It's Six Months In: Misconceptions About AI, Regulation, Context Windows and Incumbents

Asked what people get most wrong, Gil names three things, the first being that many treat this as an extension of the last decade of machine learning, the world of convolutional and recurrent networks. He says the field has shifted to two very different architectures: diffusion models for images and large language models for text. These bring new capabilities, such as chain-of-thought-style processing and the ability to act on and synthesize information, which NLP never had.

His second point is timing. ChatGPT launched about six months earlier and GPT-4 about three months earlier, so when people say enterprises aren't adopting much yet, he answers that six months is one planning cycle for a large company. He relays a conversation with a growth investor who had met 100 companies, judged the area not enterprise-ready and dropped it.

Gil predicts others will reach the same conclusion and abandon the technology, as in any wave of enthusiasm. He nevertheless expects more hype rather than less, since some products are starting to ramp revenue quickly. Early viral apps that rose and then faded were still signs of real traction and usage.

Gil's third misconception is the industry's mad rush to call for regulation, which he finds unusual and naive. He says people imagine regulators will convene a panel of industry experts, but in practice an agency is created and then acts independently. On the nuclear analogy, he notes that once the NRC existed, no new nuclear designs were approved for 50 years.

The hosts discuss nuclear power's safety record. They say existing reactors have very low death rates, that nuclear provides a significant share of power in the US, Japan and France, and that regulatory costs have meant that newer and safer reactors aren't being built. Gil says the Fukushima toll was overblown and fears the same is happening with AI.

Gil describes himself as a long-term doomer who believes there is real risk, but says that in the short run people are bundling hate speech and bias with existential risk. He notes that people in crypto, healthcare and edtech know the potential for capricious regulation. He stresses that AI could bring global equity in healthcare and education, and that misregulation could take things down a dark path.

Guo argues that models encode a large amount of public knowledge and that withholding even a flawed version from as many people as possible makes no sense, tying this to education and universal information access. She names code generation as one of her most exciting areas, since Copilot is among the few products with real workflow impact and revenue. But she says we are early: Copilot looks at only a few recent files and does function completion.

She imagines a junior-developer-style AI that takes any Jira or Linear ticket, makes a best-effort attempt with the given context and submits a pull request. The hosts mention several companies exploring the area. On context windows, Guo says the research headlines about huge windows overstate things, and that context will expand to fit the window because attention's quadratic cost keeps efficient context a continuing investment.

Her example comes from a hackathon for new-grad builders, where feeding in Kubernetes documentation alone hit the limit quickly. Gil asks where the asymptote lies; for private-data use cases, a very large window could hold all of one user's data without cross-talk. Guo says she wouldn't bet against a billion-token window, but that we aren't near the point where people stop worrying, and ordering and structure of context remain under-researched.

On China, Gil says the shift is expected: the Chinese internet has long been walled off, producing local champions like WeChat and Baidu. He cites MiniMax raising 250 million dollars at a 1.2 billion dollar valuation and Baidu announcing a 145 million dollar AI fund. It would be surprising if China didn't invest heavily in local LLM champions, and he ties this to national security and to how the US should think about regulation.

Guo adds that MiniMax's team came from SenseTime, a facial recognition company that worked with the Chinese government, and that this raises the question of what "AI with American values" would mean in contrast. On sanctions, she says Nvidia is the engine of everyone's progress, but if China can't get top GPUs with good interconnect it will invest heavily in domestic alternatives, such as Huawei's. That isn't trivial but is solvable with talent and capital.

She argues this makes some attempts to control model development misguided, since sanctions encourage a domestic industry over the long run. Separately, she sees room for innovation in making heterogeneous compute work for both training and inference, from companies like Together and Foundry to compiler-layer players.

On the flood of announcements from Microsoft and Google, one host says building on others' platforms has always carried risk, citing Facebook games, G Suite and Shopify; Klaviyo shows a business can still be built. The question is whether you create enough value on a broad enough platform. She notes Salesforce's pattern of announcing products that rarely see broad deployment and says large companies may need 12 to 18 months to show results.

Gil says the impact is overstated and expected. Incumbents sometimes fail, like Google Plus, and sometimes succeed through distribution and cross-sell, like Microsoft Teams. Startups should ask what happens if an incumbent adds the feature and whether it matters if it takes one year versus three. He was most surprised by how fast Adobe moved, saying adoption speed is almost a measure of management competence.

He also sees an opening against broad, integration-heavy systems such as Salesforce and ERP vendors. AI lets you build products that interact richly and generate connectors quickly, cutting the consulting-heavy integration work that makes rollouts of systems like SAP slow. He believes customizations could also be copied over more easily.

Gil wants to see integrated solutions for the enterprise AI stack, since companies moving between models or using open-source ones need trust and safety, prompt management and more, and the market is full of point solutions. Guo is exploring decision-making over technical enterprise data, such as security incident triage and root-cause analysis in devops. She says current SIEM products are largely useless due to false-positive and false-negative rates.

On incumbents versus startups, she says it depends on who owns the data. Large enterprises run 200-plus security products feeding the SIEM, so a new player isn't necessarily weaker, whereas Datadog collects its data natively and is well placed. She also raises annotation, synthetic data and anonymization as persistent blockers, since incumbents struggle to use customer data given privacy and contracts, though she isn't sure how large the market is.

Gil adds that bio people still emphasize data without appreciating the new capabilities. Guo closes with one company shape: automatically turning real-world economic events, such as an invoice being paid, into robust updates to financial and customer databases. She calls it a big ask but potentially very disruptive and says she has seen teams begin to try.