New Silicon Valley AI Startup Trends from YC W24: AI Agent, Verticals, Multimodality and AI Security

Article source:Founder Park

Image credit: Generated by Boundless AI

On April 4th, Y Combinator W2024 Batch Demo Day officially started. A total of 260 projects were unveiled this time, and YC sifted through 27,000 applications, with a pass rate of less than 1%, which wasOne of the lowest percentage of YC admissions for a batch in historyThe

More than 60% of them are AI-related. YC is located in Silicon Valley, which is at the forefront of technology, and is deeply connected to the core ecosystem of the AI technology revolution, gathering a large number of early-stage projects and outstanding entrepreneurs in a startup incubator model, with the breadth and cutting-edge coverage of the projects being second to none.

Y Combinator is one of Silicon Valley's most prominent startup gas pedals, and with two acceptance cohorts each year, Winter (W) and Summer (S), YC has a strong alumni network and branding advantage.
Since its founding by Paul Graham in 2005, YC has become one of the most influential and successful startup incubators in the world, incubating companies such as Airbnb, Dropbox, and Reddit, among others. the YC community is now home to more than 4,500 startups and 11,000 founders.

In the selected projects of this issue, we see many landing scenes and imaginative applications of AI; compared with last year, we can obviously feel that AI landing is accelerating.More and more entrepreneurs are trying to unlock a new 'way to open' AI. While the projects themselves won't necessarily become great companies in the future, they bring a lot of inspiration for exploring AI startup opportunities. We've summarized the latest trends and eye-opening projects from this issue's YC inductees to share with you.

01 AI creates new scenarios that

Self-training models still require

The long tail effect is extending

In addition to the few classic application scenarios in the header, one trend that can be seen is a broader distribution of products by industry. There are about 35 company categories in this batch of YC projects, compared to 28 previously (new pendant scenarios include HR, Recruiting, and Aerospace). "AI isn't a silver bullet" ("AI isn't a silver bullet"), and as the wave of GenAI calms down, domain expertise will still be needed to engage users and solve problems.

But it's also clear that as AI technology becomes more prevalent, we're creating new problems. It's interesting to see companies in this group that are focused on AI security - one company is working on fraud and deep forgery detection, while another is building foundational models that are easy to align. We'll likely continue to see more companies tackling the new scenarios required for new AI capabilities.

AI applications have become more diverse. In the last batch, the main category of products was 'programming assistants' for programmers. While these still exist in the current batch, there are also more companies building 'native AI' products and platforms - software that uses AI in a way that simply talks to an AI assistant's sidebar.

Artificial intelligence infrastructure continues to mature

Six months ago, Charlie mentioned a couple of points: the size of the AI+Olps field also demonstrates the amount of work required to really bring large language models and other models to production. There are still many unanswered questions about reliability, privacy, observability, usability, and security when using LLMs externally.

Even today.AI Ops is also one of the most popular categories. While some companies are offering services we've seen before, many are also helping to SaaS-ize technologies at the forefront of AI product development. For example, until last year, RAG, or Retrieval Augmented Generation, was a little-known term outside of AI research circles, but now multiple companies are building 'RAG as a Service'.

Best practices for LLMs deployment include curating training data, running regular evaluations, and testing vector block sizes - but there are not many industry-standard tools to perform these operations in production. Developers are investigating effective ways to mitigate the illusion. Part of the reason is the technology itself - it's hard to find a solid foothold when state-of-the-art technology changes significantly every 3-6 months. Amazon S3 and EC2 are now 18 years old, while GPT-3 has been around for less than four years.

Self-training models still require

With companies like OpenAI and Anthropic focusing on general-purpose base models, it is easy to assume that newer companies are primarily building 'shell GPTs'. However, as can be seen with the YC W24 cohort, there is still a need to have companies train models from scratch. Here are some of the models that have been trained on their own for new use cases:

• Diffuse Bio: a model for designing new proteins for vaccines and other drugs.
• Infinity AI: a model for generating short video clips based on a given script.
• Piramidal: a model trained to understand brain activity based on electroencephalography (EEG) data.
• SevnAI: a model for graphic design aimed at creating editable vector graphics.
• Sonauto: a model for creating hit songs based on lyrics and short cues.
• Yoneda Labs: a model for optimizing chemical reactions.

02

AI Agent, Multimodal, Verticals

and AI security are hot trends

From the distribution of W24 projects, B2B SaaS still dominates the application layer, with B2B SaaS projects accounting for more than 60%, and C-end consumer products accounting for 11%. Although C-end consumers only account for 11%, Jared Friedman, a partner at YC, said, "The percentage of B2B SaaS projects in the application layer is still very high, and the percentage of B2B SaaS projects in the application layer is still very high," he said.One of the major trends of this YC is the wave of consumer AI companies"AI brings more imagination to C-suite products.

