The AI industry's 'next big bet,' and the DOJ tells Google to sell Chrome | Equity Podcast

Google, Chrome, and the DOJ

• The Department of Justice has argued that Google should divest its Chrome browser as part of a remedy to break up the company's alleged illegal monopoly in online search, although this does not necessarily mean it will happen and the issue will be resolved through court hearings in the future (1m34s).
• The suggestion that the Onion buy Chrome has been jokingly proposed, with some even suggesting that the press could lend the Onion the money to make the purchase (2m4s).
• It is unlikely that Chrome will be completely disconnected from Google, but the threat of this happening could lead to Google making concessions, such as siloing Chrome and disconnecting tracking cookie technology (2m35s).
• The goal of the Department of Justice's actions is to improve search quality, which has been in decline, and to potentially break Google's monopoly in the process (3m11s).

Whiz and Dasher Acquisition

• Whiz has turned away from a possible acquisition by Alphabet and has instead acquired Dasher for $450 million to expand its cybersecurity platform, a move that is seen as smart given the complaints around Google and Alphabet's monopolistic behavior (3m39s).

Tekever and Dual-Use Drones

• A company called Tekever raised $74 million for dual-use drones, with investments from the NATO Innovation fund, and being based in Portugal is seen as a competitive advantage in the European Defense Tech scene (4m27s).
• European countries tend to prefer giving contracts to European companies, making it difficult for US-based companies to break into the market, and this trend is also seen in other regions such as Asia (4m54s).
• The development of dual-use drones is expected to lead to a convergence of features and technologies, with companies potentially adopting similar protocols and designs (5m28s).
• Tekever is developing a large drone that can carry and release a swarm of tiny drones, a technology that is already being developed by other companies (6m2s).
• The NATO Innovation fund's investment in Tekever raises questions about the fund's focus on European companies, despite NATO including the US as a member (6m17s).

Defense Tech Boom

• The US Department of Defense has a large budget for investing in dual-use technologies, including autonomous vehicles and drones (7m3s).
• There is a European Defense Tech boom, with a growing number of deals and investments in the sector, particularly in parallel with the US defense Tech boom (7m24s).
• European countries are investing heavily in the defense space, particularly in developing their own drones, due to their proximity to conflict zones and the need for local solutions (7m34s).

Converge Bio and Biotech LLMs

• Converge Bio has raised $5.5 million in a seed round to develop an "everything store" for biotech language models (LLMs), which are trained on large amounts of data but often lack specificity (8m10s).
• Converge Bio's goal is to provide post-training services for existing biotech LLMs, making them more specific and useful for companies in the industry (9m27s).
• The company's business model involves working with biotech companies to fine-tune LLMs using proprietary data, making them more useful as a tool for researchers (10m43s).
• Converge Bio's service is a mix of bespoke and automated solutions, where companies can come to them with specific needs and data, and they will fine-tune the LLMs to meet those needs (10m19s).
• The company plans to charge clients, such as pharmaceutical research companies, for their services, with prices potentially in the range of $1 million (11m4s).
• There are already companies working in this space, but Converge Bio aims to provide a unique solution for biotech companies looking to leverage LLMs (11m12s).
• Converge is different from competitors like Goo Bioworks in that it allows customers to use foundation models on their own data and own everything about it, rather than providing a full-service solution (11m33s).
• Converge thinks it can acquire earlier customers by offering a more affordable solution, such as fine-tuning a model for $100,000, rather than spending $1 million a month (12m8s).
• More than half of biotech companies have no experience with machine learning and foundation models, so Converge is trying to get in early with these companies and lock them in as customers (12m22s).

LLM Risks in S1 Filings

• The risk factor section of an S1 filing can reveal interesting information about a company's potential risks, such as the risk of damage from large language models (LLMs) (13m5s).
• Cloud service startup Service Titan included a risk factor in its IPO filing about the potential damage from LLMs, specifically naming Microsoft and Open AI (13m21s).
• This risk factor is not commonly seen in S1 filings, but it may become more standard language in the future (13m53s).
• The risk factor is not just about LLMs creating hallucinations or incorrect information, but also about the potential for security breaches and the sharing of customers' private data with third-party companies (14m13s).
• The use of AI and LLMs introduces unknown unknowns and novel ways of attacking, which makes it important to implement carefully (14m46s).
• The AI industry is expected to see more filings that mention the risks associated with new and weird technologies, even if those risks are not fully understood, as a way to cover all possible risks (15m2s).
• The inclusion of these risk factors in filings could potentially lead to the development of safer AI models, as companies may be forced to find solutions to mitigate these risks, such as using AI models designed to be safer and less prone to hallucinations (15m24s).
• Even with solutions available, companies will still include these risk factors in their filings as a precautionary measure, but this could lead to the development of standard solutions to address these risks (15m44s).
• The mention of these risks could lead to the adoption of mitigation strategies, such as retaining a pen testing startup to conduct independent testing and provide quarterly reports on vulnerabilities (16m14s).

Puerto Rico and Innovation

• Puerto Rico is being promoted as a hub for innovation, with a highly skilled and bilingual workforce, competitive tax incentives, and a vibrant ecosystem for startups and global players (17m5s).

