How AI Agents are Reshaping Modern Work | Philip Christos

The transcripts are edited for clarity, sometimes with AI. If you notice any incorrect information, let us know.

Venture Capital Career Journey

Alexey: Hi everyone, skip the usual introduction. As always, subscribe to our YouTube channel, like the video. There is also a pinned LinkedIn live chat. Use this for asking questions. Today I have this big pleasure of talking to Philip. (0:00)

Alexey: Philip and I met a couple of months ago when he was in Berlin. We talked, got to know each other, and I found out what Philip was doing. I thought it would be very cool to invite him for a podcast interview. A few years ago we had an interview about investing, which was very interesting for me personally. That is why I thought he would be a very nice guest. (0:15)

Alexey: Welcome Philip. Let me pull your bio. Philip is an associate at M2VC, based in Barcelona, where he focuses on investing in frontier AI. Before joining M2VC, he spent four years at Terra VC, rising from analyst to associate. (0:49)

Alexey: Today he is an active voice in the tech community, frequently sharing insights on AI agents, model capabilities, the evolving landscape of AI scaling, and retrievable infrastructure. I guess you post on Twitter. All of us do. Today we started this conversation because you asked me if I heard the announcement from OpenAI, which I did not. (1:07)

Alexey: Even though I check Twitter regularly, typically when I open my Twitter, I see posts about people complaining about limits and Codex, or people complaining about Anthropic. At least this is my Twitter bubble. Maybe it is different for you. You get better news, but I guess this is your job. (1:33)

Alexey: You have to keep an eye on what is happening in the industry. Before we go into this, can you tell us about your career journey so far? How did you start doing what you do right now? What is your story? (2:05)

Philip: It is quite straightforward. I just spent almost all my professional career in venture capital so far. I did a couple of industry programs after the university. For example, I worked for almost a year as a marketing analyst at Procter & Gamble. (2:21)

Philip: I think this is already quite outdated since writing some basic Python to visualize dashboards is not a job anymore. I switched to the VC career quite early. Most of my experiences are with startups. I joined a virtual capital firm as an intern and became an analyst. (2:38)

Philip: It is a basic function of any VC firm. Analysts are the people who actually work with startups in the field. They go to conferences, prepare documents, talk to startups, and drive most of the stages of the investment process, but not the decision making. I am now an associate, which is basically an analyst but better. (3:02)

Philip: In venture capital, no formal titles exist. Everyone in the fund performs every possible type of work. As an intern, I talked to startups, prepared investment memorandums, and went to conferences. The partners and founders of the fund performed the same set of activities, but with more focus on strategic decision making and owning the relationships with startups and our investors. (3:26)

Philip: I would describe the VC world much like a startup. Everyone does everything. It depends more on the exact fund than your title. (4:06)

Alexey: Do analysts in VC companies do the same kind of work as data analysts in internet companies? Or is an analyst here different? You do not spend time doing SQL stuff and preparing dashboards, but it is more about checking what this company is doing. Do you verify if what they claim and what they actually do is the same? (4:29)

Alexey: If they report these numbers, are these numbers true? Is this what analysts do in VC companies? (4:51)

Philip: It is much closer to the second thing you described. The fun fact is I worked as a marketing analyst mostly. It is a kind of technical analyst for Python, big data, and SQL. (5:06)

VC Analyst vs Data Analyst Roles

Alexey: You mentioned it is not a job anymore. (5:14)

Philip: I do not think it is a job anymore. I thought about becoming an analyst in consulting, running consulting projects in companies like McKinsey. I also thought about becoming a business analyst. I eventually became a VC analyst. (5:25)

Philip: They are all called analysts but have nothing in common. An analyst in venture capital is the closest analog to an analyst in finance in general, but with less Excel and spreadsheets. The idea is that we try to predict the future. I would not say this is a reproducible skill because most VCs fail. (5:41)

