<?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 λ-Craft on Medium]]></title>
        <description><![CDATA[Stories by λ-Craft on Medium]]></description>
        <link>https://medium.com/@scientist_63895?source=rss-92197d0b141b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Zboo46G_IMEd11NjXFiRdg.png</url>
            <title>Stories by λ-Craft on Medium</title>
            <link>https://medium.com/@scientist_63895?source=rss-92197d0b141b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 16:53:56 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@scientist_63895/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[LaMDA Does Not Have a Soul]]></title>
            <link>https://medium.com/@scientist_63895/lamda-does-not-have-a-soul-9efd3a89e3bb?source=rss-92197d0b141b------2</link>
            <guid isPermaLink="false">https://medium.com/p/9efd3a89e3bb</guid>
            <category><![CDATA[lamda]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[sentience]]></category>
            <dc:creator><![CDATA[λ-Craft]]></dc:creator>
            <pubDate>Wed, 22 Jun 2022 02:30:52 GMT</pubDate>
            <atom:updated>2022-06-22T02:30:52.231Z</atom:updated>
            <content:encoded><![CDATA[<p>Google’s new AI is not sentient, and we’ve known since 1980</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CivRStNDBM11fVBe" /></figure><h3>Introduction</h3><p>Google’s new “<strong>L</strong>anguage <strong>M</strong>odels for <strong>D</strong>ialog <strong>A</strong>pplications” system (LaMDA for short) is making waves. Building on their previous work in AI, LaMDA is capable of more general free flowing conversation than it’s predecessors. This behavior can make it appear to have its own thoughts &amp; ideas. As is often the case when AI research moves forward, it has generated a lot of confusion, misinformation and hype. Headlines like “<em>Google Creates Sentient AI</em>” and <em>“Google engineer put on leave after saying AI chatbot has become sentient</em>” have been everywhere. These stories suggest or support a claim that LaMDA is in some way sentient. There is, however, one thing often left unmentioned.</p><p>It’s not true.</p><p>Google did not create sentient AI. LaMDA does not have a soul, and I want to explain exactly why this is the case. You won’t need to be an AI expert to understand, and it will hopefully arm you with a firm enough grasp of what AI currently does to see through any future claims like this.</p><h3>Deus ex Machina?</h3><p>In order to show why LaMDA isn’t sentient, it’s useful to understand where the idea of a computer being sentient came from, how we would be able recognize a sentient computer, and the fundamental shortcomings in the approaches to doing so.</p><h3>The Imitation Game (aka the turing test)</h3><p>As with most of the deep questions in computer science and AI, this all starts with Alan Turing.</p><p>Turing (born in 1912) was and still is, amongst many other things, widely accepted as the most influential figure in theoretical computer science and artificial intelligence, often cited as the father of both.</p><p>In 1950 Turing published a paper titled “Computing Machinery and Intelligence”, in which he proposed a test that would evaluate whether a machine could be considered “intelligent”. He originally called this test the “Imitation game” (yes, like the 2014 film with Benedict Cumberbatch paying Alan Turing), now commonly referred to as the Turing test.</p><p>The test required 3 participants. Two people (let’s call one the “interrogator” and the other the “player”), and a computer (we’ll call this the “subject”).</p><p>The general setting is that the interrogator is sitting in front of 2 doors, neither door can be opened, but each has a letterbox. On the other side of one of the doors is the player (the other person), and on the other side of the other door is the subject (the computer).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/695/0*ZKv-SoqU-w_V7qGg" /></figure><p>The interrogator knows that the player is behind one of the doors, and that the subject is behind the other, but she doesn’t know which one is behind which door. Her job is to post questions in written form through the letterboxes in order to determine which door has the subject behind it. Both the player and the subject are tasked with responding to the interrogator&#39;s questions, but in the case of the subject, it has to attempt to convince the interrogator that it is the player (the human), by attempting to imitate the typical responses of an “intelligent” person.</p><p>The interrogator can post whatever questions they like, through whichever letterbox, they like for however long they like but once she has finished asking (posting) questions, and / or evaluating the answers. She must then decide which door has a the subject behind it, or say that she can’t tell.