Feb 14, 2025 · 37m · a16z
Reasoning Models Are Remaking Professional Services
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the a16z podcast, host Alex Immerman interviews Hebbia Founder and CEO George Sivulka about how reasoning models, sub-agent orchestration, and enterprise AI are revolutionizing knowledge work across finance, banking, and legal sectors. They discuss the transition from simple chat interfaces to agentic workflows, organizational change management, and the long-term economic impact of AI on capital markets.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 2.9 Guest teaching 3.3 Guest disagreement 1.7 The host pushing back 1.6
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 2:38 DeepSeek nothing burger dismissiveness
George forcefully dismisses DeepSeek as a nothing burger and criticizes China's technological transparency, taking a strong, unhedged position.
Hardest push from the host ▶ 3:25 Open source regulatory counter-argumentAlex refuses to fully accept George's dismissive framing of DeepSeek, intervening to point out its significant impact on open-source momentum and US regulatory policy.
Biggest teaching moment ▶ 10:28 Correcting the generic chatbot flawWhen Alex attributes generic chatbot failure primarily to accuracy, George explicitly corrects him, explaining that the deeper structural limitation is handling private unstructured data and complex multi-step processes.
The host holds their own ▶ 15:56 Framework for AI ROI vectorsAlex demonstrates deep domain expertise by clearly framing the AI ROI model into speed efficiency versus net new analytical discoveries, structuring the core thesis of the segment.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Title Sequence and Legal Disclaimer | 3 | 4 | 3 | 3 | Alex opens with a rapid-fire query on scaling laws and DeepSeek. George takes control with a detailed breakdown of training vs inference scaling laws, while Alex briefly pushes back on DeepSeek by highlighting open source regulatory effects. | |
| Favorite AI Tools and Agentic Deep Research | 2 | 3 | 1 | 1 | Alex prompts George on his favorite AI tools and background transitioning from a Stanford PhD to financial tech. George explains his introduction to GPT-3 and identifying market pain in junior analyst workflows. | |
| Why Generic Chatbots Fail in Professional Knowledge Work | 3 | 4 | 2 | 2 | Alex suggests accuracy is the main flaw in generic chatbots, but George reframes the issue to emphasize offline unstructured data and custom firm workflows. Alex adds context on how private data creates alpha. | |
| Beyond Chat: Orchestrating Sub-Agents and Computer Use | 3 | 3 | 1 | 1 | Alex asks about interface choices and leveraging computer use. George articulates Hebbia's design philosophy as a specialized sub-agent orchestration platform that AGI itself would utilize. | |
| Change Management and Developing AI-Native Analysts | 4 | 3 | 1 | 1 | Alex demonstrates expertise by dividing AI value propositions into speed efficiency versus net new discoveries. George elaborates with concrete financial use cases like virtual data rooms and screening SIMs. | |
| Advisory, Banking, and Legal Applications | 3 | 3 | 1 | 1 | Alex sets the stage regarding the 2025 board mandate shifting from experimental AI to ROI. George details specific time and dollar savings across advisory, legal onboarding, and private equity portfolio benchmarking. | |
| Bicycles for the Mind and Empowering AI-Native Talent | 2 | 2 | 1 | 1 | Alex asks how Hebbia designs for early-career AI-native talent. George references Steve Jobs' 'bicycles for the mind' concept and explains how adoption spreads from junior analysts up to senior executives. | |
| A 10-Year Vision for Capital Markets and AGI | 3 | 4 | 2 | 2 | Alex asks for a 10-year vision of capital markets and challenges whether incumbent mega-funds hold a permanent private data advantage. George reframes the value of historical pre-LLM data and describes building a private market Bloomberg terminal. | |
| SaaS Pricing Models for AI Agents | 3 | 4 | 3 | 2 | Alex asks about evolving SaaS pricing from per-seat to outcome-based consumption. George dismisses Silicon Valley's premature obsession with outcome pricing, stressing that adoption must precede monetization shifts outside the SF bubble. |
Statements from this episode (28)
Sivulka: AI scaling laws are fundamental mathematical properties of the universe
“And I think they're both effectively mathematical properties of the universe. I don't think they're you know, just an experimental observed thing.”
Sivulka claims Hebbia pioneered inference-time compute scaling
“And this is a technique scaling at inference that was first actually pioneered at Hebbia, and we quickly noticed another scaling law, where if we ran more models and basically more compute at inference time, you could get much better results for very complex t…”
Sivulka: China has not shown it can push frontier AI
“China has not shown that they can actually continue to play ball, ah, pushing the frontier of AI.”
Sivulka: United States will maintain global lead in AI
“America, I believe, is so far ahead, and I think that it will continue to be so far ahead and these technologies, you know, every single year are so much exponentially better than the last year that we have the head start, and I think that we'll continue to ha…”
Sivulka: Hebbia has proven out agentic deep research over private enterprise data
“And I think that this level of agentic deep research is going to be one of the most exciting things that I can't quite disclose now that, that Hebbia has proven out and is working on over private information.”
