Question 1

GPT‑4o signals a shift from monolithic models to continuously updated multimodal systems.

Accepted Answer

Friedberg explains that we’re moving away from massive, infrequently released models toward systems of smaller, specialized models stitched together (math, images, video, etc.) and tuned continuously. GPT‑4o (‘Omni’) embodies this: it ingests text, audio, images, video, and desktop context in parallel, is significantly faster and cheaper, and still benchmarks slightly better than GPT‑4 on Stanford’s MMLU—though it trails Claude 3 Opus on that specific chart. For builders, it means model identities blur into a general ‘AI service’ that’s updated all the time.

Question 2

Founders must ruthlessly decide where AI model progress ends and their product’s defensible value begins.

Accepted Answer

Sacks warns that many startups spent years building ‘AI customer support agents’ only to see GPT‑4o instantly outclass their conversational quality and speed, wiping out hundreds of millions in R&D. The lesson: if you build features that the frontier models will soon provide natively, you will be obsoleted. Durable value now lives in workflows, integrations, domain-specific data, UX, and distribution—things the model won’t commoditize away in six to eighteen months.

Question 3

Glue exemplifies ‘AI‑native’ enterprise software by rethinking the core object of communication, not just adding a bot.

Accepted Answer

Glue replaces Slack-style channels with topic-based threads—mirroring how ChatGPT organizes chats—so people subscribe to the conversation they care about instead of noisy rooms. Glue AI is embedded, not bolted on: it can query company data (like All‑In transcripts or Craft’s deal flow) via RAG, summarize positions, identify experts, draft outreach (like a tailored letter to Lina Khan), and provide sourced answers inside the main chat. Sacks’ long-term vision is ‘promptless’ AI that automatically chimes in when it can help, once models get good enough at intent detection.

Question 4

Ohalo’s ‘boosted breeding’ could massively increase crop yields and shift where food can be grown globally.

Accepted Answer

Friedberg unveils a five‑year, ~$50M R&D effort: by applying proteins that switch off the genetic ‘halving’ process during plant reproduction, parent plants each pass 100% of their DNA to offspring, doubling chromosome sets (polyploidy) in a controlled way. In Arabidopsis they saw 40–50%+ yield increases; in potatoes, two small parent lines (33g and 9g tubers) produced a boosted offspring with 682g of potatoes—far exceeding current market varieties. Beyond yield, it enables consistent, genetically stable seed in crops like potatoes that today must be propagated clonally, and it can combine traits for drought/disease resistance and adapt crops to new geographies (e.g., growing staple crops in currently marginal regions).

Question 5

AI and globalized talent are driving a deep deflation in startup and operating costs, changing company design.

Accepted Answer

The besties note that the same startup that needed $3M to build a product five years ago can now plausibly do it with $500K. Calacanis describes his largest-ever seed check into Athena, a service that pairs U.S. execs with highly trained offshore ‘EA/ops’ staff at a fraction of domestic cost—then augments them with AI. JCal’s internal ‘ADD’ framework (Automate, Deprecate, Delegate) pushes teams to use AI (automation), stop low‑value work (deprecate), and offload repeatable work (delegate). Net effect: orgs stay small, per‑employee leverage soars, and many more niche products become viable businesses.

GPT-4o launches, Glue demo, Ohalo breakthrough, Druck's Argentina bet, did Google kill Perplexity?

AI Upheaval, Super-Potatoes, and Argentina’s Radical Free-Market Bet Explained

Key Takeaways

Notable Quotes

Questions Answered in This Episode

Transcript Preview

Get more from every podcast