In this digest
- A new shape for work
- How to build a knowledge factory
- Combatting adversarial distillation
- Moving into AI Product Engineering
- What Canva’s reorg, OpenAI’s funding reset, and Adobe’s AI distribution move mean for your GTM.
- The agentic story on the eve of the 2026 holiday season
- What’s 🔥 in AI/Infra/VC #525
- Intuit's Mailchimp Sale: A $3.68B Turnaround
- -$6 Trillion In 45 Days: Is This A "Bubble" Or An Opportunity?
- Some random, unfiltered thoughts on AI and the market:
- 2026 Cutting Edges Report: A Sneak Peek
- the riemann hypothesis, from first principles
- How to Build an Agent
- Does it really matter if there’s an intelligence explosion?
- Paper Dragon Anxiety
Themes from yesterday
- From single-player chat to shared context systems: Notion, Harrison Chase, Dylan (PostHog), and Sangram all point to the same operational bottleneck. Lasting AI gains come from shared context, curated reusable skills, evals, and updated org and unit-economics structures rather than ad-hoc prompting.
- Public-market turbulence vs. real agent momentum: Even with a $6T equity drawdown, Nvidia leadership turmoil, and legacy SaaS multiple compression (Mailchimp at 2.9x), underlying numbers keep climbing: Anthropic's record $81.3B IPO, +31.9% S&P EPS growth, Netlify's agent-driven signup surge, and an 80x jump in Stripe agentic wallet spend.
- Agent distribution and commerce become contested channels: Adobe embedding workflows inside ChatGPT and Claude via MCP and A2A, OpenAI launching ChatGPT Wallet, and United moving to block booking agents signal a move from standalone apps to contested agent-mediated workflows.
- AI capability gains are outpacing security and verification pipelines: OpenAI's anti-distillation defenses, major outages triggered by rushed patches for AI-found flaws, autonomous AI pentesting (Artex), and the 175,000-file Lean formalization of a claimed Riemann Hypothesis proof all show why formal verification and automated defense can no longer lag behind model progress.
1. A new shape for work — Notion
- Why read: Notion explains how teams can move past isolated, single-player AI chat and build shared workflows that improve over time.
- Summary: Most teams see small returns from AI because individual chatbots behave like stateless computers without a hard drive. Context stays scattered, and employees keep re-explaining the same background details. Notion proposes a two-part shift: first build a shared "context system" where humans and agents read from and write to the same persistent workspace, then create shared AI "factories" maintained by a small team for the rest of the company. To get reliable results across repeated runs, operators should stop treating agents like chat assistants and treat them like programs guided by natural-language "skills," meaning reusable markdown instructions paired with templates, scripts, and tools. Because those skills keep getting refined, they turn one-off demos into auditable business infrastructure and free up people for higher-judgment creative work. Teams do not need a top-down reorganization to start; picking one shared task list or recurring workflow can produce enough visible value to bring the rest of the organization along.
- Read more
2. How to build a knowledge factory — Harrison Chase
- Why read: LangChain's CEO lays out the system architecture needed to keep enterprise agent knowledge accurate without burying teams in manual upkeep.
- Summary: Chase argues that the main bottleneck for enterprise AI is not better models or larger context windows, but turning tacit, scattered company know-how into reusable knowledge for agents. Drawing on LangChain's internal GTM, content studio, and LangSmith Engine deployments, he splits agent knowledge into three categories: procedural skills (how work gets done), domain knowledge (stable business definitions and facts), and data access guidance (where live truth lives and how to query it). Humans cannot keep all of this updated by hand, and task-executing agents should not edit shared knowledge directly because that leads to contradictions and drift. Instead, companies need a closed-loop "knowledge factory." Working agents flag missing or outdated information through a dedicated tool, and a scheduled background curator agent reviews those flags alongside raw docs and execution traces to update skills cleanly. For operators, the practical move is to stop documenting everything upfront: start with one recurring workflow, wire up trace review and issue flagging, and let actual usage drive ongoing updates.
- Read more
3. Combatting adversarial distillation | OpenAI — OpenAI
- Why read: OpenAI explains how unauthorized distillation of reasoning traces transfers dual-use cyber capabilities and details the layered defenses built to stop it.
- Summary: Unauthorized distillation is more than an IP problem; it lets adversaries copy frontier capabilities in sensitive areas like cybersecurity and biology into student models that lack safety controls. In an internal cyber transfer study, OpenAI found that fine-tuning on a teacher model's exact internal chain-of-thought traces produced stronger downstream cyber performance than training on user-visible reasoning summaries alone, especially at higher inference budgets after reinforcement learning. To protect those traces, OpenAI pairs request-level defenses (adversarial refusal training, streaming and non-streaming input and output classifiers covering tool-call arguments, and API rules binding encrypted reasoning to an unmodified conversation prefix) with campaign-level cross-account investigations and industry indicator sharing. Using automated adaptive red-teaming, OpenAI cut average attack success rates for extracting prior-turn protected reasoning from 54.0% on GPT-6 Astra down to 0.2% on GPT-6.2 Astra. For AI builders and security teams, the takeaway is that tool-call arguments and conversation-history tampering are real exfiltration paths that require layered classifiers and prefix-integrity checks rather than prompt-level guardrails alone.
