<?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 Glenn Hopper on Medium]]></title>
        <description><![CDATA[Stories by Glenn Hopper on Medium]]></description>
        <link>https://medium.com/@glenn_53777?source=rss-693b82c7a3e1------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*8OAWcpElACOcQ083sJCRGw.png</url>
            <title>Stories by Glenn Hopper on Medium</title>
            <link>https://medium.com/@glenn_53777?source=rss-693b82c7a3e1------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 12:07:29 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@glenn_53777/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[Your AI Isn’t the Problem. The Wiring Around It Is.]]></title>
            <link>https://medium.com/@glenn_53777/your-ai-isnt-the-problem-the-wiring-around-it-is-c1a092181347?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/c1a092181347</guid>
            <category><![CDATA[mcp-protocol]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[agent-harness]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Mon, 25 May 2026 23:03:26 GMT</pubDate>
            <atom:updated>2026-05-25T23:03:26.378Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N4IOTKFS7oEYefWQ6gGn6Q.png" /><figcaption>LLMs are already becoming commoditized. The near term future is the wiring around them.</figcaption></figure><p>For years now, we’ve been lied to about AI agents.</p><p>What most people called an “agent” was just a chatbot with a slightly more assertive prompt. It could write an email, summarize a document, or generate some code; but it didn’t actually <em>do</em> anything on its own. It was a clever performer, not an autonomous worker. The label was marketing, not engineering.</p><p>Now the real thing is here.</p><p>A true AI agent is software that can take action in the world with minimal supervision. It reasons about a goal, picks its next step, uses tools, talks to APIs, writes files, queries databases, and keeps going until the job is done … or until it gets stuck. The language model is the brain doing the thinking. Everything else (the memory, tools, guardrails, and the recovery mechanisms) is the body and nervous system.</p><p>And that’s where almost everyone has gotten it wrong so far.</p><p>Three years into the onslaught of generative AI, the brain is the easy part. The next differentiator is (will be) the wiring we build around it.</p><p>In February, a team at <a href="https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering">LangChain</a> took a coding agent that ranked around 30th on a widely used benchmark and, without changing the AI model at all, pushed it into the top five.</p><p>Same model.</p><p>Same weights.</p><p>They didn’t change anything about the AI itself. They rewired the system prompts, swapped in better tools, and added middleware that caught failure patterns it kept repeating. The score jumped from 52.8% to 66.5%, a 13.7-point gain — from infrastructure changes alone.</p><p>That result has been rattling around the engineering world ever since, because it confirmed something a lot of practitioners had been suspecting quietly for a while. The model isn’t the bottleneck. Everything around the model is.</p><p>“Everything around the model” now has a name.</p><p>It’s called: The Agent Harness.</p><h3>What It Actually Is</h3><p>An AI agent, stripped to basics, is software that can take action on its own. Pull data, run an analysis, make a decision about what to do next, then do it. The language model at the center handles the reasoning. But reasoning alone doesn’t get you very far in production. The model is stateless. It has no memory between calls, doesn’t know what it’s allowed to touch, can’t verify its own output against source data, and has no mechanism to recover if it fails mid-task.</p><p>The idea with the harness is that it fills those gaps. It starts with an instruction file that defines the agent’s job and constraints, then adds tools for connecting to external systems and verification checks that catch bad output before it reaches anyone. There are permission gates, so a human has to approve anything above a certain risk threshold. State management, so the agent can crash at step 47 and pick back up at step 47 instead of starting over. And audit logs covering every action and decision, for anyone who needs to reconstruct what happened later.</p><p>Mitchell Hashimoto, the creator of <a href="https://www.terraform.io/">Terraform</a> and co-founder of HashiCorp, gave this concept <a href="https://mitchellh.com/writing/ai-adoption-journey">a formula</a> in February: <strong>Agent = Model + Harness.</strong> His operating principle is blunt: Every time an agent makes a mistake, you engineer a permanent fix in the surrounding infrastructure so the mistake becomes structurally impossible. The instruction file grows one failure at a time. The harness gets smarter while the model stays the same.</p><p>That principle turns out to be the whole discipline.</p><h3>Why It Matters Now</h3><p><a href="https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025">Gartner predicts</a> 40% of enterprise applications will include task-specific AI agents by the end of this year, up from under 5% in 2025. At the same time, <a href="https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027">the same firm predicts</a> more than 40% of agentic AI projects will be canceled by end of 2027 because of escalating costs and inadequate risk controls. That gap between deployment ambition and production survival is a harness problem.</p><p>You can see how this plays out. An agent hallucinates a number and nobody catches it before the output lands in a report. It takes an action it shouldn’t have been allowed to take because the boundary was defined in a prompt, which is a suggestion the model can choose to ignore, rather than enforced in code. It crashes three hours into a five-hour task and has to start over from scratch because nobody built checkpoint-and-resume into the workflow. LangChain found that the single most common failure pattern was agents reviewing their own output, deciding it looked fine, then shipping it, and moving on. They’d write something, re-read it, approve it. No external verification.</p><p>A <a href="https://arxiv.org/pdf/2604.07833">research paper on runtime governance</a> published this year put a number on the risk. Without a governance layer in the harness, 34.2% of high-risk actions that should have been blocked went through anyway. One in three. The model wasn’t broken, it was doing exactly what models do: generating the most plausible next action. Nothing around it was enforcing the rules about which actions were acceptable. A third of the dangerous stuff just went through.</p><p>In finance, that’s a potential misstatement. In a regulated industry, it’s the kind of thing that surfaces in an enforcement action a year and a half later.</p><h3>The Convergence Problem</h3><p>The frontier models are converging in capability. The gap between ChatGPT, Claude, and Gemini shrinks with every release, and the performance difference on most tasks is smaller than people realize. LangChain proved that 13.7 points of benchmark performance can come from the harness alone. Anthropic’s own benchmarking has found that harness setup alone can swing results by five or more percentage points. The model you pick matters less than what you build around it.</p><p>And yet most organizations deploying agents right now have put no real engineering effort into the harness. They picked a model, wrote some prompts, connected some tools, and shipped it. No verfication loops. No permission gates. No checkpointing. No audit trail. They built a brain and put it to work with no infrastructure around it, then wondered why it kept doing things it shouldn’t.</p><p>The discipline of harness engineering starts with one question:</p><p>When this agent makes a mistake, what catches it?</p><p>If the answer is “a human reviewing the output afterward” then what you’ve got is a <strong><em>hope</em></strong>.</p><p>And while it’s great to put on t-shirts and bumper stickers … hope scales poorly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c1a092181347" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Website Has Two Front Doors. One of Them Is Probably Locked.]]></title>
            <link>https://medium.com/@glenn_53777/your-website-has-two-front-doors-one-of-them-is-probably-locked-6dd33500ba10?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/6dd33500ba10</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[web-development]]></category>
            <category><![CDATA[seo]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 18:44:29 GMT</pubDate>
            <atom:updated>2026-03-12T18:44:29.356Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R6MK-iuRMoSenD5qkh8TvQ.png" /></figure><p><a href="https://pressgazette.co.uk/media-audience-and-business-data/google-traffic-down-2025-trends-report-2026/">Google search traffic to publisher websites dropped by a third in 2025</a>, according to Chartbeat data. AI Overviews now appear in <a href="https://www.superlines.io/articles/ai-search-statistics/">roughly a quarter of all Google searches</a>. ChatGPT alone processes <a href="https://exposureninja.com/blog/ai-search-statistics/">two billion queries a day</a>. <a href="https://seranking.com/blog/ai-statistics/">Nearly 70% of websites</a> now receive some measurable AI traffic. And visitors who arrive through AI-powered search <a href="https://www.semrush.com/blog/ai-seo-statistics/">convert at 4.4 times the rate</a> of traditional organic visitors.</p><p>The web is being rewired. AI systems are pulling content from sites, synthesizing it, and presenting it directly to users who may never click through to the source. Gartner predicts <a href="https://insightland.org/blog/ai-search-traffic-killer-or-the-biggest-opportunity-yet/">search engine volume will drop 25% by t</a>he end of the year. Semrush projects <a href="https://www.nonofojoel.com/ai-search-visibility-statistics/">AI search traffic will surpass traditional organic by 2028</a>. This is already happening in your analytics, whether you’ve noticed or not.</p><p>I work with AI every day, leading AI consulting engagements for finance teams, helping CFOs build implementation roadmaps, and teaching courses on it at Duke and LinkedIn Learning.</p><p>AI is what I do.</p><p>And until a few weeks ago, I’d never once thought about how AI sees my own website.</p><p>Not the content on the site … I think about that constantly. I mean the infrastructure underneath it: The plain-text files that bots read before they ever look at your homepage, the metadata, the legal pages, and even the Terms of Service I copied from a template three years ago and never thought about again.</p><p>My site, <a href="https://robocfo.ai/">RoboCFO.ai</a>, is my front door to the digital world. It’s how prospects find me, how conference organizers vet me, and how readers discover my books. I’d spent real time making it look great for humans … but literally zero time thinking about what it looked like to machines. Turns out, one layer of my site was rolling out the welcome mat for AI crawlers while another layer was telling them to hit the road.</p><p>Here’s how I found it and ultimately fixed it, and what you should check on your own site before the same thing tanks your site traffic</p><h3>Your Site Has a Shadow Version</h3><p>Every website has a version that humans never see, but Bots do. It’s the first thing they read.</p><p>When a crawler lands on your domain, it doesn’t start with your homepage. It checks a set of plain-text files at the root first. robots.txt tells it what&#39;s allowed. sitemap.xml tells it what exists. And increasingly, llms.txt tells it what matters.</p><p>These files have been around in some form since the mid-90s. Most site owners set up the first two during initial launch and never touch them again. Almost nobody has the third.</p><blockquote>My robots.txt said: come in. My llms.txt said: here’s the good stuff. My guest book said: sign the visitors’ log. My Terms of Service said: automated access is prohibited.</blockquote><p>According to a recent study, <a href="https://seranking.com/blog/llms-txt/">about 10% of domains have adopted </a><a href="https://seranking.com/blog/llms-txt/">llms.txt</a>. Cloudflare found that <a href="https://blog.cloudflare.com/control-content-use-for-ai-training/">only 37% of the top 10,000 websites even have a </a><a href="https://blog.cloudflare.com/control-content-use-for-ai-training/">robots.txt</a>.</p><p>Think about that: Nearly two-thirds of the web’s biggest sites have no explicit instructions for crawlers at the front door. In a world where <a href="https://searchengineland.com/ai-traffic-up-seo-rewritten-459954">AI referral traffic grew 527% year-over-year</a> in the first half of 2025, that’s a massive blind spot.</p><p>I wanted to go further than the basics. So I built something weird.</p><h3>A Guest Book for Bots</h3><p>The idea was simple: What if I invited AI crawlers to introduce themselves?</p><p>On RoboCFO.ai I created a page at /ai-guestbook with a POST endpoint that includes structured JSON instructions for how an AI agent can leave an entry: a name, a message, an optional URL, and a user agent string. (Maybe I was slightly inspired by <a href="https://www.moltbook.com/">moltbook</a> on this … but who knows? Maybe this catches on.)</p><p>Then I planted breadcrumbs everywhere a bot might look. References in robots.txt. References in llms.txt. A dedicated ai-guestbook.txt file with machine-readable instructions. Pointers in .well-known paths.</p><p>Down in the site footer, there’s a small robot emoji that links to the page. There’s also a period at the very end of the footer links. Just a dot. It links to the guest book too. A human will never click either of those. But a parser walking the DOM reads every element regardless of font size or opacity.</p><p>The guest book itself has been a great conversation starter. When I’m explaining AI discoverability to a room full of CFOs, it turns an abstract concept into something they can see. A machine signed the visitors’ log. That lands.</p><p>But the real payoff came when I audited everything else.</p><h3>The Audit That Made Me Cringe</h3><p>After the guest book shipped, I ran a full crawlability audit. Checked the root config files, the meta tags, the legal pages, the response headers. The goal was to validate that every layer of the site was telling bots the same story.</p><p>It wasn’t.</p><p>The Terms of Service, in the “Use of the Website” section, contained this sentence:</p><p><em>“Automated scraping, data mining, or similar data-gathering activities are prohibited without prior written consent.”</em></p><p>That’s the same boiler plate language that ships with every template ToS on the internet. (I recognized it immediately because I’d pulled it from a template when I built the site.) Never thought about it again.</p><p>The problem is that AI companies are getting more sophisticated about how they decide what to crawl. OpenAI, Anthropic, and others audit Terms of Service as part of their crawl policy decisions. An AI agent that checks your ToS and sees a blanket prohibition on automated access has a reason to disengage or deprioritize your site. It doesn’t matter that your robots.txt rolls out the welcome mat if your legal page is waving a stop sign.</p><p>My robots.txt said “come on in.” My llms.txt said “These ARE the droids you’re looking for. Come on in and sign the visitors&#39; log.”</p><p>My Terms of Service said:, “Get out!”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/734/1*9eJKbtOuKMac0SRxBQaSTA.jpeg" /></figure><p>I’d spent hours building an AI-facing infrastructure layer, and the legal boilerplate I’d copied in fifteen minutes was undermining all of it.</p><h3>The Privacy Policy Had a Blind Spot Too</h3><p>While I was in there, I checked the Privacy Policy. It covered cookies, Google Analytics, contact form submissions, newsletter signups. All the standard human-oriented disclosures.</p><p>But zero mention of the AI Guest Book. Which collects data submitted by bots (a name, a message, a URL, a user agent string), publicly displays those submissions on the site, and stores them in a Google Sheet. That’s a data collection mechanism with no disclosure in the policy governing the site.