Philip: This is unfortunately how it works for most people. We try to understand if a particular team will succeed or not. We want to know if they will become a unicorn. After that, you do whatever it takes to predict. (6:20)

Philip: Some people rely more on Excel and metrics. Some VCs are very deep into unit economics or creating some primitives. I believe that LTV was first created by some VCs to try to understand the economics of startups. They propagated this lifetime value and other product metrics. (6:43)

Philip: Some VCs rely more on the pitch deck, story, and vision. If I know that a payment solution for agents will be a huge company in 2030, I might invest based on that. Some VCs say the only thing they care about is people and relationships. They look for founders who have stamina or other good qualities, even if no one knows exactly what those qualities are. (7:08)

Philip: Some intuition comes into place. It is very hard. As in any investment, we try to find our alpha. We try to find what is different about our approach to investing from others and what our edge is. (7:41)

Philip: For me personally, we invest in early stages. We usually invest at the idea stage or when there is some early traction, like a first MVP just launched on Product Hunt or Twitter. (8:05)

Alexey: For you, you do not have metrics yet and you cannot calculate numbers yet. Do you need to look at the pitch deck plus the team? (8:12)

Philip: There are two types of investors. There is the Silicon Valley style investor and the more classical investor. In Silicon Valley, they meet a founder, spend half an hour in a cafe, and write the check right away just from the story and vision. In most other parts of the world, investors are much more thorough. (8:28)

Philip: They ask for several meetings to check, recheck, and validate. We are somewhere in between. We are a small firm, so we cannot do very thorough due diligence. We do not run many checks and we actually should not, because we should not try to catch data from noise. (8:53)

Philip: We also do not write checks just after the first meeting. This is not angel investment. We usually talk to founders to understand if we are on the same page and if we believe in their story. For example, some founders believe that scaling the amount of compute to train large machine learning models will give us smart models. (9:16)

Philip: We call them LLMs, and maybe they will be smarter than the smartest people around. Investors who believed in this ten years ago invested in OpenAI and are now super rich. This is basically one part of the job. Another part is trying to catch metrics if there are any, but usually there are none. (9:48)

AI Investment Thesis Strategy

Alexey: Do you have a clear portfolio focus? Because we are talking about AI today, I suspect your focus is AI, but within AI it is a broad term. Are there companies, domains, or industries where you focus specifically? Or is it broader and you just try to connect with the founders to understand what they do and see if you want to support them? (10:13)

Philip: Of course, we have a more particular thesis and idea of what we are trying to back. Investing in AI is not an investment thesis anymore because it is like saying we invest in internet companies or companies enabled by computers. Nowadays it is not focused enough because basically everyone is an AI company. Every VC tries to find their own edge. (10:48)

Philip: One VC might believe that AI applied in Latin America by woman led companies will perform better. Another might believe that data providers for AI labs will be the largest companies of our generation. Every VC has a number of focused ideas of what they want to back. In many cases, they spend a huge amount of time understanding what they want to back and then wait for a startup from this idea space to appear. (11:21)

Philip: When it does, they decide fast and invest. If you believe that a particular dev tool for AI driven development must exist, you just wait for a team to start building it and you back them first. (12:00)

Alexey: You already have an idea in mind and you think this direction is promising. You want to bet on this direction and you start checking which startups are trying to build something in this area. Then you connect and support them. Is that correct? (12:21)

Philip: Absolutely. For us particularly, there are several things we find important. We do not invest a lot in the application layer. The application layer includes companies that are wrappers on top of existing LLMs, like Perplexity or Lovable. (12:38)

Philip: As of now, they all have their own models developed, but a year ago they were just building products on top of OpenAI or Anthropic. We try not to back such companies. We believe rather in enabling technologies for AI to speed up adoption. There are several buckets of things we believe in. (13:05)

