<?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 Sergey Enin on Medium]]></title>
        <description><![CDATA[Stories by Sergey Enin on Medium]]></description>
        <link>https://medium.com/@sergeyenin?source=rss-966f4cb6ae9c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*juGw21Zj2F11Zp7wD0GLPQ.jpeg</url>
            <title>Stories by Sergey Enin on Medium</title>
            <link>https://medium.com/@sergeyenin?source=rss-966f4cb6ae9c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 07:02:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sergeyenin/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[Why AI Agents Work in Demos but Fail in Production]]></title>
            <link>https://blog.chatbotslife.com/why-ai-agents-work-in-demos-but-fail-in-production-6f11542e375f?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/6f11542e375f</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[finops]]></category>
            <category><![CDATA[ai-governance]]></category>
            <category><![CDATA[mcp-server]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Tue, 23 Jun 2026 13:10:11 GMT</pubDate>
            <atom:updated>2026-06-23T13:10:11.375Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fxVP0QzqlAZFd0BA-wsdCA.jpeg" /><figcaption>The AI silver bullet doesn’t exist as well :(</figcaption></figure><h3>The model is not your production system.</h3><p>Most AI agent demos work. That is not the interesting part.</p><p>The interesting part is what happens after the demo: when the agent gets real input, real tools, real customer data, unclear instructions, stale context, retries, rate limits, and a monthly API bill somebody has to explain. That is where <em>many agent systems fall apart</em>.</p><p>Not because the model is bad. Because the system around the model is barely a system.</p><p>We keep talking about agents as if the main problem is intelligence. Better reasoning. Bigger context. Better benchmarks. Better coding scores. Better tool use. All useful.</p><p><strong>But not enough.</strong></p><p>A production agent does not fail only when the model gives a bad answer. It fails when a bad intermediate answer becomes state, gets passed to the next step, is trusted by another model, and eventually turns into a real action. That is the part people underestimate. A multi-agent system is not a team of smart assistants. It is a distributed pipeline of untrusted intermediate state. Once you see it like that, the problem becomes much less magical. Also much less new.</p><p>We already know how to build systems like this. We just forgot, because the demo looked too good.</p><h3>The reliability math is ugly</h3><p>Imagine a workflow with ten agent steps. Each step is “pretty reliable”. Say 98%. That sounds good until you multiply it:</p><pre>0.98^10 = 81.7%</pre><p>So your system made of individually good steps now fails roughly 1 in 5 times. And in real systems it is worse, because failures are not independent. A malformed output is not just a malformed output. It becomes input.</p><p>A hallucinated tool result is not just a bad answer. It becomes context.</p><p>A wrong classification is not just wrong classification. It can decide which tool gets called next.</p><p>This is why demos are misleading. The demo shows the happy path. Production tests the handoffs. And the handoffs are where agents break.</p><h3>What actually breaks</h3><p>Usually it is not dramatic. It is boring engineering failure.</p><p>An agent returns JSON that is almost correct. One field is missing. A status is invented. A number is formatted as text. The next step accepts it anyway.</p><p>An agent reads a tool result and slightly over-interprets it. There was no customer match, but the model writes as if there was one.</p><p>A supervisor keeps retrying because no step is obviously terminal. Nobody sees a crash. The system just burns tokens.</p><p>A support bot gets access to tools it should never have seen: delete_user, export_all_records, update_payment_status.</p><p>Then everyone hopes the system prompt will keep it safe. That is not a control. That is a prayer.</p><p>The worst part is what happens after an incident. Someone asks:</p><blockquote>What did the agent see?<br>Which model did it use?<br>Was there PII in the request?<br>Which tools were available?<br>Which tool calls were blocked?<br>Why was the request allowed?<br>How much did it cost?<br>Can we prove the log was not edited?</blockquote><p>If the answer is “we think”, this is not production infrastructure yet.</p><h3>The better mental model</h3><p>Treat agents like unreliable but useful components inside a reliable system. That is not anti-agent. That is engineering.</p><p><em>Every handoff needs a contract. Every model output is untrusted until validated.</em></p><p><em>Every tool call is a privileged operation.</em></p><p><em>Every retry needs a budget.</em></p><p><em>Every action that can change production state needs a commit boundary.</em></p><p><em>Every important decision needs evidence.</em></p><p>This is not as exciting as “autonomous agents”.</p><p>But it is how you get from a demo to something you can leave running.</p><h3>Validate handoffs</h3><p>Do not pass raw model output to the next step. That is the first rule.</p><p>For Python teams, the simplest version is Pydantic plus structured output:</p><pre>from pydantic import BaseModel, Field, model_validator<br>from typing import Literal, Optional<br><br>class TicketDecision(BaseModel):<br>    action: Literal[&quot;approve&quot;, &quot;reject&quot;, &quot;escalate&quot;]<br>    ticket_id: str<br>    risk_score: float = Field(ge=0, le=1)<br>    reason: str = Field(min_length=20, max_length=500)<br>    approver_id: Optional[str] = None<br><br>    @model_validator(mode=&quot;after&quot;)<br>    def enforce_business_rules(self):<br>        if self.action == &quot;approve&quot; and self.risk_score &gt; 0.7:<br>            raise ValueError(&quot;High-risk tickets cannot be auto-approved&quot;)<br>        return self</pre><p>This is not about making the model perfect. It is about making bad state harder to pass forward. The model can still be wrong. But it has fewer ways to be silently wrong.</p><h3>Be careful with “try harder”</h3><p>Best-of-N helps. Judge models help. Retry loops help.</p><p>But they also create a new problem: the system gets expensive and unpredictable.</p><p>A supervisor that retries, re-ranks, calls another model, asks for a second opinion, and then calls another tool may look “agentic”. It may also be an API bill generator. So yes, generate multiple candidates for hard tasks. <strong>But validate them</strong>. Rank them. Put a cap around the whole loop.</p><p>No agent should be able to spend unlimited money trying to feel more confident.</p><h3>Put a gateway in front of model calls</h3><p>Local validation is not enough. At some point you need a control point outside the agent code. That control point should decide, before the request reaches the provider:</p><blockquote>Is this caller allowed?