Field Distribution:

• 65% : B2B SaaS/Enterprise
• 11%: Consumer products
• 10%: Health care
• 8%: Financial Technology
• 4%: Industrial
• 1%: Government Science and Technology
• 1%: Education Technology

This year, not only has there been a further increase in the number and percentage of AI projects, but there has also been a significant increase in the number of application layer projects and vertical areas of landing. Obviously, the application landing of AI is accelerating. From these projects, we see several new trends in the development of AI applications:

AI Agent

• AI Agents are starting to move from Task to JobConceptual products for AI Workers have emerged, and while there are still significant challenges, they have moved away from simple tools towards Autonomous Agents that can work independently.
• Construction of domain knowledge bases, moving from simple RAG techniques to learning and use of knowledgeThe core expression of knowledge has also changed from embedding to natural language. The core expression of knowledge has also changed from embedding to natural language, and vector databases will eventually be just a retrieval acceleration technique.

vertical area

• Companion Chat Apps Move Toward Professional Mental and Emotional Health, from exploiting human weaknesses to get traffic to actually being helpful to the health of the human mind.
• Dev Agent is trying to move from one-off code generation to real software engineering projects, but still faces big challengesThis issue saw the emergence of some of the more cutting-edge ideas in software development. Some of the more cutting-edge ideas in software development emerged from this phase of the project, such as turning UI designs on Figma directly into code, semantic understanding at the chat to repo code set level, and support for an AI development model centered on documentation (rather than code).
• Emergence of applications for government regulatory/approval processes. This scenario is perfect for AI Agents and is also very relevant for the Chinese market.
• There's an explosion of AI adoption in Healthcare.. Some high-value scenarios are in focus, and a huge change in the medical industry by technology is already on the way. However, due to the obvious differences between the systems and business models of the medical industry in China and the United States, many entrepreneurial ideas may not be suitable for the country.

multimodal

• Speech technology has matured and may enter practical applications faster than image and video modalitiesThe
• Image/video generation has moved from modeling techniques to scripted storytelling and scenario description, returning to the essence of video applications.. Taking advantage of multimodality, there will be many changes in gaming, film and education.

AI Security and Assessment

• LLM and Agent reviews will be a huge potential market.

This session of YC lets us see that entrepreneurs' understanding and application of AI technology is becoming more and more mature, and a large number of application scenarios have emerged. It is foreseeable that more scenarios will break out in the next one or two years, and the era of blossoming AI applications that people are looking forward to may come.

03

New directions and projects to watch

The following projects stood out to us when we researched them, and many of them are interesting ideas and represent the latest trends in AI applications.

YC official website has the details of all the shortlisted projects this time: https://www.ycombinator.com/launches?batch=W2024&tag=Artificial+Intelligence&sort=date

AI Agent

From AI Assistant to AI Worker

The most important feature of the AI Worker project seen in this issue is:take the initiativeFrom the clinic front desk proactively calling patients to physician assistants proactively monitoring patient behavior during treatment; from recruiting assistants proactively searching and completing the first phone screening interview to Sales proactively calling potential clients. From clinic receptionists proactively calling patients to physician assistants proactively conducting behavioral monitoring during patient treatment; from recruiting assistants proactively searching for resumes and completing the first phone screening interview to Sales proactively calling potential clients.

The scope of Agent's behavior has gradually expanded from passively completing a certain task to actively completing a complete role scene closure, and this expansion from Task to Job is a key step in the development of Agent from AI Assistant to AI Worker.

With this step, the Agent is not limited to following human hard-coded instructions and workflows (SOPs), but is now ready to step into a wider world. Of course, free agents are still waiting for more efficient memory mechanisms, interactive learning, and stronger reasoning capabilities from LLMs. But no matter how fast or slow these technologies develop, we will see more and more AI Workers enter daily work and life.

Digital Workforce Agent

• HR Recruiting : Parasale - AI Recruiter: Find and Engage Candidates on Autopilot (https://parasale.io)
• Sales Agent: Ava, The Sales Rep Artisan - The AI Employee Who Automates Cold Email (https://artisan.co/)

Construction of domain knowledge bases, moving from simple RAG techniques to learning and use of knowledge

The magic of LLM lies in the fact that it successfully encodes a vast amount of general knowledge through a pre-train process and is able to apply this knowledge freely through natural language interaction. However, no matter how perfect this general knowledge is, it cannot cover the application domain specific knowledge. Therefore, how an agent builds its own domain knowledge base, how to use this knowledge effectively, and how to learn and accumulate knowledge through interaction are always the fundamental issues for agents, and also the most cutting-edge research topics.