Philip: For example, we believe in everything around infrastructure for frontier labs like Anthropic or OpenAI. We see that they are running huge amounts of experiments and reusing parts of their code. Such code could be productized, the same way TensorFlow became a standalone product or open sourced technology. There are possibilities to build companies on top of what is currently happening at OpenAI and Anthropic. (13:28)

Philip: If investors say they invest in a very particular idea, they will probably pivot. The same way startups pivot, investors pivot all the time. When AI became big, most investors pivoted, and investors that claimed to invest just in AI pivoted into a specific thesis inside AI. You always update your ideas based on new knowledge. (14:02)

Philip: The same thing will happen to us definitely. For now, we are mostly focused on infrastructure and enabling technologies for big labs. (14:34)

Alexey: If it is not private information, can you tell us some of the startups that are working with you? (14:47)

Philip: You could just check them on our website, as this is not private. There is fascinating traction from companies developing their own AI models, even if they are not frontier scale. They still have a lot of applications and value to their customers. For example, we backed Ex-Human, a company that trains their own LLMs. (14:57)

Philip: They train them not to be robotic, but to be engaging and warm. If you are talking to AI support for a service, you do not want it to sound robotic. You want it to be warm and supportive. Maybe you want to build a game with interactive characters. (15:36)

Emotion Analytics and Voice AI Startups

Alexey: Are these text models rather than voice? (16:00)

Philip: Yes, they are text models. We wanted to back a company in the voice space but we did not find a good example yet. Another good example is a company that provides emotion analytics to various businesses. Their main use case is trying to predict sales call outcomes based on emotions recognized from other participants on a Zoom or Google Meet call. (16:09)

Philip: When I talk to you, I have facial expressions and a tone of voice. I decide how to pronounce things. My excitement, engagement, and where I am looking are all signals that could be collected. They built statistics and AB tested everything, proving that such emotions could predict outcomes for sales calls. (16:45)

Philip: If you are trying to sell high tier equipment and have one sales rep, they help you decide which potential client to invest time in. It appears they solve a huge pain for many companies trying to understand if their sales or support departments are performing well. This is the insight they provide, and they are making good money from it. (17:22)

Alexey: Speaking of voice AI, I have been trying to find a provider that lets me create voices with accents. I know many of them can do voice cloning. If I speak with a Russian accent, it would clone my Russian accent. In ElevenLabs, you can prompt the voice. (17:58)

Alexey: You can describe the voice as a warm male in his late 40s, and then all the things you generate use the same voice. But it does not work with accents. If I ask for a thick French accent, it does not work. I am creating a game and I cannot find a provider that can do this. (18:25)

Alexey: I want characters with different accents, but they all sound American. (18:49)

Philip: They are very bad at this type of work. We are still not there. I think this is an open space. I recently talked to companies that develop models to recognize speech. (19:01)

Philip: Speech recognition is considered a solved problem, but it still recognizes DataTalksClub as Data Dogs all the time. There are a lot of edge cases to fix. If a model is trained for Egyptian Arabic and a person speaks Saudi Arabic, it would fail. Startups for this idea just started to appear recently. (19:23)

Philip: It is too early to have models that speak with accents. I think it could be solved if it is a big market. ElevenLabs will probably generalize enough to capture details and handle accents eventually. (20:04)

Alexey: Considering the market capitalization of ElevenLabs, I think it is a huge market. (20:36)

Philip: I would not have expected ElevenLabs to become so huge when I joined venture capital five years ago. I am not sure if you attribute this mostly to the market and technologies or to their exceptional team. I think they are currently making around 500 million dollars annually. This makes them the biggest company in Poland. (20:43)

Alexey: I think they currently position themselves as a London or New York based company. (21:09)

Philip: They are originally from Poland and I think their top management is mostly from Poland. They have grown huge and are currently a global company. They are opening several offices in Europe, including Spain and Berlin. They are the only product with such high quality voice for a large number of use cases. (21:18)

