Anthony Sutardja

Open-source models can be a strong lever for increasing gross margins at AI companies, especially those that have hefty cost structures (e.g., enterprise sales orgs that require significant customer success or implementation resources, as well forward-deployed AI services models common in healthcare and life sciences). One of many reasons to maintain a modular and flexible stack with minimal dependencies when building an AI company. More on the importance of flexibility in the AI stack: https://lnkd.in/eA2XATRk h/t Delip Rao #ai #artificialintelligence #generativeai #healthcare

50

6 Comments

LLMs will never be able to perform planning… this requires an underlying model that can be simulated (“planning” refers to a lookahead policy, also known as model predictive control). Planning usually requires live data, chess (and other games) being an exception. Think of solving a shortest path problem over a transportation network.

76

7 Comments

Mohammad Hossein Amini did a nice study of LLMs trying to play chess (it went badly). This exercise helps to highlight different methods for making decisions, and since LLMs (or any form of machine learning) only uses one of these methods, they would never be able to play a game of chess (or any strategy game). There are two primary methods used to make decisions in chess: pattern based logic (a form of PFA), and direct lookaheads (DLAs) (DLAs can also be combined with VFAs, but PFAs and DLAs are the two primary ones). PFAs are used near the beginning of a game when a player simply looks at the board and decides on a move, without thinking about future moves. Beginners only do this for a few moves, while experienced players can use this longer - the benefit is that the moves are very fast. Once a board becomes too complicated (this depends on the player), it is necessary to start planning into the future, anticipating what the opponent might do. This kind of planning (a form of tree search) requires a model of the underlying problem. This means understanding the rules of the game, which for chess is relatively simple compared to real-world problems (such as managing a supply chain). LLMs can only do pattern matching, albeit a very sophisticated form. This is why they will never be able to solve a shortest path problem through a transportation network, which requires knowing the structure of the network (one-way streets, left-hand turns, ...) as well as travel times (which can change dynamically). Chess (and other strategy games) require the ability to plan into the future, which means having an internal model of the problem. It needs to be able to simulate the different moves that each can be made with each piece, by both a player and their opponent (which requires the ability to simulate the decisions of the opponent).

82

8 Comments

 🔥 "How many engineers does it take to replicate Docusign as a software product in 2026?" Sarah Guo dropped this thought-provoking question at last week’s panel. Answers from the eng/product leaders at top companies were eye-opening: Scott Densmore: 10 engineers Romain Huet: 5 engineers Ainsley Escorce-Jones: Just 1 Lee Robinson: 1-2 engineers Say hi to a future where rewriting major software might be simpler than you think. Reach out to me if you are planning to do that! 🙋🏻‍♂️ PS: Take a look at what Gil Arditi is working on at Morph.ai 💡 #software #sdlc #softwaredevelopment #llm #ai #ml #largelanguagemodel #copilot #gpt4

29

13 Comments

Contrary to popular belief, not all friction is bad. In this interview, product leader Yue Z. shares how adding good friction actually led to higher conversion rates when onboarding users at Thumbtack. Here's the latest interview in my product leader series in Forbes: https://lnkd.in/ex842XsW

24

1 Comment

Let the wave begin! Pioneer Fund company Recall.ai *grew their revenue 10x* in the past year, has 300 enterprise customers, and just raised a $10M Series A led by Ridge Ventures. David Gu and Amanda Zhu dropped out of college at age 19 to start the company. They've grown it exponentially with a tiny team. They make it easy to build meeting bots that are cross-platform. They're aggregating all these different APIs into one unified API, the way Twilio did with cell phones. At first it "sounds" easy to build a meeting bot -- just integrate with the Zoom API. But in fact there's a huge long tail of things you need to build, security and privacy pitfalls, automations, transcription complexities, integrations, the calendar invites where someone is using Google Meet or Teams, staying up-to-date with the libraries, etc etc. Recall takes care of all of that. It's WAY easier to build on top of Recall.ai and focus on the features that make you unique. Congratulations to the Recall team!