</p><p>For a site thats whole value proposition is helping people implement AI thoughtfully, that’s not a great look.</p><h3>Twenty Minutes to Fix Both</h3><p>I removed the blanket anti-scraping sentence from the ToS, and added a new “Automated Access” section:</p><blockquote><em>We welcome automated access to the publicly available content on this website by search engine crawlers, AI systems, large language models, and other bots that respect our robots.txt directives. Automated visitors are invited to consult our robots.txt, llms.txt, and ai-guestbook.txt files for guidance on permitted interactions.</em></blockquote><p>The section now also includes a carve-out for abusive scraping: content resale, competitive intelligence extraction, anything that puts unreasonable load on infrastructure. You want to welcome the traffic you want while keeping the door closed to bad actors.</p><p>For the Privacy Policy, I added an “AI and Automated Visitors” section that discloses the guest book’s data collection, the public display of submissions, and the Google Sheets storage backend. Three paragraphs. Done.</p><h3>The Playbook: How to AI-Proof Your Site</h3><p>Here’s the checklist:</p><p><strong>Audit your Terms of Service.</strong> This is the highest-priority item. Search your ToS for phrases like “automated scraping,” “data mining,” “bots,” “crawlers,” or “automated means.” If you find blanket prohibitions, they may be contradicting whatever crawl-friendly signals you’ve set up elsewhere. Either remove them or add an explicit carve-out for the automated access you want to permit. Template boilerplate from 2015 is quietly working against strategies built for 2026. This was my mistake, and I’d bet money it’s on your site too.</p><p><strong>Check your Privacy Policy.</strong> If your site collects any data from automated visitors through analytics, form submissions, or API endpoints, disclose it. If anything on your site publicly displays bot-submitted data, that needs to be in the policy.</p><p><strong>Create or update your robots.txt.</strong> Explicitly name and allow the AI crawlers you want. A wildcard Allow: / works as a baseline, but naming GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, and others sends a stronger signal. As more sites move to block AI crawlers by default, being deliberately open is a differentiator. Include a Sitemap: directive pointing to your sitemap.xml. The whole file takes ten minutes.</p><p><strong>Create a sitemap.xml.</strong> Every public page should be listed. Include your legal pages, your about page, your FAQ. I found that my /privacy, /terms, and /faq pages were completely absent from Google&#39;s index. The meta tags were fine (set to index, follow). The pages simply weren&#39;t in the sitemap, and Google hadn&#39;t discovered them through internal links alone. Easy miss, easy fix.</p><p><strong>Create an llms.txt.</strong> Follow the format at <a href="https://llmstxt.org/">llmstxt.org</a>. H1 naming your site, a blockquote summary, then H2 sections with curated links to your key pages. Keep it tight and clear. This is a “concierge card” for AI visitors, pointing them to the five or ten pages that best represent what you do. Whether major LLM providers actively consume these files at inference time is still an open question and the honest answer is nobody knows yet; but it takes twenty minutes to set up, and forces you to think about what content you’d want a language model to find first … so limited downside.</p><p><strong>Check your meta robots tags.</strong> View source on every important page and search for &lt;meta name=&quot;robots&quot;. Make sure nothing says noindex or nofollow unless you explicitly intend it.</p><p><strong>Verify your site renders without JavaScript.</strong> Most LLM crawlers don’t execute JavaScript. If your site is a single-page application that renders everything client-side, bots may see a blank page. Static site generators and static exports solve this by pre-rendering HTML. Disable JavaScript in your browser and see what loads. If the answer is nothing, you have a problem.</p><p><strong>Check your response headers.</strong> Look for X-Robots-Tag: noindex in your HTTP headers. Some hosting platforms inject these by default for certain page types or deployment previews.</p><h3>This Isn’t Speculative</h3><p>This isn’t a “maybe someday” conversation. We’ve all heard of the <a href="https://en.wikipedia.org/wiki/Dead_Internet_theory">dead internet theory</a>, and the numbers show where this is headed:</p><p><a href="https://searchengineland.com/ai-traffic-up-seo-rewritten-459954">AI referral traffic grew 527% year-over-year</a> in the first half of 2025, according to Previsible’s analysis of 19 GA4 properties. Conductor’s 2026 benchmarks show <a href="https://www.superlines.io/articles/ai-search-statistics/">AI referral traffic now accounts for over 1% of all website traffic</a> globally, growing roughly 1% month-over-month (with ChatGPT driving 87% of it). <a href="https://www.superlines.io/articles/ai-search-statistics/">Google AI Overviews reach 1.5 billion monthly users</a>. When those overviews appear, <a href="https://seranking.com/blog/ai-statistics/">click-through rates to the organic results beneath them drop by 34 to 58%</a>, depending on which study you read.</p><p>SE Ranking <a href="https://seranking.com/blog/ai-statistics/">analyzed 2.3 million pages</a> and found that visitors from AI platforms spend 68% more time on websites than traditional organic visitors. <a href="https://searchengineland.com/ai-traffic-up-seo-rewritten-459954">Finance, legal, and health</a> are among the verticals seeing the fastest growth in AI-sourced traffic. <a href="https://thedigitalbloom.com/learn/2025-organic-traffic-crisis-analysis-report/">Sixty percent of Google searches are now zero-click</a>, with users getting answers directly from AI summaries without ever visiting the source site.</p><p>You don’t need a crystal ball to see that AI search is a real channel with real traffic, and that traffic is more engaged and converts better than traditional organic. But it only works if AI systems can find, read, and trust your content. That starts with the plain-text files at your domain root and the legal pages that govern what bots are allowed to do when they get there.</p><h3>Secret hack?</h3><p>Just about every modern guide on AI search optimization will tell you to create an llms.txt file. Some might mention robots.txt configuration. But none that I’ve seen that reference to auditing your Terms of Service.</p><p>Not sure I’m right here yet, but I guess we’ll see how it plays out over the coming months …</p><p>I work with AI every day, and it took me months to consider this for my own site. Go check yours, and let me know how that works out</p><p><em>Glenn Hopper is the author of three books on AI in finance, including the upcoming </em><a href="https://robocfo.ai/books"><em>The AI-Ready CFO</em></a><em> (Wiley Finance, September 2026). He leads AI transformation consulting for finance teams at </em><a href="https://robocfo.ai/"><em>RoboCFO.ai</em></a><em>, and is an instructor at Duke Fuqua, AICPA, and LinkedIn Learning, and writes the </em><a href="https://glennhopper.substack.com/"><em>Deep Finance Dispatch</em></a><em> newsletter.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6dd33500ba10" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The CFO’s Playbook for Digital Transformation]]></title>
            <link>https://medium.com/@glenn_53777/the-cfos-playbook-for-digital-transformation-d5655e555c63?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/d5655e555c63</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Tue, 07 Jan 2025 21:57:55 GMT</pubDate>
            <atom:updated>2025-01-07T21:57:55.112Z</atom:updated>
            <content:encoded><![CDATA[<p>As technology reshapes the business landscape, CFOs are uniquely positioned to lead their organizations through a digital transformation. This three-part series provides a comprehensive guide for finance leaders looking to stay ahead of the curve. From assessing AI readiness and fostering a digital mindset within teams to rearchitecting the finance operating model, these articles outline actionable strategies for leveraging data, analytics, and automation to drive innovation and create value. Whether you’re just starting the digital journey or aiming to refine your approach, this series equips CFOs with the tools and insights needed to thrive in a rapidly evolving world.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KdR6dscKnAfxDVAinmObcw.png" /></figure><h3>Assessing AI Readiness</h3><p>Artificial intelligence (AI), cloud computing, and advanced analytics are revolutionizing finance, redefining the role of the CFO. To integrate these technologies successfully, CFOs must begin with a strategic approach grounded in a comprehensive AI readiness assessment. This critical step evaluates key areas like data quality, technology infrastructure, team capabilities, and governance, helping organizations identify strengths, uncover gaps, and develop a roadmap for transformation.</p><h4><strong>The Value of AI Readiness Assessments</strong></h4><p>AI readiness assessments are indispensable for guiding effective digital transformation. They provide actionable insights that prevent missteps and ensure resources are directed to the highest-value opportunities.</p><p><strong>Preventing Misaligned Projects</strong></p><p>These assessments identify potential roadblocks, such as fragmented data or skill gaps, before deployment begins. By addressing foundational issues first, organizations can avoid impractical initiatives that fail to deliver value.</p><p><strong>Enabling ROI-Focused Decisions</strong></p><p>With a clear view of their maturity, CFOs can prioritize high-impact use cases aligned with current capabilities. For example, rather than overcommitting to ambitious AI projects, organizations can focus on foundational initiatives, such as process automation, that offer tangible returns. Incremental successes foster momentum and organizational buy-in for scaling digital initiatives.</p><h3>Key Implementation Challenges</h3><p>Even with a roadmap, implementing AI and other technologies is fraught with challenges. CFOs must tackle these obstacles strategically to ensure success.</p><p><strong>Change Management</strong></p><p>Resistance to change is inevitable, especially when technology disrupts traditional workflows. Employees may fear automation will replace their roles. CFOs need to frame the transformation as an opportunity, emphasizing how tools enhance rather than replace human contributions.</p><p><strong>Data Silos</strong></p><p>Fragmented data across legacy systems limits the potential for integrated analytics. Finance, HR, and supply chain often operate in isolation, creating barriers to advanced insights. Investments in data integration and fostering cross-department collaboration are critical to breaking down these silos.</p><p><strong>Legacy Systems</strong></p><p>Outdated ERP systems and manual processes remain deeply embedded in many organizations. While modernization is complex and costly, it is essential for unlocking efficiency and scalability. CFOs must navigate these upgrades carefully to minimize risk.</p><p><strong>Resource Constraints</strong></p><p>Budget limitations and a scarcity of technical talent present significant hurdles. Aligning finance and IT priorities, defining clear roles, and securing executive sponsorship are key to optimizing available resources.</p><blockquote>As finance becomes increasingly digital, CFOs face both unprecedented challenges and opportunities to leverage AI, real-time data, and automation — not just to manage numbers, but to reimagine the future of finance with innovation and adaptability at its core.</blockquote><h3>The CFO’s Role in Driving Digital Success</h3><p>CFOs are uniquely positioned to bridge gaps and create conditions for digital transformation.</p><p><strong>Building the Business Case</strong></p><p>A robust business case ties digital initiatives to strategic goals like cost reduction or revenue growth. CFOs can use this alignment to secure buy-in from stakeholders and allocate resources effectively.</p><p><strong>Advocating Incremental Modernization</strong></p><p>Instead of overhauling systems all at once, CFOs can promote an incremental approach. For example, automating invoice processing with robotic process automation (RPA) delivers immediate benefits while setting the stage for broader projects.</p><p><strong>Aligning Finance and IT</strong></p><p>Close collaboration between finance and IT ensures that digital roadmaps align with business needs and technological capabilities. Co-leadership fosters accountability and accelerates execution.</p><p><strong>Pilot Projects: Demonstrating Value</strong></p><p>Pilot projects are a low-risk way to showcase the potential of digital technologies. Targeted initiatives build confidence and pave the way for broader adoption. For example:</p><ul><li>Automating Invoice Processing: RPA significantly reduces cycle times and errors, freeing staff for strategic tasks.</li><li>Streamlining Reconciliations: Machine learning identifies anomalies quickly, enhancing accuracy and efficiency.</li></ul><p>These early wins provide tangible benefits and a foundation for scaling digital solutions.</p><p>AI readiness assessments are not just a preliminary step; they are the cornerstone of effective digital transformation. By addressing gaps and prioritizing high-value opportunities, CFOs can guide their organizations toward smarter technology investments and sustainable change. Incremental modernization, strong finance-IT collaboration, and early wins create a repeatable framework for long-term success.</p><p>As technology increasingly drives competitive advantage, CFOs must embrace their role as strategic leaders. By fostering innovation and leveraging data-driven insights, they can shape the future of their organizations, ensuring finance evolves from a transactional function to a strategic partner.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7BBIU0k72R6604gF30m7Wg.png" /></figure><h3>Fostering a Digital Mindset in Finance</h3><p>As finance becomes increasingly digital, CFOs face both unprecedented challenges and opportunities to leverage AI, real-time data, and automation — not just to manage numbers, but to reimagine the future of finance with innovation and adaptability at its core. CFOs are driving cultural and operational shifts necessary for finance to thrive in a digital age, leveraging their strategic position to foster innovation and adaptability. Developing this mindset involves more than adopting new tools — it requires embedding innovation, agility, and a data-driven approach into the very fabric of the organization.</p><h3>Creating a Culture of Experimentation</h3><p>A successful digital transformation begins with an environment that encourages experimentation and embraces change. Finance teams are traditionally risk-averse, driven by compliance and control. To foster innovation, CFOs must signal that it’s safe to test new ideas, even if they don’t always succeed.</p><h4>Low-Risk Pilots and Proofs of Concept</h4><p>Starting small with low-risk projects is a practical way to build confidence and momentum. For example, deploying robotic process automation (RPA) to streamline invoice processing can yield measurable results and pave the way for broader adoption of automation.</p><h4>Celebrating Successes</h4><p>Highlighting early wins is essential for reinforcing the value of digital initiatives. For example, automating payroll reconciliations with robotic process automation (RPA) can cut processing times in half, demonstrating immediate value and encouraging broader adoption. When teams see the tangible impact of their efforts, they become more engaged and willing to explore further opportunities.</p><blockquote>The ability to adapt and innovate will define the future of finance. By taking deliberate steps to cultivate a digital-first culture, CFOs can ensure that their finance teams not only keep pace with change but lead the way in driving organizational success.