- Read more
4. Moving into AI Product Engineering — Dylan
- Why read: A PostHog engineer shares practical lessons from moving out of traditional backend and feature-flag work into applied AI product development, focusing on evals and agent UX.
- Summary: Moving into AI product engineering still relies on classic systems habits: understanding failure modes, defining what correctness looks like, and staying close to users. The biggest new technical habit is writing evals for nondeterministic LLM outputs, such as catching regressions where an analytics agent quietly shifts from counting unique users to counting raw signup events when sliced by country. Even harder than evals is drawing the right line between human UI and agent interfaces like MCP and skills. Agents are well suited to running clearly scoped, multi-step workflows through tools so users do not have to click through menus, but humans still need clear visual previews (like blast-radius graphs before rolling out a feature flag or visual views of affected users) to decide what an agent should do. Product teams should combine idempotent, well-scoped agent tools with deliberate visual decision surfaces instead of assuming text chat should replace UI everywhere.
- Read more
5. What Canva’s reorg, OpenAI’s funding reset, and Adobe’s AI distribution move mean for your GTM. — Sangram
- Why read: Three major late-2026 tech moves are broken down into a practical 2027 operating plan covering org structure, capital discipline, and AI-native distribution.
- Summary: Canva's restructuring affected 1,271 roles and combined Product, Growth, and Platform into a single unit after the company cut revenue growth projections from 30% to 20% and delayed AI launches over serving costs. That move shows how AI inference turns product and growth choices straight into margin decisions. Meanwhile, reports that OpenAI lowered its funding target from $30B to $20B (with an ~$11B first close) make clear that operators cannot plan 2027 budgets around endless cheap capital or "monetize later" assumptions. On the distribution front, Adobe embedding CX Enterprise Coworker inside ChatGPT, Claude, Gemini, Databricks Genie, and Microsoft Copilot via MCP and A2A points to a shift away from pulling users into standalone apps and toward meeting them inside their active AI assistants. Operators should audit departmental handoffs to establish shared unit-economics ownership, tie every AI project to explicit commercial outcomes and review gates, separate AI discovery (AEO) from workflow execution integrations, and treat third-party assistants as borrowed distribution while continuing to invest in direct customer relationships.
- Read more
6. The agentic story on the eve of the 2026 holiday season — Ernie Tedeschi
- Why read: New Stripe data offers a clear look at actual consumer spending through autonomous shopping agents.
- Summary: Heading into the 2026 holiday shopping season, Stripe's Link Agent Wallet (which handles payments for personal agents including Muse, Instinct, and Grok) logged daily spend nearly 80x higher in mid-November than at the start of September on a seven-day rolling average, doubling over the past month alone. Purchases are clustered in high-friction comparison categories: more than 40% of recent volume went to airfare, travel, and lodging, about a third went to retail, and a quarter went to other categories. The usage also looks sticky rather than novelty-driven, with more than half of shoppers who complete one agent purchase coming back for at least a second one, usually in the same category. For commerce, travel, and fintech teams, that repeat buying suggests checkout compatibility, structured pricing and inventory data, and agent payment rails are fast turning from experiments into a real conversion channel.
- Read more
7. What’s 🔥 in AI/Infra/VC #525 — Ed Sim from What's Hot 🔥 in AI/Infra/VC
- Why read: Ed Sim looks at how Anthropic's sub-six-year sprint to a trillion-dollar IPO has changed venture capital, early-stage pricing, and capital requirements across tech.
- Summary: Software used to require very little upfront capital, but hyperscaler CapEx expectations are nearing $1T and frontier labs are raising enormous amounts of money: Anthropic raised over $100B in 5.8 years to reach a $1T+ IPO, whereas SpaceX raised ~$9B privately over 24 years before tapping $65B in post-IPO debt and GPU financing. That outcome has produced sharp "barbell" venture dynamics, pushing nearly 40% of priced first US AI/ML rounds to $10M+ (up from 11% in 2023) and stretching the gap between median and 95th-percentile inception valuations to 10x as mega-funds chase trillion-dollar options. The downside is heavy capital risk: Goldman estimates the AI buildout requires roughly $3.5T in annual customer revenue to pay off, while strong "tweener" startups capped at $10B outcomes face a funding squeeze. Sim contrasts legacy software companies getting repriced downward (such as Mailchimp's sale to Bending Spoons) with agent-native turnarounds like Netlify, which just reached 100,000+ daily signups on track for 25M users by year-end. Founders and investors are left with two clear paths: stay extremely capital efficient outside mega-fund consensus, or build genuine agent-native distribution that compounds organic adoption.