Sivulka: Deep Research provides the first true experience of an AI agent
“I think deep research allows you to experience What an agent can do for the first time.”
Sivulka: Meta-learning will be the most important technology of all time
“The idea and the promise of meta-learning, and what that means is teaching machines to learn how to learn was, to me, going to be the most important technology of all time.”
Sivulka: Top financial graduates perform the stupidest tasks
“And I realized that the early, like, incredibly, the smartest people in the world, incredibly smart kids, were going and doing the stupidest tasks.”
Sivulka: High-value knowledge work relies primarily on offline unstructured data
“The stuff that matters for financial services, but also for law, for all of knowledge work, is offline unstructured information.”
Sivulka: ChatGPT provides no investment alpha due to lack of custom data
“If you look at ChatGPT and the limitations of like, you give it to a bunch of investors and you ask it to draft an IC memo for a company like Hebbia, every single VC would get the same IC memo out.”
Sivulka: Hebbia platform offers an effectively infinite context window
“Hebbia's got like an infinite effective context window and then when you look at a new opportunity, you could filter to a variety of other opportunities that were similar and actually begin to piece out how it's different, how it's similar and all the pieces t…”
Sivulka: Hebbia aims to be the enterprise tool of choice for AGI
“Rather than try to build the AGI, how do, how about we build the AI platform that is so good that if AGI were to complete a task, it would choose to use Hebbia to do that task.”
Sivulka: AGI will orchestrate sub-agents instead of using massive context windows
“You wouldn't want An AGI to jam a 100,000 documents into its context window and take an infinite amount of time to, or some very, very large amount of time to process that. You'd rather use and orchestrate a bunch of sub-agents.”
Sivulka: Despite amazing demos, nobody actually knows what to use AI for
“One of the big pieces and unlocks that had the experience was coming to the realization that nobody knows what to use AI for. Like, you know, people pretend to go and create all these amazing demos, but nobody knows what to do with it.”
Sivulka: Hebbia performs almost any junior analyst task better than humans
“And you can actually look at our product today and you could argue that it could do almost any task that a junior analyst could do better than a junior analyst.”
Sivulka: Enterprise AI deployment is a sociology problem, not technology
“It's actually no longer a technology problem. It's really a sociology and change management problem.”
Sivulka: Hebbia saves 20 to 30 hours per deal process
“I mean, it saves probably 20 to 30 hours on a deal process, depending on how, how disgusting these VDRs can get.”
Sivulka: Hebbia enables firms to screen 137% more deal opportunities
“We can screen a 137% more opportunities. Like in a, yeah, in a given period of time, which is like, you know, and you, it's with the same depth.”
Sivulka: Lawyers use Hebbia to analyze contract libraries during live negotiations
“People are starting to use things like our infinite effective context window to actually look over entire libraries of formally negotiated agreements and live during a negotiation actually come with better terms or what a better understanding of what is market…”
Sivulka: Hebbia saves clients tens of thousands per deal in legal costs
“If every single credit agreement, you know, they reviewed multiple hours and it costs 2000 dollars per hour for a lawyer to review them. And now they can review them in house. They're saving on tens of thousands of dollars on a per deal basis, maybe hundreds o…”
Sivulka: Managing directors are using AI to audit junior analysts' work
“There are even some you know, MDs like the one off MD that will go in and actually check their analysts work with AI. So they'll create a matrix and like pull out red flags or inconsistencies over something their analysts sent.”
Sivulka: The arrival of AGI will trigger a massive financial market correction
“First, when AGI is here, there will be a massive correction in the financial markets. That is actually my Turing test is if AGI is here, will it actually be able to make significantly more money than a human investor? Better than human investing. And I think p…”
Sivulka: Hebbia is building the Bloomberg terminal for private market investing
“And what Hebbia is trying to build is the Bloomberg terminal for private companies. Not like a pitch book or a crunch base, which is really just for sourcing, but actually to go out and take all of the information private and public and pre-structure that. So …”
Sivulka: Pre-ChatGPT deal history won't help investors evaluate post-LLM startups
“Spending too much time thinking about the historical deals that were AI pre-ChatGPT or pre-generative AI or pre-LLMs are not actually going to help you with the deals post-LLMs and it'll, It'll be even better if you don't have that much information.”
Sivulka: Consumption pricing disincentivizes organizational AI adoption
“You need to get to the usage first before you charge on consumption, because the minute you charge on consumption or you charge on, you know, number of agents and times of salary, You actually are disincentivizing the change that is required to build an AI nat…”
Sivulka: Outside San Francisco, no one knows what AI agents are
“I think San Francisco, like everyone's, you know, talking about like different pricing strategies for AI agents, and I'm like, you know, outside of San Francisco, no one knows what an agent is at this point in time, so.”