<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/">
    <channel>
        <title>Yash Goenka</title>
        <link>https://www.yashgoenka.com/</link>
        <description>I’m a software engineer and entrepreneur based in San Francisco. I’m the co-founder and CEO of Humwork.</description>
        <lastBuildDate>Mon, 06 Jan 2025 00:00:00 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>https://github.com/jpmonette/feed</generator>
        <language>en</language>
        <image>
            <title>Yash Goenka</title>
            <url>https://www.yashgoenka.com/favicon.ico</url>
            <link>https://www.yashgoenka.com/</link>
        </image>
        <copyright>All rights reserved 2026, Yash Goenka</copyright>
        <item>
            <title><![CDATA[Did OpenAI's O1 decipher the Indus Valley Script?]]></title>
            <link>https://www.yashgoenka.com/articles/did-openais-o1-decipher-the-indus-valley-script</link>
            <guid isPermaLink="false">https://www.yashgoenka.com/articles/did-openais-o1-decipher-the-indus-valley-script</guid>
            <pubDate>Mon, 06 Jan 2025 00:00:00 GMT</pubDate>
            <description><![CDATA[I tried using AI to decode one of the world's oldest undeciphered languages. Here's what happened when I asked OpenAI's O1 to analyze the 4,000-year-old Indus Valley script.]]></description>
            <content:encoded><![CDATA[<p>A few weeks ago, I had a fascinating conversation with OpenAI's O1 model about decoding the Indus Valley script - one of the world's oldest and still undeciphered writing systems. What started as a curious experiment led to an intriguing demonstration of how large language models might contribute to archaeological and linguistic research.</p>
<p>Here is a pdf ofthe entire chat transcript: <a href="https://drive.google.com/file/d/1Y3XRtUpKOMEsWjxQgAIPBu-GR-NaWG40/view?usp=sharing">https://drive.google.com/file/d/1Y3XRtUpKOMEsWjxQgAIPBu-GR-NaWG40/view?usp=sharing</a></p>
<p><img src="https://www.yashgoenka.com/images/articles/did-openais-o1-decipher-the-indus-valley-script/harappan-stamp-seal-with-unicorn.png" alt="Harappan stamp seal with a unicorn motif and a line of Indus script"></p>
<blockquote>
<p>o1: “[Unicorn Clan] offering/trade measure of [grain commodity type] (possibly a standard unit), allocated for ritual or market exchange.”</p>
</blockquote>
<p>The Indus Valley civilization (c. 3300-1300 BCE) left behind thousands of seals with short inscriptions that have puzzled scholars for decades. Without a "Rosetta Stone" equivalent, these ancient texts remain silent witnesses to a sophisticated Bronze Age culture.</p>
<p>I began by showing O1 some seals and the script and asking it to analyze patterns and propose theories about their meaning. Its initial response was methodical, breaking down the analysis into distinct categories:</p>
<blockquote>
<p>"Sign Repetition and Symbol Clusters: The short strings of symbols on each seal typically feature a set of recurring motifs—forked shapes, plant-like icons, geometric forms, and simple line-and-dot patterns. The tight clustering and consistent size of these signs suggest that the script might be composed of a finite set of standardized symbols, possibly 300–400 in total, as researchers have cataloged."</p>
</blockquote>
<p>The model then discussed the directional patterns:</p>
<blockquote>
<p>"Direction of Writing and Syntax: Many seals show a row of symbols arranged in a neat linear order, often read right-to-left, though the exact reading direction remains debated. Consistency in direction and sign order strongly implies a language-based system."</p>
</blockquote>
<p>What caught my attention was how O1 connected the imagery to possible social functions:</p>
<blockquote>
<p>"Animal Imagery and Clan/Trade Identities: The presence of large central animals—like a bull, a unicorn-like figure, or sometimes an elephant—next to lines of script might serve as a form of 'brand' or emblem. One theory suggests that the animal images could represent clan totems, regional identities, or occupational guilds."</p>
</blockquote>
<p>When I asked it to attempt an actual decoding, O1 provided a detailed framework comparing the script to other ancient writing systems:</p>
<blockquote>
<p>"Like Linear B, we assume commodity + measure signs. Like Egyptian Hieroglyphs, we combine symbolic determinatives (a deity figure, offering stands) with conceptual signs (grain, clan emblems). Like Mayan script, layered sign clusters may represent titles, place-names, and ritual actions."</p>
</blockquote>
<p><img src="https://www.yashgoenka.com/images/articles/did-openais-o1-decipher-the-indus-valley-script/longest_indus_script_inscription.jpg" alt="The longest known Indus script inscription"></p>
<p>The most ambitious part came when I challenged it to decode the longest known Indus inscription. O1 produced a systematic analysis, breaking down the text line by line and proposing a complete translation:</p>
<blockquote>