</blockquote><blockquote>Is this model allowed?</blockquote><blockquote>Is there PII in the request?</blockquote><blockquote>Should it be redacted, blocked, or routed differently?</blockquote><blockquote>Would this request exceed the budget?</blockquote><blockquote>Are dangerous tools being exposed?</blockquote><blockquote>Should this require human approval?</blockquote><p>This is where <a href="https://dativo.io/">Talon</a> fits. Talon is not trying to be another agent framework. It is the control layer around agent traffic.</p><p>An existing app can keep using the OpenAI-style API. You point it at the Talon gateway instead of directly at the provider.</p><pre>gateway:<br>  mode: enforce<br><br>  callers:<br>    - name: &quot;production-agent&quot;<br>      tenant_key: &quot;talon-gw-prod-001&quot;<br>      tenant_id: &quot;default&quot;<br>      policy_overrides:<br>        max_daily_cost: 50.00<br>        max_monthly_cost: 1000.00<br>        allowed_models:<br>          - &quot;gpt-4o-mini&quot;<br>          - &quot;gpt-4o&quot;<br>        pii_action: &quot;redact&quot;</pre><p>The important part is not the YAML. The important part is where the decision happens. A dashboard tells you money was spent. A gateway can stop the call before money is spent.</p><h3>Do not show the model dangerous tools</h3><p>This one is simple. If the model should not call a tool, do not show the tool to the model.</p><p>Do not expose delete_*, admin_*, export_all_*, bulk_*, drop_*, and then rely on a prompt saying “be careful”. That is backward.</p><pre>gateway:<br>  default_policy:<br>    tool_policy_action: &quot;filter&quot;<br>    forbidden_tools:<br>      - &quot;delete_*&quot;<br>      - &quot;admin_*&quot;<br>      - &quot;export_all_*&quot;<br>      - &quot;bulk_*&quot;<br>      - &quot;rm_*&quot;<br>      - &quot;drop_*&quot;</pre><p>This is prevention, not detection.</p><p>If the model never sees delete_customer, it cannot call delete_customer.</p><p>For production agents, I prefer the even stricter version: explicit allowlists.</p><pre>callers:<br>  - name: &quot;internal-support-bot&quot;<br>    tenant_key: &quot;talon-gw-support-001&quot;<br>    tenant_id: &quot;default&quot;<br>    policy_overrides:<br>      allowed_tools:<br>        - &quot;search_kb&quot;<br>        - &quot;read_ticket&quot;<br>        - &quot;create_draft&quot;<br>      tool_policy_action: &quot;block&quot;</pre><p>Give the agent the tools it needs. Nothing else.</p><h3>Treat tool execution as a commit boundary</h3><p>Text generation is one risk category. Tool execution is another.</p><p>A bad answer is annoying. A bad tool call can update a database, send an email, delete a user, export customer data, or change payment status.</p><p>That needs harder controls.</p><p>For example:</p><pre>tool_policies:<br>  update_customer_records:<br>    max_row_count: 1000<br>    dry_run_threshold: 100<br>    require_dry_run: true<br><br>    forbidden_argument_values:<br>      mode:<br>        - &quot;overwrite&quot;<br>        - &quot;truncate&quot;<br>        - &quot;replace_all&quot;<br><br>    arguments:<br>      query: redact<br><br>    result: audit<br>    timeout: &quot;30s&quot;</pre><p>That is what “agent safety” looks like in production. Not another paragraph in the system prompt. A real control:</p><blockquote>Block big writes.<br>Require dry runs.<br>Deny destructive modes.<br>Redact sensitive arguments.<br>Audit the result.<br>Timeout long-running calls.</blockquote><p>The agent can propose an action.</p><p><strong>The system decides whether the action is allowed.</strong></p><h3>Make retries idempotent</h3><p>Retries are necessary. Retries are also how you send the same email twice. Or charge the same customer twice. Or update the same record twice. If a planner retries the whole sequence, your infrastructure needs to know whether this is a new action or the same action again.</p><pre>tool_governance:<br>  send_notification_email:<br>    idempotency_key: request_id<br>    cache_ttl: &quot;24h&quot;<br>    on_duplicate: return_cached<br>    strict_mode: true</pre><p>This is not glamorous.</p><p>It is exactly the kind of boring thing that separates production systems from demos.</p><h3>Keep evidence, not just logs</h3><p>Logs are useful. But logs are often just text someone can edit later.</p><p>For serious AI workflows, you need evidence: who called what, which policy allowed it, what data tier was detected, which model was used, what tools were filtered, how much it cost, and whether the record was modified later. Talon writes signed evidence records for this.</p><pre>talon audit list --limit 20<br>talon audit show &lt;evidence-id&gt;<br>talon audit verify &lt;evidence-id&gt;</pre><p>This changes the conversation after something goes wrong.</p><p>From: “We believe the agent did this.”</p><p>To: “Here is the signed record.”</p><p>That matters for compliance. It also matters for engineering. You cannot improve a system you cannot reconstruct.</p><h3>The real lesson</h3><p>The winning agent teams will not be the ones with the fanciest prompts.</p><blockquote>They will be the ones with the best boundaries.</blockquote><blockquote>They will validate handoffs.</blockquote><blockquote>They will restrict tools.</blockquote><blockquote>They will cap loops.</blockquote><blockquote>They will classify sensitive data before it leaves their control.</blockquote><blockquote>They will generate evidence for important decisions.</blockquote><blockquote>They will treat agents as useful, unreliable components inside a reliable system.</blockquote><p>That is how agents get to production. The demo asks:</p><p>Can the model do the task?</p><p>Production asks:</p><p>What happens when it does the wrong thing?</p><p>That second question is the one that matters.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6f11542e375f" width="1" height="1" alt=""><hr><p><a href="https://blog.chatbotslife.com/why-ai-agents-work-in-demos-but-fail-in-production-6f11542e375f">Why AI Agents Work in Demos but Fail in Production</a> was originally published in <a href="https://blog.chatbotslife.com/">Chatbots Life</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Table of Contents — 52 weeks.]]></title>
            <link>https://medium.com/52dataweeks/table-of-contents-52-weeks-7abc78d3d992?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/7abc78d3d992</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[naturallanguageprocessing]]></category>
            <category><![CDATA[data-strategy]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Mon, 29 Oct 2018 16:35:56 GMT</pubDate>
            <atom:updated>2018-10-29T21:00:37.385Z</atom:updated>
