shyamal's space


Applied AI @ OpenAI • Startups GTM • On Deck Fellow • Proud Son • Duke + Wisconsin Alum • Building for impact • Venture Scout • Neo Mentor • Duke AI Advisory Board

15 March 2024

frontiers, knowledge work, 2024++

by Shyamal Anadkat

Share on:


Illustration: Shyamal × DALL·E

At the time of writing this, gpt-4 marked its first anniversary. Simultaneously, the world witnessed the demo of the first AI software engineer by Cognition Labs. It was incredible to see. I’m genuinely humbled by gpt-4’s impact - how it shifted worldviews and landscapes, sparking new opportunities for developers, startups, students, and economies globally while driving critical progress on AI alignment and safety. We realized intelligence is almost an emergent property of matter. Language models have become more intelligent, multi-modal, faster, fine-tunable, interpretable, open-source, safer, and in some cases, well, meme-worthy. In less than two years, we’ve gone from the days of 4k context-length to 1M context-length LLMs. Deep learning advancements are not slowing down anytime soon, and scaling laws continue to hold. The neural architecture + model capabilities might just be good enough to help us evolve from the AI assistants/copilots that helped us write some working code and summarize email drafts to the ones doing the tasks that probably take us 10 minutes or longer - like updating your website, completing a driver’s license application, or drafting your tax returns.

2024 will be the year for startups that are redefining how we think and work. Founders thinking deeply about knowledge work from the first principle basis and ones chasing to accelerate that vision by deploying knowledge work assistants are worth taking a bet on. I would bucket these startups roughly into three categories - (1) ones that help distribute knowledge, (2) ones that build tools for creating new knowledge, and (3) the ones developing copilots/assistants for knowledge work. We’re realizing that shipping AI without calling it AI is underappreciated. What’s interesting about that is that it brings back focus on users and value; no hype that’s driven by an artificial AI strategy, just utility that users gradually rely on.

The knowledge work x AI momentum is very real. Just a few weeks back, we saw Klarna announce its AI assistant which has had 2.3 million conversations, two-thirds of Klarna’s customer service chat, doing the equivalent work of 700 full-time agents (let that sink in) and is on par with human agents regarding customer satisfaction (CSAT) score. More importantly, customers now resolve their errands in less than 2 mins compared to 11 mins previously. This is wild. Prosumer applications in the general/ knowledge search and discovery have seen impressive adoption. Perplexity for example, is reimagining search (redesigning what search originally should be) and is creating value by offering real-time web searching, source integration, human-like explanations, advanced summarization, and citation for source verification. Unlike traditional search engines, they focus on understanding user intent and providing accurate and contextually relevant answers by analyzing queries and user interactions. This is going to change how researchers, analysts, and clinicians find information and create research plans. Personalization and agent-based browsing is going to be the next frontier here. ChatGPT and its enterprise offering (disclaimer: I work at OpenAI), on the other hand, is building a truly personal work assistant that will revolutionize work. It’s akin to an iron(wo)man suit for every employee. Companies adopting ChatGPT enterprise have been building customGPTs/assistants across functions in marketing, sales, finance, ops, R&D, and HR. And these are very early days. ChatGPT admin will be quite a fun role (actually, we’ll have a GPT for that too).

On a day-to-day basis, knowledge work assistants (KWAs) will eventually automate more of the routine knowledge tasks, such as data entry, document processing, and information retrieval. By offloading these tedious tasks, hopefully, we can focus more on higher-value activities that require human judgment, creativity, and critical thinking. Glean is building solutions for enterprise search and knowledge discovery. Julius is an AI Data Scientist that can analyze and visualize massive datasets, perform complex analyses like forecasting and regression, and even train ML models. Superhuman is building an assistant for emails. Class companion is building an AI assistant to help teachers give instant, personalized feedback on written assignments. Canva has awed us with the virtual design partner that can create custom, on-brand, and attention-grabbing designs in seconds. Glass Health is empowering clinicians with their AI platform for developing differential diagnoses and drafting clinical plans. Abridge and Deepscribe are building AI scribes that transform patient-clinician conversations into structured clinical notes in real-time using custom models. These tools aim to assist and transform knowledge work in sectors ranging from education to law and medicine, enhancing cognitive abilities, and learning effectiveness, and reducing cognitive overload from information overload. AI will not replace human workers but augment their capabilities, enabling them to work smarter and more efficiently. On the enterprise front, tools like Salesforce Einstein Copilot are being used to truly understand customer data and relationships, augmenting the work of GTM teams in various industries; one of Salesforce’s customer quotes: Einstein has helped our agents be more efficient and confident in their work, without losing the human connection we pride ourselves on as a company. Harvey is revolutionizing legal with AI (and I hope they expand to more professional services over time). In their own words, Large language models have the potential to give knowledge workers superpowers by solving complex tasks efficiently and accurately. The best part: these solutions don’t leave knowledge workers without work; they allow them to focus on the quintessential aspects of their professions—strategy, advice, and judgment—the exact work that makes the long hours, in school and at the job, worthwhile.