86

7 Comments

The VC/DNVB honeymoon is over. 🥀 Here's why: both sides have realized that DTC brands aren’t really tech companies. Recently, the "DTC darlings" have taken a hit: Bonobos was resold for $75M after Walmart bought it for $300M 📉 Unilever parted ways with Dollar Shave Club seven years after buying it for $1 billion 💸 And that's just the start—other brands like Casper, Outdoor Voices, and Smile Direct Club have faced similar fates. These brands thrived during the early 2010s VC boom. 🚀 Since 2006, venture capital investments skyrocketed from $30 billion to over $340 billion in 2021. With limited SaaS options to invest in, VCs turned to startups with similar growth potential. DTC brands fit the bill perfectly, wooing investors with three key points: 1️⃣ Tech-savvy teams: In the early 2010s, selling online required a strong dev team, and competition for talent was fierce. 2️⃣ Cutting out the middlemen: Facebook and Google allowed these brands to reach customers directly at a low cost, bypassing traditional distribution channels. 3️⃣ Attractive business models: Gross margins were high, similar to software, and recurring revenue from subscriptions offered an impressive LTV/CAC ratio thanks to low distribution costs via Facebook. But several changes exposed flaws in this model: 1️⃣ E-commerce tech became accessible to everyone through platforms like Shopify and Klaviyo, removing the unique advantage of in-house tech. 2️⃣ Facebook and Google increased ad costs, replacing one middleman with another. Then, iOS 14 made targeted advertising even harder. 3️⃣ Post-COVID, e-commerce growth slowed, and to sustain their growth, vertical SaaS platforms raised their prices. 4️⃣ Rising raw material costs squeezed margins further. 📈 5️⃣ VCs realized that big companies like Unilever and P&G (potential buyers) didn’t care about LTV/CAC ratios since multiple sales channels (including retail) complicated these metrics. This undermined the growth acceleration strategy VCs had envisioned. In the end, e-commerce is more about logistics, product, and marketing than pure tech. 📦🛍️📢 All this makes it a less appealing investment for venture capitalists.

2

VC funding in healthcare has been more resilient than in other industries. Here are some key takeaways from recent Carta data: 👉 At the priced seed, Series A, and Series B stages, dilution for health startups has noticeably decreased, while dilution for SaaS startups has stayed flat or increased slightly. 👉 In priced seed rounds, median dilution for healthcare companies dipped to 14.2%, compared to 20% for SaaS.  👉 The frequency of bridge rounds surged in the health sector in Q1, jumping to 44.3% from 36.9% in Q4 2023. This is the first significant increase in bridge rounds in healthcare over SaaS since early 2022. 👉 More founder-friendly terms for health startups may result from enduring VC interest since COVID. Healthcare has remained strong through economic turmoil, enabling higher valuations with smaller round sizes. What are the implications? The healthcare sector might see a rise in new ventures due to favorable investment conditions and lower dilution rates. The surge in bridge rounds suggests that certain sectors within healthcare face unique challenges that require interim financing solutions. Founders and investors in the healthcare space- what are your thoughts on this? I’d love to hear from you. If you know a founder innovating in healthcare, tag them in the comments! The Artemis Fund is especially interested in groundbreaking care tech companies. You can check out Carta’s full blog linked in the comments. If you haven’t read our latest blog on care tech, you’ll find that there too. #healthcare #caretech #fundraising

21

3 Comments

Ross Haleliuk posted the fantastic article below. This is the world where I live day-to-day. At TechOperators, we invest mostly in early-stage cyber, but we do things a little differently, and we believe there are ways to invest successfully outside of pure Power Law math. The article argues that many security problems are too small for VC. I agree. I often try to convince bootstrapped founders not to raise venture because doing so can turn a successful, slow-growing bootstrapped company into a failed venture-backed company because, despite a large infusion of capital, it couldn't double every year. VC is not monolithic--not by stage, strategy, or style. Venture is often equated with "Tier 1 Venture". Ross argues that VC is not always great for early-stage cyber--and he is right. Bootstrapping AND VC work when incentives are aligned. Does it work for an early-stage VC with a <$200M fund to invest in several early companies at reasonable valuations, setting up the conditions for reasonable exits that pay off for both investors and founders? Yes. Does it work for $800M funds investing in seed stage at $100M+ valuations? Well, that depends! Power law says it does (for VC), but the unfortunate externality is that these rounds destroy companies and founder equity more often than not. There is a role for patient capital in this ecosystem to fuel successful companies that retain exit optionality as they scale--driving exit value for both founders and investors.