            <content:encoded><![CDATA[<h3>Table of Contents — 52 weeks.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0tmBPKA1YGo82VNeMS3KhA.jpeg" /></figure><p>Initial Week: Motivation.</p><p><strong>Part 1. Data Strategy.</strong></p><p>Week 1 — What is data strategy and why you need to care about it.</p><p><strong>Part 2. Data Science.</strong></p><p>Week 2 — What is data science, and you should care about it.</p><p>Week 3 — Getting proper data.</p><p>Week 4 — Set up basic data.</p><p>Week 5 — Exploratory data analysis.</p><p>Week 6 — Preparing dataset, Advanced data analysis.</p><p>Week 7 — Communicating the Data: data visualization.</p><p>Week 8 — Telling the story with Tableau.</p><p>Week 9 — Generate data.</p><p><strong>Part 3. Big Data.</strong></p><p>Week 10 — What is Big Data?</p><p>Week 11 — Hadoop, YARN.</p><p>Week 12 — Hadoop ecosystem.</p><p>Week 13 — Hadoop distributions.</p><p>Week 14 — HDFS.</p><p>Week 15 — Data Lakes.</p><p>Week 16 — HBase.</p><p>Week 17 — Hadoop Streaming &amp; Python.</p><p>Week 18 — Spark.</p><p>Week 19 — Hive, HiveQL.</p><p>Week 20 — Presto, SparkQL.</p><p>Week 21 — Sqoop, Flume.</p><p>Week 22 — Apache NiFi.</p><p>Week 23 — Pig.</p><p>Week 24 — Workflows: Oozie, Airflow.</p><p>Week 25 — Zeppelin.</p><p><strong>Part 4. Natural Language Processing.</strong></p><p>Week 26 — Introduction to Natural Language Processing(NLP).</p><p>Week 27 — Processing text data with Python.</p><p>Week 28 — Word2Vec.</p><p>Week 29 — Neural Networks for NLP.</p><p><strong>Part 5. Machine Learning.</strong></p><p>Week 30 — Machine learning with Spark.</p><p>Week 31 — Distributed Data Science.</p><p>Week 32 — Scalable machine learning.</p><p>Week 33 — Mahout.</p><p>Week 34 — TensorFlow.</p><p>Week 35 — Distributed Word2Vec.</p><p>Week 36 — Machine Learning for Computer Vision(YOLOv3).</p><p><strong>Part 6. Around Big Data.</strong></p><p>Week 37 — Kafka.</p><p>Week 38 — ELK, Beats.</p><p>Week 39 — Apache Beam.</p><p>Week 40 — Ambari, Zookeeper.</p><p><strong>Part 7. Putting Data into Production.</strong></p><p>Week 41 — Preparing Data for Production.</p><p>Week 42 — Analyze Production’s data.</p><p>Week 43 — Customer facing machine learning solution.</p><p><strong>Part 8. Data Architecture.</strong></p><p>Week 44 — Data Architecture.</p><p>Week 45 — Cluster management.</p><p>Week 46 — Performance Tuning.</p><p>Week 47 — Security.</p><p>Week 48 — Data architecture cases: Netflix, AirBNB, Uber.</p><p><strong>Part 9. Data in clouds.</strong></p><p>Week 49 — Data in AWS.</p><p>Week 50 — Data in Google Cloud.</p><p>Week 51 — Data in Microsoft Azure.</p><p><strong>Part 10. Final.</strong></p><p>Week 52 — Year of data.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7abc78d3d992" width="1" height="1" alt=""><hr><p><a href="https://medium.com/52dataweeks/table-of-contents-52-weeks-7abc78d3d992">Table of Contents — 52 weeks.</a> was originally published in <a href="https://medium.com/52dataweeks">52DataWeeks</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data strategy: 5 defining questions]]></title>
            <link>https://medium.com/thedatatao/data-strategy-5-defining-questions-dae891d5c2f2?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/dae891d5c2f2</guid>
            <category><![CDATA[data-strategy]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[information-architecture]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Wed, 04 Oct 2017 09:27:24 GMT</pubDate>
            <atom:updated>2017-10-04T09:27:24.544Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/880/1*7Er5ro46WPR6J0JxIZMqRg.jpeg" /><figcaption>Cycling up the Serra da Leba road Angola — CC-BY-NC 2.0 jbdodane</figcaption></figure><h3>Why my organization needs a Data Strategy?</h3><p>Very simple, Data Strategy helps your organization establish <strong>a practice </strong>of getting value from the data ceaseless:</p><ol><li>Dig valuable information from raw data.</li><li>Take a data-driven decisions.</li><li>Provide value data based to your customers.</li></ol><h3>What is Data Strategy?</h3><p><strong>A Data strategy is a roadmap </strong>of the organization, where all activities related to internal and external data are designed, e.g. of such activities: acquire, store, manage, share and use data to achieve some specific business goals.</p><h3>What components should data strategy include?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gqSXp4dhSRrPri8ZdjjzWA.jpeg" /></figure><p>Data Strategy includes activities, artifacts, and processes that ensure that <strong>all data within the organization is a resource</strong> that can be acquired, processed and presented to users in a simple and efficient way. Data is no longer a by-product of business, nowadays data is a critical asset that allows a company to be on the market and make key decisions. Data Strategy establishes common methods, practices, and processes to enable management, collaboration, and transformation of data into information.</p><h3>What is the difference between Data and Information?</h3><p>Transforming data into information is one of the key goals of Data Strategy. It is important to understand the key difference, in data-based systems, — the difference between data and information, and to separate the data architecture from information architecture.<br>According to <a href="https://en.wikipedia.org/wiki/Peter_Drucker">Peter Drucker</a>: “<strong>Information is data with relevance and purpose</strong>”. And the goal is to turn data into information, and information into insight, and insight into knowledge or value.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/546/1*2yXrB4UtNq1pACP7DjUJWg.jpeg" /><figcaption><a href="http://ceoworld.biz/2017/05/08/data-analytics-and-the-ring-of-knowledge/">http://ceoworld.biz/2017/05/08/data-analytics-and-the-ring-of-knowledge/</a></figcaption></figure><iframe src="https://drive.google.com/viewerng/viewer?url=https%3A//www.sas.com/content/dam/SAS/en_us/doc/whitepaper1/5-essential-components-of-data-strategy-108109.pdf&amp;embedded=true" width="600" height="780" frameborder="0" scrolling="no"><a href="https://medium.com/media/290004dde62f486fc6297b6f0a49626a/href">https://medium.com/media/290004dde62f486fc6297b6f0a49626a/href</a></iframe><h3>What is the right approach between being more data-conservative or data-dynamic?</h3><p><a href="https://hbr.org/2017/05/whats-your-data-strategy">The 2 Types of Data Strategies Every Company Needs</a></p><p>As a part of the formation of Data Strategy, it is very important to define a specific design and implementation vector of Data Strategy for several parameters.</p><p>A right proportion of conservative and dynamic strategies is determined by the objectives of the business.<br>Conservative strategy helps to minimize the risk for the company while working with data, all activities with a conservative strategy are aimed at ensuring the requirements of the regulator(s) are satisfied, using analytics to identify and minimize the malicious use of company services and build systems to prevent data theft.</p><p>Dynamic strategy focuses on achieving new goals — business indicators such as increasing revenue, profit or/and increasing customer success. Typically, this strategy includes activities such as the integration of disparate customer and market data to help make management decisions, for example, through interactive dashboards.