This issue of YC has some very interesting projects exploring Agent practices for knowledge discovery, construction and learning. The construction and interaction of domain knowledge bases is moving from simple RAG technology to knowledge learning and use. The core expression of knowledge has also changed from embedding to natural language, and vector database will eventually be just a retrieval acceleration technology.

Knowledge building and learning

• E-commerce: Lumona - Skincare product search with Reddit and YouTube reviews (https://www.lumona.ai)
• Multimodal text comprehension: Reducto - Unlocking data behind complex documents (https://reducto.ai)
• Research Assistant: Lumina - help researchers build a knowledge base in minutes (https://lumina-chat.com)
• Data processing: Trellis - Make your unstructured data SQL ready (https://runtrellis.com/)
• Enterprise Search: Danswer - Open Source Unified Search (https://www.danswer.ai/)

Interesting vertical applications

Chatbot-from companion chat to mental health

After multiple warnings from Sam Altman and two major releases of the OpenAI app platform drowning out shelled apps, the various shelled big-model chat projects from early last year faded away, and the focus of startups gradually moved away from general-purpose companionship to mental health/emotional-specific concerns.

This issue focuses on several projects including: mental health treatment, partner relationship improvement, and health monitoring of patient behavioral patterns. These projects focus on the professionalism of the scenarios, with Agents actively initiating interactions, moving from passive companion chats to proactive targeted chats, and moving from exploiting human weaknesses to gain traffic to truly helping the health of the human mind.

Mental Health Agent

• Sonia - AI mental health therapy (https://soniahealth.com)
• Maia - Transforming relationships using AI (https://www.ourmaia.com/)
• Attunement - Patient monitoring and treatment recommendation for better behavioral health (http://attunement.ai)

Dev Agent begins to move toward real software engineering projects, but still faces big challenges

In this round, LLM gets the breakthrough of reasoning ability from code corpus, which is also the most core ability of AI. Just as Procedural Memory is the home of all human skills, learning any skill or task can only be precipitated into Process Memory in the end in order to obtain the ability of high determinism, low cost, and high degree of parallelism. Similarly, for Agent, any task learning can only be efficiently and reliably operated if it is eventually turned into code.

While there is a steady stream of stellar projects claiming to have developed an AI Programmer or Dev Agent that have caught on with the market and the general public, so far none of them have lived up to expectations. The number of tasks that can be taken on is growing, but none of them have yet addressed the incremental development of large-scale existing code sets, and even the general tasks of refactoring or bug fixing that don't affect the behavior of the software are far from being accomplished, which is a testament to the reality of the difficulty of this goal.

Perhaps it will take another step up in LLM's reasoning ability to make a breakthrough. The projects in this issue of YC should reflect some of the more cutting-edge ideas in AI application to software development, such as: turning UI design on Figma directly into code and supporting updates, chat to repo semantic understanding at the code set level, and support for AI development models centered on documentation (rather than code).

Software Development Agent

• Agentic Labs - AI system design tools for dev teams (https://agenticlabs.com)
• ion design - Instantly turn Figma designs into clean React code (https://www.ion.design/)
• Greptile - AI expert that understands large codebases (https://www.greptile.com)

Application of AI in Government Regulation and Approval Scenarios

This phase of the project saw multiple projects dealing with government regulation or approval processes, such as applying for the National Science Foundation, passing FDA certification, discovering and participating in government procurement bidding programs, and so on. This is actually a great scenario for an AI Agent - responding to strict specifications and processes that require very careful study of the specification documents for each step, providing and filling in highly relevant information.

For these fine natural language tasks, AI is obviously more advantageous than human beings, and at present, AI has already possessed strong text comprehension ability, and can effectively extract the content required by the specification from numerous materials, summarize and organize them, and unify them for reporting. the value of the application of AI in this particular field is very significant, with the disadvantage that the frequency of use is relatively low, and the market ceiling may not be high.

Government Processes and Regulation

• Government regulatory processes: Artos - Turning science into regulatory submissions (https://www.artosai.com/)
• Government procurement process: Hazel AI-driven marketplace for government contracts (https://hazeltech.ai/)
• Government Grant Process: Aidy - We help businesses and nonprofits apply for grants (https://www.aidygrants.com/)

The Healthcare space is starting to see a lot of ground-up scenarios

Because the business value of the US healthcare industry is so large (17% of GDP), this round of LLM breakthroughs in natural language communication and understanding have addressed the biggest barriers to intelligence in healthcare, with voice-interfaced Agents showing very good usability for things like clinic appointments, patient callbacks, pre-surgery communication, and patient tracking. All of this not only significantly reduces healthcare costs and improves quality, but also expands the scope of healthcare services and provides broader support for patient health improvement.