34

2 Comments

In this week's episode of the Mobile Dev Memo podcast, I speak with notable X personality Trung Phan about the evolving social media landscape, and how to tame the mayhem and pandemonium of X. https://lnkd.in/dVHMzbui

1

Key takeaways from VC Goat - Vinod Khosla podcast. - The firm cares more about working on interesting problems with large technical solutions rather than just maximizing IRR. The team is there because they believe in the mission. - Khosla assumes they've lost the money the day they invest, and then maximizes for the upside opportunity. He calls it "option value investing" rather than IRR investing. - Khosla has contrarian bets in AI (neuro-symbolic computing, probabilistic approaches), biotech, robotics, crypto, and more. - He believes aviation fuels, fusion energy, new transit systems, and other contrarian areas ignored by most investors will be huge opportunities. - Most large innovations come from outsiders, not industry insiders. Khosla looks for founders who can learn a business rapidly rather than have deep domain expertise. The Future Impact of AI - AI will be deflationary and increase productivity growth to 4% annually vs the typical 2% forecast. This will cause great abundance but also increasing income inequality. - Bipedal robots will take over most manufacturing and manual labor jobs within 20-25 years. This will free humans to be more creative and pursue their passions. - Education will shift from job training to creativity. Uniquely human elements like taste and curation will be most valued in an AI-enabled world. https://lnkd.in/gqBsmbks

66

5 Comments

Did you catch Elon’s speech at Tesla’s recent shareholder meeting? He shared an update on FSD (Full Self-Driving), where the miles driven under FSD are not only increasing rapidly but also, the time between human interventions is also increasing. This means that only when an intervention happens do they know what new model is best to roll out to users. So what has changed? In August 2023, Tesla announced version 12 of FSD, where they started using Neural Networks for training and navigating. They are leveraging the array of cameras around the car to record millions of human-driven miles to train their models. There is no code written; it is simply the AI that learns. For example, when the car sees a stop sign, it should slow down and come to a halt. Every mile driven is essentially improving the experience and capability of their product. The flywheel is in process, so expect a parabolic curve to appear. See the second chart below; it shows the amount of NVIDIA H100s, giving them the AI infrastructure to process, compute, and train their models. They can handle the complexity of an ever-changing environment. I find this exciting because in our world of investing in NatureTech, just like Tesla, we are trying to understand an ever-changing environment and how to operate as efficiently as possible. Everyone is excited about AI; some say it is a bubble, but we are only getting started. Profound change is ahead. Companies like Tesla are a great example of how we can leverage AI to operate efficiently in our ever-changing world. This is the type of NatureTech business I’m looking to find and support, capturing streams of primary data and generating understanding and intelligence on the complex multiple systems that we call Nature. #Tesla #Naturetech #AI #intelligence

20

5 Comments

There's a new breed of CEO in town. The AI-Native CEO. These people are either rebuilding archaic companies or building in "adjacent possibilities," they're creating truly durable businesses with real customer value, and importantly, they're building their companies in an AI-native way; more efficient than ever, stitching together technologies that once required high-overhead, and focusing their hiring on people who are autodidacts, the ones who are inhaling knowledge on how to improve their own functions leveraging the newest tooling for the future of their jobs. A few of my favorite examples... Valley, backed by General Catalyst, getting 30-50% response rates for salespeople using conversational selling instead of archaic and annoying drip campaigns. AminoChain, backed by a16z, turning biobanks into to shoppable specimen centers so that healthcare institutions and researchers can purchase biomaterial like blood an stem cells and so donors can earn and track their participation in science. HIFI, backed by Blockchain Founders Fund, using stablecoin for B2B payments, creating instant settlement times and eliminating fee extortion. Dune Security, backed by Craft, using AI to test, identify, and train employees on their own risky actions, automatically reducing the biggest risk any business has--the behavior of its own employees. https://lnkd.in/eJUnuM4z #venturecapital #ai #artificialintelligence Antler