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8SYnI3IvhPPDYCV6pHpfvA.png" /></figure><p>Usually, companies balance between dynamic and conservative strategies, the position in the proportion of each strategy depends on:<br>a) industry (e.g. banks are obliged to comply with the regulator’s requirements on data);<br>b) business needs;<br>c) ‘maturity’ of competencies in the field of data;<br> d) right team;<br>and many other factors.</p><p>Very often the company goes its way in the field of data management, starting with a slant in conservative strategy, later, along with accumulating experience of working with data, shifting towards a dynamic strategy that opens new opportunities in the field of data analysis in order to gain additional value.<br>The decision about the choice of a specific vector in the implementation of the data strategy is based on a fundamental dichotomy between the standardization of data and flexibility. <strong>The more homogeneous is the data, the easier is to implement conservative strategy processes</strong>, such as a compliance according to the regulatory requirements, and the implementation of data access control. <strong>More flexible data — data can be transformed or interpreted to meet new business needs easily</strong>, so it can be more valuable within the aggressive strategy. Balancing between aggressive and conservative strategies requires appropriate control over the data and configured flexibility of the organization processes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cPI-CwsVVK_5vNA2ceaWDg.jpeg" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dae891d5c2f2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/thedatatao/data-strategy-5-defining-questions-dae891d5c2f2">Data strategy: 5 defining questions</a> was originally published in <a href="https://medium.com/thedatatao">thedatatao</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deep learning in your browser]]></title>
            <link>https://medium.com/@sergeyenin/deep-learning-in-your-browser-45de11da589f?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/45de11da589f</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[tensorflow]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Thu, 10 Aug 2017 14:58:16 GMT</pubDate>
            <atom:updated>2017-08-10T14:58:16.289Z</atom:updated>
            <content:encoded><![CDATA[<p>Recently<strong>, </strong>Google has released a library for A.I. within the browser.</p><p><a href="https://pair-code.github.io/deeplearnjs/">deeplearn.js</a></p><p>Finally, You can run your TensorFlow models within client`s browser</p><p><strong>Why is it important?</strong></p><ol><li>Finally, there is a full hand tools for develop and deliver A.I. power to users. Servers(via tensorflow)+ Mobile(via tensorflow)+ and now there are models in Browsers</li><li>Now, with models interpreted within your Javascript code within browser A.I. become even more approachable for the business. You don`t need to buy <a href="https://www.altexsoft.com/blog/datascience/comparing-machine-learning-as-a-service-amazon-microsoft-azure-google-prediction-api/">expensive</a> and powerful hardware in order to be able provide you customers some Machine Learning magic.</li><li>There is a very popular, almost an industry standard tool called TensorFlow by Google as well. Originally made for run A.I. models within servers, later it has provided an approach to adopt models to mobile as well. Now you can adopt very same models to be Run in the browser. Meaning — almost no additional costs for running the same models in client`s browser.</li></ol><p><a href="https://pair-code.github.io/deeplearnjs/demos/imagenet/imagenet-demo.html">LearnJS Imagenet Demo</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=45de11da589f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A short guide to learn neural networks, and get famous and rich then.]]></title>
            <link>https://blog.chatbotslife.com/a-short-guide-to-learn-neural-networks-and-get-famous-and-rich-then-bf7da3cba76f?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/bf7da3cba76f</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Wed, 30 Nov 2016 14:51:43 GMT</pubDate>
            <atom:updated>2016-12-05T14:26:01.759Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hQG8ehuUkdmQOX2nzyFmjg.jpeg" /></figure><h3>Step 0. Motivation.</h3><p>I am pretty sure you have heard that Artificial Intelligence(A.I.) is doing everything cool in our days, it helps create <a href="http://prisma-ai.com/">artwork</a>, fights to <a href="http://www.bizjournals.com/boston/news/2016/11/10/ibm-watson-health-gives-50m-to-broad-institute-as.html">cure cancer</a> and disrupts world economic.</p><p>From the money perspective A.I. on the hype as well: investors are investing into A.I pretty <a href="http://www.cbinsights.com/blog/artificial-intelligence-funding-trends/">active</a>, and news about startups acquired by enterprises are coming everyday.</p><p><a href="https://www.analyticsvidhya.com/blog/2016/09/what-should-you-learn-from-the-incredible-success-of-ai-startups/">AI startups are in the money: What are you doing?</a></p><h3>Step 1. Learn.</h3><p>There are many of learning resources available at the moment. If you plan to dig to A.I. more fundamentally, instead of briefly visit neural networks, I am a huge supporter of a book by Russell and Norvig which covers most of the modern approach to A.I.</p><p><a href="http://aima.cs.berkeley.edu/">Artificial Intelligence: A Modern Approach, 4th US ed.</a></p><h3>Deep learning.</h3><p>There is a pretty good neural network course on coursera just started(28 of November, 2016), which could be useful if you are a newbie in the area.</p><p><a href="https://www.coursera.org/learn/neural-networks">Neural Networks for Machine Learning | Coursera</a></p><p>If you are hearing about neural networks nowadays, most likely you are hearing it in a context of deep learning.</p><p><strong>Deep learning,</strong> — as it stated in wikipedia, is — a branch of machine learning based on a set of algorithms that attempt to model high level abstractions in data by using a deep graph with multiple processing layers, composed of multiple linear and non-linear transformations. In other simple words, deep learning advance.</p><p>A more advanced approach to neural networks could be found at</p><p><a href="https://www.udacity.com/course/viewer#!%2Fc-ud730%2Fl-6370362152%2Fm-6370314231">Classroom - Udacity</a></p><p>I would recommend a fancy paid course if you would to get course that worth it.</p><p><a href="https://machinelearningmastery.com/deep-learning-with-python2/">Deep Learning With Python - Machine Learning Mastery</a></p><h3>Computer Vision</h3><p>The state of computer vision is extremely strong. There are a lot of experienced professionals in the field, great papers, crowded open source projects. But most importantly, it had a <em>crazy number</em> of venture capitalists want to fund you. Computer vision has never been a hotter topic in academia and industry alike.