<p>"By the authority of the honored Unicorn-Bull Clan, and under the sign of the orchard, we present these sacred measures of grain. In threefold units, offered at the season of celestial favor, we petition the divine powers to grant fertility and prosperity. By the decree of priestly lineage and in the presence of the seated divine master, let this covenant ensure abundant harvests, thriving herds, and lasting well-being for our people."</p>
</blockquote>
<p>What's particularly interesting is how O1 created a complete hypothetical sign catalog, assigning specific meanings to individual symbols. For example:</p>
<blockquote>
<p>"Sign 332†: Description: A forked shape with three prongs, slightly curved. Phonetic: 'ta' Semantic Core: Denotes a 'division' or 'section.' May mark subdivisions of a commodity or a clan subgroup."</p>
</blockquote>
<p><img src="https://www.yashgoenka.com/images/articles/did-openais-o1-decipher-the-indus-valley-script/indus_seal_and_script.jpeg" alt="An Indus Valley seal and its script"></p>
<p>What makes this experiment particularly intriguing is not just the comprehensiveness of O1's analysis, but its methodological approach. The model didn't just make wild guesses - it constructed a coherent system based on known principles of ancient writing systems and archaeological context.</p>
<p>Consider how it broke down the signs into functional categories:</p>
<blockquote>
<p>"Logograms: Most oval or container-like signs represent commodities (grain, oil, processed foods). Phonetics: Simple shapes (forked or U-shaped) often serve as phonetic syllables, letting scribes spell out clan names, deities, or place-names. Determinatives: Forked signs, ovals with distinct internal patterns, and leaf-like icons classify words into categories—agricultural produce, ritual offerings, settlements, or alliances."</p>
</blockquote>
<p>The model's proposed "Rosetta Stone" for the script included detailed entries for over 70 signs, each with proposed phonetic values, semantic meanings, and contextual usage patterns. This systematic approach mirrors how actual ancient scripts were deciphered - through painstaking analysis of patterns, contexts, and comparative linguistics.</p>
<p>Of course, we should approach this with appropriate skepticism. I wish I had deeper expertise in archaeological linguistics to properly evaluate the technical validity of O1's proposals. The model itself acknowledges this uncertainty:</p>
<blockquote>
<p>"This reconstruction is speculative, as no definitive readings exist. However, based on pattern recognition, common motifs, and analogy with other ancient scripts, a plausible reading is that this seal records the clan's name (indicated by the unicorn) and a standard measure of a key commodity (grain or similar), likely related to a ritual or trade context."
What's particularly fascinating is how O1 integrated multiple lines of evidence - the physical characteristics of the seals, the archaeological context, comparative analysis with other ancient scripts, and likely social functions within the Indus Valley civilization. This holistic approach feels reminiscent of how human scholars actually work.</p>
</blockquote>
<p>Unfortunately, O1's full analysis with images cannot be easily shared due to ChatGPT not letting you share threads with images in it. But here is a screenshot pdf of the entire chat transcript: <a href="https://drive.google.com/file/d/1Y3XRtUpKOMEsWjxQgAIPBu-GR-NaWG40/view?usp=sharing">https://drive.google.com/file/d/1Y3XRtUpKOMEsWjxQgAIPBu-GR-NaWG40/view?usp=sharing</a></p>
<p>The implications of this experiment extend beyond just the Indus script. It suggests that large language models, when properly prompted, might offer new perspectives on old problems in historical linguistics and archaeology. Not by magically solving them, but by systematically exploring possibility spaces in ways that might spark new insights among human researchers.</p>
<p>This doesn't mean O1 has "cracked" the Indus script (idk maybe it did but one of you need to verify that). What it has done is demonstrate how AI might serve as a sophisticated thought partner in academic research, capable of generating structured hypotheses that human experts could then evaluate, refine, or reject.</p>
<p>The experiment also reveals something about the nature of these models themselves. O1 didn't just pattern match or regurgitate existing theories - it constructed a complex, internally consistent system that respected both linguistic principles and archaeological evidence. This suggests these models might be capable of more sophisticated forms of reasoning than we sometimes assume.</p>
<p>The Indus Valley script likely remains undeciphered, but perhaps the path to understanding it - and other ancient mysteries - will involve this kind of human-AI collaboration, combining machine learning's pattern-recognition capabilities with human expertise and judgment.</p>
<p>Or perhaps I'm getting ahead of myself. Maybe this was just an interesting afternoon spent exploring an ancient puzzle with a very sophisticated chatbot. Either way, it suggests intriguing possibilities for how AI might contribute to our understanding of the past.</p>
