CHAPTERS
- 0:00 – 0:28
Episode 100 milestone & why DeepSeek dominates the news cycle
Sarah opens the 100th episode and immediately tees up DeepSeek as the main topic amid a busy few weeks in AI. They frame the discussion around what’s genuinely new versus what’s predictable on existing trend lines.
- •100th-episode context and rapid-fire AI news backdrop
- •DeepSeek introduced as the main story of the past month
- •Set up: separating hype/narrative from technical and economic reality
- 0:28 – 2:25
DeepSeek’s significance: open-source SOTA, cost claims, and the ‘mystery’ narrative
Elad breaks down why DeepSeek captured attention: strong reasoning performance, open-source release, and headlines about surprisingly low training cost. He adds a third factor—speculation about who’s behind it—and argues parts of the reaction were overstated.
- •Three drivers of attention: capability, openness, and alleged low cost
- •Novel RL techniques in the paper and real technical contribution
- •Cost narrative: final run vs total R&D/iteration spend
- •Market reaction (e.g., NVIDIA drop) seen as unwarranted
- •Speculation about origins/actors adds to hype
- 2:25 – 3:32
Why the market ‘noticed’ late: base model vs post-training and the R1 reasoning moment
Sarah contrasts DeepSeek V3’s quieter release with the later R1 reasoning-model moment that made the public pay attention. They emphasize that post-training and instruction-tuning are what turn base models into breakthrough user experiences, echoing the GPT-3.5/ChatGPT pattern.
- •DeepSeek V3 existed earlier without the same market shock
- •Post-training/instruction tuning drives real user utility
- •R1 as a parallel to OpenAI o1 triggered the “breakthrough” perception
- •People don’t value raw next-token prediction as a product
- 3:32 – 4:41
US–China tech dominance and what the cost ‘violation’ implies about entry barriers
They connect DeepSeek to the broader geopolitical narrative of US vs China technology leadership. Even if the $6M figure is incomplete, Sarah argues the bigger shock is that competing may not require multi-billion-dollar budgets, challenging assumptions about barriers to entry.
- •DeepSeek as evidence Chinese labs can catch up quickly
- •Cost may be misstated, but experimentation and end-to-end costs are large
- •Still, total cost may be far below “multi-billion” perceptions
- •Lower perceived entry price rattles markets and incumbents
- 4:41 – 5:38
The cost-collapse trend: training and inference getting dramatically cheaper
Elad argues DeepSeek is consistent with an ongoing cost collapse in AI, not an outlier. He cites steep reductions in both training and inference costs for equivalent capability over the last 18 months, reframing the event as trend-confirming rather than trend-breaking.
- •Training GPT-4-level capability is far cheaper than two years ago
- •Inference cost per token for equivalent models down ~180x in 18 months
- •DeepSeek’s efficiency matters, but it’s within an existing curve
- •Implication: continued downward pressure on model pricing
- 5:38 – 6:58
Are models commoditizing? Benchmark convergence and narrowing gaps
Prompted by Sarah, Elad assesses commoditization through the lens of independent benchmarking. He notes performance across providers is converging across many domains, though leapfrogs can still happen with major breakthroughs.
- •ArtificialAnalysis.ai as a lens for normalized, retested benchmarks
- •Models converging across reasoning, math, coding, multilingual, and cost/perf
- •Gaps are narrower than 12–18 months ago
- •Frontier leaps may still create temporary dispersion
- 6:58 – 7:53
What frontier leadership buys you: distribution, stickiness, and bootstrapping better models
Elad outlines why being best-in-class can still matter even in a converging field. Advantages include market share and switching costs, plus using superior models to accelerate data generation, labeling, tooling, and potentially faster self-improvement cycles.
- •Frontier quality can drive adoption and retention via tooling/prompt optimization
- •Better models can help build the next model (labeling, synthetic data, post-training)
- •Model strength boosts adjacent workflows (e.g., coding tools)
- •Speculative “liftoff” argument: better models may bootstrap faster iteration
- 7:53 – 9:16
DeepSeek’s consumer spike: capability vs curiosity and the role of synthetic data as a leveler
Sarah adds that high-quality models enabling synthetic data generation can level the playing field—an under-discussed dynamic. They debate DeepSeek’s app-store rise: whether consumers felt a meaningful capability/price advantage or whether geopolitics and curiosity drove trials.