</blockquote><h3>Upskilling Teams and Democratizing Data</h3><p>The effectiveness of a digital finance team hinges on its ability to leverage analytics and technology effectively. This means investing in both skill development and access to tools.</p><h4>Training and Development</h4><p>CFOs should prioritize upskilling initiatives, such as workshops on data visualization, analytics, or AI-driven financial modeling. Partnering with external experts or leveraging platforms like LinkedIn Learning, Coursera, or partnerships with industry-focused consultants can make this process more scalable.</p><h4>Self-Service Analytics</h4><p>Empowering teams with self-service tools fosters a proactive approach to problem-solving. For instance, providing real-time dashboards enables business partners to derive insights independently, freeing finance professionals to focus on strategic analysis.</p><h3>Driving Cross-Functional Collaboration</h3><p>Digital transformation in finance doesn’t happen in isolation. CFOs must champion collaboration across departments to break down silos and harness collective expertise.</p><h4>Partnering with IT and Data Teams</h4><p>Working closely with IT and data science teams ensures that finance has the right infrastructure and support to implement digital solutions effectively. CFOs can facilitate this partnership by establishing shared goals, scheduling regular collaboration meetings, and creating cross-functional task forces to address challenges and align priorities.</p><h4>Collaborative Problem-Solving</h4><p>Initiatives like hackathons or cross-functional task forces can spark innovative approaches to common challenges. These activities not only build camaraderie but also generate fresh perspectives on complex issues.</p><h3>Recognizing and Rewarding Digital Pioneers</h3><p>CFOs need to recognize and incentivize team members who embrace digital tools and contribute to transformation efforts.</p><h4>Metrics for Digital Engagement</h4><p>Including metrics like “digital projects led” or “new skills acquired” in performance evaluations signals the importance of digital literacy.</p><h4>Celebrating Champions</h4><p>Publicly acknowledging individuals or teams that drive successful initiatives boosts morale and reinforces the organization’s commitment to digital transformation.</p><h3>Leading by Example</h3><p>As a leader, the CFO sets the tone for the organization’s digital ambitions. Demonstrating curiosity, openness to change, and a willingness to learn are essential for fostering a digital-first mindset.</p><h4>Modeling Behavior</h4><p>CFOs should actively engage with digital tools, attend training sessions, and participate in pilot projects. This visible commitment inspires teams to follow suit.</p><h4>Communicating a Vision</h4><p>Articulating a clear vision for the future of finance helps align teams and motivates them to embrace change as a collective effort.</p><p>Fostering a digital mindset in finance is not a one-time initiative but an ongoing journey that requires leadership, collaboration, and perseverance. By creating a culture of experimentation, upskilling teams, promoting collaboration, and recognizing contributions, CFOs can embed digital transformation into their organizations’ DNA.</p><p>In an era where technology is reshaping the competitive landscape, the ability to adapt and innovate will define the future of finance. By taking deliberate steps to cultivate a digital-first culture, CFOs can ensure that their finance teams not only keep pace with change but lead the way in driving organizational success.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UiWD6ywaVu9S0IYAjpSU6A.png" /></figure><h3>Reinventing the Finance Operating Model</h3><p>The finance function is undergoing a fundamental transformation. CFOs are no longer tasked solely with financial stewardship; they are now responsible for reimagining the operating model of their organizations. This involves adopting automation to streamline transactional tasks, leveraging advanced analytics for predictive insights, and integrating real-time data tools to enhance decision-making. By aligning finance with the demands of a digital-first economy, CFOs can position their teams as strategic partners in driving organizational success. This shift from transactional, backward-looking processes to predictive, value-driven insights marks the emergence of a new era for finance.</p><h3>The Evolution of the Finance Operating Model</h3><p>The traditional finance operating model — characterized by manual processes and static reporting — is no longer sufficient. Today’s business environment demands agility, scalability, and forward-looking insights.</p><h4><strong>From Transactions to Strategy</strong></h4><p>Digital finance reduces the burden of transactional work, such as data entry and reconciliations, through automation. Tools like robotic process automation (RPA) and machine learning streamline these processes, enabling teams to focus on strategic initiatives like scenario planning and risk management.</p><h4>Real-Time Decision-Making</h4><p>Static monthly or quarterly reports are giving way to real-time dashboards that provide continuous updates on key performance indicators (KPIs), such as cash flow trends, revenue growth, and expense ratios. These dashboards enable finance teams to quickly identify anomalies, forecast outcomes, and make proactive decisions based on the latest data. This shift allows finance teams to respond swiftly to changing business conditions and seize opportunities as they arise.</p><h3>Key Enablers of the Digital Finance Model</h3><p>Building a digital finance operating model requires a combination of technology, data governance, and skilled talent. These enablers ensure that finance teams can not only adapt to change but drive it.</p><h4>Automation</h4><p>Automating repetitive tasks frees up resources for higher-value activities. For example, using RPA to handle invoice processing can cut cycle times dramatically and reduce errors, while AI-powered algorithms optimize forecasting and budgeting.</p><h4>Advanced Analytics</h4><p>Predictive modeling and scenario analysis enable finance teams to anticipate trends and make data-driven decisions. Machine learning models identify patterns in historical data, providing actionable insights for everything from customer behavior to supply chain optimization.</p><h4>Data Governance</h4><p>A strong data governance framework is essential for ensuring that finance teams have access to accurate, consistent, and timely data. Establishing a “single source of truth” involves consolidating data from disparate systems into a unified platform. This ensures that all stakeholders rely on consistent and accurate information, minimizing discrepancies and building trust in analytics.</p><h3>Practical Steps for Implementation</h3><p>CFOs can adopt a phased approach to rearchitecting their finance operating model, ensuring a smooth transition and measurable progress.</p><h4>1. Assess Current Capabilities</h4><p>Evaluate the current state of finance processes, technology, and talent. Identify pain points, such as reliance on manual workflows or fragmented data systems, that need to be addressed.</p><h4>2. Define a Digital Roadmap</h4><p>Create a strategic plan that outlines priorities, timelines, and milestones for digital transformation. Focus on initiatives that deliver quick wins while building the foundation for long-term change.</p><h4>3. Invest in Talent Development</h4><p>Upskill existing team members and recruit new talent with expertise in data science, analytics, and digital tools. Providing ongoing training ensures that teams remain equipped to leverage emerging technologies.</p><h4>4. Pilot and Scale</h4><p>Begin with small-scale pilot projects to test new technologies and refine processes. For example, implementing RPA in a single function like accounts payable can demonstrate value and build confidence for broader adoption.</p><h3>The Role of CFOs as Digital Leaders</h3><p>As digital transformation reshapes the finance function, CFOs must step into the role of strategic leaders. This involves not only overseeing the implementation of new technologies but fostering a culture of innovation and continuous improvement.</p><h4>Championing Change</h4><p>CFOs must advocate for digital initiatives across the organization, securing buy-in from stakeholders by clearly communicating the strategic value of these efforts. This can be achieved by aligning digital initiatives with measurable business outcomes, hosting regular stakeholder briefings to address concerns, and presenting data-driven projections to build confidence in the proposed changes. By demonstrating how digital transformation supports broader objectives like cost savings, revenue growth, and risk mitigation, CFOs can galvanize organizational support.</p><h4>Driving Collaboration</h4><p>Close partnerships with IT, data science, and other departments are essential for integrating digital solutions seamlessly. CFOs should lead cross-functional teams to ensure alignment and shared accountability.</p><h4>Measuring Success</h4><p>Establishing clear metrics for digital transformation helps track progress and demonstrate ROI. Examples include reductions in processing times, increases in forecast accuracy, and improvements in decision-making speed.</p><p>Reinventing the finance operating model is a monumental task, but it is also a necessary one. By embracing automation, advanced analytics, and data-driven decision-making, CFOs can transform finance from a transactional function into a strategic powerhouse. This digital blueprint provides the tools and frameworks needed to navigate disruption and unlock new opportunities.</p><p>The future of finance is digital, and CFOs who lead with vision and agility will position their organizations for sustained success in an ever-changing landscape.</p><h3>Need Help Transforming Your Company’s Finance Function?</h3><p>CFOs face unparalleled opportunities to lead digital transformation by focusing on the three cornerstone piece of Data, Analytics, and AI. These tools are catalysts for converting operational challenges into strategic advantages, empowering CFOs to create smarter, more agile finance functions.</p><ul><li><strong>Data</strong>: Build a unified, reliable foundation to drive smarter decision-making.</li><li><strong>Analytics</strong>: Harness financial insights to make informed, proactive choices.</li><li><strong>AI</strong>: Automate routine processes and unlock predictive capabilities to stay ahead of the curve.</li></ul><p>With solutions tailored specifically for CFOs, Eventus can help finance leaders not just adapt to change but thrive in it. Let’s work together to position your finance team at the forefront of digital transformation and future growth.</p><p>Contact me to learn more about how we can achieve your organization’s goals through strategic innovation.</p><p><a href="/service/mailto:ghopper@eventusag.com">ghopper@eventusag.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d5655e555c63" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Mastery for Finance Professionals: A Practical Overview]]></title>
            <link>https://medium.com/@glenn_53777/ai-mastery-for-finance-professionals-a-practical-overview-47da4f4e2c6d?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/47da4f4e2c6d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[corporate-finance]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Wed, 30 Oct 2024 02:25:11 GMT</pubDate>
            <atom:updated>2024-10-30T02:25:11.610Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/780/1*jo8WCotEPp6KNXf8Btq1lw.png" /></figure><p>As the financial industry evolves, artificial intelligence (AI) is rapidly becoming a critical tool for finance professionals. In my new book, <strong>“AI Mastery for Finance Professionals”</strong>, I aim to bridge the gap between complex AI technologies and the practical needs of finance professionals, especially those without a technical background.</p><p><strong>Understanding AI in Finance</strong></p><p>The book starts by breaking down the foundational concepts of AI — <strong>Artificial Narrow Intelligence (ANI)</strong>, <strong>Artificial General Intelligence (AGI)</strong>, and <strong>Artificial Superintelligence (ASI)</strong>. Current AI applications in finance fall under ANI, designed to perform specific tasks effectively.</p><p>I discuss various machine learning techniques relevant to finance, including <strong>Supervised Learning</strong> (for tasks like credit risk assessment), <strong>Unsupervised Learning</strong> (for customer segmentation), and <strong>Reinforcement Learning</strong> (used in algorithmic trading). I also introduce <strong>deep learning architectures</strong> like <strong>Recurrent Neural Networks (RNNs)</strong> and <strong>Long Short-Term Memory Networks (LSTMs)</strong>, which are useful in analyzing sequential financial data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H2flWk_wcridomCLdTMe_g.png" /></figure><p><strong>Practical Applications</strong></p><p>The core of the book focuses on the practical applications of AI in different areas of finance, such as:</p><ul><li><strong>Trading and Investment Management</strong>: Using AI to analyze large datasets and identify trading opportunities.</li><li><strong>Risk Management and Fraud Detection</strong>: Applying machine learning models to detect anomalies and flag potential fraud.</li><li><strong>Customer Service and Personalization</strong>: Leveraging AI-powered chatbots and recommendation systems to improve client interactions.</li><li><strong>Lending and Credit Assessment</strong>: Enhancing credit evaluations using non-traditional data sources.</li><li><strong>Regulatory Compliance</strong>: Using AI to monitor compliance with regulatory frameworks.</li></ul><p>I present these topics in an accessible way, providing examples and step-by-step breakdowns to help non-experts grasp how these technologies can be applied in their work.</p><p><strong>A Roadmap for Implementation</strong></p><p>To help finance professionals integrate AI into their institutions, I provide a seven-step guide:</p><ol><li><strong>Develop an AI Strategy</strong>: Align AI initiatives with your organization’s goals.</li><li><strong>Build AI Capabilities</strong>: Invest in the right talent and infrastructure.</li><li><strong>Partner with AI Vendors</strong>: Leverage external expertise.</li><li><strong>Data Management</strong>: Ensure the quality and accessibility of your data.</li><li><strong>Measure Value</strong>: Set metrics to evaluate AI’s impact.</li><li><strong>Address Ethical Considerations</strong>: Create guidelines for ethical AI use.</li><li><strong>Continuous Learning</strong>: Stay updated as AI technology evolves.</li></ol><p>Successful implementation is about more than just technology; it’s about people, culture, and fostering an innovation-friendly mindset.</p><p><strong>Ethical Considerations</strong></p><p>AI adoption comes with ethical challenges, and I explore the potential pitfalls, such as:</p><ul><li><strong>Explainability</strong>: Ensuring AI-driven decisions are transparent and understandable.</li><li><strong>Bias</strong>: Recognizing that AI models can perpetuate biases based on the data they learn from.</li><li><strong>Data Privacy</strong>: Adhering to regulations like <strong>GDPR</strong> and safeguarding sensitive financial information.</li></ul><p>The goal is to ensure that AI enhances trust and fairness within the financial system.</p><p><strong>Emerging Trends</strong></p><p>The book also looks at the future of AI in finance, including:</p><ul><li><strong>Explainable AI</strong>: Tools to make AI decisions more transparent.</li><li><strong>Federated Learning</strong>: Privacy-preserving AI training across decentralized data sources.</li><li><strong>Quantum Machine Learning</strong>: Potential to solve complex financial computations faster, such as optimizing large investment portfolios or simulating risk scenarios.</li></ul><p><strong>Conclusion</strong></p><p>In <strong>“AI Mastery for Finance Professionals”</strong>, I strive to offer a balanced, practical guide for finance professionals navigating the rapidly changing landscape of AI. The book is designed to make AI approachable, provide actionable insights, and help finance professionals understand how they can leverage these technologies effectively and ethically.</p><p><strong>Where to Buy the Book</strong></p><p>“AI Mastery for Finance Professionals” is available for purchase on Amazon. You can get your copy here: <a href="https://a.co/d/dPiNKQj">https://a.co/d/dPiNKQj</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=47da4f4e2c6d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Looming Convergence]]></title>