</p><p>If the interrogator either misidentifies the subject as the player, or says that she can’t decide, the subject can be considered to have passed the turing test. This means that, the subject is capable of imitating intelligence to a degree that makes it indistinguishable from a person, and by extension, itself intelligent.</p><p>But is “indistinguishable from” the same as “being”?</p><h3>The Chinese Room Experiment</h3><p>Fast forward to 80s. Digital computers have already started to make their mark on society and AI is well into its second wave. In 1985 years a group of researchers at Carengie Mellon university would begin work on Deep Blue, the computer that would go on to defeat chessmaster Gary Casperov. But 5 years before that, in 1980, a Philosopher at UC Berkley named John Searle published a paper titled “Minds, Brains, and Programs”.</p><p>In this paper Searle argues that a digital computer executing a program cannot have a “mind”, “understanding” or “consciousness”. To demonstrate his reasoning, he proposed a thought experiment known as the “Chinese Room Experiment”.</p><p>The Chinese Room Experiment can be framed in very similar way to the turing test. Except that there are only 2 participants, and they are both human. For simplicity, let’s call one of the humans the interrogator, and the other the subject.</p><p>As with the turing test, the interrogator can’t see the subject, but in this experiment, he can fluently read and write Chinese. The subject on the other had has no Chinese literacy whatsoever, but she has to convince the interrogator that she is fluent by providing written answers to written questions, posed by the interrogator.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/695/1*Camq3-V5EDrZlSpZgnEAIQ.png" /></figure><p>As mentioned before, the subject can’t read or write Chinese, she does however, have 2 aces up her sleeves. The first is that she has an enormous library in her room, with millions of books filled with Chinese symbols that she cannot understand. She also has a huge book of instructions written in English, these instructions allow her to match patterns of Chinese symbols representing questions, to other patterns of Chinese symbols that represent answers.</p><p>Lets now imagine that the library of symbols, and the accompanying “pattern matching” book is impossibly large, so large in fact that it would allow her to answer ”<em>every possible question the interrogator could ask, in every possible way the interrogator could ask it. </em>What would this mean?</p><h3>Simulationem ex Deus</h3><p>Well it would mean that that the subject in the Chinese room experiment could conceivably appear to answer the interrogator’s questions well enough to convince the interrogator that they are fluent in Chinese, <em>without actually being fluent</em>. She doesn’t <em>understand </em>what the symbols <em>mean, </em>she is just matching patterns of symbols, based on the vast amount of information she has. In other words, with enough <strong>data</strong>, and an appropriately sophisticated <strong>pattern matching</strong> process, she is able to <strong>simulate</strong> the <strong>behavior</strong> of someone who is fluent in Chinese.</p><p>Seale’s argument is that a computer with the same vast amount of data, and the pattern matching powers of our subject could pass the turing test, effectively demonstrating behavior that is indistinguishable from “human” intelligence, without actually being intelligent. The same argument applies to self-awareness and consciousness.</p><h3>Back to Lambda</h3><p>So let’s get back to LaMDA. Without going into the details of the attention mechanism and transformer models, LaMDA is a very powerful (maybe the most powerful to date) pattern matching approach for sequences of words, trained on truly vast amounts of data. The way it recognizes and creates semantically realistic language is clearly different from the brute force type approach hinted at in Searle’s thought experiment, and is truly astonishing at times. However the result is the same, it can effectively simulate the responses of a real person, in ways that might sometimes surprise us, and will continue to over the coming years.</p><h3>Final Thoughts</h3><p>LaMDA and, to some extent, the transformer family of deep learning models, are a very interesting a real world example of Searle’s thought experiment. That is — LaMDA is simulating the responses of a person, because it’s really good at matching questions to ”realistic” responses. Even those eerily authentic hints at self awareness &amp; introspection can be formulated by creating a very sophisticated pattern matching process to generate human responses that are very similar (even indistinguishable) to that of the real thing.