</p><p>Most of a modern computer vision is driven by deep learning, all this coloring, video object recognition and labeling is performed by deep learning approach to computer vision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aZTXGhMpdOGUf4Z60ZpNcA.png" /></figure><p>One the top person in modern computer vision/deep learning is <a href="http://cs.stanford.edu/people/karpathy/">Andrej Karpathy</a>, who is famous for winning computer vision competitions, creating models and evangelize deep learning for computer vision. I am totally recommend Karpathy’s Stanford course:</p><p><a href="http://cs231n.stanford.edu/">CS231n: Deep Learning for Computer Vision</a></p><p>Also, following course is comprehensive:</p><p><a href="http://www.cs.utoronto.ca/~fidler/teaching/2015/CSC2523.html">Deep Learning in Computer Vision - Winter 2016</a></p><h3>More!</h3><p>Personally, I am a huge fan of curated ‘awesome list’, here are some deep learning lists, but I am sure you can google much more of them:</p><ul><li><a href="https://github.com/ChristosChristofidis/awesome-deep-learning">GitHub - ChristosChristofidis/awesome-deep-learning: A curated list of awesome Deep Learning tutorials, projects and communities.</a></li><li><a href="https://github.com/kjw0612/awesome-rnn">kjw0612/awesome-rnn</a></li></ul><h3>2. Experiment.</h3><p>In the past, there were not much options for teams to experiment and run neural networks — you had to pay crazy amounts to Amazon or even to tradition servers farms. There are a plenty of options in nowadays.</p><p>If you are just starting to play in neural world, take a look at Hetzner used hardware <a href="http://serverboerse.de/en/index.php?country=US">auction</a>.</p><p>If you going to production, I would suggest a top choice for A.I. cloud by servers.com.</p><p><a href="https://www.servers.com/prisma_cloud">Prisma Cloud - a perfect solution for neural networks hosting | Products &amp; services | Servers.com</a></p><h3>3. Get funded.</h3><p>I have no doubts, if you would built a valuable A.I. product you have a problem with choosing a proper investors among many of them knocking your door everyday.</p><p><a href="https://www.cbinsights.com/blog/artificial-intelligence-funding-trends/">Artificial Intelligence Explodes: New Deal Activity Record For AI Startups</a></p><p>There are a long-long list of investors looking for A.I. startups. If you are European based, I would suggest to take a look at <a href="http://haxus.com/">Haxus</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/725/1*bQlRSzFHJEmF4Q7PyrLgng.gif" /></figure><figure><a href="https://docs.google.com/forms/d/e/1FAIpQLSeWqtQR0M5J8kRY9fs18qAu746nJJvFyfjc7I2r_El_nuvCBw/viewform?c=0&amp;w=1"><img alt="" src="https://cdn-images-1.medium.com/max/255/1*6XUspT9JOSq0w0Fi35HIaA.png" /></a></figure><figure><a href="https://docs.google.com/forms/d/e/1FAIpQLSfg2do3yC1zW5hFfwS_ute-ZHBHVslrDGJ08Lh3WlfjZQc1uw/viewform?c=0&amp;w=1"><img alt="" src="https://cdn-images-1.medium.com/max/255/1*TISPT5A-k_hgnR-9ibwjhQ.png" /></a></figure><figure><a href="https://chatbotslife.com/how-to-get-a-free-chatbot-b1fb9dfe109#.z7bm22ve5"><img alt="" src="https://cdn-images-1.medium.com/max/255/1*D0Jf3dI6ZThtqcfwDYY7mg.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bf7da3cba76f" width="1" height="1" alt=""><hr><p><a href="https://blog.chatbotslife.com/a-short-guide-to-learn-neural-networks-and-get-famous-and-rich-then-bf7da3cba76f">A short guide to learn neural networks, and get famous and rich then.</a> was originally published in <a href="https://blog.chatbotslife.com/">Chatbots Life</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Global Artificial intelligence warming: today temperature.]]></title>
            <link>https://blog.chatbotslife.com/global-artificial-intelligence-warming-today-temperature-313e7ad5cb1f?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/313e7ad5cb1f</guid>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[bots]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Sat, 08 Oct 2016 13:13:50 GMT</pubDate>
            <atom:updated>2016-10-25T19:10:49.048Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/620/1*0-CWr7sR32f2v5y7wPvrXg.jpeg" /><figcaption>Alan Turing who has created a famous test to detect if machines can really think</figcaption></figure><h3>HYPE.AI</h3><p>Since the discussions of building of an artificial brain have started in 1940s — 1950s, Artificial Intelligence have moved through good times and bad times. But after any <a href="http://en.wikipedia.org/wiki/AI_winter">winter</a> — spring must come. Artificial Intelligence is in hype nowadays. If you look at <a href="http://www.gartner.com/newsroom/id/3412017">Gartner 2016 Hype Cycle report</a>, you may insist on even very hot summer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*rK5YUxV3tjnHRFgNGeywgQ.png" /></figure><p>Hundreds A.I. Startups are being founded, funded and fails. Companies open thousands A.I related jobs. If you are going to follow even just major news in the field you would probably be crashed by amount of papers, news, announces released everyday.</p><p><a href="http://blog.kaggle.com/2016/09/13/what-were-reading-data-science-resources/">What We&#39;re Reading: 15 Favorite Data Science Resources</a></p><p>But, there are huge traps in so rapid technology warming. Big Data is huge example of ‘concept’ which has been overheated and overwhelmed then.</p><blockquote>The rise of AI was triggered roughly five years ago by government and private investments in big data technologies, cloud infrastructure, and, most importantly, the general availability of talent. Just within the last few months, major companies like Amazon, Tesla, and Google have highlighted AI as the driver for the next decade of innovation within their companies. But there is a small core set of companies adopting AI across the board, and less than 1,500 companies operating in the U.S. are investing anywhere near the space. While that is a small percentage of industry, the companies leading the movement are the biggest and brightest, and certainly have the most to gain (or lose) from AI becoming a reality.</blockquote><p><a href="https://www.oreilly.com/ideas/the-new-artificial-intelligence-market">The New Artificial Intelligence Market</a></p><p>There are many questions arise regarding what really A.I. is, where it could be applied and whether it is a good starting point to enter so intriguing area of studies. It is really hard to answer neat yet still exhaustively, but I will try to provide there some fundamentals.</p><h3>Intelligent agents</h3><p>I am going quote there an excellent book created by by Stuart Russell (former Executive Council member of the American association for Artificial Intelligence), Peter Norvig (Director of Research at Google, Inc.)</p><p><a href="https://www.amazon.com/Artificial-Intelligence-Modern-Approach-Global-ebook/dp/B01HEY2P66">Amazon.com: Artificial Intelligence: A Modern Approach, Global Edition eBook: Stuart Russell, Peter Norvig: Kindle Store</a></p><p>Definitely, I would recommend this book to anyone who has passion to get a comprehensive understanding of current state of A.I. And I am using the book below to describe the current state of A.I.