Bold calls, on the record

Every falsifiable prediction the smartest podcasts made, dated and quoted. 154 tracked so far. We grade them as their horizons pass, and this record cannot be backdated.

• There will be a rapid shift towards agent-native identity, where agents will use specific 'personas' with segmented access.
  On Latent Space · within a few years (e.g., by 2029) · source ↗
• AI agents will transform mechanistic interpretability (mechan) from an ad hoc process into a true science by automating hypothesis testing.
  On Latent Space · within 5 years · source ↗
• Coding agents will enable the routine generation of highly secure code that would be too complex or tedious for human engineers.
  On Latent Space · within 5 years · source ↗
• Companies that move fast and break things with AI infrastructure without community buy-in will invite future scrutiny and regulatory hurdles.
  On Latent Space · in the future · source ↗
• The primary engineering task for AI products will shift to removing complexity to improve user experience.
  On How I AI · within 3 years · source ↗
• Continuous Integration (CI) will become a strategic investment for velocity in AI product development.
  On How I AI · within 3 years · source ↗
• The AI product development cycle will become a constant loop of refining evaluations.
  On How I AI · within 3 years · source ↗
• Engineers who offload tasks below the agent line will consistently enter a 'maker schedule'.
  On How I AI · Once engineers offload tasks below the agent line using AI. · source ↗
• It will become significantly harder for US AI companies, including ambitious startups, to recruit and retain the highest-quality international talent.
  On TBPN · within 2 years · source ↗
• An AI oligopoly will form, dominated by hyperscalers (Amazon, Microsoft, Google) acting as trusted, regulated gatekeepers.
  On All-In Podcast · by end of 2031 · source ↗
• AI CEOs' public 'doom trolling' will inadvertently cause increased government intervention, leading to centralized control of AI by larger entities.
  On All-In Podcast · by end of 2031 · source ↗
• Future market leaders in semiconductors will not win by being generalists.
  On No Priors · by 2030 · source ↗
• Future market leaders in semiconductors will hyper-focus on one niche area, forge smart strategic partnerships, and deliver complete, full-stack solutions, from hardware to software.
  On No Priors · by 2030 · source ↗
• Intel will transform into an organization where AI is embedded across design, engineering, and every part of the business.
  On No Priors · by 2028 · source ↗
• By 2030-2032, people will begin to understand the significant product potential of Intel.
  On No Priors · By 2032 · source ↗
• The AI infrastructure boom will undergo a massive, sustained expansion.
  On 20VC with Harry Stebbings · within the next 3-5 years · source ↗
• Compute demand will grow exponentially.
  On 20VC with Harry Stebbings · within the next 3-5 years · source ↗
• The amount of AI compute needed will increase by tens or hundreds of times.
  On 20VC with Harry Stebbings · in the coming years · source ↗
• Cheaper AI intelligence will increase consumption, rather than reduce it.
  On 20VC with Harry Stebbings · going forward, as AI intelligence becomes cheaper · source ↗
• Enterprises will require ten times more data storage within the next three years.
  On All-In Podcast · by 2029-06-14 · source ↗
• Within the next five years, "system of work" SAS will undergo a radical re-engineering.
  On All-In Podcast · within 5 years · source ↗
• Google will spend a billion dollars a month with XAI.
  On My First Million · monthly, ongoing from Q3 2026 · source ↗
• SpaceX will build data centers in space.
  On My First Million · in the future, as part of their space compute vision · source ↗
• Starship will achieve unprecedented rapid reusability.
  On My First Million · in the future, enabling the space compute vision · source ↗
• OpenAI and Anthropic's combined compute capacity is projected to reach 10 gigawatts by the end of 2026 and 20 gigawatts by the end of 2027.
  On TBPN · by end of 2026 and by end of 2027 · source ↗
• Data center moratoriums would "overnight" cede global AI leadership to China.
  On TBPN · overnight (immediately after moratoriums are implemented) · source ↗
• Data center bans would halt US GDP growth.
  On TBPN · soon after bans are implemented · source ↗
• A tiny incremental cost will transform multi-billion dollar facilities, such as data centers, into welcome community assets.
  On My First Million · within 5 years · source ↗
• AI's next frontier will be in the physical world, encompassing robotics, manufacturing, and industrialization.
  On Lenny's Podcast · short to medium term · source ↗
• The shift of AI into the physical world will lead to unprecedented demand for physical components.
  On Lenny's Podcast · as the shift of AI into the physical world occurs · source ↗
• Memory prices for consumer hardware, robotics, and physical AI will significantly increase.
  On Lenny's Podcast · within the next 1-3 years · source ↗
• Early advancements in AI for CAD will more probably come from hobbyists or on-premise AI training systems than from large, established incumbents.
  On Lenny's Podcast · within 5 years · source ↗
• An LLM inference system operating without batching user requests will incur a per-token cost that is approximately one thousand times higher than an optimized system that utilizes batching.
  On Dwarkesh Podcast · Indefinite · source ↗
• Any future LLM service that experiences only a few sporadic user requests will inherently operate at smaller batch sizes, necessitating a sacrifice of cost efficiency for instant response.
  On Dwarkesh Podcast · Indefinite · source ↗
• Pipelining will dramatically reduce the memory capacity needed per rack for storing model weights.
  On Dwarkesh Podcast · within 1 year · source ↗
• Pipelining will not significantly reduce the memory footprint for the KV (Key-Value) cache.
  On Dwarkesh Podcast · within 1 year · source ↗