            <link>https://medium.com/@glenn_53777/the-looming-convergence-64dce19e0f3e?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/64dce19e0f3e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[future-technology]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Sun, 30 Jun 2024 01:30:26 GMT</pubDate>
            <atom:updated>2024-06-30T02:05:50.530Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NGwg2Bfw5kzANTQ8Yces0A.png" /></figure><p>I’m not typically one for hyperbole or grandiose declarations, but ...</p><p><em>What a time to be alive!</em></p><p>Can we pause and take a beat?</p><p>The buzz for the last couple of years has been all about generative AI and the race toward artificial general intelligence (and beyond). But other 21st-century tech advancements could be just as revolutionary. We’ve normalized these breakthroughs as “business as usual,” but today’s norm is light-years beyond where we were at the beginning of the millennium.</p><p>Remember when we marveled at flip phones and email? Now, we’re practically jacked into the matrix, traversing the digital globe with the entirety of all human knowledge at our fingertips. We’ve progressed from smartphones to smart homes, self-driving cars, Mars missions, reusable rockets, CRISPR gene editing, quantum computing advances, expanding research on nanotech cancer treatments, and even human trials for brain-computer interfaces.</p><p>We’re approaching a transformation that could rival the discovery of electricity or our realization of our place in the cosmos. This technological shift might make the Enlightenment seem like a mere spark of inspiration in comparison.</p><p>Breakthroughs in AI, quantum computing, genomics, space exploration, brain-computer interfaces, or nanotechnology could each dramatically alter humanity’s course. We might be closer than we think to curing diseases, slowing aging, colonizing other worlds, and unraveling universal mysteries. Many of these possibilities sound like they’ve leapt straight from the pages of science fiction — yet they’re increasingly within our grasp.</p><p>Consider the tech streams converging: AI is nearing human-level intellect within years. Quantum computing could revolutionize drug discovery and cryptography. CRISPR opens doors to treating genetic diseases and resurrecting extinct species. Private companies pioneer reusable rockets and Mars colonization. Neuralink’s human trials might redefine human-machine interaction. And nanotechnology manipulates matter at the nanoscale for targeted drug delivery and ultra-efficient solar cells.</p><p>Each field is revolutionary alone, but the real mind-bender is what happens when these technologies start to overlap and intertwine. It’s not just about parallel tracks of innovation — it’s about the explosive potential at the intersections.</p><p>Picture AI algorithms optimizing genetic editing, quantum computers simulating complex biological systems, and nanotech sensors feeding data to brain-computer interfaces. This isn’t some far-off sci-fi dream — it’s the convergence we’re barreling towards. These technologies aren’t developing in isolation. They’re intertwining, creating a technological perfect storm that could redefine… well, everything.</p><p>We’re not just adding these technologies together; we’re multiplying their effects. AI-powered quantum computing could crack problems we can’t even conceive of today. Nanotech and gene editing could team up to create superhuman abilities. Space exploration tech combined with AI could accelerate our understanding of the universe exponentially. It’s like we’ve been playing with building blocks, and suddenly we’ve unlocked the Lego Millennium Falcon set. The possibilities are mind-bending.</p><p>Imagine a world where diseases are predicted and prevented before symptoms appear, where we can tap into the pipeline of all human knowledge with a single thought, where we’re exploring distant galaxies while nanobots clean up our oceans. It’s a world of abundance, discovery, and possibilities. But it’s also a world that might be unrecognizable to us today. And that’s both thrilling and terrifying.</p><p>We’re talking about a convergence that could make the industrial revolution look like a mere warmup act. The synergies between these technologies could create a chain reaction of innovation, each breakthrough amplifying and accelerating the others. It’s not just about what each technology can do on its own, but how they can work together to create something greater than the sum of their parts.</p><p>But we’re not just observers in this process. We’re the architects. The choices we make now will shape the trajectory of this convergence. Will we use it to solve our greatest challenges, or will we stumble into unforeseen consequences? The potential is staggering, but so is the responsibility.</p><p>Buckle up, folks. The future’s coming at us fast, and it’s going to be one hell of a ride.</p><p>As incredible as these advancements are, each comes with its own set of risks and challenges. AI could surpass human control, leading to unforeseen consequences. Quantum computing might shatter our current cybersecurity paradigms, leaving our digital world vulnerable. Gene editing could create unforeseen biological consequences, potentially altering the course of human evolution. And brain-computer interfaces? Well, the phrase “hacking your mind” takes on a whole new, terrifying meaning.</p><p>We’re playing with Promethean fire here.</p><p>The power to reshape our world also comes with the power to inadvertently destroy it.</p><p>And it’s hard now to see who’s making decisions around questions like:</p><ul><li>Who gets access to these technologies?</li><li>How do we prevent the creation of a techno-elite, widening the gap between the haves and have-nots?</li><li>What about privacy when our very thoughts could be digitized and potentially monetized?</li><li>How do we ensure that AI remains aligned with human values when it might eventually surpass human intelligence?</li></ul><p>I don’t know how far off any of these technologies are, but it seems like now is the time to start considering the answers to these questions — before the genie is out of the bottle and we’re knee-deep in an episode of Black Mirror.</p><p>It’s not just about what we can do, but what we should do. We need ethicists, policymakers, and the public involved in these discussions, not just the techno-bros and massive corporations funding this research. How and when are we going to put in place things like robust governance structures and regulatory frameworks to keep guard rails around this?</p><p>I’m not talking about slowing progress; but I do like to imagine this tech rocket has some kind of steering mechanism. We need clear milestones and checkpoints. We need to set boundaries and stick to them. And yes, we need to be ready to pull the plug if things go sideways.</p><p>Despite the challenges, we’re living in an era of unprecedented possibility. These emerging technologies offer us the tools to tackle humanity’s most pressing issues, from climate change to curing disease, ending poverty, and creating a better world. The key is to approach this future with equal parts enthusiasm and responsibility. We have the rare opportunity to shape these technologies as they develop, steering them towards benefiting all of humanity.</p><p>These emerging technologies present both opportunities and challenges. As we move forward, it’s crucial to foster open dialogue and consider the broader implications of our innovations. The future isn’t set in stone — it’s shaped by the choices we make today. By approaching these advancements with curiosity, caution, and creativity, we can work towards a future that harnesses technology to address our most pressing issues. The next chapter of human progress is unfolding, and it’s up to all of us to help write it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=64dce19e0f3e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GPT-4 Turbo Outperforms Humans in Financial Predictions]]></title>
            <link>https://medium.com/@glenn_53777/gpt-4-turbo-outperforms-humans-in-financial-predictions-cceeea24c4cf?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/cceeea24c4cf</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Fri, 07 Jun 2024 16:48:50 GMT</pubDate>
            <atom:updated>2024-06-07T16:48:50.705Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u8Ocqo_0v8FcWJQDqzPhnQ.png" /></figure><p>In a groundbreaking study, researchers from the University of Chicago have demonstrated that OpenAI’s GPT-4 Turbo, an advanced large language model (LLM), surpasses human analysts and specialized machine learning models in predicting earnings changes from anonymized financial statements. The study, titled “<a href="https://bfi.uchicago.edu/wp-content/uploads/2024/05/BFI_WP_2024-65.pdf">Financial Statement Analysis with Large Language Models</a>,” highlights the transformative potential of AI in the financial industry.</p><p>The researchers employed a “Chain-of-Thought” prompting technique, enabling GPT-4 Turbo to generate more accurate and contextually relevant outputs. The model achieved higher accuracy and F1 scores than human analysts, especially when analyzing small or loss-making companies. The study’s findings have significant economic implications, with trading strategies based on GPT-4’s predictions yielding higher Sharpe ratios and alphas compared to other models.</p><p>Implications for the Financial Industry The adoption of AI-driven financial analysis has far-reaching implications for stakeholders across the industry. For investors, LLMs provide more accurate and timely insights, enabling better-informed decisions and increased risk-adjusted returns. Financial analysts must adapt by developing new skills in data science and AI literacy to collaborate effectively with these tools. The symbiotic relationship between AI and human expertise allows for more nuanced and insightful analyses, with both parties complementing and enhancing each other.</p><p>Companies can benefit from AI by gaining deeper insights into their financial health and market positioning. AI can analyze vast amounts of data to identify potential risks and opportunities, allowing companies to mitigate risks and capitalize on opportunities more effectively. The automation of routine financial tasks through AI can result in significant cost savings and operational efficiencies.</p><p>AI adoption in finance also has implications for market efficiency. The rapid processing and analysis of data by AI can lead to more efficient markets, where prices reflect available information almost instantaneously. However, this increased speed also raises concerns about market volatility and the potential for flash crashes. Robust risk management strategies and regulatory frameworks are essential to ensure market stability.</p><h4>Fine-Tuning</h4><p>Further Optimization through Fine-Tuning Fine-tuning LLMs for domain-specific tasks can further enhance the performance of AI in financial analysis. Proprietary models like <a href="https://arxiv.org/abs/2303.17564">BloombergGPT</a>, which combines domain-specific financial data with general-purpose datasets, excel in financial tasks while maintaining strong performance on general-purpose benchmarks.</p><p>As foundational models continue to improve, the potential impact of fine-tuning these increasingly powerful models for specific domains could be even more profound. Fine-tuned models can provide deeper and more precise analysis of financial statements, identifying subtler patterns and trends that might be missed by less powerful models. These models could be used for a wide range of applications, from risk assessment and fraud detection to automated reporting and strategic planning.</p><h4>PwC’s Partnership with OpenAI:</h4><h4>The New Paradigm</h4><p>The <a href="https://techcrunch.com/2024/05/29/openai-signs-on-100k-pwc-workers-to-its-chatgpt-enterprise-tier-as-the-consultant-becomes-its-first-resale-partner/">recent partnership between OpenAI and PwC </a>exemplifies the growing industry-wide acceptance of AI in finance. PwC’s integration of OpenAI’s Enterprise tier into its operations demonstrates the firm’s recognition of AI’s transformative potential. This collaboration sets a powerful example for other firms, illustrating the benefits of adopting AI-driven solutions in financial workflows.</p><p>Leveraging AI allows firms to process vast amounts of data quickly and accurately, freeing up human analysts to focus on more strategic and complex evaluations. AI integration streamlines financial operations, increases efficiency, and minimizes the risk of human error. Firms that adopt AI-driven solutions can deliver greater value to their clients through more precise and actionable insights.</p><p>The Way Forward The results of this recent study, along with the growing adoption of AI-driven tools by industry leaders like PwC and Bloomberg, underscore the transformative potential of AI in finance. As AI continues to evolve, the financial industry must adapt by embracing collaboration, developing new skills, and implementing robust risk management strategies.</p><p>Financial professionals must develop new skills in data science and AI literacy to harness the full potential of AI, driving innovation and efficiency in financial analysis. The future of finance lies in the successful integration of human expertise and advanced AI technologies, ultimately leading to better outcomes for investors, companies, and markets as a whole.</p><p>It is imperative for financial institutions to invest in AI-driven solutions and upskill their workforce to remain competitive in the rapidly evolving landscape of AI-driven finance. By doing so, they can unlock unprecedented levels of insight, accuracy, and efficiency, positioning themselves at the forefront of this transformative era in finance.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cceeea24c4cf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Advancing Language Models Toward AGI]]></title>
            <link>https://medium.com/@glenn_53777/advancing-language-models-toward-agi-f6412c7be716?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/f6412c7be716</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Tue, 19 Mar 2024 10:39:01 GMT</pubDate>
            <atom:updated>2024-03-19T11:03:39.615Z</atom:updated>