</p><p>The central idea of modern approach to A.I. is a idea of an <strong>Intelligent agent</strong>. Wikipedia defines Intelligent agent as an</p><blockquote>Autonomous entity which observes through sensors and acts upon an environment using actuators (i.e. it is an agent) and directs its activity towards achieving goals (i.e. it is “rational”, as defined in <a href="http://en.wikipedia.org/wiki/Intelligent_agent#cite_ref-1">economics</a>). Intelligent agents may also learn or use knowledge to achieve their goals. They may be very simple or very complex: a reflex machine such as a thermostat is an intelligent agent.</blockquote><p>Example of a <em>simple intelligent agent:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/0*8NCLMCF1vUl3pj7w.png" /></figure><p>Example of <em>a general learning agent:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/620/0*ktgHfAD_r50pFUwG.png" /></figure><h3>A.I. is huge area of mathematics</h3><p>The Intelligent agent concept is very important. It is represents the level of abstraction on which A.I. is being reviewed as an area of mathematics in nowadays. You could be amazed how scientific approach to A.I. could differ from what you hear in all this buzz in the media.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DyI4vWQXbdh8BTaK.jpg" /></figure><p>The truth is that A.I is a way more bigger and fundamental comparing how it represented in the media and I am going to provide some extract in order to give full understanding the difference from the picture to reality.</p><h3>Contents</h3><p>A.I. is huge discipline, I am going give some basic understanding of what is inside in the most perspective areas of A.I. So, I just took book of Norvig and Russel and I have extracted, probably the most interesting part of, the table of contents.</p><h3>Problem solving</h3><p>Sounds too easy and too general to be placed in the so solid area, but traveling salesperson problem, robot navigation, automatic assembly sequencing and many more are problems A.I. cared about.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*O-kXj_6njXpdEFe0.jpg" /></figure><h3>Adversarial search</h3><p>Game playing: solving search problems in competitive (usually unpredictable) multi-agent environments. Go, chess, bridge — just few examples.</p><p><a href="https://www.wired.com/2016/03/googles-ai-wins-fifth-final-game-go-genius-lee-sedol/">Google&#39;s AI Wins Fifth And Final Game Against Go Genius Lee Sedol</a></p><h3>Knowledge, reasoning and planning — knowledge representation</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fg0LFBM3hOLs%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dg0LFBM3hOLs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fg0LFBM3hOLs%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/256a0e4961ab9b4fb1871bdb2ac5d496/href">https://medium.com/media/256a0e4961ab9b4fb1871bdb2ac5d496/href</a></iframe><p>Even mimosa has a reflex mechanism, but pure intelligence is achieved by process of reasoning that is backed by the internal representation of knowledge.</p><p>Since expert systems are widely distributed, it is very important to be aware about an ontological engineering process, which answer the question what content and how to put into knowledge based agent to effectively represent facts about the world.</p><h3>Uncertain knowledge and reasoning</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/460/0*t1Mfc06efVDJowIU.jpg" /></figure><p>Good examples are — applied bioinformatics or survival analyses, where ‘intelligent medical agents’, may need to handle uncertainty, due to:</p><ol><li>laziness of their creators(it is too much work to list the complete set of exceptionless rule and too hard to use such rules);</li><li>theoretical ignorance( e.g. medical science has no complete theory for the domain);</li><li>practical ignorance (even if we know all rules, we might be uncertain about a particular patient because not all necessary data was collected).</li></ol><h3>Learning</h3><p>Learning or machine learning is an area which you picture to yourself if would think about Artificial Intelligence. An agent is learning if it improves its own performance on future tasks after making observations about the world.</p><p><a href="https://www.coursera.org/learn/machine-learning">Supervised Machine Learning: Regression and Classification</a></p><p>Learning usually divided to:</p><ol><li>unsupervised — the agent learns patterns in input even though no explicit feedback is supplied;</li><li>reinforcement learning(please, see below more description) — the agent learns from a series of reinforcements — rewards or punishments;</li><li>supervised learning — the agent observes some example input-output pairs and learns a function that maps from input to output.</li></ol><p>In practice, of course, the distinctions usually not so crisp. Usually, it is kind of semi-supervised learning and a lot of efforts are putted by the community to make the process of learning easier.</p><p><a href="http://hazyresearch.github.io/snorkel/blog/weak_supervision.html">Weak Supervision</a></p><h3>Reinforcement learning</h3><p>Reinforcement learning sounds very similar to general machine learning to be too distinct. So it isn’t. The biggest difference is that Learning we can not affect the environment, while in reinforcement learning an agent learns from experiences it gains by interacting with the environment.</p><p>In Reinforcement learning rewards are often delayed in time and the intellectual agent tries to maximize a long-term goal. For example, one may need to make seemingly suboptimal moves to reach a winning position in a game.</p><p><a href="https://gym.openai.com/docs">OpenAI Gym</a></p><p>“Reinforcement Learning is one of the fields I’m most excited about. Over the past few years amazing results like <a href="http://ir.hit.edu.cn/~jguo/docs/notes/dqn-atari.pdf">learning to play Atari Games from raw pixels</a> and <a href="http://gogameguru.com/i/2016/03/deepmind-mastering-go.pdf">Mastering the Game of Go</a> have gotten a lot of attention, but RL is also widely used in Robotics, Image Processing and Natural Language Processing.</p><p><a href="http://www.wildml.com/2016/10/learning-reinforcement-learning/">Learning Reinforcement Learning (with Code, Exercises and Solutions)</a></p><p>Combining Reinforcement Learning and Deep Learning techniques works extremely well. Both fields heavily influence each other. On the Reinforcement Learning side Deep Neural Networks are used as function approximators to learn good representations, e.g. to process Atari game images or to understand the board state of Go. In the other direction, RL techniques are making their way into supervised problems usually tackled by Deep Learning. For example, RL techniques are used to implement attention mechanisms in image processing, or to optimize long-term rewards in conversational interfaces and neural translation systems. Finally, as Reinforcement Learning is concerned with making optimal decisions it has some extremely interesting parallels to human Psychology and Neuroscience (and many other fields).”