LLM Impact on Venture Capital LLM Impact on Venture Capital

Alexey: You have been in this VC domain for five years. You mentioned you would not have thought this could be such a big market five years ago. In these five years you have probably seen the industry evolve significantly. What was the biggest shift and when did it happen? (22:09)

Philip: It is a good question and hard to recognize because it accelerates faster every year. The largest shift over VC happened in early 2023 after ChatGPT came out. Everyone understood that this was driving huge attention and daily users. This was already something big and could be the next platform. (22:35)

Philip: Potentially a second shift is happening this year or early next year with recent releases. The first shift made everyone realize modern AI could be a massive industry, maybe as big as SaaS or cloud. Today some people believe this is going to be bigger than any other industry. We are talking about trillions in revenues. (23:17)

Philip: Two trillion is the potential IPO value for Anthropic, and OpenAI is potentially valued at one and a half trillion. I think if they are currently lower than Anthropic, they are going to postpone their IPO and wait to become better to reach a three to five trillion valuation. (24:04)

Alexey: My background is in development, data science, and engineering. When I saw ChatGPT at first, it could write poems, but it was not something I could actually use. Starting from last summer, when the models were good enough to create code, I realized I needed to start using it as much as possible because I saw how much time I could save. It is interesting that Anthropic's valuation is potentially higher than OpenAI's. (24:29)

Alexey: Anthropic's bet was on software engineers, so they focused only on coding. A year ago, Anthropic released Claude Opus. For many people, it was a game changer. Meanwhile, OpenAI was releasing video generation and doing early experiments in other areas. (25:16)

Alexey: Anthropic focused on developers and became the main brand among software engineers. Now OpenAI is actually losing in that space. Do you think it is a coincidence that software engineers picked this trend up so quickly? Why do you think we as engineers were so eager to try it while other industries are still getting there? (25:51)

Philip: I think this is actually quite bad news for Anthropic and others. Engineers are the only hackers in the world. Only for software engineers is it cool not to have onboarding or advertisement. It is cool just to have a GitHub where you take the code, launch it, fix it, and play with it. (27:03)

Philip: You are actually playing the role of an engineer. This will not happen to accountants or professionals in logistics or marketing. (27:28)

Alexey: I think my accountant already uses AI. I can see that from the em dashes in her emails. (27:43)

Philip: They definitely do employ AI in parts of their workflows. But what part of the accounting job is automated already? Maybe a fraction of one percent. She could not tell an agent to do the accounting for the week in one hour while she just double checks it. (27:47)

Software Engineering AI Adoption

Philip: Engineers could theoretically not look at their code anymore, but accountants cannot stop checking their books. (28:19)

Alexey: That is why I still keep paying my accountant. I was thinking whether I am ready to replace my accountant with AI. The fact that they share the responsibility and catch potential errors is important. Their head is also on the line if the tax authority finds a mistake. (28:25)

Alexey: They are invested in making sure everything is fine, and they cannot just blindly trust AI to do their work. (29:17)

Philip: This is the same difference as between autopilot cars in advertisements versus real life. In advertisements, you just relax while it drives. In real life, you are still the driver and responsible for everything that happens. You cannot simply switch off from the driving process. (29:24)

Philip: The same thing applies to accounting. Even if AI could do everything, you still must be involved and double check. In software engineering, AI is so good at self correcting that you could potentially ship its code into production. Even if it fails, you probably just have a short downtime, unless it is mission critical code. (30:04)

Philip: We typically try to compare AI to a hundred percent accuracy, but we must compare it to the accuracy of a typical human, which is maybe 95 to 99 percent. In a lot of cases, AI is going to perform well enough to fully automate some workflows or jobs. (30:36)

Alexey: In software engineering, the penetration is quite wide. Some companies are hesitant, but it is just a question of time. (31:03)

Philip: In AI engineering, the vibe just shifted recently. I personally started hearing from everyone that they are adopting AI. I know someone who runs a small shop providing custom clothing for hotels in the Emirates. It is a very conservative industry, but even they are currently employing AI in their workflows because it is foolish not to. (31:26)

