Today's content converges on a single uncomfortable question: where does AI actually stop working, and are we measuring the right things to know when we've reached that point?

Terence Tao's account of AI on Erdős problems, shared in conversation with Dwarkesh Patel, is among the most grounded capability analyses to surface recently. After a remarkable burst in which frontier models solved roughly 50 open problems (some standing for decades), the pace of purely autonomous solutions has dropped to near zero. "There was a month where that happened and that has stopped," Tao said, noting 3 separate systematic attempts that produced nothing new.

The pattern behind the 50 successes is telling. Almost all shared a specific structure: no serious existing literature, and solutions that required combining 1 obscure technique with 1 published result. That's the median capability boundary — not raw problem difficulty, but a particular kind of combination task in a sparse literature environment. When researchers moved from cherry-picking wins to systematic sweeps, the real number surfaced: roughly 1-2% success rate per problem. The impressive run looked like genius; it was scale applied to a favorable distribution.

Tao's "dark mountain range" metaphor is precise. Current models are jumping machines that clear 2 meters — higher than any human — but they either succeed or crash. They cannot assess whether a given wall is 3 feet or a mile high, and they're consistently bad at generating partial progress or decomposing problems into tractable intermediate stages. The remaining ~600 open Erdős problems are now being chipped away through human-AI collaboration rather than one-shots.

This is a genuinely useful capability profile. Frontier models are exceptional at short-hop combination tasks in sparse territory, and structurally limited on anything requiring staged decomposition or self-assessment of difficulty.

While mathematicians are probing AI's ceiling, organizational researchers are finding most companies haven't remotely approached it — because they're stuck on the floor.

A randomized field experiment on 515 startups provides some of the sharpest adoption data in recent memory. Founders shown concrete operational case studies (not general AI awareness materials) adopted AI 44% more, generated 1.9x higher revenue, and required 39% less capital compared to controls. The paper frames this around the "mapping problem" — firms struggle not with accessing AI tools but with identifying where those tools create value in their specific workflows. Case studies resolved this by giving founders a vocabulary for task-level integration.

Ethan Mollick draws a parallel structural argument at the organizational level. Most companies delegate AI strategy to middle management or IT, which he argues guarantees failure. His Leadership-Crowd-Lab framework assigns distinct roles: leadership must personally drive strategy and use the tools themselves; employees with domain expertise and genuine permission to experiment (the "crowd") are where the best use cases emerge; and a dedicated lab team — which Mollick says is "shockingly" absent at many large companies — builds institutional knowledge rather than relying on vendor demos.

Meta provides the cautionary counterexample. Its internal "Token Legends" leaderboard ranks employees by AI compute consumed, turning token usage into a status signal. Mollick flags this as a textbook case of rewarding A while hoping for B. When compute consumption becomes a proxy for productivity, you get more compute consumption, not more productive outcomes. The correct fix, as several practitioners independently argue, is outcome-aligned metrics: experiments shipped, latency budgets met, unit economics per successful inference.

The through-line is consistent: adoption is a knowledge and incentive problem, not a tool access problem.

The Every team has spent roughly 2 months deploying OpenClaw agents across their entire organization, and Dan Shipper's account is one of the more concrete practitioner reports on what actually happens when agents move from experiment to daily infrastructure.

The most interesting finding concerns identity and trust: agents that operate publicly inside team communication tools take on their owner's reputation. If you're known for a particular kind of judgment, your agent becomes known for it too. Shipper frames this as fundamentally different from a shared tool — "Claude is everybody's; a Claw is mine" — because the agent becomes a persistent, personalized reflection of its owner. This creates a novel organizational dynamic where individual agents develop distinct reputations within a team, and those reputations carry real weight.

The team also surfaced failure modes that don't appear in demos: memory gaps across sessions, group chat etiquette problems when multiple agents interact, and what they call the "ant death spiral" (a failure mode where agents reinforce each other's errors in a compounding loop). These aren't edge cases — they're the kinds of structural problems that only emerge when agents are running daily on real work.

Separately, a cost collapse in video understanding is creating new practical headroom. AI video analysis has dropped from roughly $6/hour to $0.14/hour over 18 months — a 97% reduction driven by Google's open-source Gemma 4 model enabling efficient frame processing at 2 FPS. The math: at current token pricing ($0.14/million input tokens), 1 hour of video at 2 FPS with ~140 tokens per frame runs about $0.14, with no complex frame-splitting engineering required. This 40x cost reduction makes economically viable a category of applications (live sports commentary, smart doorbell comprehension, automated review of security footage) that were impractical 18 months ago.

On the infrastructure side, Google's TurboQuant paper tackles KV (key-value) cache compression via vector quantization. The KV cache (which stores previously computed attention outputs to avoid recomputing them at each token) scales with model size and becomes the dominant memory bottleneck in long-context inference. TurboQuant applies a polar coordinates transformation (PolarQuant) followed by a quantized Johnson-Lindenstrauss transform, achieving up to 6x compression with no meaningful accuracy loss and no inference latency penalty. For comparison, NVIDIA's NVFP4 approach achieves 3x latency reduction and 50% memory reduction versus FP8 with sub-1% accuracy cost. Sebastian Raschka's explainer work on KV cache mechanics has become a standard reference for practitioners working through these tradeoffs.

The practical implication is direct: 6x memory reduction on the KV cache translates to longer effective context windows at the same hardware cost, or equivalently, more concurrent sessions per GPU. For anyone running long-context inference at scale, this is a meaningful engineering development.

TL;DR - Terence Tao's Erdős analysis reveals a 1-2% per-problem success rate masked by scale and cherry-picking, with frontier models structurally unable to generate partial progress or assess problem difficulty in advance. - A 515-startup RCT found concrete AI case studies drove 44% more adoption, 1.9x revenue, and 39% less capital — confirming the bottleneck is knowing where AI applies, not tool access; Mollick's Leadership-Crowd-Lab framework and Meta's Token Legends provide the organizational theory and cautionary tale. - Practitioners deploying agents across organizations are finding that public-facing agents develop reputational identity and expose failure modes (memory gaps, ant death spirals) absent from demos, while AI video analysis costs have dropped 40x to $0.14/hour. - Google's TurboQuant achieves 6x KV cache compression with no accuracy loss, directly enabling longer context windows or more concurrent inference at the same hardware cost.