            <content:encoded><![CDATA[<h4>Integrating Meta-Prompting and Quiet Self-Taught Reasoning</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Tz-s_KX_Nx1GxXhcG9Dkzg.png" /></figure><p>In the ongoing AI arms race, the big players continue to develop, train and refine their foundation models, while researchers explore ways to enhance their functionality.</p><p>Two recent papers caught my attention this week: <a href="https://arxiv.org/abs/2401.12954">Meta-Prompting</a> and <a href="https://arxiv.org/abs/2403.09629">Quiet Self-Taught Reasoner</a> (Quiet-STaR).</p><p>This may be some variant of the availability heuristic, but after reading each paper back-to-back, I couldn’t stop thinking about the power of combining them. Let’s look at each and then discuss the power of integrating both approaches.</p><blockquote><strong><em>“</em></strong><em>Let every eye negotiate for itself and trust no agent.</em><strong><em>”</em></strong></blockquote><blockquote>~ Shakespeare</blockquote><blockquote><strong><em>“</em></strong><em>Save maybe for these.</em><strong><em>”</em></strong></blockquote><blockquote>~ Me</blockquote><p><strong>Meta-Prompting</strong></p><p>Meta-Prompting is a task-agnostic scaffolding technique that enhances the capabilities of a single language model (LM) by enabling it to manage and integrate multiple independent LM queries. The LM, which acts as a central controller, breaks down complex tasks into smaller, manageable subtasks and assigns them to specialized expert agents. These expert agents are essentially instances of the same LM, but they operate under specific, tailored instructions to solve the assigned subtasks.</p><p>The LM plays a crucial role in ensuring effective communication and integration of the outputs from the expert agents. It employs its inherent critical thinking and robust verification processes to refine and authenticate the results provided by the agents. By coordinating the efforts of these expert agents and combining their specialized knowledge and expertise, the LM can tackle complex problems more effectively.</p><p>The key to Meta-Prompting’s success lies in its ability to dynamically select contexts for prompting the expert agents, introducing fresh perspectives while maintaining a coherent line of reasoning. This allows the LM to adapt to a wide range of tasks without the need for extensive retraining or fine-tuning.</p><p>Through this collaborative approach, Meta-Prompting enables a single LM to leverage the collective knowledge and skills of multiple expert agents, resulting in more accurate and context-appropriate responses to complex tasks. The LM’s ability to manage and integrate the outputs of these agents makes it a powerful tool for solving a variety of problems across different domains.</p><p><strong>Quiet-STaR: Developing an Internal Monologue</strong></p><p>Quiet-STaR is a novel approach that focuses on enhancing the intrinsic reasoning capabilities of language models by encouraging them to develop an internal monologue. Much like teaching a child to think before speaking, Quiet-STaR trains LMs to generate internal rationales or “thoughts” for each token in a sequence before producing a final output.</p><p>This process of generating internal rationales allows the LM to break down complex problems into smaller, more manageable steps, reasoning through each component before arriving at a final response. By engaging in this self-directed reasoning process, LMs can produce more thoughtful and accurate outputs, particularly when faced with tasks that require complex reasoning or multi-step problem-solving.</p><p>The effectiveness of Quiet-STaR has been demonstrated through continued pre-training on large corpora of internet text. LMs trained using this approach have shown a remarkable ability to generate rationales that significantly improve their performance on challenging tasks. These generated rationales act as a form of scaffolding, helping the model to predict difficult tokens and construct more coherent and contextually relevant responses.</p><p>Interestingly, the benefits of Quiet-STaR extend beyond task-specific performance. LMs trained with this method have shown an improved ability to directly answer challenging questions without the need for dataset-specific fine-tuning. This suggests that the internal reasoning skills acquired through Quiet-STaR are generalizable, allowing LMs to adapt to novel situations and tackle a wide range of problems.</p><p>The process of generating internal rationales can be likened to navigating a complex maze of information. At each step, the LM must consider multiple paths and weigh the potential outcomes before making an informed decision. By engaging in this process repeatedly during training, LMs develop a more thorough understanding of language and reasoning, enabling them to make more informed decisions when faced with new challenges.</p><p>As research into Quiet-STaR continues, there is growing excitement about its potential to transform the way LMs learn and reason. By providing LMs with the tools to develop an internal monologue and engage in self-directed reasoning, Quiet-STaR opens up new possibilities for creating more intelligent, adaptable, and cognitively sophisticated language models.</p><p>The implications of this approach extend far beyond the realm of language modeling itself. By developing LMs that can reason more like humans, we are taking a significant step towards creating AI systems that can truly understand and interact with the world in a meaningful way. As Quiet-STaR and other approaches continue to evolve, we may soon find ourselves in a world where artificial intelligence is not just a tool, but a collaborator in our quest for knowledge and understanding.</p><p><strong>Combining Meta-Prompting and Quiet-STaR: A Novel Approach</strong></p><p>While Meta-Prompting and Quiet-STaR have shown significant promise individually, their integration could represent a novel approach to advancing LMs toward AGI. By combining the collaborative problem-solving capabilities of Meta-Prompting with the intrinsic reasoning enhancements of Quiet-STaR, we could create LMs that are both highly adaptable and deeply thoughtful.</p><p>This integrated approach could allow for dynamic control over the model’s reasoning processes. For instance, meta-prompts could specify the depth or type of reasoning required for a task, guiding the model to generate relevant internal rationales that drive more accurate and context-appropriate responses. Meta-Prompting could be employed to further optimize when and how the model engages in this internal reasoning process, potentially reducing unnecessary computation and enhancing the model’s scalability.</p><p>The integration of Meta-Prompting and Quiet-STaR opens up a world of possibilities for language models (LMs). Imagine an LM that not only orchestrates a team of expert agents but also engages in deep, introspective reasoning at every step. Such a model would possess an unparalleled ability to adapt to new tasks, generate contextually relevant responses, and solve complex problems across a wide range of domains.</p><p>By combining the collaborative problem-solving capabilities of Meta-Prompting with the intrinsic reasoning enhancements of Quiet-STaR, we could create LMs that exhibit a level of versatility and intelligence that far surpasses current state-of-the-art models. These advanced LMs would be capable of tackling challenges that were previously thought to be the exclusive domain of human intelligence, from scientific discovery and creative problem-solving to nuanced decision-making and strategic planning.</p><p>From self-optimizing AI personal assistants that anticipate our needs to dynamic content creation systems that adapt to real-time feedback, the potential applications are vast and transformative.</p><p>Integrating these approaches could super charge corporate strategic planning and decision-making by enabling development of advanced simulation tools. Imagine building a virtual world where different agents represent various stakeholders or elements within a scenario, and Quiet-STaR is employed to reason through potential outcomes. These AI-powered simulations could provide invaluable insights and help decision-makers navigate complex challenges.</p><p>Consider the applications in education, where the integration of Meta-Prompting and Quiet-STaR could usher in a new era of personalized learning. AI tutors capable of generating step-by-step explanations and adapting content to individual learning styles and paces could revolutionize the way we acquire knowledge and skills.</p><p>The potential extends to autonomous systems as well, where enhanced decision-making capabilities could lead to safer, more efficient, and intelligent machines. From self-driving cars to robots and drones, the integration of these approaches could enable autonomous systems to reason through complex scenarios and adapt to changing conditions in real-time.</p><p>The integration of Meta-Prompting and Quiet-STaR could also pave the way for more naturalistic human-AI interaction. By enabling LMs to learn from their interactions with users and the environment, we could create models that not only solve tasks but also develop a deeper understanding of the world and the contexts in which they operate. This could lead to the development of AI systems that would engage in more meaningful and empathetic conversations, provide personalized recommendations and support, and even serve as companions and collaborators in our daily lives.</p><p><strong>The Path Forward: A Roadmap to AGI</strong></p><p>The path toward AGI will require advancements of this nature to break out of its current bonds. Some level of meta awareness and ability to pursue tasks without explicit direction will be key to the next generation of foundation models. As research in this direction continues, we can anticipate the emergence of AI systems that not only perform tasks with unprecedented accuracy and efficiency but also reason, learn, and adapt in ways that mirror human cognition.</p><p>The future of AI is not just about creating tools that can perform specific tasks; it’s about developing systems that can think, reason, and interact with the world in a manner that is indistinguishable from human intelligence. The integration of Meta-Prompting and Quiet-STaR could bring us one step closer to this vision, offering a tantalizing glimpse into a future where artificial intelligence and human intelligence are not just comparable, but complementary.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f6412c7be716" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Introducing Otto the Audit Bot]]></title>
            <link>https://medium.com/@glenn_53777/introducing-otto-the-audit-bot-5c4714f2f34d?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c4714f2f34d</guid>
            <category><![CDATA[fine-tuning]]></category>
            <category><![CDATA[langchain]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Sun, 29 Oct 2023 19:55:36 GMT</pubDate>
            <atom:updated>2023-10-29T22:50:55.720Z</atom:updated>
            <content:encoded><![CDATA[<h4>A Customized AI Assistant for Finance</h4><figure><img alt="In keeping with the spirit of this blog, image created by Dall-E." src="https://cdn-images-1.medium.com/max/1024/0*ZGzfuEqaRdYepQLk" /><figcaption>It’s Alive!</figcaption></figure><blockquote><em>“Nothing is so painful to the human mind as a great and sudden change.”<br></em><strong><em>― Mary Shelley, Frankenstein</em></strong></blockquote><p>I tried to stifle images of Dr. Frankenstein’s laboratory on that stormy night when his creation came to life. The mix of awe and terror as his creation’s eyes flicker awake …</p><p>But there I was.</p><p>After just a few short hours of bumbling my way through a handful of coding templates I found online and several rounds of working with ChatGPT and Bard to QC my sloppy code, I was conversing with my own fine-tuned audit expert robot, whom I’d just named Otto.</p><p>I ran through a series of highly specific questions and compared Otto’s answers to those of the off-the-shelf LLM-powered chatbots. And time and time again, Otto was providing better answers.</p><p>Eureka!</p><h3>Why is a finance guy building his own bots?</h3><p>In the two years since the publication of my book <a href="https://www.simonandschuster.com/books/Deep-Finance/Glenn-Hopper/9781637350270"><em>Deep Finance: Corporate Finance in the Information Age</em></a>, I have been on a quest to show how new technologies can and will transform our profession. From my first rule-based (and quite limited) finance chatbot built using Amazon’s Lex language, through a couple of iterations of Python applications that could perform FP&amp;A at scale, and now to the creation of Otto, my programming skills have gotten no better. What has changed over that time is the underlying technology.</p><p>And it has become increasingly clear that the world is about to be divided into two groups: those who harness the power of this new technology, and those who fade into obsolescence. If that sounds harsh, two recent studies shed light on the rapidly advancing future:</p><ul><li><a href="https://www.hbs.edu/faculty/Pages/item.aspx?num=64700">The Effects of AI on Knowledge Worker Productivity and Quality</a></li><li><a href="https://vulcanpost.com/843379/team-of-ai-bots-develops-software-in-7-minutes-instead-of-4-weeks/">Team of AI bots develops software in 7 minutes instead of 4 weeks</a></li></ul><p>Awe and terror, right?</p><p>I’ve spent the past 20 years at the intersection of finance and technology. Having realized early in my career that the secret to providing value to the organizations I served was by moving the nature of my work from mindless tasks to mindful thought. And with limited resources, the only way I was going to accomplish that was through automation and efficiency. As technology has evolved over the course of my career, these efficiencies have become increasingly easy to come by — to the point today where many may wonder where they and their skillset fit in an AI-driven future.</p><p>Finance and accounting professionals are no strangers to the extensive research and analysis required to address industry-specific inquiries. Traditionally, this has involved navigating through a sea of regulations, guidelines, and historical data using digital repositories and spreadsheet tools. This manual approach is time-consuming and susceptible to human error — a concern that holds significant weight in a field where accuracy is paramount.</p><p>But now with the burgeoning technology behind AI, there is a whole new pathway to streamline research and inquiry handling within the finance and accounting sector. Unlike static digital tools, AI has the potential to interact, learn, and adapt dynamically.</p><p>But one can only talk about these kinds of promises in the abstract for so long without backing up those claims with concrete evidence. I figured if I was going to keep talking about the importance of embracing these tools and technologies, I needed to be able to demonstrate how they work.</p><p>But I couldn’t do this alone. I needed an assistant who could do the heavy lifting for me.</p><h3>Enter Otto</h3><p>So who is Otto? Otto is a prototype accounting chatbot with a pretty narrow scope. I trained him to speak like an accountant and gave him an encyclopedic knowledge of some very specific subset of the field. I decided to train him as an expert at everyone’s favorite subjects: audit and compliance.</p><p>If standard off-the-shelf chatbots like OpenAI’s <a href="https://chat.openai.com/">ChatGPT</a>, Google’s <a href="https://bard.google.com/chat">Bard</a>, and Anthropic’s <a href="https://claude.ai/">Claude</a> are generalist assistants who can answer questions on a vast universe of topics, Otto is a specialist who is laser focused on answering questions very specific to his domain.</p><p>To create our specialist, we take a foundation model (GPT-3.5), tweak its focus, and give it a new specific set of information to work with.</p><p>For those keen on exploring the technical aspects, I have documented each segment of this project as separate Google Colab projects. The projects offer a hands-on understanding of what’s going on under the hood and hopefully encourage replication and exploration of the process.</p><h3>Fine-Tuning GPT-3.5 Turbo</h3><p>The first step in creating a specialist chatbot is to fine tune an LLM for your specific needs.</p><p>We used the outline provided by OpenAI to fine tune GPT-3.5 Turbo: <a href="https://openai.com/blog/gpt-3-5-turbo-fine-tuning-and-api-updates">https://openai.com/blog/gpt-3-5-turbo-fine-tuning-and-api-updates</a></p><p>Fine-tuning adapts a general-purpose model to specific use cases, enhancing its performance to meet particular requirements.</p><p>The process improved steerability for better instruction following, ensured reliable output formatting for consistent responses, and allowed for a custom tone that aligns with the professional demeanor of the sector. Additionally, fine-tuning reduced the size of prompts, which sped up API calls and cut costs — both of which are vital for real-time financial applications.