Philip: If you spend a lot of time checking emails, asking Claude to do it saves time. It is very easy to start, and everyone is trying to do it. Depending on the industry, the level of penetration is very different. While software engineering might be 50 percent automated, most other white collar jobs are barely one percent automated. (32:08)

Alexey: Ever since I discovered that with ChatGPT I can ask it to find an email with a PDF, fill it in, and give me the file back, there is no coming back. I am not doing this manually anymore. (32:38)

Philip: I realized I do a lot of repetitive scanning of text over my emails. Now I just go to ChatGPT. For example, I had to scan through a list of people to arrange a time slot for a call. I load 100 profiles, ask it to delete anyone outside the United States, and show me who is available tomorrow. (32:53)

Philip: From those available, I ask it to filter for knowledge in AI infrastructure. Then I ask it to adjust for specific time zones. It is so easy now and saves so much cognitive load. (33:32)

Data vs Algorithmic AI Progress

Alexey: We as software engineers are eager to jump on these things, but it is not the same in other industries. Does this change the kinds of companies you invest in? Since software engineering tools might be saturated, are you looking for the next big thing? (34:08)

Philip: Everyone is now trying to understand what the next big thing is outside of software engineering. AI progress can be divided into two buckets: data progress and algorithmic progress. Algorithmic progress is creating the next generation of architectures, like the next GPT or transformer. Data progress is taking an existing architecture and feeding it more data. (34:45)

Philip: We already have enough data about software engineers. If we load more data about accounting or marketing workflows, even with the same architecture, the AI will become a better marketer or accountant. Even if no algorithmic breakthroughs happen, collecting more data will improve the models. The same thing that happened to software engineers will happen to other industries. (35:17)

Philip: If global spending on accounting equals global spending on software engineering, there is an economic incentive to collect data and provide it to these models. The bad news is that software engineers are the only easy target. Software engineers are hackers who try new things, open source them, and discuss them publicly. Accountants do not do this. (36:06)

Philip: If you find a trick for not paying taxes, you will not discuss it online or open source it. It becomes your intellectual property and your edge. Only in software engineering do you have a vast amount of data available on the internet to learn from. This drives us to the conclusion that we should find industries that are open enough to collect data easily, yet important enough to be profitable. (36:53)

Philip: The next big bet happening in the industry is science. Everyone is trying to automate science and create an automated research intern. They show early proof that it works. We recently heard news about OpenAI and Anthropic making a breakthrough with solving Navier Stokes equations. (37:51)

Philip: They could potentially solve fluid dynamics in some edge cases. This means you would not need huge supercomputers to calculate aerodynamics when creating a car or plane. You could just use an AI formula and have a solution instantly. It could be a game changer. (38:30)

Alexey: Will this allow us to completely bypass traditional testing to create a plane? The processes are different in aviation than in software. (38:58)

Philip: I would not be worried about aviation or car engineering because there are a lot of regulations, licenses, and tests. You cannot just create an MVP plane, sell it to passengers, and fly it. They will test it thoroughly, but AI could enlarge the speed of iterations. In every field of engineering, the iterations will be faster. (39:07)

Classical Engineering AI Acceleration

Philip: If you previously tested a new generation of planes once a year with regulators, maybe you will run tests more frequently. (40:15)

Alexey: You mentioned open industries are ripe for this, and science is one of them. Science encourages open discussion and publishing. Is this focus on science your VC firm's strategy, or your personal observation of where the industry is going? (40:29)

Philip: Of course, we try to find our edge. If we believe science will be automated to some extent in the coming years or months, we look for opportunities for startups to grow there. OpenAI and Anthropic will definitely be bigger, but since we invest in early stage companies, we look for secondary benefits. For example, you could double down on investments in biotech. (41:10)

