Image Credits: TechCrunch
It’s that time of year again: the week when startups from Y Combinator’s latest batch present their products to media — and investor — scrutiny. Over the next two days, a total of 217 firms will present, slightly smaller than last winter’s 235-firm group, as VC enthusiasm subsides a bit.
In the first half of 2023, VCs backed nearly 4,300 deals worth a total of $64.6 billion. That may sound like a lot. But deal value shows a 49% decline in H1 2022 while deal volume is down 35% year-on-year.
On a brighter note, one segment — driven by equal parts hype and demand — is vastly outperforming the rest: AI.
According to Crunchbase, nearly a fifth of the total global venture funding from August to July came from the AI sector. And the enthusiasm is evident in this summer’s Y Combinator cohort, with double the number of AI companies (57 vs. 28) compared to the winter 2022 batch.
To understand which AI technologies are driving investment these days, I dug deep into the summer 2023 batch, rounding up the YC-backed AI startups that I thought stood out the most — or showed the most promise.
AI infrastructure startups
Many of the startups in the Y Combinator cohort aren’t focused on what AI can achieve, but on the tools and infrastructure it needs. to tie A.I. from the ground up
For example, there is ShadeForm, which provides a platform to enable customers to access and deploy AI training and predictive workloads to any cloud provider. Founded by data engineers and distributed systems architects Ed Goode, Ronald Ding, and Zachary Warren, Shadeform aims to ensure AI jobs run on time and at “optimal cost.”
As Good notes in a blog post on the Y Combinator website, the explosion in demand for hardware to develop AI models, particularly GPUs, has created a shortage of capacity. (Microsoft recently warned of service disruptions if it doesn’t get enough AI chips for its data centers.) Smaller providers are coming online, but they don’t always deliver the most predictable resources—making them hard to outrun.
Shadeform solves this problem by letting customers start AI jobs anywhere in a public cloud infrastructure. Leveraging the platform, companies can manage GPU instances from a single pane of glass on each provider, configure “auto-reserve” when the machines they need are available or deploy in a server cluster with one click.
Image Credits: Shadeform
Another interesting Y Combinator startup tackling challenges in AI operations is Cerelyze, founded by former Peloton AI engineer Sarang Zambare. This is Cerelize Zambare’s second YC go-around after leading the AI team at cashier-less retail startup Capper.
Cerelyze takes AI research papers — typically found on open access archives like Arxiv.org — and translates the math they contain into functioning code. Why is it useful? Well, many papers describe AI techniques using formulas but don’t provide links to the code used to put them into practice. Developers are typically left to reverse engineer the methods described in the documentation to create working models and apps from them.
Cerelyze seeks to automate execution through a combination of AI models that understand language and code and a PDF parser “optimized for scientific content.” From a browser-based interface, users can upload a research paper, ask Serialize natural language questions about specific parts of the paper, create or modify code, and run the resulting code in a browser.
Now, Cerelyze cannot translate everything Code to paper – at least not in its current state. Zambre admits that the platform’s code translation currently only works for “a small subset of documents,” and that Cerealize can only extract and analyze equations and tables from documents, not figures. But I still think it’s a fascinating concept, and I hope it grows and improves over time — and is a worthwhile investment.
AI Dev Tools
Still developer-centric but not an AI infrastructure startup, Sweep autonomously handles smaller development tasks like high-level debugging and feature requests. The startup was launched this year by William Zheng and Kevin Lu, both veterans of the video-game-turned-social-network Roblox.
“As software engineers, we find ourselves shifting from exciting technical challenges to mundane tasks like writing tests, documentation, and refactors,” Zeng wrote on Y Combinator’s blog. “This was frustrating because we knew that larger language models (like OpenAI’s GPT-4) could handle this for us.”
A sweep can take code errors or GitHub issues and plan how to fix them, Zeng and Lu say — by writing code through pull requests and pushing it to GitHub. It can also address comments from code maintainers or owners on pull requests – a bit like GitHub Copilot but more autonomous.
“Sweep started when we realized that some software engineering tasks were so simple that we could automate the entire change,” Zeng said. “It does this by writing the entire project request with the sweep code.”
Given the AI’s propensity to make mistakes, I’m a little skeptical of the sweep’s reliability in the long run. Fortunately, Zeng and Lu are too — Sweep doesn’t automatically apply code fixes by default, which requires a human review and edit before pushing them to the master codebase.
AI apps
Moving away from the tooling subset of AI Y Combinator startups this year, we have Today, which bills itself as an “AI co-pilot for corporate event planning.”
Anna Sun and Amy Yan co-founded the company in early 2023. Sun was previously at Datadog, DoorDash and Amazon while Yan held various roles at Google, Meta and McKinsey.
Most of us haven’t had to plan a corporate event – certainly not this reporter. But Sun and Yan describe the test as difficult, unnecessarily tiring and expensive.
“Corporate event planners are bombarded with endless calls and emails when planning an event,” Sun writes in a Y Combinator blog post. “Strained by tight schedules, planners are paying for full-time assistants or tools that cost more than $100,000 a year.”
So, thought Surya and Yan, why not offload the most painful parts of the process to AI?
Enter Nowadays, which – providing upcoming event details (eg dates and number of attendees) – can automatically reach out to venues and vendors and manage relevant emails and phone calls. These days can also account for personal preferences surrounding events, such as amenities near a given location and activities within walking distance.
I should note that it’s not entirely clear how it works behind the scenes these days. Is AI really answering and placing phone calls and returning emails? Or are humans involved somewhere along the way — say for quality assurance? Your guess is as good as mine.
Nevertheless, it’s a very cool idea with a potentially large addressable market these days ($510.9 billion by 2030 according to Allied Market Research), and I’m excited to see where it goes next.
Image Credits: Nowadays
Another startup trying to abstract away traditionally manual processes is FleetWorks, the brainchild of former Uber Freight product manager Paul Singer and Quang Tran, who previously worked on Moonshot projects at Airbnb.
FleetWorks targets freight brokers – the essential intermediaries between shippers and carriers. Designed to sit alongside a broker’s phone, email and transportation management system (TMS), FleetWorks can automatically book and track loads and schedule appointments with shipping facilities that don’t have a booking portal.
Typically, loads that are not automatically tracked require brokers to reach out to drivers and shippers via phone or email for updates on shipment status. At the same time, they have to juggle calls from trucking companies interested in booking loads and negotiate prices, as well as set up appointment times for unscheduled loads.
Singer and Tran claim that FleetWorks can lighten the load by triggering calls and emails and pushing (no pun intended) all relevant information to a TMS or email. Apart from sharing load details, the platform can discuss pricing and book carriers, even call the driver and update the account team on cropping issues.
“FleetWorks helps freight operators focus on high-value work by automating routine calls and emails,” Singer wrote in a Y Combinator post. “Our AI-powered platform can leverage email or use a human-like voice to make tracking calls, cover loads and reschedule appointments.”
If it works as advertised, it sounds really useful.