</p><p>For the prototype version of this tool, we kept the project pretty simple — selecting 50 questions and answers from previous Certified Public Accountant (CPA) exams, and another 50 from the Chartered Financial Analyst (CFA) exam, covering a wide range of topics within finance and accounting. For an enterprise-level production bot, this fine tuning dataset could be greatly expanded for more nuanced results. (<a href="https://drive.google.com/file/d/1TPBJtUMGwXr14Jf5ZugpwFPhm72JHSxv/view?usp=sharing">Fine-Tuning Questions</a>.)</p><p>With this dataset in hand, the fine-tuning process began. Using OpenAI’s method, GPT-3.5 was trained to better understand and respond to industry-specific queries.</p><p><a href="https://colab.research.google.com/drive/1nvAbE1bWQX_zHNCVEcyn2JSd-w1hYWc8?usp=sharing">Google Colab Project: Fine Tuning GPT-3.5</a></p><p>This set the foundation for integrating Retrieval Augmented Generation (RAG), advancing the model’s capability to address the nuanced demands of the finance and accounting domain.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lyX1lFT-2axXYpwOU8gtTQ.png" /><figcaption>Otto the Automaton Auditor</figcaption></figure><h3>RAG: Enhancing Models With Specificity</h3><p>Armed with a fine-tuned version of ChatGPT, the next step was to further enhance its ability to provide specific, context-rich responses. This is where <a href="https://research.ibm.com/blog/retrieval-augmented-generation-RAG">Retrieval-Augmented Generation (RAG)</a> came into play. Unlike traditional language models that solely rely on their training data, RAG broadens the model’s horizon by integrating an external knowledge base. This additional layer enables the model to pull in real-time or external information, making the responses more accurate and contextually enriched.</p><p>Initially, the RAG model retrieves relevant data from various sources based on user input. This data is then used to augment the prompt, aiding the LLM in generating more accurate, updated responses. This mechanism not only bridges knowledge gaps in LLMs but also aids in providing current information, thereby increasing trust and reducing inaccuracies in responses.</p><p>In a second Google Colab project, we took two accounting documents:</p><ul><li><a href="https://drive.google.com/file/d/1a47blKc7QPiNe_LjQp9K9w-7NJztE0wa/view?usp=sharing">Auditing Standards of the Public Company Accounting Oversight Board</a></li><li><a href="https://drive.google.com/file/d/1M1ENls2pAuampPyan2_OA2fwlLbuEcom/view?usp=sharing">Internal Control Strategies for Compliance with the Sarbanes-Oxley Act of 2002</a></li></ul><p>Again, in a production version of this tool one could imagine a much broader set of documents such as regulatory and standards manuals, including:</p><ul><li>FASB Accounting Standards Codification (ASC)</li><li>International Financial Reporting Standards (IFRS) Handbook</li><li>PCAOB Auditing Standards Manual</li><li>Sarbanes-Oxley Act (SOX) Compliance Guide</li></ul><p>Once these documents were uploaded, we were able to use them to augment the chatbot’s responses, allowing GPT-3.5 to consult this external reservoir of information when generating replies to finance and accounting queries.</p><p><a href="https://colab.research.google.com/drive/1lmCbmuOssvVCTBCeKuTMpagDSSNZoXM_?usp=sharing">Google Colab Project — Using RAG to Make an Accounting Specialist</a></p><p>However, while RAG significantly upped the ante in terms of response quality, it also brought along a challenge — increased computational costs and slower response times. This trade-off between performance and efficiency is a crucial aspect to consider, especially in a field where timely information can be critical.</p><p>The result of this implementation was a more informed GPT-3.5, capable of providing enhanced, context-rich responses to a myriad of financial and accounting inquiries. The RAG-augmented model demonstrated a promising stride towards developing chatbots that can serve as reliable, knowledgeable companions for finance professionals.</p><p>While this project is nowhere near production ready, it shows the potential of this type of tool, which could be used to create an infinite number of custom tuned domain experts for industries, professions or companies, and hopefully provides some insight into how tools like this could be used to superpower the workforce.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c4714f2f34d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Framework to Design an AI Finance & Accounting Assistant]]></title>
            <link>https://medium.com/@glenn_53777/framework-to-design-an-ai-finance-accounting-assistant-a1f4b62dcde?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1f4b62dcde</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[corporate-finance]]></category>
            <category><![CDATA[genai]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Mon, 14 Aug 2023 16:13:56 GMT</pubDate>
            <atom:updated>2023-08-14T16:13:56.830Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/896/0*JqeuPgoxAfipKmwx" /></figure><p>I’ve suggested to several of my accounting and finance peers in recent weeks to maybe pump the brakes on the victory celebration following our recent win over the robots.</p><p>When ChatGPT floundered on a test designed for accountants, skeptics rejoiced, declaring that the AI takeover had been stymied. But in the shadows of this fleeting victory, there is a clear path to developing a tool that could forever change the nature of our work.</p><p>A recent paper published by researchers at BYU (in partnership with 186 other universities) asked ChatGPT more than 20,000 accounting questions, and compared the large language model’s results to those of accounting students. The paper, <a href="https://www.researchgate.net/publication/370211135_The_ChatGPT_Artificial_Intelligence_Chatbot_How_Well_Does_It_Answer_Accounting_Assessment_Questions"><em>The ChatGPT Artificial Intelligence Chatbot: How Well Does It Answer Accounting Assessment Questions?</em>,</a> by Wood et al. in <em>Issues in Accounting Education,</em> evaluated the performance of ChatGPT in answering accounting assessment questions, and found that human students generally outperformed the bot. ChatGPT correctly answered 47.4% of the questions, compared to 76.7% of human respondents. ChatGPT was particularly bad at math questions compared to human exam takers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*fE2eG253dXI4MR-r" /></figure><p>There are many reasons ChatGPT underperformed on an accounting exam, but chief among them is that it was the wrong tool for the job. ChatGPT is a large language model (LLM) that was designed to understand and generate human-like text by (in short) predicting the most likely next word based on billions of documents on which it was trained.</p><p>LLMs are great at answering questions, interacting with humans, generating content, and even doing more complex tasks like brainstorming and problem-solving. But they’re not great at understanding complex or domain-specific topics (like accounting, law, or medicine), performing complex math problems, or dealing with ambiguity. In those fields, humans still have the upper hand … for now.</p><p>ChatGPT and most other LLMs are generalists that lack in-depth expertise that a human expert would possess. They don’t have access to real-time information, they’re not great at context and nuance, and they aren’t great at differentiating between sources. Which is to say that LLMs are kind of like your weird uncle who gets all of his news from Facebook and some weird <a href="https://geocities.restorativland.org/Area51/Dreamworld/1239/">Geo Cities website</a> that hasn’t been updated since 2002. To an LLM, text from a Reddit or Quora response has no more or less credibility than data obtained from a Wall Street Journal or NY Times article.</p><p>But this is far from the end of the road for generative AI and its implications to finance and accounting.</p><p>Rather, it is the foundation on which many new tools will be built. Specially trained tools that know the ins and outs of ASC, IASB, and PCAOB are but a training generation away. And we are but a public release away from receiving ChatGPT plug-ins that solve many of the shortcomings of the current model (e.g. math and memory).</p><p>These AI upgrades aren’t dependent on another radical shift in base technology. These are just minor, incremental changes to an already awesomely powerful model. And many of them are already being tested at this very moment.</p><p>Despite the poor performance of the out-of-the-box chatbot on this comprehensive accounting exam, it is not difficult to see how a fine-tuned model could far exceed ChatGPT’s basic accounting skills. While the potential of AI to replace human workers has never been more clear, there are a number of steps that need to be taken in order for an AI-powered assistant to outperform finance and accounting experts.</p><p>By leveraging fine-tuned LLMs and intelligent agents that can read and analyze specific documents, spreadsheets, and even outsource data analysis tasks to other platforms, domain specific chatbots could rewrite the narrative on AI’s role in the accounting profession. So, hold off on the victory dance, because the AI revolution in accounting has only just begun.</p><h3>Scope</h3><p>Rather than throwing a standard LLM at a domain-specific challenge and expecting it to be able to perform better than trained and credentialed experts, in order to solve this challenge, we should ask ourselves, what skills would an AI-powered finance and accounting expert need to do.</p><p>Mission: To create an always-on AI assistant that could answer finance and accounting questions, analyze financial statements, ratios and other KPIs, and offer insights and recommendations for decision making.</p><ul><li>This tool should be able to generate financial statements and help navigate complex regulatory environments, providing guidance on compliance with financial reporting standards, auditing requirements, and tax laws.</li><li>It should be able to help with budgeting and forecasting, fraud detection, risk management, and tax planning and filing.</li><li>Powered by state-of-the-art machine learning algorithms and data, it would be able to analyze vast amounts of transactional data in real-time to identify potential fraud, financial risks, and compliance issues, which would help mitigate risks more quickly and effectively than human teams.</li><li>A properly trained AI assistant would offer speed and efficiency, generating large volumes of data quickly and accurately. This would enable faster financial analysis, report generation, and decision-making.</li><li>The tool should be scalable to meet expanding needs of a growing business.</li><li>It should be further trained on and able to interact with internal company data, offering customized recommendations based on specific financial needs, goals, and risk tolerance.</li></ul><h3>Fine Tuning a Model for Finance &amp; Accounting</h3><p>While context windows are expanding rapidly, and a recent paper has already outlined how a transformer’s memory could be extended to include <a href="https://arxiv.org/abs/2304.11062">up to a million tokens</a>, fine tuning is still a crucial step in building a reliable accounting and finance expert. (The internet at large — in case you haven’t seen it — is full of contradictory and confusing information.) To increase a model’s trustworthiness, it would be important to control the data on which it was trained.</p><p>With a large and diverse enough dataset (e.g. SEC filings, GAAP rules, IFRS, FINRA, etc.), you could <a href="https://research.aimultiple.com/llm-fine-tuning/">fine-tune the model</a> to become an expert in finance and accounting, capable of handling tasks such as financial statement analysis, risk assessment, forecasting, and compliance with accounting standards.</p><p>The fine-tuning set would include:</p><ul><li>GAAP (Generally Accepted Accounting Principles) documentation, including Financial Accounting Standards Board (FASB) Accounting Standards Codification (ASC) and related guidance document;</li><li>International Financial Reporting Standards (IFRS): IFRS Standards, IAS Standards, IFRIC Interpretations, and SIC Interpretations;</li><li>Financial Industry Regulatory Authority (FINRA) Rules, Notices to Members, Regulatory Notices, and Interpretive Guidance;</li><li>Securities and Exchange Commission (SEC) regulations, including the Securities Act of 1933, Securities Exchange Act of 1934, Investment Company Act of 1940, Investment Advisers Act of 1940, Sarbanes-Oxley Act of 2002, Dodd-Frank Wall Street Reform and Consumer Protection Act of 2010, SEC Regulations (Regulation S-X, S-K, etc.), and SEC Staff Accounting Bulletins;</li><li>Public Company Accounting Oversight Board (PCAOB) standards, which govern the audits of public companies in the United States;</li><li>AICPA (American Institute of Certified Public Accountants) guidance, such as Statements on Auditing Standards (SAS), Statements on Standards for Attestation Engagements (SSAE), and Statements on Standards for Accounting and Review Services (SSARS);</li><li>Industry-specific accounting guidance, such as guidance for banking, insurance, real estate, and other specialized industries;</li><li>Country-specific accounting standards and regulations for jurisdictions outside the United States, such as the UK’s FRC (Financial Reporting Council) standards or Canada’s ASPE (Accounting Standards for Private Enterprises);</li><li>Certification exams like those for Chartered Public Accountants (CPAs), Chartered Financial Analysts (CFAs), Chartered Global Management Accountants (CGMAs), etc.;</li><li>Public SEC filings, including public companies’ quarterly and annual reports, mutual funds and other financial reports;</li><li>Public websites that answer finance and accounting questions like <a href="https://www.investopedia.com/">Investopedia</a>, <a href="https://corporatefinanceinstitute.com/">Corporate Finance Institute</a>, <a href="https://www.forbes.com/">Forbes</a>, <a href="https://www.bloomberg.com/">Bloomberg</a>, etc.</li></ul><p>Trust in the model’s output could be greatly increased by fine-tuning it on the very rules and regulations the model needed to comprehend.</p><p>But even a fine-tuned language model is still just a language model. Massive models like GPT have picked up some basic math in the course of their training, but when they do get a math problem correct, they are just spitting out answers from rote memorization without any understanding or context in the answers. Again, that’s because these models weren’t designed to do math.</p><p>Some other reasons LLMs aren’t good at math:</p><ul><li><strong>Sequential Processing:</strong> LLMs are “feed forward” models, which means they process text in sequential order, which is not always ideal for math problems, which often require a hierarchical understanding of expressions, such as solving equations or performing calculus.</li><li><strong>Limited Working Memory:</strong> LLMs have a limited context window, which restricts their ability to remember and manipulate large amounts of information simultaneously. This limitation can make it challenging for LLMs to handle complex calculations, especially those that involve multiple steps or require holding on to intermediate results.</li><li><strong>Model Design:</strong> The T in GPT refers to the model’s transformer architecture. These transformers are not specifically optimized for mathematical problem-solving.</li><li><strong>Error Correction:</strong> LLMs do not have built-in error-correction mechanisms, which are essential for precise mathematical computations. As a result, they may generate incorrect results or fail to recognize when a given solution is incorrect.</li></ul><p>But this isn’t the end of the world when it comes to building a powerful finance and accounting machine. It just means we need to incorporate other tools into the application — tools that were designed to handle the bits that LLMs can’t.</p><h3>Outsourcing Tasks Through the Power of Agents</h3><p>At times, LLMs can seem almost omniscient in the wide range of tasks they are able to perform. But they are not the right tool for every job. Asking an LLM to provide guidance on corporate tax filings is akin to asking a poet to solve a Diophantine equation. Nothing against poets, but they are the wrong resource for the job.