AI Scaling Laws and Test Time Compute

• The current state of AI scaling laws is showing diminishing returns, forcing AI labs to change course, and this will be discussed further (17m37s).
• Scaling laws, such as Moore's Law, refer to the idea that certain technologies will improve at a predictable rate, enabling new applications and advancements, and AI has its own scaling law related to the number of parameters in AI models (17m50s).
• Puerto Rico is being promoted as a hub for innovation, with a highly skilled and bilingual workforce, competitive tax incentives, and a vibrant ecosystem for startups and global players (17m5s).
• The AI industry has been scaling AI models by increasing the number of parameters and compute power, but it's unclear whether this approach will continue to yield exponentially better results or eventually plateau, with some people experiencing only 1% better AI despite increased compute power (19m6s).
• The industry is running out of data and facing high compute costs, making it difficult to achieve significant performance gains by simply making models bigger (19m32s).
• Researchers are exploring alternative methods to improve AI models, including "test time compute," which involves allowing models to "think" about answers before responding, similar to how humans break down complex questions into smaller steps (19m52s).
• This approach has shown promise in improving the reliability and accuracy of AI models, and Open AI's 01 model is an example of this concept in action (20m38s).
• The idea of test time compute is being marketed as the AI industry's next big bet, with potential applications in various fields (21m23s).
• The shift towards incremental improvements in AI could have implications for the perceived AI arms race between the US and China, potentially rendering the idea of an AI cold war obsolete (22m2s).
• Some experts, including those in the Venture side, are exploring the implications of this shift and how it may affect the AI industry and related fields, such as foreign policy and defense (21m36s).
• The AI industry's success is being questioned, with concerns that it may be solely dependent on the amount of money and resources invested, specifically in building data centers and acquiring high-performance computing hardware like H100s, rather than research and ingenuity (22m9s).
• If the success of AI is just a matter of capital investment, it becomes a capital race, and there's no guarantee that any particular country or company can stay ahead of others, such as China, which is reportedly about six months behind in AI development (22m50s).
• The concept of "test time compute" is unclear, and its scalability is uncertain, which could lead to a significant advantage for those who can figure it out before their adversaries (23m16s).
• If test time compute fails, there is a possibility that AI chips specializing in high-speed inference could become more prominent, with one A16 partner suggesting that they could see a dramatic increase in demand (23m25s).

Network States

• A network state is an internet community seeking to establish a physical presence, often by buying land and creating special economic zones, with the ultimate goal of forming their own country (23m59s).
• One notable example of a network state is the project led by Balaji Srinivasan, who has raised over $500 million in funding to build a city and has been attempting to convince countries to allow him to establish a city, with his most recent attempt being in Greenland (24m39s).
• This is not the first attempt at creating a network state, with previous waves of activity centered around cryptocurrency, and it is unclear whether this is the second, third, or fourth wave of such attempts (25m24s).
• Balaji Srinivasan has made previous attempts to buy uninhabited land, although it is unclear whether he has tried to do so in Greenland before or in other areas (25m31s).
• A person has been traveling to around 30 countries, spamming local government officials, and offering to build a city in exchange for permission, with the Mediterranean and Caribbean being potential locations, and a recent event in the Dominican Republic possibly hinting at the chosen location (25m44s).
• These projects are seen as "resorts for rich people" with features like Bitcoin vending machines, rather than actual utopias, as they prioritize the wealthy over the working class and do not address the needs of the lower levels of society (26m26s).
• To create a true utopia, it is suggested that one should start by overpaying and overcompensating the ground-level workers, rather than catering to the rich, as the current approach is seen as merely creating luxury resorts or hotels (26m45s).
• The concept of "Network states" is mentioned, where rich individuals buy luxury land and create exclusive communities, which is seen as being more capitalist than utopian (27m5s).
• A conversation with Jen is mentioned, where he expressed feeling emboldened by the Trump presidency and its potential to help these Network states, despite them being located abroad (27m31s).
• The existence of successful Network states is questioned, with one example being Prospera in Honduras, which allows businesses to register and choose from 36 countries for regulation, but is seen as a "total scam" or a tax haven catering to the rich (28m0s).
• Prospera is also mentioned to be involved in medical testing, with a notable example of a person getting a Tesla key implanted (29m4s).
• Some people have embedded devices in their bodies, such as Tesla keys in their arms, and there are businesses that offer implantation services, although these are not FDA approved (29m10s).
• The rise of crypto may lead to more "Network States," which are communities that form around a shared interest or goal, often facilitated by blockchain technology (29m27s).

Equity Podcast

• Rebecca Balan had a conversation with the founder of blockchain analysis startup Chainalysis about the state of crypto after the election, which may provide interesting insights (29m46s).
• The podcast "Equity" will be back next week and can be found on social media platforms under the handle @EquityPod (30m16s).
• The podcast is produced by Teresa Consolo, edited by Kell Bryce, and illustrated by Durban, with thanks to the audience development team and Henry Pick who manages TechCrunch audio products (30m25s).