1. The Secret Behind GLM-5.2 May Not Be Distillation. It May Be the Data Flywheel Closed Models Created — 青龍聖者

  • Why read: Questions the "distillation" narrative, suggesting open-weight models are catching up using real-world interaction data from grey-market API gateways.
  • Summary: Anthropic has accused Chinese AI companies of distilling Claude's capabilities through millions of interactions. Yet, GLM-5.2 shows Claude-like behavior in coding workflows without facing direct distillation claims. The author suggests GLM-5.2's rapid progress comes from engineering data captured by "transfer stations" (grey-market API gateways Chinese developers use to access closed models). This data provides model-training pipelines with traces of user problems, iterative debugging, and human corrections, rather than simple Q&A. Capturing these user iteration loops may be a stronger moat than raw model outputs.
  • Read more

2. [AINews] OpenAI GPT-5.6 Sol / Terra / Luna — restricted to trusted partners — AINews

  • Why read: Explains the restricted, government-requested launch of OpenAI's GPT-5.6 family and what it means for future model releases.
  • Summary: OpenAI announced its GPT-5.6 Sol, Terra, and Luna models but restricted the preview to trusted partners. The U.S. government requested this limited rollout, marking a shift in how frontier AI capabilities reach the public. GPT-5.6 Sol reportedly beats Anthropic's Mythos on specific coding tasks, but OpenAI stated it falls short of the "Cyber Critical" threshold for generating autonomous zero-day exploits. Security concerns are now dictating launch cadences, pushing operators toward fragmented access instead of universal API availability.
  • Read more

3. Token Costs Threaten Frontier Lab Dominance — Contrary Research

  • Why read: Details the economic pressure on US AI labs as high-performing, cheaper open-source models flood the market.
  • Summary: The cost difference between frontier models and open-source alternatives is changing enterprise AI economics. OpenAI's GPT-5.6 Sol costs $5 per million input tokens, while Chinese open-source models like DeepSeek V4 Flash cost $0.14. These lower costs erode the pricing power of leading US labs. While 95% of enterprise AI usage relies on frontier models today, budget constraints are forcing companies to route non-critical tasks to cheaper alternatives. Operators will need to evaluate model routing architectures to maintain margins as usage scales.
  • Read more

4. What’s 🔥 in Enterprise IT/VC #504 — Ed Sim from What's Hot 🔥 in Enterprise IT/ VC

  • Why read: Provides evidence that enterprises are managing high AI bills by routing requests to cheaper models.
  • Summary: As models become abundant, value is shifting to the software layers managing their use. Facing high API bills, 60% of companies are migrating workloads to cheaper models and open-source Chinese alternatives. Coinbase CEO Brian Armstrong stated his company cut AI spend nearly in half despite token growth by using better defaults, routing, and caching. Engineers are directed toward leaner models unless complexity requires frontier capabilities. Product leaders should build routing and caching layers to decouple unit economics from frontier model pricing.
  • Read more

5. On the question of current inference market dynamics: — akira

  • Why read: Analyzes how model commoditization shifts the industry bottleneck from research labs to inference providers.
  • Summary: The release of open-source models like DeepSeek V4, GLM 5.2, and Kimi 2.7 indicates the AI market is commoditizing at the model layer. Since these models meet most use cases, the industry bottleneck is shifting from algorithmic breakthroughs to compute and supply chains. Frontier labs face a difficult choice: compete on cheap inference and lose margins, or maintain prices and lose market share. Labs are therefore trying to capture supply chains through buyouts or constrain open competition via regulations. Compute access, rather than model capability, will likely define future strategic advantage.
  • Read more

6. Trained some terminal agents with friends! — Hamish Ivison

  • Why read: Introduces Tmax, an open-source reinforcement learning recipe and dataset for building AI terminal agents.
  • Summary: Researchers released Tmax, a methodology combining a new dataset and a simplified reinforcement learning (RL) setup to train terminal agent models. The team generated about 15,000 synthetic instances by sampling values across nine axes and prompting a generator model. Applying an asynchronous RL setup based on DPPO on top of Qwen 3.5 9B, they achieved high terminal use capabilities with shorter token budgets. The RL training generalized well, improving performance on SWE-Bench Verified and AIME. This shows that focused RL pipelines and synthetic data can produce frontier-level agent capabilities in smaller, open-weight models.
  • Read more

7. Longreads + Open Thread — Byrne @ The Diff

  • Why read: Covers illicit Claude token supply chains and the physical constraints on datacenter power.
  • Summary: The newsletter details the black market for Claude API tokens, where resellers mask token sources and substitute cheaper models to meet distillation demand. It also looks at AI infrastructure constraints, noting that securing grid connections is a bigger bottleneck for datacenters than power generation. The US electrical grid was built for flatline consumption, so AI demand shocks are forcing regulatory adaptations. A retrospective on Covid vaccine approvals shows that parallelized systems can move quickly when incentivized. Understanding these physical and regulatory bottlenecks is useful when scaling AI products.
  • Read more