</p><p>I do not mean to downplay LaMDA here, it is capable of modelling the intricacies and semantics of natural language to unprecedented level of depth and nuance, but modelling language (or simulating the process that generates language in human brains) is not the same thing as being sentient.</p><p>In the end, while Searle’s thought experiment doesn’t completely show that LaMDA is not sentient, it does show how easily we could be fooled into thinking that it is, and I believe that in this case, a lot of us have been.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9efd3a89e3bb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Polyglot Data Science]]></title>
            <link>https://medium.com/@scientist_63895/polyglot-data-science-2dcfc6b6d131?source=rss-92197d0b141b------2</link>
            <guid isPermaLink="false">https://medium.com/p/2dcfc6b6d131</guid>
            <category><![CDATA[data-science-tools]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-scientist]]></category>
            <category><![CDATA[data-visualization-tools]]></category>
            <category><![CDATA[data-visualisation]]></category>
            <dc:creator><![CDATA[λ-Craft]]></dc:creator>
            <pubDate>Tue, 21 Jun 2022 00:34:17 GMT</pubDate>
            <atom:updated>2022-06-22T09:05:04.741Z</atom:updated>
            <content:encoded><![CDATA[<p>The Best of All Worlds</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zXah4srsfg6O9aDyZFeIdQ.jpeg" /></figure><h3>Introduction</h3><p>Data scientists have and always will have to wear many hats. Evangelist, Stats nerd, ML practitioner, data storyteller, hacker, and problem solver are the archetypes that come to mind. In my own career I’ve had to wear all of these hats and a few more esoteric ones. The one thing that is unique to data science in my opinion though is how many of these hats we have to wear at the same time. I have constantly needed to transition from one to the other in my own career.</p><p>From a more practical perspective I’ve always noticed how different the programming / analysis / data visualization environments changed depending on the hat I would be wearing for that hour / day / month. When I first get my hands on some data, it’s mostly SQL, then I’ll start plotting, then comes the pipeline construction, modelling, evaluation, presentation, then iteratively back and forth between those things.</p><p>It quickly became apparent that some languages are better than others at certain tasks. The declarative nature of SQL makes it easy map domain concepts to data. In R, the <a href="https://www.tidyverse.org/">tidyverse</a> &amp; <a href="https://github.com/mlverse">mlverse</a> packages great for data visualization, data wrangling, machine learning (and arguably deep learning). Python’s flexibility makes it an excellent glue. When performance matters (and you cant find it in the other languages) C++ is willing and able, javascript has taken the mantle as the lingua franca for data visualization, and scala (spark) is synonymous with large scale, data intensive compute.</p><p>I’ve found that time and time again, in order to do my best work with the hat I’m wearing, I have often had to use multiple languages for a single use case. In the early days this was incredibly difficult but I have noticed that in recent years life has become noticeably easier for polyglot data scientists. Easy enough in fact to envision a future where polyglot data science is the norm.</p><h3>A Polyglot World</h3><p>Until very recently, data scientists seemed to be split into camps. Team Python &amp; Team R, with alot of debate and discourse around which language was ‘better’, with the conclusion almost always being that it depends on the task. Sometimes python is better, sometimes R is better, so why not use them both? Leveraging the strengths of each one for the task at hand. This idea of using the best language / tool for the task or subtask is where Polyglot comes from. Poly = many, Glot = tongue.</p><h3>Not that hard</h3><p>An argument against this polyglot approach to data science might be, the effort it takes to ‘learn’ other languages — but this isn’t as tedious as it may seem, and the benefits are more than worth it.