1. Towards infinite context windows: neural KV cache compaction — Charlie O'Neill

  • Why read: Understand the technical breakthrough move from expensive RAG or lossy fine-tuning to a high-fidelity "intermediate" working memory for LLMs.
  • Summary: This technical deep-dive introduces STILL, a perceiver bottleneck that compresses LLM KV caches by 8x in milliseconds while maintaining over 85% accuracy. Current models suffer from a "memory problem" where they either store everything verbatim (linearly scaling costs) or lose detail through compression. By amortizing the optimization process—similar to how Sparse Autoencoders (SAEs) work for dictionary learning—this method allows for continual learning at forward-pass speeds. For operators, this represents a path toward "intern-style" LLMs that accumulate knowledge across multi-day projects without massive compute overhead. The practical implication is a shift toward models that can manage their own "working memory" layer between short-term context and long-term weights.
  • Link: https://twitter.com/oneill_c/status/2039389427462770815/?rw_tt_thread=True

2. The Economics of Generative AI: Two Years Later — Apoorv Agrawal

  • Why read: A sobering 2026 reality check on where the actual money is flowing in the AI value chain.
  • Summary: Two years after the initial AI boom, the "physics" of the industry remains starkly inverted: the semiconductor layer (mostly NVIDIA) captures 79% of all gross profit, while the application layer struggles at ~33% margins. Despite a 5x growth in total ecosystem revenue to $435B, NVIDIA's incremental revenue alone is triple the size of the entire application layer. The "Apps" layer is currently a two-player game dominated by OpenAI and Anthropic, who command 75% of that segment's revenue. For founders, the takeaway is that "selling shovels" remains the only consistently high-margin strategy, as the application layer has yet to flip the economics in the way the cloud transition did. Strategic success now requires finding moats that don't depend purely on commoditized compute.
  • Link: https://twitter.com/apoorv03/status/2039407534768087400/?rw_tt_thread=True

3. My top takeaways from @clairevo on all things 🦞 (OpenClaw) — Lenny Rachitsky

  • Why read: Practical, "boots-on-the-ground" advice for deploying autonomous agents in a personal or professional workflow.
  • Summary: Claire Vo shares a masterclass in agent management, treating OpenClaw deployment like hiring an employee rather than installing an app. Key tactics include running agents on dedicated hardware (like a $500 Mac Mini) to isolate security risks and creating "Claws" for highly specific roles—Sales, Family, Podcasts—to prevent context dilution. The "Soul + Heartbeat + Jobs" framework allows agents to maintain a consistent personality while checking for tasks every 30 minutes autonomously. A major insight for operators is using the "yappers API" (voice notes) as a high-bandwidth communication channel, letting the agent handle the structuring. This shift from "perfect prompting" to "management-style onboarding" is the secret to reclaiming 10+ hours per week.
  • Link: https://twitter.com/lennysan/status/2039498785693540534/?rw_tt_thread=True

4. software 3.0 — kache

  • Why read: A provocative thesis on why hand-written code (2.0) is being replaced by "mystery bags of floats" (3.0) that act as predictors.
  • Summary: We are moving toward a world where neural networks aren't just writing code, they are the software. Traditional 2.0 software trades off fuzziness and robustness for interpretability, but modern problems—from driving control systems to fraud detection—require degrees of freedom that human-written logic cannot match. Software 3.0 consists of non-interpretable information transformers produced by "mining the space of possible programs" rather than manual engineering. This shift means the role of the developer moves from "architect of logic" to "curator of predictors." For product leaders, the implication is that the speed of "hill-climbing" to find a solution now outpaces the speed of writing interpretable instructions.
  • Link: https://twitter.com/yacineMTB/status/2039344583687934053/?rw_tt_thread=True

5. The Trillion Dollar Loop B2B Never Had — Jaya Gupta

  • Why read: Strategic insight into how B2B companies can finally build compounding data moats similar to consumer giants like TikTok or Netflix.
  • Summary: For two decades, B2C companies have compounded "behavioral signals," while B2B companies only recorded "end states" (e.g., a final discount price, not the negotiation reasoning). AI now allows for the capture of "decision traces"—the sparse, fragmented reasoning embedded in Slack threads, document comments, and meeting transcripts. By structuring these into a "context graph," enterprise software can finally create a learning loop where every decision improves the system's future recommendations. This commoditizes the "feature layer" and shifts the value to the proprietary decision data a company owns. Operators should focus on instrumenting "reasoning" rather than just "outcomes" to build durable moats.
  • Link: https://twitter.com/JayaGup10/status/2039441705586602134/?rw_tt_thread=True