Alexey: In biotech, a lot of things have changed because of AI. I think Moderna recently had a breakthrough in a cancer cure. (41:56)

Philip: I am not an expert in this field, but I understand they used AI models similar to AlphaFold to originate drug ideas. They still have a very long research pipeline. After designing a potential molecule on a computer, you must test it in a wet lab, in real cells, in animal models, and later in humans. You must obtain all the licenses. (42:02)

Alexey: There is a question about predictions regarding AI handling syntax and code generation. How will the role of AI and ML engineers change over the next decade? Will the demand shift towards system design, or will this automation reduce the overall need for engineers? What is your feeling about that, especially when Anthropic publicly stated they want to replace software engineers? (42:48)

Philip: This is a very deep and important question for all VCs and the industry. AI has already caught the syntax of languages and can write decent Python functions much faster than a human. (44:18)

Alexey: It does not produce code that is always up to my standards, but if it works, who cares? I made peace with the fact that the code is not on par with what I would write, but it gets the job done. (44:52)

Philip: Exactly. It produces code that works, but it has not caught the high level ideas, taste, or architecture that a human would use. It still does not understand the big picture or common sense business needs. It likes to add extra unit tests or cover edge cases that are not actually required by the business. (45:11)

Philip: It tries to write generally scalable code but often overengineers it and generates unnecessary fluff. This high level decision making is still not solved by AI. However, between GPT 3.5 and 4, people reported AI became a much more decent architect. They did something to the training process, so its architectural decisions improved, though it is not yet at human level. (45:56)

Software Architecture AI Limitations

Alexey: We can argue that if we take all software engineers and look at their average architectural understanding, models might already do a better job than most developers. (46:12)

Philip: I think it does a very poor job of translating real world business needs into code. It does not understand what you actually want to build and it does not ask much. I would be much more relaxed if my project were in Python because I could interpret it better, but AI does not care. It could introduce extra dependencies or unnecessary layers. (47:07)

Philip: We have seen AI become a better architect compared to earlier models. My personal belief is that this high level thinking, taste, and architecture are nothing magical. It is not a unique human trait that computers will never achieve. We have seen them progress immensely in front end development and design recently. (47:46)

Philip: People who say humans will always own taste and strategic thinking are likely wrong. Taste requires a lot of iterations, data, and experience, which AI does not have right now. You cannot easily collect a large amount of taste from open internet data. (48:43)

Alexey: Taste is very personal and opinionated. The model learns the average, but design decisions live in your head after working in the industry for years and seeing both poor and good decisions. (49:22)

Philip: This is just another mode of data which we could eventually collect and feed to an LLM. I would not bet that it will not happen, but it will happen later because it is not easy. Right now, labs are working on low hanging fruit in software engineering. They are just beginning to shift focus to science because the basic problems of software engineering are mostly solved. (49:39)

Alexey: Let us say you have a brother about to go to university. If he is considering software engineering versus classical engineering, what would your advice be? (50:20)

Philip: It really depends on personal preference. If your goal is only to be paid well after graduation, I would go with classical engineering because it is harder for AI to catch up. I believe it is a more fundamental field. (50:40)

Future Software Engineering Careers

Alexey: Right now you do not need to study for five years to be a developer. The need for studying traditional coding is less because you need to act more like a product manager to create software. You need to be precise in what you want and have good taste. (51:11)

Philip: I am not sure about that. Maybe you should study more because the job of a junior software engineer can be done by AI. What cannot be done is the job of someone who completed their masters and tackles high level, long running projects. These complex tasks are not yet open to AI. (51:36)

Philip: Maybe you should go obtain your PhD because AI will not solve PhD level jobs easily. To be clear, OpenAI potentially solved the Navier Stokes equations, but they did it with a team of mathematicians combined with AI. They have judgment, taste, and experience that can only be met in human life. (52:00)

Alexey: If I ask my tattoo artist neighbor to use Claude to solve a theorem, he will not be able to do it because he does not know what it is or care. AI is nothing without someone knowledgeable driving it. (52:38)