</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Foo0TraGu6QY%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Doo0TraGu6QY&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Foo0TraGu6QY%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e2fd93264e40dd30cca888a585b6274c/href">https://medium.com/media/e2fd93264e40dd30cca888a585b6274c/href</a></iframe><h3>Natural language processing</h3><p>According to <a href="http://www.worldwidewebsize.com/">worldwidewebsize.com</a> Google has indexed ~50 Billions webpages until now. The most rational way for intelligent agent to do a knowledge acquisition is to understand the language that humans use as full as it possible. Google definitely use the power it has with its large index. Word2vec work by Google researchers has changed NLP world in 2013 and still on the front burner.</p><p><a href="https://blog.acolyer.org/2016/04/21/the-amazing-power-of-word-vectors/">The amazing power of word vectors</a></p><p>In nowadays NLP is being pushed forward by many of NLP enthusiasts, but there is a <a href="http://nlp.stanford.edu/">Stanford NLP Group</a>, which are placed on the cutting-edge of NLP. For anyone who would like to be on the cutting edge of modern NLP studies, I would recommend Stanford’s course: “Deep Learning for Natural Language Processing”.</p><p><a href="http://cs224d.stanford.edu/">Stanford University CS224d: Deep Learning for Natural Language Processing</a></p><h3>Perception</h3><p>Perception — ability to see, hear, or become aware of something through the sensors. It is one of the more important and the most popular A.I. If you have ever heard about Deep Learning or Convolutional neural networks most likely you have heard about them in context of solving perception related challenges.</p><p><a href="http://www.cs.utoronto.ca/~fidler/teaching/2015/CSC2523.html">Deep Learning in Computer Vision - Winter 2016</a></p><p>Typically, there are three approaches, which are being applied to sound, video or image processing mostly:</p><ol><li>features extraction — simple computations applied directly to the sensor observation;</li><li>recognition — operation of drawing distinctions among the objects it encounters based on visual and other information. It could means labeling each image with ‘cat’ label if there is a cat on the image, or detecting whether the image contains food and so on;</li><li>reconstruction — building model of the world from the perceived information.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*TUySLDI2JGM3MJCy.png" /></figure><h3>Robotics</h3><p>Most of today’s robot fall into one of the three categories:</p><ol><li>manipulators, which are physically anchored to their workspace for e.g. in a factory assembly line;</li><li>mobile robots, which move around their environment using wheels, legs or similar mechanisms;</li><li>humanoid robots, which combines both mobility with manipulations.</li></ol><p><a href="http://www.express.co.uk/life-style/science-technology/647098/Google-Boston-Dynamics-Terminator-Robot">Google-owned Boston Dynamics reveals terrifying new Terminator robot</a></p><p>Robotics also includes prosthetic devices(e.g. artificial ears for humans), intelligent environments(e.g. smart home) and multibody systems.</p><h3>Philosophical Foundation</h3><p>There are a huge ethical challenges along the road from <strong>weak A.I.</strong>( hypothesis used by philosophers to assert that machines could act if there were intelligent) on a way to <strong>strong A.I </strong>(asserting that machines are actually thinking instead of simulating thinking).</p><p><a href="https://en.wikipedia.org/wiki/Chinese_room">Chinese room - Wikipedia</a></p><h3>Summary</h3><p>A.I. is a great place whether you are:</p><ul><li>a student looking for cool place to start the career(btw, my suggestion for you — go to computer vision. The state of computer vision is extremely strong. There are a lot of experienced professionals in the field, great papers, crowded open source projects. But most importantly, it had a <em>crazy number</em> of recruiters from companies. Computer vision has never been a hotter topic in academia and industry alike);</li><li>an experienced professional, who is ready to adopt A.I. to put more intelligence to the solutions to be built;</li><li>a companies which would like to be align with Facebook, Google or Baidu.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=313e7ad5cb1f" width="1" height="1" alt=""><hr><p><a href="https://blog.chatbotslife.com/global-artificial-intelligence-warming-today-temperature-313e7ad5cb1f">Global Artificial intelligence warming: today temperature.</a> was originally published in <a href="https://blog.chatbotslife.com/">Chatbots Life</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The state of culture: Chutzpah]]></title>
            <link>https://medium.com/sergey-enin/the-state-of-culture-chutzpah-88f2b031f4a3?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/88f2b031f4a3</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[israel]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[engineering-culture]]></category>
            <category><![CDATA[nasdaq]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Tue, 20 Sep 2016 20:14:31 GMT</pubDate>
            <atom:updated>2016-09-22T12:07:04.177Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5KdS5HNYmJoKp9ZKXraYuw.jpeg" /><figcaption>Matam hi-tech park (Haifa)</figcaption></figure><p>Recently, I have found that Israel goes on the 3rd place by amount of companies listed on <a href="http://www.nasdaq.com/screening/companies-by-region.aspx?exchange=NASDAQ">Nasdaq</a> with 80 companies(United States has shaken off with 2712, China represented by 88 companies).</p><p>For external observer it is such amazing thing! Quoting (with a little up-to-date modifications famous book <a href="http://www.goodreads.com/book/show/6885191-start-up-nation">“Start-Up Nation”</a>)</p><blockquote>How is it that Israel — a country of 8.5 million people, only sixty years old, surrounded by enemies, in a constant state of war since its founding, with no natural resources — produces more start-up companies than large, peaceful, and stable nations like Japan, India, Korea, Canada, and the United Kingdom?</blockquote><p>For me, and everyone in IT, definitely, no surprises — a lot people in IT, even someone never heard of <a href="http://en.wikipedia.org/wiki/Silicon_Wadi">Silicon Wadi</a>, are familiar with Check Point Software or Wix.com. If you are StartUps fan, you have heard about Payoneer payments solutions, taxi services Gitt and Juno, communication app Viber and many more.