CHAPTERS
- 0:00 – 0:30
Rapid model releases reshape expectations for 2024
Elad and Sarah open with a host-only check-in on how quickly the AI landscape is changing. They discuss the recent wave of model launches and why this year is likely to end with multiple GPT-4–class options available to builders.
- •Host-only episode to synthesize recent AI shifts
- •Many new model releases (e.g., Claude, Grok, DBRX) changing the narrative
- •Expectation of several GPT-4–level models by year-end
- •Implications for builders as capability becomes more widely accessible
- 0:30 – 2:10
Scaling laws, cheaper training, and the “Mosaic’s Law” viewpoint
Sarah argues the cost-to-capability curve is improving fast, citing Databricks’ perspective that comparable models should get dramatically cheaper each year. They contrast how much can be done at “GPT-4 level” versus what remains uncertain at the true frontier.
- •Open-source and smaller-compute efforts are closing gaps
- •Databricks/DBRX as evidence of strong capability with less compute
- •“Mosaic’s Law”: capability requiring ~1/4 the capital each year (claimed)
- •Frontier scaling may still remain compute-dominated and concentrated
- 2:10 – 3:10
Value capture shifts: frontier oligopoly vs a broad model ecosystem (and clouds win)
Elad frames a likely near-term oligopoly at the very cutting edge due to capital needs and bootstrapping advantages from strong models. Underneath that, many viable models will proliferate, and cloud platforms may capture significant value by hosting and monetizing usage.
- •Frontier advantage compounds as models help build next-gen models
- •Below-frontier diversity of models expands options for enterprises
- •Clouds potentially capture margin by hosting many competing models
- •Value capture beyond model labs is under-discussed
- 3:10 – 5:18
Microsoft–Inflection deal: product leadership, hedging, and compute sponsorship
They analyze Microsoft’s Inflection move as both a push for AI-native product/research leadership and a strategic hedge against reliance on external model providers. The discussion highlights how next-generation model spend may require hyperscaler-level backing.
- •Microsoft wants AI-aware leadership to drive Copilot and broader products
- •Copilot traction exists, but ambitions extend across productivity + search
- •Deal reads as a hedge to reduce dependence on outside labs
- •Future model scale may require deep-pocket sponsorship vs independence
- 5:18 – 7:04
Voice cloning and the release dilemma: capability vs deepfake risk
Sarah raises OpenAI’s voice cloning and what happens if top labs have strong voice/video/image models. Elad explains why many players may be holding back due to regulatory and societal risks, and suggests possible verification/attestation mechanisms.
- •Voice cloning seen as imminent commodity capability among top labs
- •Primary blocker: deepfake risk and societal/regulatory concerns
- •Potential mitigations: attestation and verification flows
- •Surprising limited competition given open-source and prior art
- 7:04 – 8:04
Who funds foundation models now: VCs vs hyperscalers and strategic capital
Elad argues most real scale funding comes from hyperscalers and big tech rather than traditional venture capital. Cloud providers have strong incentives to keep investing because AI drives substantial cloud revenue and utilization.
- •VC checks can be big, but hyperscalers supply the true billions
- •China mirrors the same dynamic: incumbent internet giants fund models
- •Cloud ROI flywheel: AI workloads directly drive cloud revenue growth
- •Expect continued funding where it increases platform utilization
- 8:04 – 9:34
Beyond language models: the next frontier of specialized foundation models
They predict fewer brand-new general LLM entrants from pure venture funding, but accelerating investment into foundation models across other modalities and scientific domains. Elad lists areas like music, video, biology, materials, physics, and robotics as emerging arenas.
- •Fewer new general LLMs from VC/angels going forward
- •More funding for music, TTS, image/video generation
- •Scientific/industrial models: biology, materials, physics, robotics
- •Pattern: early VC rounds followed by strategic/industry backers
- 9:34 – 12:16
Picking domains that matter: missing capabilities, time series, robotics, and biotech
Sarah outlines a framework: look for capabilities still missing (e.g., richer reasoning over time series) and domains where foundation-model approaches are suddenly showing leading results. She emphasizes data collection, simulation, and value-chain choices in robotics and other hard industries.
- •Time series + reasoning as a major unsolved, high-value capability
- •Applications: anomaly detection, monitoring, security, healthcare, behavior
- •Robotics/biotech seeing renewed momentum as smart teams converge
- •Key bottlenecks: embodied data, simulation vs real-world collection, vertical vs horizontal approaches
- 12:16 – 16:54
Whitespace and memetic startup behavior: why some obvious areas stay empty
Elad describes a paradox: many teams swarm the same “hot” areas while other promising opportunities remain underserved. He attributes it to memetic dynamics in startup waves and the challenge of finding the right product packaging even when the core meme is correct.
- •Large perceived whitespace despite crowded clusters of teams
- •Memes guide what founders build; often correct but incomplete
- •Historical analogies: many photo apps before Instagram; many search engines before Google
- •Hard domains require stronger product substantiation and execution
- 16:54 – 20:26
Video AI as a product and control problem (not just a modality)
Sarah argues commercial video is less about ‘solving video’ and more about controllability, interfaces, and production workflows. She uses HeyGen’s avatar-driven growth to illustrate how narrowing the creative degrees of freedom can unlock user adoption and practical utility.
- •Demand for high-quality, shareable video is effectively unbounded
- •Key gaps: controllability, length, quality, and workflow integration (A-roll/B-roll)
- •Commercial viability depends on deep product + targeted research
- •HeyGen example: avatar video, cloning/spokesperson use cases, editing/replace-speech workflows
- 20:26 – 22:27
Agentic user experiences: Devin-style transparency and human-in-the-loop control
Elad explains why agent UIs are shifting beyond chat/autocomplete toward transparent, multi-pane experiences that show plans, actions, and artifacts. He predicts this is an intermediate phase: today’s agents need supervision like a junior intern, but UI may recede as reliability improves.
- •Devin popularized a multi-tab agent UI (plan/shell/code/chat)
- •Users want visibility and steering, not ‘black box’ waiting
- •Human-in-the-loop supervision as the near-term operating model
- •Long-term: better reasoning/base models may make UI less necessary
- 22:27 – 26:02
Prosumer-first AI adoption: why early winners look like consumer tools
They argue the first big wave of application AI is structurally prosumer because individuals adopt faster than enterprises. Sarah cites Canva’s revenue mix and notes that strong AI products can generate their own distribution, later expanding into professional and enterprise use cases.
- •Enterprise adoption is slower due to security, process, and risk constraints
- •Prosumer apps deliver immediate $10/month-type value and scale quickly
- •Examples: ChatGPT, Perplexity, HeyGen, Suno (mentioned)
- •Prosumer products can ‘grow upmarket’ into larger professional markets