- •Synthetic data generation from strong base models can be a major equalizer
- •Consumer adoption: “best/cheapest wins” vs “novelty/geopolitics drives testing”
- •Sarah leans toward curiosity as the primary driver in this case
- •Frontier capability still can enable new consumer experiences
- 9:16 – 11:02
OpenAI Deep Research: a new bar for knowledge work—and its limits
They shift to OpenAI’s Deep Research product, which Sarah finds highly valuable for certain research and survey tasks. However, she warns that in domains where she has expertise, the system’s implicit authority ranking and source quality judgments require careful auditing.
- •Deep Research raises the baseline for analyst/intern-style tasks
- •Most helpful for fast surveying and mapping expert landscapes
- •In expert domains, outputs can contain questionable rankings/claims
- •Users must audit and verify rather than accept summaries as authoritative
- 11:02 – 12:46
Gell-Mann amnesia for AI: trust, sourcing, and agentic opacity
Elad connects Deep Research’s issues to “Gell-Mann amnesia”: people notice errors in domains they know, then trust the system elsewhere. They discuss how AI as an agent that searches and reports back increases opacity, reshaping how people form beliefs and evaluate knowledge.
- •Definition of the cognitive trap: distrust locally, trust globally
- •AI’s role overlaps with search while making sourcing less visible
- •Agent workflows add opacity: you don’t see the steps taken
- •Big implication: how society will consume ‘primary’ information
- 12:46 – 13:31
Propaganda, censorship, and why a multi-model world + open source matters
Elad argues AI systems could become an ultra-powerful interface combining search and social media, amplifying risks around controlled narratives. He expresses relief that multiple competing models—and open source—can provide checks and preserve civil liberties.
- •AI as ‘Google + Twitter + Facebook’ in one interrogable interface
- •Control of outputs becomes politically and socially potent
- •Risk surface includes propaganda, censorship, and narrative shaping
- •Competition and open source as partial mitigations
- 13:31 – 15:38
Stargate and the capital/compute question: does scale still win?
Sarah unpacks Stargate into underlying questions: the strategic value of massive infrastructure, the depth of capital markets, and sovereign involvement. Her core view is uncertainty about which scaling vectors dominate, but she can’t imagine top labs declining the biggest cluster if capital is available—implying pre-training still matters.
- •Stargate raises questions: infrastructure advantage, financing depth, sovereign roles
- •Shift from ‘risk’ to ‘uncertainty’ given emergent capabilities
- •Algorithmic efficiency vs scale vs test-time scaling remains unclear
- •If free capital exists, frontier labs will still pursue maximum compute
- 15:38 – 17:41
Predictions for 2025: consolidation, vertical AI ops, autonomy in vehicles, consumer resurgence, early robotics
They pivot to 2025 forecasts: Elad predicts consolidation in foundation-model adjacencies and some LLM tiers, plus growth in science applications. He also expects vertical AI ops at scale, more self-driving attention, renewed consumer experimentation, and early signals in agents/robots.
- •Market consolidation across modalities and some foundation models
- •New ‘races’ in biology, physics, materials alongside continued scaling
- •Vertical AI ops (legal, support, scribing, code) expand and add agents
- •Self-driving and robo-taxis accelerate attention and adoption
- •Consumer experiments scale; early glimmers in agents and robotics
- 17:41 – 21:12
Sarah’s 2025 view: agents get real, edge vs web distribution, robotics proofs, and domain-specific data strategies
Sarah agrees on broader agent adoption, emphasizing multi-step task success and action-taking, enabled by better reasoning and better product handling of failures/state. She expects more capable small/low-latency models to unlock consumer experimentation, anticipates technical breakthroughs (not deployments) in robotics/generalization, and highlights a mispriced trend: smarter domain-specific data generation driven by new cohorts of AI-native domain experts.
- •Agents defined pragmatically: multi-step tasks + real actions in user environments
- •Key enablers: reasoning gains, state management, and failure handling
- •Small/low-latency models unlock more consumer product iteration; on-device vs web debate
- •Robotics/generalization: proofs this year, deployments later
- •Domain-specific data innovation (bio/materials/health) as a major lever
- 21:12 – 22:32
Episode 200 banter: beards, AI-hosted podcasts, and ‘RLHF farm vs Ibiza’ futures
They close with light banter about Elad’s facial hair and a playful question about what the world looks like by episode 200. Sarah jokes that two agents might host the show, then offers a tongue-in-cheek forecast: they’ll be either doing RLHF labor or relaxing in post-abundance Ibiza.
- •Humorous detour: beard/hat era discussion
- •Speculation: agents may replace hosts as knowledge intermediaries
- •Playful scenarios for the future of AI and work
- •Wrap-up and listener call-to-action