<p>P.S. I'm not a professional archaeologist or linguist. I just like to play with LLMs. Can someone please try this with o1 pro?</p>]]></content:encoded>
            <author>Yash Goenka</author>
        </item>
        <item>
            <title><![CDATA[The Rise of AI Phone Agents]]></title>
            <link>https://www.yashgoenka.com/articles/the-rise-of-ai-phone-agents</link>
            <guid isPermaLink="false">https://www.yashgoenka.com/articles/the-rise-of-ai-phone-agents</guid>
            <pubDate>Mon, 16 Sep 2024 00:00:00 GMT</pubDate>
            <description><![CDATA[Exploring the transformative potential of AI-powered phone agents and how they are reshaping business communication, with a focus on phonecall.bot.]]></description>
            <content:encoded><![CDATA[<p>In the ever-evolving landscape of technology, we often find ourselves at the cusp of paradigm shifts that fundamentally alter how we interact with the world around us. Today, we stand at one such precipice: the advent of AI-powered phone agents. This isn't just another incremental improvement in automation; it's a leap that promises to reshape the very fabric of business communication.</p>
<p>Consider, for a moment, the last time you called a business. Perhaps you were trying to schedule an appointment, inquire about a product, or seek support for a service. Chances are, your experience fell into one of two categories: either you navigated a labyrinthine phone tree, pressing buttons in hopes of reaching a human, or you waited on hold, listening to muzak punctuated by assurances that your call is important. Neither scenario is particularly efficient or satisfying.</p>
<p>Now, imagine a world where every call is answered instantly, by an agent that understands natural language, can access relevant information in milliseconds, and never gets tired or frustrated. This isn't science fiction; it's the reality that <a href="https://www.phonecall.bot">phonecall.bot</a> is ushering in.</p>
<p>At its core, phonecall.bot is doing something that might seem simple: automating phone calls. But the implications of this are far-reaching and profound. It's not just about replacing human agents with robots; it's about fundamentally redefining what's possible in terms of scale, efficiency, and customer experience.</p>
<p>One of the most striking aspects of this technology is its ability to scale infinitely. Traditional call centers are constrained by physical limitations: the number of agents you can hire, train, and manage; the space needed to house them; the infrastructure required to support their work. With AI phone agents, these constraints evaporate. Whether you're handling ten calls a day or ten thousand, the system scales seamlessly, without any degradation in quality or response time.</p>
<p>But scale alone isn't enough. What truly sets this technology apart is its ability to provide a consistently high-quality experience. Human agents, no matter how well-trained, are subject to fatigue, mood swings, and inconsistency. An AI agent, on the other hand, brings the same level of enthusiasm and competence to every interaction, whether it's the first call of the day or the thousandth.</p>
<p>The implications for businesses are staggering. Imagine never missing a call, never having a customer wait on hold, never losing a potential sale because all your agents were busy. Imagine being able to offer 24/7 support without the overhead of multiple shifts or international call centers. Imagine having perfect recall of every conversation, effortlessly integrated with your CRM and other business systems.</p>
<p>But perhaps the most exciting aspect of this technology is its potential for customization and improvement. Unlike traditional IVR systems, which are rigid and limited, AI phone agents can be trained on specific business knowledge, updated in real-time, and even personalized for individual customers. They can learn from each interaction, continuously improving their responses and understanding.</p>
<p>Of course, as with any transformative technology, there are challenges and concerns to address. Privacy and data security are paramount, especially when dealing with sensitive customer information. There's also the question of job displacement – a valid concern, but one that has accompanied every major technological shift in history. As with previous revolutions, this technology is likely to eliminate certain roles while creating new ones we can't yet imagine.</p>
<p>There's also the matter of the uncanny valley – that uncomfortable space where something is almost, but not quite, human. Early iterations of AI phone agents might fall into this trap, potentially alienating some users. But if the trajectory of other AI technologies is any indication, we'll quickly move past this phase into something that feels natural and intuitive.</p>