On the software engineering front, while Devin has had success passing practical engineering interviews from leading AI companies, Github copilot and Anysphere are integrating AI into coding workflows, assisting developers with code suggestions and automation, and streamlining the programming process. In Cursor’s own words - while Github Copilot is tremendously helpful for eliminating low-entropy keystrokes while writing new code, it does not help you save low-entropy keystrokes when you need to make small, simple changes to existing code blocks. The biggest successes have not just come from innovations on the model side but even more difficult decisions and innovations on the UX. Building a really good AI software engineer is a hard UX problem and without the right UX could fail - even with the most capable models and sophisticated tree of thought / chain of thought++ prompting. Often the best solutions hide complexity behind simple interfaces that provide long-term value.

For startups in this space, the focus should be on building AI that bridges the knowledge gap between experts and non-experts. Imagine the assistants that can not only automate tasks but also translate complex knowledge into actionable insights for everyday knowledge workers. eg: AI tutors that personalize training for specific domains, or AI co-pilots that assist knowledge workers in real-time during decision-making processes. By democratizing access to expertise, these solutions have the potential to revolutionize how we live and work. Let’s briefly talk about the impact this will have on the economy. I’ll offer one perspective which is that automating knowledge work tasks will certainly catalyze a shift from a knowledge-based economy towards a resource-based economy (I expect this shift to happen entirely in 2035). Here’s how this transition might unfold over the next 5-7 years:

1/ democratization of knowledge: As AI systems/agents become more capable of performing various knowledge tasks, such as research, software engineering, analysis, and content creation, the barriers to accessing and utilizing knowledge will be lowered. Knowledge, which was previously a scarce and valuable resource concentrated among highly skilled professionals, will become more widely available and democratized.

2/ shift towards resource optimization: As knowledge becomes more readily available and certain knowledge-based tasks become commoditized, the competitive advantage will shift towards the ability to optimally allocate and utilize resources. Companies and individuals will need to focus on identifying and leveraging the most valuable resources, whether they are physical resources (e.g., raw materials, energy), human resources, or technological resources (e.g., chips, data).

3/ rise of the “resource economy”: In the shift towards a resource economy driven by AI and knowledge work, success will be determined by the ability to acquire, manage, and effectively utilize scarce and valuable resources. Individuals will increasingly take on the role of managers or orchestrators of AI agents and systems. We will need to develop expertise in task decomposition, breaking down complex problems into smaller, more manageable tasks that can be efficiently handled by AI agents. We’ll also need to learn how to effectively monitor and evaluate the performance of AI agents, identify areas for improvement, and provide feedback or adjust parameters as needed. Companies that can efficiently integrate AI with human expertise and other resources will gain a competitive edge.

The question then is “If generative AI makes knowledge workers more productive in finite bursts, it may simply raise expectations and pile on ever more work to fill the time savings. What should I do about it?” And for those whose core competence involves activities an AI can now match or exceed, career prospects will dim as the value of their services is driven down by fierce competition from tireless digital laborers. The wisest path for knowledge workers may be to quickly become experts at using AI tools themselves to amplify their skills, to stake out new high ground in roles that make the most of human strengths like empathy, strategy, and managing ambiguity. It’s also reasonable to say that the companies that will thrive in this AI-powered future will be those that encourage relentless experimentation to discover the ideal division of labor between human and machine intelligence. The mediocre companies will blindly add AI tools without rethinking workflows, only to find their productivity gains erased by rising costs and a workforce paralyzed by status anxiety.

The whole point is that as AI takes over more cognitive tasks, it will democratize access to knowledge and free up human beings to focus on what truly matters – creativity, innovation, and solving the world’s most pressing challenges. This transition towards a resource economy is not something to fear but to embrace. Its paradigm shift will usher in a new era of abundance, where scarcity is no longer a limiting factor.

In the next phase of building KWAs, simply having capable AI models is not enough - the real challenge lies in designing complete user experiences + product workflows that truly deliver sustained value to knowledge workers. The key will be developing “AI-first” interfaces, editing experiences, and agentic systems that make AI a complementary tool enhancing human capabilities rather than just an impressive but fleeting technology demo. Companies solving these UX challenges around prompt engineering, model customization, and embedding AI assistants into productive workflows will be the winners. Another key thing to keep in mind is that the trojan horse for AI is not anthropomorphism but intelligence augmentation. Rather than building agents/integrations/KWAs that aim to pass as humans, the best builders will create tools that feel like natural extensions of our abilities - leveraging computation to enhance what humans already do well. And this is what will help unlock the full potential of human ingenuity and create a better future for all.

tags: Startups - AI - Large Language Models - Knowledge Work - AGI