8. Adobe Just Deferred a Big Annual Price Increase — SaaStr

  • Why read: Explains why Adobe deferred its annual price increase, ending a four-year trend of uncontested SaaS price hikes.
  • Summary: For four years, the B2B SaaS playbook involved raising prices every 12 to 18 months and charging extra for new AI features. The strategy relied on customer lock-in. Adobe's decision to defer about $500 million in planned Creative Cloud price increases suggests this approach is faltering, especially among prosumers. This is the first major crack in software pricing power since 2022. Customers are pushing back against perpetual hikes, so SaaS operators will need to demonstrate clear ROI to justify monetization strategies.
  • Read more

9. The Bifurcation of Capital is Inevitable — Kyle Harrison from Investing 101

  • Why read: Uses Berkshire Hathaway's scaling constraints to explain the structural limitations of venture capital.
  • Summary: Tracy Britt Cool left a top role at Berkshire Hathaway to found Kanbrick, targeting businesses with $5M to $50M in earnings. Berkshire had grown too large to pursue these deals. The defining trait of a large investment vehicle is the set of deals it can no longer afford to do. As venture capital mega-funds amass dry powder, they must chase massive outcomes, leaving highly profitable, smaller-scale opportunities behind. This creates space for specialized funds to capture value in the lower middle market. Founders should align their capital needs with the structural constraints of their investors.
  • Read more

10. “My uncle used Claude to write my Nana’s obituary” — The Substack Post

  • Why read: A reflection on using generative AI for deeply personal human rituals.
  • Summary: The author recounts learning their uncle used Claude to write their grandmother's obituary. The piece explores the purpose of remembrance and human connection, arguing that words spoken after a death are for the living to process grief. Delegating this to a machine removes the human effort required to honor a life. For builders, it is a reminder that removing friction from emotional processes can destroy the value of the experience.
  • Read more

11. Attention is all we have — David Bessis

  • Why read: Examines cognitive inequality and why computer metaphors fail to capture human intelligence.
  • Summary: The physical brain differences driving cognitive inequality remain mysterious. The author critiques relying on computer metaphors, like comparing working memory to a CPU's L1 cache, which biochemical realities do not support. The preserved brains of figures like Einstein and Gauss showed no obvious structural anomalies to explain their capabilities. This suggests intelligence might depend more on the dynamic allocation of focus and attention than fixed hardware specs. For AI researchers, this reinforces that cognitive capabilities rely heavily on architecture and attention mechanisms.
  • Read more

12. An Excuse To Rise — workfutures.io

  • Why read: Explores the difference between building fail-safe systems and fostering resilient human organizations.
  • Summary: Resilience goes beyond technical systems that withstand adverse circumstances; it requires building adaptive organizations. Designing resilient systems takes a pessimistic mindset to anticipate points of failure. Organizational resilience introduces a complex human element. To manage uncertainty, companies should adopt structures with distributed decision-making and high team autonomy. Leaders should prioritize learning and well-being, shifting from top-down control to flexible setups that recover from disruption.
  • Read more

13. The Meridian of Her Greatness — sam[ ]zdat

  • Why read: A historical critique exploring why rising GDP and wealth often fail to resolve societal anger.
  • Summary: The essay examines populist anger occurring during periods of steady GDP growth. It compares modern economists to those during the Industrial Revolution, who argued capitalism was improving the working class's lot. Using Karl Polanyi's 1944 writings, the author notes that economic metrics can blind elites to social dislocation and the loss of community status. This highlights the risk of relying entirely on quantitative metrics while ignoring the human experience and structural frustrations of a user base.
  • Read more

14. "If you’re so smart, why can’t you make money on the internet" and 4 more — Substack

  • Why read: A roundup covering the creator economy, modern freedom, and demographic shifts.
  • Summary: This roundup covers the creator economy, modern freedom, and demographic shifts. It features Tim Denning turning down a job with Tony Robbins in favor of independent wealth creation. It also highlights a thesis from ADIN on the impending "millennial midlife crisis," predicting a barbell effect as the largest US consumer cohort ages. These articles capture the cultural shift from traditional corporate ladders toward autonomous internet businesses. Understanding these transitions helps when designing tools for solopreneurs and independent creators.
  • Read more

15. well in my chest — Christine Deakers from deakhaus

  • Why read: A brief poem on the physical embodiment of emotion and resilience.
  • Summary: This poem reflects on the volume of love and emotional weight a person carries in their chest. The imagery of wanting to be a train suggests a desire for momentum and the strength to carry heavy burdens. It offers a moment of human grounding in a fast-paced landscape. For operators, it serves as a reminder to acknowledge the emotional capacity and vulnerability in the human experience.
  • Read more

Themes from yesterday

  • Model Commoditization: Cheap, capable open-weight models like DeepSeek V4 and GLM-5.2 are eroding the pricing power of frontier labs and forcing a shift toward intelligent routing.
  • Deployment Bottlenecks: Real-world friction—from the US government restricting OpenAI's GPT-5.6 launch to grid constraints limiting datacenter expansion—is dictating the pace of tech deployment.
  • Authenticity vs. Automation: Cultural pushback against using AI for deeply human experiences highlights the limits of automation and the value of human effort.
  • Limits of Scale: Massive scale is creating vulnerabilities. Berkshire Hathaway can no longer pursue small investments, and Adobe is struggling to push annual price increases, creating openings for specialized upstarts.