- Read more
8. Intuit's Mailchimp Sale: A $3.68B Turnaround — SaaStr
- Why read: Intuit's exit from Mailchimp puts a public price tag on mature, zero-growth B2B SaaS at the exact moment Anthropic's IPO shows what hyper-growth AI economics look like.
- Summary: Five years after acquiring Mailchimp for $12 billion (~15x revenue when it was profitable and growing 20% at $800M), Intuit agreed to sell the business to Bending Spoons for $3.68 billion ($3.31B in cash and $368M in stock), roughly 30 cents on the dollar and 2.9x revenue on 0% growth. SaaStr notes that many B2B software companies are similarly large, profitable, and stalled, but until now the market lacked a clean large-scale transaction to price them. Mailchimp sets that floor. In the same issue, SaaStr compares this repricing with Anthropic's IPO figures: $33 billion in nine-month revenue, 900% net revenue retention, and a 12x one-year jump in $100K+ customers. For SaaS executives and boards, the message is direct: profitability without durable AI-era expansion compresses multiples sharply, widening the valuation gap between static workflow incumbents and rapidly growing AI platforms.
- Read more
9. -$6 Trillion In 45 Days: Is This A "Bubble" Or An Opportunity? — The Kobeissi Letter
- Why read: A detailed macro breakdown explains the overlapping drivers behind the $6 trillion market selloff and why underlying AI earnings still look very different from crisis-level sentiment.
- Summary: Over 31 trading sessions since early October, US stocks shed $6 trillion amid a Q3 Big Tech CapEx shock ($201.9B combined across Amazon, Alphabet, Meta, and Microsoft, up 356% since Q1 2024), aggressive Fed rate hikes, a Democratic midterm sweep ending the "Trump Put," leadership turmoil at Nvidia after Jensen Huang's departure, and a large liquidity drain from Anthropic's oversubscribed IPO. Yet beneath the panic, where the VIX doubled and tech forward P/E fell to March 2020 lows (with Nvidia trading at 9x despite ~$750B FY2028 revenue guidance), S&P 500 companies delivered +31.9% YoY EPS growth in Q3, beating estimates for a 13th straight quarter. Kobeissi argues the drop looks more like a liquidity and sentiment correction than a popped AI bubble, pointing out that extreme regulatory fears such as a US AI ban are unlikely given geopolitical competition with China, while falling oil prices could ease Fed pressure in 2027. For operators and allocators, the lesson is to separate short-term market liquidity shocks and free-cash-flow timing worries from real enterprise earnings momentum when planning hiring, fundraising, or capital deployment.
- Read more
10. Some random, unfiltered thoughts on AI and the market: — BuccoCapital Bloke
- Why read: A blunt look at the widening split between Wall Street's infrastructure panic and how Silicon Valley is restructuring companies around AI.
- Summary: The divide between "New York" and "San Francisco" has rarely been wider. Public markets are repricing perfection after shocks at Nvidia and OpenAI, while on the ground AI is changing how big a company needs to be, illustrated by Anthropic (~6,000 employees) and Meta (67,000 employees) both hovering near $1.5T valuations. Greenfield startups are building from scratch around cheaper, fast-improving intelligence and scaling rapidly, while Fortune 500 incumbents still struggle to get even basic agents into production at scale. Problems at individual bellwether vendors do little to slow real-world AI adoption; broader economic softness can even speed it up because tokens are more flexible and fungible than headcount. The market is moving into phase two of the AI buildout, where more value shifts to the application layer even though long-term defensibility remains unsettled. Operators and investors should spend less time predicting data-center counts or Washington policy shifts and more time identifying leaders who can actually rebuild their companies around abundant intelligence.
- Read more
11. 2026 Cutting Edges Report: A Sneak Peek — Contrary Research
- Why read: Contrary rounds up the week's biggest signals across power bottlenecks, mega-IPOs, AI-driven cybersecurity failures, and agentic-commerce friction.
- Summary: Alongside a preview of its upcoming 2026 Cutting Edges Report on global power interconnection queues and regulatory bottlenecks, Contrary tracks a dense week of shifts across tech and policy. In infrastructure and capital markets, Anthropic raised $81.3B at a ~$1.76T fully diluted valuation on Nasdaq, the Senate passed a permitting bill requiring large data centers to pay more grid costs, and Nvidia halted H200 sales to China as Commerce closed overseas cloud-access loopholes. Technical operators also face two urgent security warnings: Wall Street executives and the NSA traced major outages at Microsoft, JPMorgan, Apple, and the FAA to rushed patches for AI-discovered vulnerabilities moving faster than safe deployment pipelines, while the FBI warned hackers are using the open-source AI pentesting tool Artex to breach networks in under three hours. Finally, OpenAI's launch of biometric-approved ChatGPT Wallet alongside United Airlines moving to block agent bookings shows where commercial friction over agentic transactions is building. Taken together, these developments point operators to harden patch-management pipelines, factor grid and compute costs into budgets, and prepare for both defensive and offensive agent disruption.
