Latent Space

Anima Anandkumar

Enterprise voice deployments still favor cascaded STT→LLM→TTS over speech-to-speech, despite Smallest AI calling S2S “the eventual future.”Against a customer-support market tied to hundreds of billions in call-center spend, reported switches from GPT-4o/4.1 realtime to Electron point to latency, cost, and API reliability as key wedges, while hybrid S2S routing remains the unresolved path.

Vibhu × swyx × Joon Sung Park

Simile reports an early glimpse of a simulation scaling law, with more human data and compute producing predictable performance gains.A validated 1,000-person study reached 85% behavior-and-attitude replication versus frontier models' 20–30% on niche populations, while preregistered-experiment post-training delivered significant gains; current deployments aim to shape decisions, though TAM and future foundation-model-scale costs remain unresolved.

RJ Honicky × Matt McPartlon × Neil Patil

Chai Discovery is commercializing a neutral modeling and product layer for pharma rather than developing its own drugs, with partnerships including Eli Lilly, Pfizer, Novartis, and Genentech.Chai-2 designed antibodies against 50 targets, finding binders for about half with an average binding hit rate of around 20%, while enabling GPCR agonists and multispecific formats traditional screening struggles to produce.The opportunity depends on improving developability and epitope prediction, but compute scarcity, validation latency, and talent shortages remain structural constraints as Chai scales its platform.

swyx × Vibhu × Philip Kiely × Ali Taha

Inference remains an early optimization market: on identical hardware, quantization, caching, speculation and traffic tuning can typically deliver 2–4X gains, while production support requires weeks of debugging.The economics move customers from pay-per-token trials to dedicated capacity, as reliability, isolation and workload-specific tuning justify self-saturating infrastructure; faster interconnects and training-integrated optimization remain the next catalysts.

swyx × Vibhu × Akshay Nathan

ChatGPT Work has reached 10 million users by extending Codex’s agentic capabilities beyond developers, but it remains paid-only and not ChatGPT’s default.OpenAI is standardizing one harness across Codex and Work, while Sites, artifacts, and persistent context move the product above traditional applications.The next catalyst is broader distribution into knowledge work and personal workflows; permissions, trust, and misleading productivity metrics remain unresolved risks.

swyx × Vibhu × Eiso Kant

Poolside’s claimed moat is a model factory that turns checkpoints into repeatable launches, running 10,000–20,000 experiments monthly with fewer than 70 researchers and roughly 35 engineers.Laguna S suggests persistence and verification can offset parameter scale—118B total, 8B active—while open research could widen competition, though Poolside still lacks a complete business model and must scale with frontier rivals.

Bo Wang × Ci Chu

Xaira links protein design, X-Cell, and patient-representation models around a causal drug-discovery loop, arguing that observational expression profiles cannot answer intervention questions.Seven genome-wide Perturb-seq campaigns across 16 contexts and 25 million filtered cells enabled X-Cell to transfer predictions from resting to activated T cells, held-out iPSC types, and primary donors, while patient, organ, and longitudinal dynamics remain unresolved.

Andy Beam × Rafa Gómez-Bombarelli

Lila is betting that controlled experiments can become AI’s next internet-scale training corpus, with nature providing verifiable rewards and an information-gain-driven lab turning experiments into a compounding model moat.Lila reports a six-month, two- or three-person in vivo CAR-T program versus roughly six years and $100 million, but clinical translation, scale-up, regulation, and 5–6% model FLOPs utilization remain key watchpoints.

swyx × Vibhu × Akshat Bubna

Modal is repositioning infrastructure around agent experience, using decorators, CLI observability, and a 17-provider footprint instead of owning data centers.Its investment case rests on bursty workloads and elastic orchestration: speculative decoding may deliver 2× to 4× speedups, while batch pricing and reliability determine whether compute planning converts into margins.

Brandon Anderson × Evan Feinberg × Sergey Edunov

Genesis Molecular AI is applying diffusion and inference-time scaling to 3D molecular design, targeting roughly 1 Å protein-ligand accuracy because 2 Å can hide chemically fatal errors.Its prospective edge comes from Incyte and Insitro programs that connect models to synthesis and ADMET feedback, while OpenBind showed stronger unseen-target generalization and GPUs remain the scaling bottleneck.