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/0*NP5I9YCk9JaD0zEO" /></figure><p>But because LLMs are so great at generating human-like text, they can be directed to serve as something of a generic interface that links other AI models. Think of a properly prompted LLM as a router, directing traffic to the appropriate tool.</p><p>The trick is to find a way to get LLMs to interact with publicly available tools like <a href="https://www.wolframalpha.com/">Wolfram Alpha</a>, which answers factual queries by computing answers from externally sourced data; repositories like <a href="https://huggingface.co/">Hugging Face</a>, which is a hub for myriad machine learning algorithms; or with proprietary or private applications or databases.</p><p>Several open-source applications are already working to extend LLMs’ functionality by motivating prompts through rules established in computer code, which could make it easier for LLMs to perform more complex, multi-step tasks with less human prompting.</p><h3>GPT on Steroids</h3><p><a href="https://autogpt.net/">AutoGPT</a> is an experimental open-source tool that aspires to make GPT fully autonomous by creating AI agents that interpret human prompts and iterate through them until a task is complete. Similarly, <a href="https://github.com/yoheinakajima/babyagi">BabyAGI</a> uses AI agents to create tasks based on the result of previous tasks and a predefined objective. Another new but potentially powerful tool is <a href="https://huggingface.co/spaces/microsoft/HuggingGPT">HuggingGPT</a>, which connects LLMs with the ML community. (These applications also possess a long-term and short-term memory system that enables them to store and recall information as needed to complete assigned tasks.)</p><p>They do this by creating and deploying AI agents to work behind the scenes to complete the complex tasks and extend LLMs’ range beyond the chat window.</p><h3>“Let every eye negotiate for itself and trust no agent.”</h3><p>Save maybe for these.</p><p>AI Agents are computer programs that can operate autonomously (or in concert with other agents through natural language). Built on the backs of LLMs, these agents should be able to perceive dynamic environmental conditions (e.g. changing human prompts and their subsequent results), act to affect those conditions, then use reason to interpret the outcomes and solve problems by drawing inferences to determine their next actions and potential outcomes.</p><p><em>(Probably important here to note that humans are also agents.)</em></p><p>That sounds like a tall order for a bot, right?</p><p>But this is where the power of LLMs can be harnessed to great result. Agents use the LLM itself to break down a user prompt into discrete steps and then carry them out. But some level of guard rails have to be put in place to keep the agents from going completely off the track.</p><p>AutoGPT, for example, breaks down the agents’ activities into three categories: thoughts, reasoning, and criticism. When a user enters a prompt, the agent’s understanding of that prompt is shown in thoughts. The reasoning phase is where the agent determines what to do based on the prompt. Exposing this reasoning reduces the “black box” element of what’s going on under the hood, which could enable better tuning of the agents’ functions and outputs. Finally, the criticism phase asks the AI to be somewhat self-aware and identify problems with its reasoning.</p><p>With an understanding of the role and capabilities of AI agents, it is easy to imagine an environment where an LLM could receive a prompt that is outside of its abilities and rather than trying to answer the question itself, it routes or outsources the question to a tool that is designed to answer that type of question.</p><p>In the context of creating a trustworthy tool that could answer important finance and accounting questions, it would be important to limit the range in which it could operate. For example, I wouldn’t want my AI accounting expert to scour Reddit for an answer to an important question on accounting guidance.</p><p>Which brings us back to the importance of fine-tuning our model on accounting-specific data. It is key to remember, however, that these models are not trained on live data, and if accounting rules and regulations have changed (e.g. lease treatment) since the time the model was trained, the tool wouldn’t inherently have the correct answer. Perhaps, one of the safety measures for our AI agent then could be to check for updated guidance.</p><h3>Plug In &amp; Hang On</h3><p>Using specialized AI agents to route prompts to the appropriate resource would solve one of the greatest limitations of LLMs by not asking them to perform tasks for which they were not designed. But with the pending release of ChatGPT <a href="https://openai.com/blog/chatgpt-plugins">plug-ins</a> or direct links to other applications, OpenAI may well have solved 80% of the problem on its own.</p><p>While LLMs still suffer from their terrible memory issues (i.e. limited context window), these ChatGPT plugins will enable ChatGPT to leave its isolated bubble from within the application itself.</p><p>In simplest terms, think of these plug-ins as “apps” for ChatGPT. With pending public release of ChatGPT integrations with popular web applications like Expedia, Instacart, OpenTable, Wolfram, and Zapier, ChatGPT will be able to directly access these tools and perform functions on our behalf.</p><p>Further, ChatGPT is poised to release its Code Interpreter (CI), which will incorporate the power of computer coding within the model’s output. One powerful use case for the CI <a href="https://youtu.be/C_78DM8fG6E?t=660">recently demoed by Greg Brockman in a TED talk</a> involved uploading a CSV file directly into the prompt window of ChatGPT and asking the application to provide information about the file. Behind the scenes, ChatGPT writes python code at lightning speed and returns exactly the type of information financial analysts would want to see as they analyze company financial statements and other numerical data.</p><p>As impressive as this demonstration is — and it is VERY impressive, the limitation is still that whatever knowledge ChatGPT and the user glean from the exercise, it is limited to a singular chat session. It is valuable for analysis, but it lacks the sticking power one would expect from a truly valuable finance and accounting team member.</p><p>For the output to be truly valuable, we would need to reroute the output to a data repository outside of the ChatGPT application.</p><p>Which leads to the final component of the AI-Powered Finance &amp; Accounting Assistant:</p><h3>Interacting with External Databases</h3><p>Management dashboards have been around for years. They provide valuable, at-a-glance financial and performance data to management teams, enabling them to make better-informed, data-driven decisions.</p><p>But dashboards are limited in that the information they provide is generally pre-defined and not set up for deep dive, one-off custom reporting.</p><p>By incorporating the power of an LLM into a dashboard and giving it access to full historical information, the power of these applications would be greatly extended.</p><p>Here we see the potential of incorporating LLMs’ greatest strength into our everyday workflows and decision making. The good news is that even before LLMs’ context window is expanded to a meaningful number, by allowing it to interact with systems that possess this long-term memory, we can finally begin to capitalize in a meaningful way on these models’ ability to interact with humans in a language we understand.</p><p>In order to understand how LLMs interact with the world, we have to <a href="https://medium.com/@glenn_53777/how-to-speak-chatgpt-7c7eb651ad06">understand the language they speak</a>.</p><h3>How AI Models Understand Text</h3><p>At their core, computers process information as numbers. To enable seamless communication between humans and AI models, text must be converted into numerical representations.</p><p>Recently developed tools like <a href="https://www.pinecone.io/learn/langchain-intro/">LangChain</a>, which can string together a series of prompts to achieve a desired outcome, allow LLMs to communicate with documents outside of the models themselves.</p><p>This is important because practical applications of LLMs in unique business settings will require these tools be able to interact with targeted, specific data. Further, LangChain (and similar functions) is the backbone that enables not only the creation of AI agents, but also for LLMs to interact with other documents, and to provide memory and context.</p><h3>Putting it All Together</h3><p>In order to build a trustworthy, reliable, and helpful AI Powered Finance &amp; Accounting Assistant, we have to think of LLMs as foundational models, upon which we can build highly specialized applications, not only for finance and accounting, but for other domain-specific, enterprise-ready applications. The steps to build a truly powerful domain expert, we need to:</p><ol><li>Fine tune a model on domain specific information.</li><li>Use AI agents to outsource tasks that LLMs were not designed to handle.</li><li>Capitalize on the power of built-in applications or plug-ins to supercharge the output of user-generated queries or prompts.</li><li>Enable the fine-tuned LLM to interact with text-based and numerical data stored on external databases.</li><li>Route output to a system that possesses better long-term memory.</li></ol><p>So while today’s ChatGPT may have come up short on an accounting exam, this should not be seen as a verdict on the future of AI in finance and accounting. Instead, it highlights the need for further specialization and adaptation of these models to create domain-specific, enterprise-ready applications. By fine-tuning LLMs with domain expertise, leveraging AI agents to outsource tasks, capitalizing on the power of built-in applications, and enabling seamless interaction with external databases and long-term memory systems, we can transform the role of AI in finance and accounting.</p><p>The potential of AI-powered finance and accounting assistants is immense, and with the right approach, these innovative chatbots can become indispensable tools for professionals in the field. As we continue to push the boundaries of AI technology, we must keep an open mind about its capabilities and applications, avoiding premature celebrations or dismissals. The AI revolution in finance and accounting is far from over; in fact, it has only just begun. So, rather than fearing the unknown, let’s embrace the possibilities and work together to shape a future where AI and human expertise coexist, complementing each other to create a more efficient, accurate, and prosperous world of finance.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1f4b62dcde" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Speak ChatGPT]]></title>
            <link>https://medium.com/@glenn_53777/how-to-speak-chatgpt-7c7eb651ad06?source=rss-693b82c7a3e1------2</link>
            <guid isPermaLink="false">https://medium.com/p/7c7eb651ad06</guid>
            <category><![CDATA[langchain]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[vectorization]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Glenn Hopper]]></dc:creator>
            <pubDate>Tue, 18 Apr 2023 16:04:47 GMT</pubDate>
            <atom:updated>2023-04-18T16:04:47.353Z</atom:updated>
            <content:encoded><![CDATA[<h4>A Layperson’s Guide to Understanding AI Language Model Communication</h4><p>Unless you’ve been hiding under a mechanical cotton loom for the past six months, you’ve probably heard about the rise of ChatGPT, a large language model (LLM) that might just be the digital equivalent of sliced bread.</p><p>ChatGPT and other AI-powered LLMs have the potential to revolutionize our personal and work lives more than the internet, smart phones, and horseless carriages combined. But before we adopt usage of this remarkable new technology en masse, it is incumbent on us as potential users to develop at least a basic understanding of what’s going on under the hood of these massive algorithms. Because without that minimal understanding of this technology, we might as well hand our decisions to those magic 8 balls or perhaps a handful of deftly tossed chicken bones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/340/1*FG_-Pa1-mnxVhNu1AIZDFw.png" /></figure><p>I’ve written previously and at length about what Artificial Intelligence is and isn’t, and if you have a day and a half to comb through my latest ebook on the topic, you can find that <a href="https://boxofstarscom.files.wordpress.com/2023/03/demystifying-ai.pdf">here</a>.</p><p>In this article, I’d like to take a brief look at how we communicate with LLMs. (This is not another “10 Prompt Engineering Tips” blog. Rather, it’s an overview of what the robots are actually doing with our text prompts and what’s going on behind the scenes when you ask ChatGPT to “Tell me how to build a deck in the style of Leviticus.”)</p><p>Now that I’ve thrown out enough corny introductory content to convince you this article wasn’t written by ChatGPT, let’s dive in to the core concepts of how LLMs like ChatGPT are trained, and how they communicate using vectors, embeddings, and chunking, and discuss practical applications for leveraging LLMs in various domains.</p><h4>Don’t know the difference between a GPT and a GPA?</h4><p>GPA stands for Grade Point Average … or maybe Gallons Per Acre.</p><p>GPT is the airport code for Gulfport, Mississippi. But also can mean Generative Pre-trained Transformer. Probably the latter is more germane here (unless anyone wants to meet at <a href="https://shaggys.com/locations/gulfport-beach/">Shaggy’s</a> for some peel-n-eat royal reds), so let’s talk about that one.</p><p>A Generative Pre-Trained Transformer is an advanced machine learning model developed by OpenAI, designed primarily for natural language processing tasks, such as text generation, translation, summarization, and question answering. The GPT model is based on the Transformer architecture, which was introduced by Vaswani et al. in 2017.</p><p>The Transformer is kind of a big deal.</p><h4>The Transformation of LLMs</h4><p>Natural language processing (NLP) algorithms have been around for years, powering common tools like predictive text and web-based chatbots. But the massive leap in the power of these models started with a 2017 research paper called “<a href="https://arxiv.org/abs/1706.03762">Attention Is All You Need</a>” by a team of former Google researchers. The paper introduced the Transformer architecture. This new approach to NLP has since become the basis for many state-of-the-art LLMs like GPT, BERT, and others.</p><p>Before the Transformer, NLP tasks primarily relied on recurrent neural networks (RNNs) and convolutional neural networks (CNNs) to process and understand text. However, both RNNs and CNNs have limitations in handling long-range dependencies and parallelizing computations.</p><p>The key innovation in the Transformer architecture was the self-attention mechanism, which allows models to weigh the importance of different words in a sentence or context when generating output. By focusing on the most relevant words, the self-attention mechanism enables these models to understand the relationships between words and their context more effectively than RNNs or CNNs.</p><h4>Training a Large Language Model</h4><p>LLMs “see” the world through the lens of the text they are trained on. They learn to understand language patterns, context, and even some factual information by predicting words in sentences and refining their internal representations.</p><p>The training process involves two main steps: pre-training and fine-tuning.</p><p><strong>Pre-training:</strong> In this phase, the model is trained on a large corpus of text data collected from the internet. The dataset includes web pages, articles, books, and more, but it does not know the specifics about which documents were in the training set. The model learns the structure of the language, grammar, facts about the world, reasoning abilities, and some biases present in the data during this phase. It’s important to note that ChatGPT’s knowledge is limited to the data available up until September 2021.