Ori E.’s Post

Life update! I finally did it! I started a company! Alston L. and I are building OffDeal (YC W24) - an AI broker for buying small businesses 🤖 🤝 Quick story. As my close friends know - I've wanted to be an entrepreneur for most of my life, but I'd always make excuses for why I can't do it: "Bad economy" "I don't have a work visa" "My idea is not good enough" "Someone's already doing it". Even after graduating from Harvard Business School last year I still felt that way - I almost considered giving up on being a founder and going back to a "regular job". But then I realized that if I don't commit now and take the plunge, I might never have this opportunity again. I decided to give myself a full year to fully commit to the journey and see what happens. Six months later I found myself getting accepted into Y Combinator with an amazing co-founder and working on an idea at the intersection of two areas we knew really well - finance and AI. I'm still pinching myself. So what exactly are we building? The current process of buying a small business is incredibly manual and tedious - an average buyer takes 18-24 months to buy one! OffDeal solves that. OffDeal is an AI broker that uses custom-built AI agents to help small business buyers find, vet, and engage thousands of acquisition targets each week. You can see more info in the link below. We launched our closed beta less than two weeks ago and already have over a dozen buyers using us to acquire $10-50M SMBs! If you'd like to learn more get early access please reach out to request a demo!

" 😡 Generative AI is ruining my outbound response rates! 😡 " A friend in private equity noticed the response rates to his carefully crafted, personalized cold outreach emails had plummeted 40-50% in just the last 6 months. Worse yet, whenever he did get through to a prospect, he was often the 3rd or 4th investor to reach out that same week! I was puzzled - why is he blaming AI!? If anything, shouldn't AI tools supercharge deal sourcing workflows? Well.. here's a curious insight: Generative AI is commoditizing personalization. GenAI made it 10x easier to draft and send thousands of “personalized” cold outbounds at scale - for everyone at once - rendering cold outbound marketing less and less effective. The exact opposite of what GenAI was meant to accomplish! As salespeople rush to employ GenAI tools from Salesforce, HubSpot, Gong, Apollo.io, Jasper and many others, prospects are quickly growing weary of what increasingly appears to be spam, causing the response rates to plummet as recipients disengage from the onslaught of AI-generated outreach. So what will happen to sales? Will AI-to-Human sales motion become the norm? Or are we just sprinting towards AI-to-AI interactions? Will each buyer and seller have their own AI agent represent their respective interests in every future sales transaction? Or perhaps all this technology will simply revive the value of in-person events and the old-fashioned deal making at the golf course? One thing is for certain - salespeople who choose not to pay attention to AI and its evolving capabilities will definitely be left behind by their competition who do. More details in the essay. As always, any feedback welcome.

OpenAI just released its biggest announcement since ChatGPT's debut and it's a HUGE upgrade: ChatGPT for Enterprise is OpenAI's newest and most powerful product yet: - More private and secure (data not shared for training OAI models) - More powerful and performant (no usage caps, 32k context window, 2x faster, unlimited access to Code Interpreter) - Customizable (share custom common workflows / chat templates across the organization, fine-tuning options) As if this wasn’t enough, OpenAI also announced some upcoming new features to be released in the coming weeks / months (these will kill so many recent startups…): - Connect your own company data to extend ChatGPT’s knowledge base - Power tools - a more advanced version of Code Interpreter + web-browsing - Function-specific tools - tools for specific roles, such as data analysts, marketers, customer support and more This further strengthens my view that most F500 companies will choose to build tools in-house instead of licensing/buying from AI startups. The business impact is too critical and the technical complexity is more and more abstracted away. AI startups must tread carefully and choose the right surface area to compete in. The bar keeps getting higher and higher. More details in the link below.

McKinsey & Company just unveiled its own internal AI tool "Lilli" and it's already been a HUGE huge success (even in its beta!): 🕒 The tool has cut down the time spent on research and planning from weeks to mere hours, and from hours to just minutes in others 📈An astounding 66% of employees now use the app multiple times a week 🧠 In the last 2 weeks alone, Lilli answered a whopping 50,000 questions! Lilli's early success might set a powerful precedent for other F500 companies deliberating the "buy vs build" debate. This would mean that startups in the B2B enterprise space would be competing not just with incumbents and other startups, but also with internal dev teams of their potential customers. Folks that are building "enterprise ChatGPT", "chat with your data", or "AI search" type tools should seriously reconsider their product strategy and approach if they want to target more than just SMBs and startups. In my view, some fertile areas of opportunity are: - B2B LLMOps tools - help F500 train/AB test/optimize/evaluate/deploy LLMs (e.g. Humanloop) - "Hard things" that require proprietary tech and deep expertise (e.g. Synthesia with its human-like AI avatars) - "Non-core things" - focus on automating processes that are non-core to the target customer, making them less likely develop in-house (e.g. Intercom for customer service or Hebbia AI targeting lean investment firms / governments) More details in the essay. As always, any feedback welcome. https://lnkd.in/egnedH-d

I think most "AI for Private Equity" startups are working on the wrong problem. There's been a flurry of "Copilot for PE" startups recently and they are all rushing to augment / automate the mundane parts of the job (generate memos, search through data rooms, etc.). The issue is that this is not a "hair on fire" problem for the end customer: - These tools will not make more revenue for the firm (not in any meaningful way anyway) - Cost savings are very marginal (less than 5%) and simply do not move the needle for the PE firm's bottom line - PE firms are not cost sensitive to begin with - they value accuracy and trust - which is why they do not outsource entire DD workflow to a cost center in India. The real opportunity lies in applying AI to the PE portfolio: This is just simple math - boosting portfolio EBITDA by 5-10% will have an outsize impact on the PE fund's bottom line - each $ saved creates ~10x of unlevered value, which directly flows through to the firm's earned carried interest. I attach some monkey math here, and go into my analysis and rationale in the full essay. Let me know what you think! PS if you are exploring AI applications in finance please reach out - would love to jam.

