1. Great Models Aren't Enough — X (formerly Twitter)

  • Why read: Meta's CTO explains why raw model intelligence is now table stakes.
  • Summary: Andrew Bosworth argues AI's lasting value comes from product, distribution, and consumer relationships. Meta is betting on "personal super intelligence" built into daily habits via wearables and assistants. Since most tasks only require a bounded level of intelligence, the future favors cheaper, faster models routed efficiently over a single massive one. For developers, the competition is shifting from model leaderboards to user experience. The lesson: obsess over the human side of AI instead of swapping out models.
  • Read more

2. The new rules of context engineering for Claude 5 models — X (formerly Twitter)

  • Why read: How to prompt the new Claude 5 models without holding them back.
  • Summary: Claude 5 requires a new approach to prompting. Instead of strict rules, heavy guardrails, and explicit examples, developers should rely on the model's own judgment. The best approach is to provide simple context and use progressive disclosure to offer information and tools only when needed. Stripping out most old system prompts and focusing on clean tool design reduces conflicting instructions and boosts performance. Stop writing exhaustive negative rules and start designing intuitive workflows.
  • Read more

3. How to design an AI agent — X (formerly Twitter)

  • Why read: A mental model for designing AI agents that customers actually understand.
  • Summary: AI agents often fail because they seem capable of doing anything, which leaves users confused about what they actually do. The fix is to give every agent a "mind" and a "body," using the body to intentionally restrict what the agent can do. By grouping specialized skills into a tight unit, you narrow the options until the agent's purpose is obvious. This constraint makes the product easy to understand and separates it from generic chatbots. Developers need to define clear edges, like specific domains and identities, so users immediately grasp the agent's limits.
  • Read more

4. The Broader the Agent, the Finer the Pricing Unit — beehiiv.com

  • Why read: How to price AI agents as they handle more complex tasks.
  • Summary: As AI agents take on broader workflows, their costs become unpredictable. Single-job AI features are easy to price per action because compute costs are stable. But when agents iterate, debug, and switch contexts, a flat rate either undercharges power users or scares off casual ones. Vendors like n8n are moving to granular billing models based on tokens or inference credits. As AI products become more open-ended, companies need to tie pricing directly to compute consumption.
  • Read more

5. Myths About Distillation — X (formerly Twitter)

  • Why read: The technical reality of AI model distillation and why it isn't inherently theft.
  • Summary: Distillation transfers capabilities from a large model to a smaller one. Critics argue small datasets pulled from rival models shouldn't matter next to trillion-token pretraining, but these targeted datasets are highly effective during mid-training or when bootstrapping reinforcement learning. A teacher model can also act as a grader for a reward model, transferring "taste" without putting its original text into the training data. Small, high-quality synthetic datasets steer AI pipelines significantly and have a massive impact on performance.
  • Read more

6. Run a team of AI Employees — X (formerly Twitter)

  • Why read: A playbook for shifting from an individual contributor to managing a team of AI agents.
  • Summary: As AI agents improve, founders and operators need to learn how to manage digital workers. Moving your development environment to the cloud is a must so you can run multiple agents in parallel without local conflicts. To handle the massive increase in decision volume, batch your reviews and stick to a strict schedule instead of reacting instantly. Any recurring task, like testing or data checks, should go to specialized agents on scheduled loops. This shift increases output but requires strict project management and delegation.
  • Read more

7. 10 Ways Clay's GTM Engineers Use AI to Accelerate Sales — Substack

  • Why read: How top Go-To-Market teams use AI to build custom sales campaigns.
  • Summary: Sales teams are moving past basic automation and using AI to enrich accounts and write personalized outreach at scale. By connecting Claude to CRM and platform APIs, reps get timed insights and drafted messages for target accounts directly in Slack. Sales engineers can build custom mini-apps in minutes for proof-of-concepts, skipping the usual design delays. This replaces manual research and standard slide decks with assets tailored to each buyer. To cut through the noise, sales teams need to adopt this engineering mindset.
  • Read more

8. Practical multi-agent orchestration in Codex — X (formerly Twitter)

  • Why read: How to organize multi-agent setups to improve efficiency and cut compute costs.
  • Summary: Multi-agent systems work best when you assign specialized roles based on the reasoning required. Use a "coordinator" agent for task delegation, and send lightweight "scout" agents to gather information. Limit context inheritance so agents get fresh, focused tasks instead of drowning in long chat histories. To prevent overlapping work, set strict boundaries and stop lower-level agents from spawning their own sub-agents. Good coordination saves expensive compute for when you actually need it.
  • Read more