6. The Icarus Problem in AI: When "Ship Fast" Starts to Break the Stack — Pranav Ramesh

  • Why read: A critical warning on the compounding risks of automated compliance and "move fast" supply chain vulnerabilities.
  • Summary: The "move fast and break things" mantra has become a dangerous gamble as velocity outruns verification in the AI stack. Recent incidents involving Delve (automated compliance) and LiteLLM (supply-chain malware) illustrate how "small approximations" compound into massive security breaches. LiteLLM's rushed SOC 2 certification didn't prevent a credential-stealing compromise that eventually hit Mercor and led to a 4TB data exfiltration by LAPSUS$. The practical lesson is that automated "security theater" provides no protection against actual supply-chain attacks. Companies must balance the thrill of frictionless updates with rigorous, human-in-the-loop verification to avoid "melting their wings."
  • Link: https://twitter.com/pranavramesh25/status/2039161958650536417/?rw_tt_thread=True

7. Tokenmaxxing — Tomasz Tunguz

  • Why read: Learn how to 20x your productivity by maximizing "electricity-to-useful-work" through massive parallel agent execution.
  • Summary: Tomasz Tunguz introduces "tokenmaxxing," the deliberate practice of maximizing token consumption to increase output. By structuring a daily plan that allows multiple agents to work simultaneously—pulling git history, querying logs, fact-checking, and building slide decks—he burned 250 million tokens in a single day. The key is moving from 1-hour autonomous windows to the 12-hour windows now possible with modern models. This "productivity ceiling" is much higher than most operators realize, requiring a shift from sequential tasks to parallelized agentic flows. The practical outcome is an unbundling of complex projects into background streams that execute while the human focus remains elsewhere.
  • Link: https://www.tomtunguz.com/tokenmaxxing/

8. Skill chaining and why skills should be actions — akira

  • Why read: A reframing of agent "skills" from static prompts to dynamic, contextual behaviors.
  • Summary: Most agent systems treat skills as static prompts the agent "reads," but Slate implements them as "actions" the agent does. Using an execution model of "continuations," Slate allows threads to execute partially, return state, and be resumed later, mirroring human contextual behavior. This approach addresses the "knowledge overhang"—the gap between what a model knows and what it actually chooses to do. By injecting context-specific skills only when needed ("progressive disclosure"), the system avoids flooding the context window with thousands of irrelevant tokens. For developers, this means building agents that learn "behaviors" rather than just parsing instruction sets.
  • Link: https://twitter.com/realmcore_/status/2039382343581147414/?rw_tt_thread=True

9. Skip the boat. Sell the catch. — Emergence Capital

  • Why read: Discover the "AI-Native Services" business model that is replacing traditional SaaS.
  • Summary: A radical shift is occurring: AI companies are moving from "selling the boat" (software tools) to "selling the catch" (owning the outcome). This "AI-native services" model collapses timelines and shifts ownership of workflows entirely to the provider. Examples like Hanover Park (AI-native fund administration) show how rebuilding a service from first principles with AI makes the old manual benchmarks unacceptable. For founders, the goal is no longer to build a "better UI on a known process," but to stop selling software and start owning the result. This transforms the line between software and services into a single, outcome-oriented business model.
  • Link: https://emcap.substack.com/p/skip-the-boat-sell-the-catch

10. Your website still matters. Here's what to prioritize now. — Emily Kramer

  • Why read: Re-evaluate your GTM strategy for a world where LLMs are the primary research interface.
  • Summary: While less traffic is hitting B2B websites directly, your site is more important than ever because it "feeds the LLMs." What you publish shapes the training data and search context that models use to describe your product. Since prospects now arrive with half-formed opinions from AI chats, conversion on the remaining direct traffic is higher stakes. You must prioritize high-fidelity details that both humans and LLM scrapers can parse easily. Avoid "vibe-coded" chaos and focus on regular updates to ensure AI models don't serve stale information to potential buyers.
  • Link: https://newsletter.mkt1.co/p/website-examples-2026

Themes from yesterday

  • Outcome-Based Business Models: A decisive shift from selling SaaS tools ("the boat") to selling completed results ("the catch"), particularly in B2B and professional services.
  • The Management of Agents: Moving away from "one prompt fits all" toward "organizational design" for agents, including separate hardware, specific roles, and "onboarding" processes.
  • Economic Inversion: The realization that while the AI ecosystem is growing 5x, the vast majority of profit is still being captured by the hardware layer (NVIDIA), with application margins under pressure.
  • The "High Agency" Cult: A cultural and professional pivot toward "agentic" behavior and individual resourcefulness as the primary survival skill in an automated world.
  • Capture of Reasoning: The technical transition from recording "what happened" to instrumenting "why it happened" (decision traces) to create proprietary B2B data loops.