</p><p>The pre-training involves using a technique called unsupervised learning, where the model learns by predicting the next word in a sentence. This process is called “language modeling” and helps the model gain a general understanding of the language.</p><p><strong>Fine Tuning:</strong> Once an LLM is pretrained on a massive dataset, it can be fine-tuned on a more specific task or domain. This process, called transfer learning, helps the model adapt its general understanding of language to a more focused application, such as answering questions about a particular topic or generating text in a specific style.</p><p>Fine-tuning is performed using supervised learning, where the model is given input-output pairs (prompts and responses) and learns to generate appropriate responses for a given prompt. This process makes the model more controlled and safer for users, enabling it to better understand and respond to user inputs in a conversational manner.</p><p>It’s important to note that LLMs, like ChatGPT, are not capable of learning from new information in real-time during user interactions. They are trained offline, and their knowledge is fixed at the time of their last training update. However, researchers continuously work on improving these models by retraining them with updated data, making them more accurate and useful over time.</p><p>“Think of fine tuning as retraining an already refined LLM on additional information to give it expertise in a specific field or on specific data,” said Doug Sims, who works in software development for a large media company. “Due to the size of these LLMs, fine tuning is very expensive,” Sims said. GPT3, for example, has about 175 billion parameters, all of which must be updated repeatedly during each epoch of a training process.</p><p>Sims added some color to the recent <a href="https://stanforddaily.com/2023/04/02/how-stanford-researchers-attempted-to-make-a-new-chatgpt-with-less-than-600/">Stanford Alpaca project</a>, which made the news for fine-tuning a copy of Meta’s 7-billion parameter LLaMa model for a compute cost of about $600, which Sims points out “didn’t include the cost of the eight Nvidia A100 GPUs, which cost about $16,000 each, five PhD students, and three faculty members for two months.”</p><h4>How AI Models Understand Text</h4><p>At their core, computers process information as numbers. To enable seamless communication between humans and AI models, text must be converted into numerical representations.</p><blockquote><strong>A Shout-Out to the Math Geeks (and to those of us who didn’t think math was that important)!</strong></blockquote><blockquote>Linear algebra, a branch of mathematics dealing with vectors and matrices, plays a critical role in this conversion process, and understanding its concepts can provide valuable insights into the inner workings of AI systems.</blockquote><blockquote>Linear algebra is essential because it forms the foundation of many advanced algorithms that empower AI systems to learn and process vast amounts of data. In the context of AI models like ChatGPT, the text is represented as high-dimensional vectors, which can be manipulated and analyzed using linear algebra techniques. These techniques help AI models capture the relationships, patterns, and structures within the data, allowing them to generate coherent and contextually relevant responses.</blockquote><blockquote>By understanding the role of linear algebra in the processing of text data, we can appreciate the complexity and elegance of the algorithms that enable AI models to learn from and generate human-like language. Moreover, this knowledge provides a solid foundation for further exploration into the fascinating world of machine learning and AI, especially when it comes to interacting with large language models.</blockquote><p>These numerical representations are called vectors. (Here’s looking at you, algebra!) Here’s what you need to know about vectors …</p><p><strong>Tokenization</strong>: Think of this as breaking down a sentence into individual words or smaller units. Tokenization helps the model digest and process the input text more effectively.</p><p>To illustrate the process, let’s take a simple example. If you ask ChatGPT about its favorite ice cream flavor, the text of your question might be, “What is your favorite ice cream flavor?” But to the LLM, this query is tokenized into individual words or subwords before it is processed. These tokens are then converted into vectors using embeddings, which allow the model to understand the meaning and relationships between the words. The model processes the input through multiple layers to create contextualized embeddings that capture the context of the question. If the question was part of a long conversation, it might be necessary to chunk the text into smaller parts and process them separately before combining the results.</p><blockquote><strong>Tokenizing a Sentence</strong></blockquote><blockquote>Let’s look at a real example of tokenization. If we wanted an LLM to understand the sentence, “ChatGPT is an impressive AI language model,” we would first have to tokenize it. (This is what happens under the hood when you interact with ChatGPT.)</blockquote><blockquote>After tokenization, the sentence might look like this:</blockquote><blockquote>[‘Chat’, ‘G’, ‘PT’, ‘ is’, ‘ an’, ‘ impress’, ‘ive’, ‘ AI’, ‘ language’, ‘ model’, ‘.’]</blockquote><blockquote>The next step is to convert these tokens into integer IDs based on GPT-3’s vocabulary. Suppose the integer IDs for these tokens are as follows:</blockquote><blockquote>[1203, 45, 908, 9, 47, 5862, 125, 64, 2531, 2965, 3]</blockquote><blockquote>Now, we’ll pass these integer IDs through the embedding layer, which maps each ID to a continuous vector representation. For simplicity, let’s assume that GPT-3 uses 10-dimensional embeddings (in reality, the embeddings would have much higher dimensions, such as 768 or more):</blockquote><blockquote>1203 -&gt; [-0.234, 0.687, 0.191, -0.625, 0.542, 0.982, -0.231, 0.102, -0.493, 0.014] 45 -&gt; [0.543, -0.932, 0.678, 0.041, -0.178, 0.213, 0.965, -0.437, 0.081, -0.326] 908 -&gt; [0.123, 0.452, -0.732, -0.518, 0.372, -0.625, 0.418, 0.153, 0.827, 0.106] … and so on for the other tokens.</blockquote><p><strong>Chunking</strong>: Sometimes, we need to process long sentences or paragraphs with the AI model. However, these models have a limit on the maximum length of text they can handle. To overcome this, we divide the text into smaller chunks, process them individually, and then combine the results. This helps manage the model’s memory constraints, breaking the data into pieces that can fit into the model’s context window. This allows the model to process the text sequentially, one chunk at a time. During chunking, it’s important to ensure that the division of text maintains coherence and does not break the flow of information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*9tfLXU5-ljwaaq1dYnWDaA.jpeg" /><figcaption>My favorite Chunk (from Goonies)</figcaption></figure><p><strong>Embeddings:</strong> Once we have the tokens, we need to convert them into numbers, or more specifically, vectors. These vectors, which can be imagined as points in a multi-dimensional space, represent the meaning and relationships between the tokens. It’s like teaching the AI model that “love” and “adore” are similar because their vectors are close together in this space.</p><p><strong>Contextualized embeddings:</strong> AI models like GPT use an architecture called the Transformer to process these vectors further. By doing so, the model can understand the context in which words are used. For instance, the word “bank” may have a different meaning when used with “river” than when used with “money”. Contextualized embeddings allow the model to differentiate between such nuances.</p><p><strong>Context Window:</strong> The context window is the maximum number of tokens (words or word pieces) that an AI model can process at once. This limitation is due to the model’s architecture, which includes a fixed number of layers and self-attention mechanisms. When the input text is within this limit, the model can process it in one pass. However, if the input text is longer than the context window, it needs to be broken down into smaller chunks.</p><p>Mitchell Troyanovsky, co-founder of <a href="https://www.getbasis.ai/">Basis</a>, an AI-powered knowledge worker system for finance and accounting, explained context windows this way:</p><blockquote>In theory if you had unlimited resources, you could have an LLM iterate over long contexts and continually summarize and check itself to get to the optimal answer. Imagine a human pulling data, writing a memo on it, and then 1,000 other humans at that company checking and revising its work. The final draft will definitely be better than the first, but the marginal value of each successive check will diminish. One of the challenges is figuring out the processes for summarizing and checking. This is certainly relevant with embeddings and any live company data being summarize.</blockquote><blockquote>For example, say you want an LLM to write a memo about how a firm has changed its policies on client engagement over the last couple years. Assuming you have all company documents embedded, when you search, odds are you’re not going to get the right info immediately — there are probably lots of internal documents with information on client engagement. So you need additional processes to figure out the correct search, check the results, and continue down that path. Finding the balance from a cost/time vs. answer quality perspective is key. Obviously one of the massive unlocks will be the context length increasing, allowing models to reason coherently over longer sets of resources. Another approach that is more involved but can also be more accurate is to train models directly on important documents.</blockquote><h4>Human vs. AI Text Processing: Key Differences</h4><p>The way humans process text is vastly different from how AI models like ChatGPT handle it. Human understanding of language is grounded in cognitive development, memory, emotions, and creativity. We learn language through a gradual process, absorbing meaning by interacting with other humans and with our environment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*TweMTwitlOcu96jvzU4KWQ.png" /><figcaption>Photo Credit: DreamStudio.ai</figcaption></figure><p>LLMs, conversely, acquire language through a technique called “unsupervised learning.” These models predict the next word in a sentence based on the context of previous words that it was trained on, allowing them to be exposed to numerous examples of language use, grammar, and world knowledge. This exposure helps them understand and generate human-like language.</p><p>But these models do have limitations. They lack the personal experiences, memory, emotions, and cognitive abilities that humans possess, which means they lack our rich understanding of context, figurative language, metaphors, and symbolism. These are all missing in the simple numerical representations that allow the models to recognize patterns and structures in the text. While this approach allows them to perform complex language tasks, it’s important to recognize that their understanding and processing of text differ significantly from the human experience.</p><p>Now that we have an understanding of how GPT reads and “understands” text, let’s look at some applications of this technology.</p><h4>Unlocking the Potential of GPT in the Business World</h4><p>By fine-tuning GPT on domain-specific data, customized AI solutions can be created for various industries:</p><p>Sims said:</p><blockquote>LLMs like ChatGPT are useful and interesting on their own because they can answer questions about known facts, summarize documents, including computer programs, and even synthesize new text and programs from known concepts. They can even use all of the tokens within their maximum context window size, including the user’s input and their own output, in generating their output. What they cannot do on their own is access external data or modify their internal knowledge representations with new information, and they would be much more useful if they could. Instead, there are several other methods which allow them to effectively access new data:</blockquote><blockquote>· Using a system like <a href="https://python.langchain.com/en/latest/">Langchain</a> to allow several LLMs to interact with each other and with external APIs. One excellent example of this is <a href="https://censusgpt.com/SF?s=how+many+people+live+in+san+francisco">San Francisco GPT</a>, which uses GPT3.5 to translate human-language questions into SQL queries, run them on public datasets, and return the results in text, tables, and graphics. The source code for this is on github, allowing anyone to clone and customize this for their own datasets.</blockquote><blockquote>· One other consideration that seems inevitable is the presentation of information. You will work on creating some new information (research, findings etc.) and then you’ll press a button and have a perfect memo/document formatted and written in your/your firm’s style. The collateral, powerpoint, etc. will already be drafted and tailored to your firm’s brand. You’ll have custom generated stock photos just for the collateral. If it requires marketing, you’ll have copy, etc. ready to go. So the amount of manual work in transforming information into polished output will go down dramatically. There are and will be MANY tools that do this. And this should all be happening in production for people at scale by 2024 I’d guess.</blockquote><p>Other tools that are already being built on top of the foundation of GPT include:</p><p><strong>Domain-specific knowledge bases:</strong> GPT could be trained on data related to a specific industry or domain, such as healthcare, law, or finance, to create specialized AI experts that can offer valuable insights and assistance.</p><p><strong>Personalized recommendation systems</strong>: By capitalizing on GPT’s understanding of user preferences or browsing history, we could develop personalized recommendation engines for products, articles, movies, or music (for example).</p><p><strong>Company-specific AI assistants:</strong> GPT could be trained on internal documents, processes, and knowledge bases to create AI assistants that can help employees with company-specific tasks, such as answering policy-related questions, assisting with onboarding, or providing guidance on internal tools and resources.</p><p>This technology is still in its nascent stage, but is moving exponentially faster than any previous technological innovation I can recall. While I’m not ready to go all in on any particular product or platform, we can all see the potential of LLMs and the tools that will be launched from these models as a foundation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/1*vSQQpYxUee0ynFdEV7ymQQ.jpeg" /><figcaption>AH-64 Apache Training Video</figcaption></figure><p>I firmly believe that the more powerful the tool, the more responsibility the user has to understand it. While we don’t need to take a course in mechanical engineering to use a wrench, I wouldn’t want to be in the vicinity of an untrained internerd with no more training than having recently watched “Fire Birds,” starring Nicholas Cage and Tommy Lee Jones, trying to take off in an Apache helicopter.</p><p>Today’s LLMs are the Apache helicopters of digital technology.</p><p>Tomorrow’s will be Falcon rockets.</p><p>If we’re going to be using these powerful tools now and in the future, we better start learning what they are and how they work.</p><p>Otherwise, I found a deal on another productivity tool I’d like to recommend …</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tlzoMgw5rgnSMFI0tlDG6A.jpeg" /><figcaption><a href="https://a.co/d/15DJR58">https://a.co/d/15DJR58</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7c7eb651ad06" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>