9. 1.6 Million Datasets, 12 Survived — Substack

  • Why read: The reality of finding reliable, free B2B datasets on public repositories.
  • Summary: Platforms like HuggingFace and Zenodo host millions of public datasets, but most are unstructured AI training data useless for building sales lists. After filtering 1.6 million datasets for quality and origin, only 12 were trustworthy enough for B2B sales. Many popular datasets are just commercial scraping operations using academic sites for free distribution, with no data validation. This shows how rare high-trust public data is and why you need automated grading to filter the noise. Trusting dataset titles without checking their source will ruin your outreach campaigns.
  • Read more

10. Turn And Face The Strange — X (formerly Twitter)

  • Why read: Why AI coding agents are forcing cloud providers to change their product strategies.
  • Summary: Fly.io notes that traditional cloud infrastructure, built for strict CI/CD pipelines, can't handle a future of AI-generated software. As coding agents make software creation easier, the number of custom apps will explode, demanding highly flexible infrastructure. In response, Fly.io is launching "Sprites"—lightweight, agent-managed compute instances built for dynamic orchestration. Sticking to rigid, older cloud architecture means betting against the speed of AI development. Companies need hosting environments that support rapid, agent-driven deployments.
  • Read more

11. [AINews] Black Forest Labs FLUX 3 - Multimodal Flow Models — Substack

  • Why read: Updates on open-weights media generation and a massive new open-code dataset.
  • Summary: Black Forest Labs launched FLUX 3 Video, an open-weights model for text-to-video, image-to-video, and native audio that rivals top proprietary models. At the same time, Hugging Face released The Stack v3, offering about 5 trillion deduplicated code tokens to train the next generation of open-source coding models. Together, these releases show open source is catching up to proprietary models in both media generation and code training. As base models become commodities, companies will have to build their competitive advantage in the application layer, rather than in raw model capabilities.
  • Read more

12. Weekly Dose of Optimism #203 — Substack

  • Why read: Travis Kalanick's startup Atoms raised $1.7B from a16z to automate physical industries.
  • Summary: Travis Kalanick has been quietly building Atoms, a company applying software engineering to physical sectors like food, real estate, transportation, and mining. Backed by $1.7 billion from a16z, Atoms runs robotics, autonomous hauling, and ghost kitchens under one roof. Kalanick treats manufacturing as the CPU, real estate as storage, and transportation as the network for physical automation. This funding round marks a major shift of venture capital into applying AI and automation to heavy industry.
  • Read more

13. I Helped Build The ABM Category. 78% of Programs Are Still Failing. So I'm Back. — Sangram (Email Newsletter)

  • Why read: A creator of Account-Based Marketing explains why it's broken and how to fix it.
  • Summary: Account-Based Marketing has 80% market adoption, but 78% of programs fail to meet expectations. The main problem is the ongoing disconnect between sales and marketing, making it hard to scale past early tests or prove ROI. The programs that actually work rely on signal-driven data instead of static lists, which speeds up deals and increases contract sizes. Companies need to drop rigid ABM playbooks and use AI-assisted intent tracking to match their messaging with when buyers are actually ready to purchase.
  • Read more

14. Few things going on — X (formerly Twitter)

  • Why read: The fragile financing and bottlenecks behind the AI hardware boom.
  • Summary: Building AI infrastructure is expensive, leading to a shaky ecosystem of middlemen and off-balance-sheet entities trying to grab hardware quickly. Hardware manufacturers are funding new startups to avoid relying entirely on major tech companies. This creates extreme speculation and aggressive financing structures, like mezzanine debt. If the market turns, these untested credit setups could easily collapse. Investors and operators in AI infrastructure have to balance the need for speed against the risk of these fragile financial dependencies.
  • Read more

15. Beyond Tech’s Power: On Status, Sacrifice, and the Search for Legitimacy — Michael Dempsey: Blog

  • Why read: How founder status is moving from financial metrics to societal legitimacy.
  • Summary: Startup status symbols used to be fundraising totals, then ARR, and now revenue-per-employee. But these metrics only matter to insiders; they don't establish a founder's legitimacy in the real world. As AI makes operational efficiency common, true differentiation will come from what founders sacrifice for their vision and how they handle issues outside of tech. To build something lasting, founders have to step out of the tech bubble, show they understand real-world problems, and earn public trust.
  • Read more

Themes from yesterday

  • Managing AI Agents: Software development is moving from writing code to managing autonomous agents by setting clear boundaries and allocating compute.
  • Model Commoditization: As open-weight models and distillation techniques improve, raw intelligence is becoming a given. The real value now lies in user interfaces and tight workflow integration.
  • A Split AI Economy: The startup world is dividing into two camps: capital-heavy infrastructure projects built on risky credit, and lean application teams using AI to increase output.
  • Go-To-Market Shifts: Traditional outbound and ABM playbooks are failing. Successful teams are switching to intent signals and custom mini-apps built by AI.