Frederick Brooks's 1975 observation that adding manpower to a late project makes it later remains relevant in the age of AI agents. The bottleneck has shifted from human onboarding to agent orchestration, unclear requirements, and sequential tasks. Nine agents still can't deliver a baby in one month, and adding agents to ambiguous projects accelerates wrong answers.

Cloudflare disclosed a cross-request KV cache exposure in its GLM-5.3 Workers AI model, affecting 155 responses to prompts over 681,280 tokens between Aug 27 and Sep 9, 2026. The issue stemmed from a configuration mismatch and an SGLang engine bug. Cloudflare resized the KV pool, patched the engine, and released hardening updates. No evidence of deliberate exploitation was found.

OpenAI and Anthropic have cut flagship LLM prices by up to 80% in four months, driven by open-weight competition, falling infrastructure costs, and IPO timing. Models like GPT-5.6 Luna and Claude Opus 5.5 are now dramatically cheaper, reshaping AI economics and enterprise adoption.

A dermatologist's study reveals that a 10-year-old CNN (Inception v3) outperforms frontier LLMs like GPT-6 Astra and Claude Fable 5.1 in skin cancer diagnosis. The CNN achieved 62% accuracy vs. 42% for LLMs. The test used 200 biopsy-proven lesions, highlighting that specialized models may still hold an edge in medical imaging tasks.

Cerebras has revealed how it runs trillion-parameter AI models on its wafer-scale WSE-3T chips, achieving 1,000-2,000+ tokens per second. Using the open-source Kimi K3 (3T parameters) as an example, the company splits models across 59 wafers, pipelines requests, and relies on 44GB on-chip SRAM per wafer to keep weights stationary. This approach, validated by OpenAI's $20B deal, challenges GPU-based inference.

Apollo chief economist Torsten Slok reports that despite massive AI investment, the technology has delivered zero financial payoff to corporate America outside of tech. Revenue per worker and profit margins remain flat for the S&P 493, while the Magnificent 7 see gains. Token spending is increasingly viewed as a cost center, and productivity data shows no AI-driven boost. The AI story's outcome depends on the S&P 493, not the Mag 7.

AI image generation has evolved from diffusion models to multimodal agents that plan images like sentences. This deep dive covers how models work, the real costs per image (from $0.003 to $0.134+), prompting techniques that avoid the 'AI look,' pixel art challenges, and integrating generation into code and agents. Key takeaways: pay for the canvas, not content; describe how it was made, not how it looks; and hybrid pipelines beat pure AI for structural tasks.

Block's research lab unveils Dessa, a full-duplex speech-to-speech AI trained from scratch on raw human conversation, without any text or pretrained models. The model natively processes emotion, sarcasm, laughter, and interruptions, running locally on a MacBook Pro M4 Max. This breakthrough challenges conventional voice AI pipelines and opens new research frontiers.

NVIDIA has released OpenShell, an open-source runtime designed to securely govern fleets of autonomous AI agents. It uses kernel-level enforcement, formal verification, and hardware-level integration with BlueField DPUs to control agent access to files, networks, and credentials. This addresses the critical security challenges of deploying autonomous agents at scale.