68

1 Comment

The VC funding market has literally flipped for AI venture funding! Early-stage AI startups in Q1 `24 commanded higher median valuations than their late-stage counterparts. It begs the question: is there a dearth of truly groundbreaking early-stage AI companies, or is there an influx of too much capital chasing the next big thing? Likely both. Here’s why: AI Hype: There's no doubt AI is a transformative technology. But with so much excitement, some investors might be throwing money at anything with "AI" in the name, regardless of long-term viability. Limited Proven Winners: While there are some incredible AI success stories, the field is still young. This lack of established, late-stage AI companies with strong returns could be making investors hesitant to invest in them at high valuations. Limited enterprise adoption. Startups are betting that the tech giants, esp. AI tech giants, will have limited appeal to large enterprise. But the solutions are getting more and more compelling, and security and data privacy is being addressed. Meanwhile, the majority of capital is still going to a few players like Anthropic, Mistral AI, xAI, just like last year it went to OpenAI, Anthropic, and Inflection AI. The future of AI is bright, but for venture to succeed, sustainable growth requires a focus on companies that can clearly differentiate themselves in the market fray. What are your thoughts on the AI VC market? Where do you think the opportunities are? #AI #VentureCapital #Investment 

26

4 Comments

What do VCs actually look for in pre seed deals, particularly given the recent wave of AI startups? As a B2B SaaS focused investor, I have the opportunity to evaluate thousands of software businesses, and invest in 100+ each year. One of my favorite parts of my job is meeting exceptional founders and learning about problem spaces that are completely new to me, but the investment team at Forum Ventures spends a lot of time discussing focus areas and opportunities where we see the most promise and are actively looking to invest. In the coming weeks, I'll be sharing a series of market opportunities where I am looking to back exceptional founders solving challenging problems.  * Automation of complex, legacy workflows within enterprise, particularly across regulatory and compliance use cases. * Application of AI across the defense and broader government sectors, specifically related to operational efficiencies, the infrastructure stack required to leverage cutting edge AI advancements, and the development and adoption of autonomous systems. * Manufacturing/hardware design and production, spotlighting modern manufacturing - robotics, aerospace, alternative energy, medical devices, etc. * Complex manufacturing supply chains- from extraction to processing and refining, and ultimately through manufacturing and distribution.  If any of the above sounds like something you are building, let’s connect. #GenerativeAI #VentureCapital #EarlyStageStartups #Innovation #TechTrends #AI

39

3 Comments

Please media types writing about the cost of AI: recognize there is a difference between huge GPU/TPU clusters needed to train foundation models vs. inference costs to run most apps on top of foundation models. The two are not synonymous, so you can dial back some of the doom and gloom about AI startups and your commingling of the two different AI business types.

11

How does David beat Goliath aka early-stage startups close candidates also interviewing at FAANG / larger companies? Generally, they don’t, but here’s what @SingleSprout's data (80% offer close rate in last 6 months) says: 1) $30k: If the offer from a big company exceeds yours by over $30k, you're very likely not going to get that candidate. 2) # of Competitors: A candidate considering one large company might still be within reach. However, the more large companies they entertain, the less likely you'll get them. 3) Motivation: Understanding why a candidate genuinely prefers your company is crucial. Look for and test for interest in impact, equity, and the entrepreneurial journey. Unfortunately, most startups hiring on their own can't get this level of information from a candidate, and sadly waste a lot of precious time :(

23

10 Comments

As a B2B SaaS VC 👩‍💻 and a huge data nerd 🤓, fewer things bring me more joy than a truly statistically significant analysis (600+ companies surveyed) of B2B SaaS metrics across early stage companies. Great work by our team at Emergence Capital! See comments for full report, and please let us know how we can make this better in the future!

266

15 Comments

We're launching the inaugural edition of Beyond Benchmarks, a deep dive into the metrics and trends observed across the early-stage enterprise cloud market. This is the first look at some of the business data around GenAI in SaaS. Some initial observations: 1) 60% of SaaS companies have already integrated GenAI into their offerings (60% of these as part of their current product and 40% as a new product), and an additional 20% are planning to do so this year. Almost half aren't monetizing these features. Those that are are primarily experimenting with both usage-based and flat fee pricing models. 2) Most companies are utilizing OpenAI as their primary LLM, but many are experimenting with multiple models - and we are noticing an early trend towards intelligently routing GenAI inference requests to different models based on cost, performance, and security. 3) Though we are in the initial innings of GenAI, we are seeing some positive signals - companies that implemented GenAI had 7% higher NDR than those that did not. Still early days, but we'll be doing this benchmarking report regularly to track and share how things progress in this new era. Link to report in comments.

104

6 Comments