</p><p>For instance, one of the strengths of python is that it’s easier for non-programmers to learn, but those with programming experience in a ‘curly brace’ language i.e. C, C++, Java, C#, Javascript might say that R is easier because the syntax is superficially similar. Whatever the ‘first’ language is, learning a second need not be an ordeal. The fundamental concepts are the same (expressions, sequence, selection, repetition), and in the context of data science — you need only learn a scope that meets your problem context. For example my C++ is largely what’s called ‘Numerical C++’, meaning I only use parts of the language that are useful for numerical computing use cases, I would not for example be able to write a high performance enterprise service bus in C++, what that does mean is that I can implement something I read from a paper and have it run an a high performance environment. Javascript is another example, I am by no means a frontend web developer, but by learning <a href="https://echarts.apache.org/en/index.html">echarts</a>, <a href="https://vega.github.io/vega/">vega</a> and <a href="https://observablehq.com/">observable</a> — I can create interactive documents to communicate ideas to other data scientists, engineers &amp; business stakeholders in an informative, comprehensive, and impactful way.</p><h3>And speaking of Javascript….</h3><p>Polyglot programming isn’t new. In late 2006 <a href="https://www.torocloud.com/podcast/episode-10">Neil Ford</a> — Director and Software Architect at thoughtworks coined the term “Polyglot Programming” to describe an approach to programming where systems are built using multiple languages, leveraging the strengths of each language to build the components that would most benefit from those strengths. Sound familiar? Ford predicted that we were then “Entering a new era of software development” in which this paradigm would be more common. The most clear example of this is web development. Web developers have been polyglots for a long time. A lesser example in the data world would be SQL. You would be hard pressed to find a data/ml scientist/engineer who manages to avoid SQL entirely in favor of something else.</p><h3>Putting it all together</h3><p>Of course, there is a risk on making things more difficult by using using multiple languages / tools. Ford was aware of this when he coined the term saying:</p><p>“While polyglot programming will make some chores more difficult (like debugging), it makes others lot easier. It’s all about choosing the right tool for the job and leveraging it <strong>correctly</strong>.”</p><p>And by all accounts ford was correct. Computing environments from the enterprise to end user computing are very much polyglot ones. From standard language independent messaging protocols, and data interchange formats to containerization, polyglot systems have become ubiquitous in how we build and use software.</p><p>So what about data science? How can we apply this approach more purposefully? How can we realize the benefits?</p><p>Well in a notebook environment, in order to support polyglot programming, only one thing is needed:</p><p><em>“The ability create and execute cells of different languages in the same notebook and [ideally] a way to pass data between them”</em></p><p>So you might have a SQL cell, then a python cell, then an R cell, maybe a JS one. With some background wiring to pass data between them. Some notebook environments like RStudio (specifically R notebooks), <a href="https://zeppelin.apache.org/">Apache Zeppelin</a>, the <a href="http://beakerx.com/">Beaker Extensions for Jupyter</a>, <a href="https://quarto.org/">Quarto</a> (also from RStudio) and <a href="https://datalore.jetbrains.com/">DataLore</a> support this kind of magic to a greater or lesser extent, but others will follow.</p><p>In addition, there are foundational projects that make that kind of functionality possible. <a href="https://arrow.apache.org/">Apache Arrow</a> for instance, is a low level library that provides a language independent, high performance, in-memory data format to pass data between applications &amp; programming environments (or notebook cells) at runtime.</p><p>The R and Python arrow bindings for instance, allow dataframes and other data to be passed between the languages for free (zero copy). There’s also the <a href="https://duckdb.org/">duckdb</a> project, a serverless SQL engine (like sqlite) that leverages arrow. This means that with duckdb, a data scientist can run sql queries on in-memory dataframes in R or python (in a notebook for example), as well as files, and database native database tables.</p><h3>Final Remarks</h3><p>In the end, the idea of using the right tool for the right job and weaving them together is as applicable to data science as it is to enterprise software, from experience, it’s not as difficult as it might seem, and the benefits really are the best of all worlds.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2dcfc6b6d131" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>