><p>If the developer were allowed to specify, in his response to the platform, that the response is a function of only the query and the context, then the platform can cache the response. Future responses can then be delivered from the platform, directly to the user, without having to place an additional request to the developer.</p><p>The developer would also include a time-to-live with his response, specifying for how long the response should be cached.</p><h3>Method for decreasing latency by forming direct connections</h3><p>Currently, platforms such as Actions on Google and Amazon Alexa route all user queries through a central platform, pre-process those interactions, and then forward processed information about those queries to the fulfillment server of the third party.</p><p>Response time can be further reduced by allowing the hardware device to directly request information from the third-party fulfillment server. This reduces the distance that the information needs to travel, and the number of servers that interaction must pass through.</p><h3>Method for decreasing latency via alternate protocols of data transmission</h3><p>Currently, the method for transmitting data to the third-party’s fulfillment centers is either to send an HTTPS post request to a URL endpoint, or to directly invoke a server-less function, (AWS Lambda or Google Functions, or Google Firebase, etc).</p><p>Allowing either the hardware or the platform to connect over a faster protocol, such as gRPC, and transmit information using a compressed format, such as protocol buffers, would reduce latency.</p><h3>Applications for iOT, mobile apps</h3><p>Each of the methods listed above can be applied not just to platforms, but to internet-of-things (IOT) devices, and to mobile applications that by-pass platforms such as Alexa and DialogFlow</p><h3>Method for hybrid NLP using local and remote processing</h3><p>To save mobile device battery, decrease bandwidth usage, and speed up responses to user, a device could perform text-to-speech, intent recognition, entity extraction, and other NLP locally.</p><p>After such processing is performed, a certainty-level is assigned to the output. If the certainty level is above a specific amount, a locally cached responses is immediately served, or a response is fetched from the cloud. No NLP is performed remotely.</p><p>If the certainty level is below the specified amount, then the original response from the user is sent to the cloud for processing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abdc755bd04c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>