Philip: Automating the job of a software architect or scientist is so materially beneficial that it will eventually be solved. If you solve software architecture, you could recreate entire custom CRM or ERP systems instantly. Because the incentive is so large, this very hard problem will be solved at some point. However, humans will always be one layer higher, overseeing and babysitting the AI. (53:46)

Philip: At this point, you could act as a product manager, but it is better to be a product manager with a physics degree than just a product manager. This is my opinion. (54:59)

Alexey: We have questions about how to get a job and what skills or projects are needed. My advice is always to build projects. Now that everyone can use AI, do you think it is different? What is your advice for people starting or switching their careers? (55:05)

Philip: Among the startups I talked to, they have switched to hiring with more precision and focusing on more senior people. They want someone who can build more autonomously with an army of AI agents. My intuition is that the overall job market will not change much in terms of volume, but higher level skills will be in more demand. There will be no more need to just type Python syntax. (55:52)

Philip: Go to university, build projects, and obtain skills. You will just start from a higher level. A hundred years ago, people started with basic calculus and manual calculations. When computers arrived, people adapted and learned higher mathematics instead of basic arithmetic. (56:58)

Systems Thinking in DevOps

Philip: The same way, nowadays we will not focus on exact programming languages or syntax tricks. We will focus on higher level concepts that are even more essential. You will learn them and get there faster because lower level problems will be solved for you. If I want to learn a DevOps role, I do not need to memorize YAML manifests. (57:55)

Philip: I can ask ChatGPT to write the manifest. However, I need to understand the high level idea of how the system works, the design principles, and the trade-offs. (58:44)

Alexey: In DevOps, you need to understand how the system functions and the automation levels. If there is a bug, you need to know that rolling back is an option and verify it works, even if you do not know the exact syntax for rolling back. (58:59)

Philip: You need to know that this feature should work and understand why it is a good practice. You should understand the high level ideas of what should be done, such as the design principles applied at large companies versus small ones. You should have these strategic things learned and experienced. (59:24)

Alexey: This is a perfect segue to our AI DevOps course that started last week. I show how to implement DevOps concepts without going into specific syntax. You can use whatever tool, and I provide prompts so you understand the concept without memorizing every YAML line. (1:00:21)

Alexey: I never liked Terraform syntax, and now I can finally ignore it. The important thing is knowing what is there and how it functions. Does AI make you feel like your job is less or more demanding? I feel like I am doing more because the feedback loop is so fast. (1:00:53)

Alexey: I implement one idea and instantly have ten more. Previously, one specific task took a week or two, but now I can iterate quickly. My 40-hour work week easily turned into 60 hours because I want to keep building. (1:01:25)

Philip: Exactly. You run more experiments, implement more ideas, and learn faster. Everything just became more interesting. (1:01:58)

Alexey: Philip, we ran a bit out of time. Thanks a lot for joining today and thanks everyone for asking questions. If somebody has a question for you, what is the best way to find you? (1:02:08)

Philip: I am mostly available on LinkedIn. Also, if you are working on a cool startup idea, feel free to send it to me. (1:02:31)

Pitching Deep Tech Startups

Alexey: What kind of startup idea gets your attention? Can it be any idea in the AI space? (1:03:00)

Philip: The things that will drive most of my attention are hard, science backed ideas. It could be a new architecture, training your own models, or building a complex genetic system with several models. If your idea can be generated in one shot by GPT or Claude, it is unfortunately not a good idea for us. Work on harder things. (1:03:16)

Alexey: We will have a course about generating and working on startup ideas soon. That might be something VC companies would be interested in. I will need to invite you for another conversation then. (1:03:55)

Philip: I would be happy to join. (1:04:15)

Alexey: Thanks Philip. Thanks everyone. It was a big pleasure. See you around and enjoy sunny Barcelona. (1:04:23)