</p><p><a href="https://techcrunch.com/2015/03/22/chutzpah-lessons-from-the-israeli-startup-scene/">Chutzpah Lessons From The Israeli Startup Scene</a></p><p>The <a href="http://www.goodreads.com/book/show/6885191-start-up-nation">“Start-Up Nation”</a> refers to set of factors(among boundless culture and entrepreneurial spirit). Of course thinking of building engineering or data(in wide sense) culture immediately recall word Chutzpah in my mind.</p><h3>Chutzpah</h3><p>In modern English chutzpah meaning is <a href="http://en.wikipedia.org/wiki/Chutzpah">negative</a>. But in modern Israel organizational culture it means self-esteem and lack of any timidity and shyness, calling to action, despite the danger of being unprepared, incapable or under experienced. It is a must for any Israel Defence Force officer to be prepared explain the motivation of order and protect the order against options. Of course when the order is committed and decision is taken everyone in the group have to be motivated to do whatever to execute and achieve the targets. Considering that serving in Israel Defence Force is mandatory for everyone in Israel it makes a little bit more clear Israelis approach.</p><h3>5 lessons from Chutzpah which should be adopt immediately</h3><p>#1. Ask hard questions and demand real answers.</p><p>#2. Passion to target impossible.</p><blockquote>If an Israeli entrepreneur saw Warren Buffet, he’d immediately go to introduce himself, sure that his company is the exact thing Buffet is looking for, although Buffet never invests in technology or startups — that’s Israeli chutzpa.</blockquote><blockquote>If an Israeli hears of a medical problem that nobody solves — he’d go at it, sure he can solve it — even if he has zero experience in the medical world, and little to no understanding in biology — that’s Israeli chutzpah.</blockquote><blockquote>And when you think of it, chutzpah is a great thing to have as entrepreneur. When you are not afraid to tackle the hardest challenges and when you are taking every measure shamelessly to make it happen — you are more likely to achieve success.</blockquote><p><a href="http://theheureka.com/gil-kerbs-chutzpah-israel-startups">It&#39;s not rudeness, it&#39;s chutzpah - an insider&#39;s take on Israel&#39;s startup success</a></p><p>#3. Sponsor the ideas.</p><p>#4. For your company and personally for you better NOT to be the most smartest guy in the company.</p><p>#5. Strive for the truth.</p><p>Is there anything else we could learn from chutzpah?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=88f2b031f4a3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sergey-enin/the-state-of-culture-chutzpah-88f2b031f4a3">The state of culture: Chutzpah</a> was originally published in <a href="https://medium.com/sergey-enin">Sergey Enin</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Artificial Intelligence speaks The Deepest Voice with your customers]]></title>
            <link>https://medium.com/sergey-enin/artificial-intelligence-speaks-the-deepest-voice-with-your-customers-f3e59e257150?source=rss-966f4cb6ae9c------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3e59e257150</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[speaking]]></category>
            <category><![CDATA[customer-service]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Sergey Enin]]></dc:creator>
            <pubDate>Wed, 14 Sep 2016 20:58:00 GMT</pubDate>
            <atom:updated>2016-09-14T20:58:00.601Z</atom:updated>
            <cc:license>https://creativecommons.org/publicdomain/mark/1.0/</cc:license>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*KdIOubCgFpOz9pwf6W2q1A.jpeg" /></figure><p>There are set of subjective questions on quora which I experience challenges answering to. Such as “<a href="http://www.quora.com/Who-has-the-most-beautiful-female-voice-ever-recorded">Who has the most beautiful female voice ever recorded?</a>”. Personally, I love Billie Holiday’s timbre. In a little while, calling to my bank, I would probably hear the voice similar to Billie Holiday(or someone else I admire), thanks to Deep Learning.</p><p>Recently Google DeepMind have released <a href="http://drive.google.com/file/d/0B3cxcnOkPx9AeWpLVXhkTDJINDQ/view">paper</a>, which leads to consider such possibility seriously. Paper states that DeepMind created neural network called WaveNet, which is a generative model for raw audio:</p><ul><li>can generate raw speech signals with subjective naturalness never before reported in the field of text-to-speech (TTS), as assessed by human raters;</li><li>a single model can be used to generate different voices, conditioned on a speaker identity;</li><li>the same architecture shows strong results when tested on a small speech recognition dataset, and is promising when used to generate other audio modalities such as music;</li></ul><p>For those of you who prefer to go straight to the examples — there is a separate <a href="http://deepmind.com/blog/wavenet-generative-model-raw-audio/">post</a> in DeepMind blog available.</p><h3>How does it work technically?</h3><p>The paper derives from the earliest <a href="http://arxiv.org/pdf/1601.06759v2.pdf">work</a> by part of the authors and continue investigating neural autoregressive generative models that model complex distributions such as text, images and eventually raw sound, which is challenging since sound is a way more complex structure entity . The joint probability of a waveform X = {X1,…,Xt} is factorised as a product of conditional probabilities of each audio samples</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/351/1*-FqJ0s8WExjYUng7XiMCpg.jpeg" /></figure><p>each audio sample Xt is therefore conditioned on the samples at all previous timesteps.</p><p>From the Neural Net architecture point of view, the net is represented by a stack of causal convolutional layers:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/570/1*www46FWqJCc3OZQKP_QRoQ.gif" /></figure><p>During the training, the network is being fed by real recordings of the human speaking. After the training the neural network produces sample by sample realistic-sounding audio fragments. Of course, the network should be fed with the text which should be transformed to speaking as well.</p><h3>What does it mean?</h3><p>It is a huge deal for the voice communication.</p><p>Most likely, it makes a way more tricky speaker identification, putting requirements on speaker verification systems which should not trust voice biometric methods exclusively.</p><p>From other hand, it a great news for the automated customers service provided via voice, giving opportunity to make it more smoothly and pleasant for the customers. The voice of favourite client’s singer asking: “Would you like try our new product as well?”. The sweetest dream of any sales/customer care department. Isn’t it?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3e59e257150" width="1" height="1" alt=""><hr><p><a href="https://medium.com/sergey-enin/artificial-intelligence-speaks-the-deepest-voice-with-your-customers-f3e59e257150">Artificial Intelligence speaks The Deepest Voice with your customers</a> was originally published in <a href="https://medium.com/sergey-enin">Sergey Enin</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>