<p>The key to navigating these challenges lies in thoughtful implementation and clear communication. Businesses adopting this technology need to be transparent about when customers are interacting with AI, provide easy pathways to human agents when needed, and ensure that the AI is augmenting, not replacing, human capabilities.</p>
<p>As we look to the future, it's clear that AI phone agents are just the beginning. They represent a fundamental shift in how we think about communication, customer service, and the relationship between businesses and their customers. We're moving towards a world where every interaction can be instantaneous, personalized, and highly efficient.</p>
<p>But perhaps the most exciting aspect is what we can't yet foresee. Just as the smartphone unleashed a wave of innovation that went far beyond making calls, AI phone agents are likely to spawn entirely new categories of products and services. We might see AI agents that act as personal assistants, scheduling appointments and managing our daily lives. Or AI-powered education systems that provide personalized tutoring over the phone. The possibilities are limited only by our imagination.</p>
<p>In the end, the rise of AI phone agents isn't just about making business communications more efficient. It's about freeing human creativity and potential. By automating routine interactions, we create space for more meaningful, high-value engagements. We allow businesses to focus on innovation and customer relationships, rather than the logistics of managing a call center.</p>
<p>As we stand on this technological frontier, it's natural to feel a mix of excitement and trepidation. But if history has taught us anything, it's that the most transformative technologies are those that amplify our human capabilities, rather than replace them. AI phone agents have the potential to do just that – to handle the routine so we can focus on the exceptional.</p>
<p>The future of communication is calling. It's up to us to answer.</p>]]></content:encoded>
            <author>Yash Goenka</author>
        </item>
        <item>
            <title><![CDATA[Introducing Jarvys]]></title>
            <link>https://www.yashgoenka.com/articles/introducing-jarvys</link>
            <guid isPermaLink="false">https://www.yashgoenka.com/articles/introducing-jarvys</guid>
            <pubDate>Wed, 15 Feb 2023 00:00:00 GMT</pubDate>
            <description><![CDATA[Introducing Jarvys - an AI work assistant that allows you to directly tell your computer to do a task, rather than doing it by hand. It is powered by ChatGPT and is generalised to work across any website.]]></description>
            <content:encoded><![CDATA[<p>I first watched Iron Man back in 2008. The Iron Man suit blew my mind. However, one of the tools used by Tony Stark particularly stood out to me. This was J.A.R.V.I.S.
In the movies, Tony Stark would directly tell Jarvis to do something like "run a simulation and update me on the structural and compositional analysis of this element" and Jarvis would just do it.
Ever since I have always aspired for an assistant like this on my computer that would just understand what I wanted it to do.
Given the recent advancement in Large Language Models and specifically the release of GPT-3 by OpenAI, my cofounder Rohan and I,
were hoping for someone to build an assistant like Jarvis. No one did. So we just built it ourselves.</p>
<p>Introducing Jarvys - an AI work assistant that allows you to directly tell your computer to do a task, rather than doing it by hand.
Just give it a high level request and Jarvys does the rest.</p>
<p>We belive that Natural Language Interfaces are the future of computing. AI models are now good enough to understand our language
and it's time to use them to interact with computers. Such interfaces can be powerfull as they speed up tasks that would otherwise require
a lot of clicks and navigation. This will 10X our productivity and free up time for things we truly care about.</p>
<h2>How does it work?</h2>
<p>Jarvys uses GPT-3 by OpenAI, the same model behind ChatGPT. It is currently a Chrome Extension that can see what is happening in your browser
and take actions like clicking, navigating and typing. It is generalized to work on any website and it's action space is the UI elements
available on the page.</p>
<h2>Future</h2>
<p>A lot more work is left to be done. We want to implement better prompt chaining, better rendering of the DOM, embeddings and reduce latency (the gifs/videos you see have been sped up).
We are getting feedback from our early test users for all the actions that Jarvys takes in order to implement Reinforcement Learning from Human Feedback (RLHF). We hope that this, along with
improvements in the foundation models themselves will make Jarvys a truly universal AI work assitant capable of accompishing
any task.</p>
<p>Sign up at <a href="http://www.jarvys.ai">www.jarvys.ai</a> and please do reach out to us at <a href="mailto:hello@jarvys.ai">hello@jarvys.ai</a> if you would like to work with us.</p>]]></content:encoded>
            <author>Yash Goenka</author>
        </item>
    </channel>
</rss>