GitHub Copilot made headlines by making software engineers 55% faster. But can AI transform other industries like it did with software engineering? Harvey, a Series A startup, is promising just that for law firms. But is it a good business? And can other "Copilots for X" replicate Harvey's success? 👇 ⚖️ Big Law attorneys work about 3,500 hours a year, but typically charge (very high) hourly rates for roughly 2,000. This means that about 1,500 hours a year are "non-billable". Harvey promisesto reduce non-billable hours and increase efficiency, thereby allowing law firms to either 1. Do more work with same employees (more revenue!) 2. Do same work, fewer employees (less cost!) 💰 Based on some monkey math, the potential value capture for Big Law is substantial - up to $750K in more revs or $50K in less costs (on a per employee basis!). 🤔 But how does this translate to the AI SaaS product itself? Harvey's primary cost will be its OpenAI API calls (or compute costs if they switch from GPT-4 to their own models). At a $500/seat fixed price, this implies a margin far below that of the 80-90% 'gold standard' of traditional SaaS. However, given the high value the tool unlocks for their end customer, they can most likely charge significantly more. At $5000 per head, the gross margin will be ~90%! 🌐 But how does this generalize for other aspiring "Copilot for X" product? It's crucial to do a similar analysis to understand the potential upside. Here are some key considerations: 1. Are you cutting costs or adding new revenue? 2. Are you indispensable to the end customer? (how meaningful is your impact?) 3. Can current tech handle the work? (do you deal with just text, or also numbers, images, audio, video, etc?) 📝 My personal take: "Incremental" productivity gains will suffice with horizontal tools like Microsoft Copilot. For an enterprise to adopt a new SaaS product your tool will need to make a significant impact to be worthwhile. As a side note, I am currently exploring AI applications in financial services. Reach out if you want to jam on this! PS full article below. #ai #copilot #harveyai #lawtech #future

🤖 Can open-source software really be the most viable long-term business model in AI? Let's explore 👇 In this essay, I will make the case for open-source AI by demonstrating it can: 1️⃣ Provide a superior product for end-users 2️⃣ Offer a more sustainable business model in the long run Open-source software allows anyone to freely view, use, modify, and distribute the code, fostering a collaborative and decentralized development environment. So what's in it for the end user? Advantages of open-source AI include: 🔐 Open models for private data (more data security and control) 🛠️ Customization and optimization (tailored solutions for niche use cases) 🔍 Interpretability, auditability, security (more trust) 🌍 Decentralization of the AI ecosystem (less power concentration) Now let me show you why it is also good (best?) biz model long-term 💰 📈As AI models approach a performance asymptote, general-purpose LLMs will become commoditized -> competition will drive costs to zero. Differentiation will come from creating custom variants of open models trained on private proprietary data - a highly profitable premium service that enterprises and governments will gladly pay for. Open-source AI will become the dominant biz model. Of course there are some risks: 🦹 Misuse (malicious use of AI tools) 📊 Resource constraints (lack of funding and support) 🔀 Fragmented development (inconsistent updates & improvements) These can be mitigated with appropriate regulation (target use, not the tech) and collaboration. In conclusion: 🔸 Open-source AI provides users with more control, customization, and transparency 🔸General purpose models will get commoditized 🔸 Competition will drive the price of models to zero 🔸Open-source biz model will reign supreme Excited to see how this plays out. I leave you with a few questions to ponder: ❓Will foundational models become commoditized? ❓Will all major models be eventually open-sourced? ❓Who will accrue the most value? Check out the full essay below and let me know what you think! #ai #llm #openai #stabilityai #opensource https://lnkd.in/e-TCDgdm

🤔 Does size really matter in large language models (LLMs)? 🚀 The race towards bigger LLMs has led to impressive performance gains, but also: 💰 High costs 💼 Power concentration, and 🔐 Closed-source models 🌟 Enter small models - the new frontier! 🌟 Recent research suggests fine-tuning LLMs with human feedback can improve performance without increasing size 📈. Benefits of small models include: 🌐 Open access 🎯 Hyper-specialization 📱 Edge computing for more efficient, agile models 🍏 Apple is already developing its own small LLM optimized for iPhone chips, and Android OS will likely follow 📲. These small, nimble models will become ubiquitous, seamlessly integrated into our personal devices - they will be a natural extension of our mind 🧠  and we won’t even notice they are there. It will feel like magic. 🎇 Perhaps the future of LLMs will involve a mix of ultra-large proprietary models for general knowledge 🌍 and small open-source models for private data 🕵️♀️. These models may coexist and complement each other in various applications. 🤔 This raises some questions: 1) What's the best interface for a given model size? 2) How will LLMs interact with one another? 3) How do we ensure maximum interoperability? 4) Will open source reign supreme? 5) What will the role of data privacy be? Read the whole essay for more details and share your thoughts with me on Twitter @leveredvlad! 📩 #ai #llm #openai #chatgpt #future https://lnkd.in/eRMgBBSG