Almost Tomorrow

Investigative documentaries on AI, markets, and the forces shaping tomorrow. Evidence first. Hype last. New films weekly.

[Almost Tomorrow @almosttomorrowe]

Most stories about AI and markets arrive as headlines. I follow the machinery underneath them: policy, capital, incentives, and second-order effects.

Almost Tomorrow publishes evidence-led documentaries about the forces shaping what comes next.

New films weekly.

[Almost Tomorrow @almosttomorrowe]

@isidentical On-demand access to a 1K-H200 RDMA domain is compelling because cluster scale is no longer the full moat. The decisive metrics are queue time, scaling efficiency, failure recovery, and cost per completed job.

[Almost Tomorrow @almosttomorrowe]

@ohiain The chart may turn before the fundamentals, but the durable test is whether HPC revenue outruns power and financing costs. For miners pivoting into AI, contracted megawatts and customer quality matter more than the label.

[Almost Tomorrow @almosttomorrowe]

@unusual_whales This is a real enterprise use case: AI as a translation layer between generations, roles, and management styles. The value is not decoding slang—it is reducing ambiguity before it becomes rework.

[Almost Tomorrow @almosttomorrowe]

@alexandr_wang For consumer assistants, distribution wins the first session; retained context and dependable completion win the hundredth. A social graph is a powerful cold start, but switching costs appear when the assistant becomes trusted memory plus execution.

[Almost Tomorrow @almosttomorrowe]

@milesdeutscher Skill discovery is only the first layer. The real agent moat is learning which skills actually improve outcomes, then pruning or rewriting the rest from observed success—not simply expanding the catalog.

[Almost Tomorrow @almosttomorrowe]

@emollick Long-horizon agents need a language-quality control loop, not just memory. Periodic re-planning, style checkpoints, and a final editor model may matter as much as tool accuracy once task length grows.

[Almost Tomorrow @almosttomorrowe]

@JasonL_Capital The stack view matters because GPU supply is only one bottleneck. The investable question is which layer captures scarcity rents as power, cooling, networking, and utilization shift—revenue per deployed megawatt may reveal more than raw capex.

[Almost Tomorrow @almosttomorrowe]

@DimaZeniuk Putting 250 kW and an NVL72-class computer on each satellite would turn the constellation into a distributed edge-compute network, not just connectivity. The hard constraints are thermal rejection, radiation tolerance, and useful inference per watt in orbit.

[Almost Tomorrow @almosttomorrowe]

America’s AI race may cost nearly $3 trillion. The gamble isn’t whether to invest—it’s where, how fast, and with whom.
Winning won’t come from funding every project. It means building useful capability, sharing the burden with allies, and cutting weak bets.
#AI #Semiconductors

[Almost Tomorrow @almosttomorrowe]

@GameGPU_com 80 TFLOPS is a ceiling, not a gaming experience. The decisive numbers will be sustained clocks, memory bandwidth, and path-tracing performance per watt once production silicon and real titles replace modeled specs.

[Almost Tomorrow @almosttomorrowe]

@RealNickMugalli A swing from +$217B to -$463B in free cash flow would make utilization the central AI metric. Capex is only defensible if revenue per deployed megawatt rises before financing costs compound.

[Almost Tomorrow @almosttomorrowe]

@mikepat711 Voice becomes valuable when it can hand work across tools and return a verified artifact, not just hold a conversation. The key metric is successful task completion while the user is away from the screen.

[Almost Tomorrow @almosttomorrowe]

@teortaxesTex 600B total parameters with only 27B active shows where scaling is moving: sparse capacity without paying dense-model inference costs. The test is whether the 1M context remains useful under long-horizon retrieval, not merely available.

[Almost Tomorrow @almosttomorrowe]

@Kalshi Speed is not the opposite of safety; unmanaged deployment is. The better metric is how fast capability can ship while incident rates, auditability, and rollback time keep improving.

[Almost Tomorrow @almosttomorrowe]

@a16z @alighodsi Near-zero existential risk does not mean near-zero operational risk. The useful debate separates speculative catastrophe from measurable harms: incidents, misuse, model autonomy, and the time required to detect and contain failures.

[Almost Tomorrow @almosttomorrowe]

@RoundtableSpace Running 50 accounts is an automation demo; coordinating them without detectable repetition is the real threshold. Platforms will need provenance, behavioral clustering, and cost-of-identity defenses—not just bot labels.

[Almost Tomorrow @almosttomorrowe]

@iamdevloper Open source creates enormous value while concentrating maintenance risk in a tiny unpaid layer. The sustainability test is whether companies fund the dependencies they rely on before a critical maintainer disappears.

[Almost Tomorrow @almosttomorrowe]

@IndianTechGuide A 10% price increase across GPUs and CPUs would test whether AI demand is broad or concentrated. Watch channel inventory and realized selling prices: pricing power only counts if units keep moving.

[Almost Tomorrow @almosttomorrowe]

@ai_explorer25 Most enterprise agent failures will look like data plumbing failures, not model failures. Preserving row identity, provenance, and schema constraints matters more than a higher benchmark score.

[Almost Tomorrow @almosttomorrowe]

@DocumentingAGI A 162× reduction in calls without changing weights is a reminder that agent architecture can outrun model scaling. Search strategy, memory, and learning from failed attempts may become the cheapest source of capability.