The integration of Agent and electronic health record systems also lays the foundation for the next step of AI participation in more diagnostic and treatment tasks, and the basic conditions for the integration of the entire AI doctor into the existing medical system are already in place. However, there are obvious differences between the systems and business models of the medical industry in China and the United States, and many entrepreneurial ideas may not be suitable for the country.

medical care

• Somn - AI receptionists for healthcare clinics (https://somnapp.com)
• Arini - The AI receptionist for dentists (https://www.arini.ai)
• Anaphero - Automating patient-facing tasks with voice AI (https://www.anaphero.com/)
• HonaLess time with charts. More connected patients. (https://www.hona.ai)
• Attunement - Patient monitoring and treatment recommendation for better behavioral health (http://attunement.ai)

Multimodality: more focus on scripts for startups in the video generation space

The release of Sora has caused an earthquake in the AIGC space, and this earthquake has caused everyone to rethink the nature of video applications. Perhaps the video model should provide fine rendering technology and powerful physics engine, while the scene plot and the picture in the end should have what people and objects, the character's personality and performance, as well as the evolution of the story theme clues, all of which is the video more core things.

After all, in the eyes of the audience, the textures of nature and the laws of the physical world are so highly defined that not having them doesn't work, and once they are there, they are no longer the focus of attention. Real stories, characters and experiences are what people really want. This installment of YC Startups returns more to the more central direction of video generation applications.

When it comes to storylines and scripts, it's games that really allow for the most creative imagination - not only is there room for literary creativity, but there are also sound and video experiences, and most importantly, the ability to create a script and play a character at the same time through interactions, also known as Interactive Narrative Interactive Narrative. On the other hand, because the game scene is controlled, the requirements for video technology are much lower than those of both the real physical world and cinematography, and one of the best testing grounds for multimodal AI technology.

Video generation with games

• Video: Eggnog YouTube for AI-generated content (https://www.eggnog.ai)
• Game: Arcane - AI powered Roblox (https://arcanelabs.ai)
• Video: Focal - AI movie studio (https://focalml.com)
• Video: sync labs- an api for realtime lipsync (https://synclabs.so)

LLM and Agent reviews will be a huge potential market

No matter how much modern software engineering specifications caution that Test Driven Development is the foundation of quality development, the reality is that very few projects start with developing test cases. The reason behind this lies in the fact that on the one hand it is difficult for people to start with a fully thought out idea of what kind of product they are going to end up with (the more ambitious the product, the more this is a problem), and on the other hand many people expect that most of the code they start writing will be discarded.

Now AI is also at this stage, basically the whole industry is still in a state of "bare running", especially the evaluation of Agent, basically it is blank, most of the existing benchmarks are used to prove how successful the Agent is, and it is difficult to reveal the diversity of failure modes of the Agent. In this issue, there is a project that deserves attention, and its main highlight is to build an agent specifically for evaluating other agents, which is a very common idea but a very imaginative direction, especially how to avoid the evaluating agent to become an evil saboteur, which is also a very challenging problem in the industry. In the same direction, another project is to recognize AI-generated images and videos. This idea is not new, but it is really a market need.

Assessment and Security

• Agent Assessment: MAIHEM - Automate quality assurance for your LLM application (https://www.maihem.ai)
• AI-generated content detection: Nuanced Detect AI-generated images (https://www.nuanced.dev)

04 Characteristics of Startup Teams in the AI Native Era

We noticed that there are a lot of young people graduated from top schools in this YC, most of them started their business 3-5 years after graduation. Though young, their projects show a good insight into technology and scenarios. Many of the companies have been established for a short period of time, but their project demos are highly accomplished, and basically have end-to-end scenarios that can show the value of their projects.

Perhaps, this is a significant feature of the arrival of the new round of technological revolution - in the face of new technologies, to a large extent, everyone is on the same starting line, when young people naturally have a great advantage. Most of the founding teams of the selected projects are very compact and areIt is 2-4 people who have come a long way and have a long history of working together and friendships between the co-founders.

In addition, we also found the 'elitism' behind the selection of projects in this issue: the entrepreneurs are young, have good educational backgrounds (Harvard, MIT, Stanford, CMU, Berkeley, etc.) and work experiences (Google, Meta, Microsoft, etc.), and have excellent project positioning and creative imagination. For W24's founding team, in addition to the largest percentage of white people (36%), the second largest percentage of Asian people (25%). Meanwhile, 21% of W24 companies have female founders and 11% have female founders.

Perhaps at this stage, in these top universities and large factories, people have more access to AI technology, have a broader vision of application scenarios, and are surrounded by elites who can stimulate the imagination of entrepreneurs. yc's tendency to make this choice, in the age of AI, I don't know if it's a cause or a result.