- Read more
12. the riemann hypothesis, from first principles — karan ganesan
- Why read: Along with a clear visual explanation of math's most famous open problem, this post examines Anthropic researcher Levent Alpöge's claimed Claude-assisted Lean proof and what verifying it actually takes.
- Summary: Ganesan builds up to the Riemann Hypothesis from basic primes and Euler's zeta function, explaining its core claim that every non-trivial zero lies on the ½ critical line, which keeps prime-count fluctuations tightly bounded near √x. He then examines the November 2026 claim by number theorist Levent Alpöge that he and Claude proved the hypothesis by manufacturing high vanishing multiplicity inside automorphic forms using a chain of ~√x iterated theta lifts (a chain length Alpöge credits to Claude), formalized across roughly 175,000 Lean files referencing mathlib's standard definition. Rather than accepting the claim on faith, Ganesan outlines a strict verification checklist: checking whether the archive builds cleanly against unmodified mathlib using only Lean's three standard axioms, with no smuggled definitions, hidden gaps, or kernel bugs. Even if independent verification holds up, human-readable exposition still lags behind machine-checkable proof, pointing to a major shift in AI-assisted research. For technical leaders, the broader lesson is how formal verification tools like Lean let AI systems move past human intuition while keeping correctness auditable without blind trust.
- Read more
13. How to Build an Agent — ampcode.com
- Why read: A hands-on guide takes the mystery out of coding agents by building a working file-reading and editing agent from scratch in under 400 lines of Go.
- Summary: Stripping away the hype around modern coding agents, this walkthrough shows that an autonomous agent is just an LLM, a conversation loop, and structured access to tools that read or change state outside the context window. Starting with a 90-line terminal chat loop built on the Anthropic SDK, the guide shows how tool use works under the hood: you supply each tool's name, behavioral description, JSON input schema, and execution function, and the host loop runs the model's requested calls and feeds the outputs back into the stateless conversation history. Implementing `read_file` with Go struct reflection demonstrates how little glue code is needed before an LLM starts chaining actions, recovering from errors, and working directly on a local codebase. For product engineers and technical leaders evaluating agent frameworks, building a minimal loop like this shows where real product differentiation lives: not in exotic orchestration layers, but in tool design, clear descriptions, error handling, and tight feedback loops.
- Read more
14. Does it really matter if there’s an intelligence explosion? — Noah Smith
- Why read: Noah Smith argues that today's steady, fast rise in AI capability is already disruptive enough to make debates over a sudden Singularity mostly beside the point.
- Summary: Smith maps out the current split between skeptics like Ramez Naam and Theo Jaffee (who point to diminishing returns, jagged agent workdays, and OpenAI's own data showing incomplete automation of short research tasks) and accelerationists like Dwarkesh Patel, who expect verifiable AI research to repeat math's sudden leap from struggling on arithmetic to tackling Millennium Prize problems. Interesting as that debate is, Smith argues the "intelligence explosion" question is a weak foundation for policy or planning. First, if a Singularity threshold exists, "pacing the frontier" only delays it by a few years rather than stopping it; global competition makes a permanent halt politically unrealistic, and superexponential progress would outpace human alignment work anyway. Second, steady non-explosive progress is already set to make AI superhuman across most cognitive domains, and increasingly physical ones, within a short timeframe, bringing threshold-based economic adoption shocks whether or not recursive self-improvement goes vertical. Operators and policymakers should plan for rapid, ongoing spread of superhuman capabilities instead of waiting for a binary verdict on a sci-fi Singularity.
- Read more
15. Paper Dragon Anxiety — Kyle Harrison from Investing 101
- Why read: Harrison explains why smart analysts reach opposite conclusions on China, and why its industrial capacity, robotics, and software-defined manufacturing still matter for Western tech and defense operators.
- Summary: Responding to growing "China bear" arguments from analysts like Noah Smith and academics like Stephen Brooks, Harrison contends that both sides are reading different scoreboards. On the commercial-profit scoreboard, the US leads comfortably, taking roughly 38% of global corporate profits versus China's 16%, while China deals with overstated GDP numbers and steep demographic decline (7.9M births in 2025). On the physical-capacity scoreboard, however, China accounts for ~31% of global manufacturing value added (versus ~15% for the US), built over 53% of the world's commercial ships in 2024 (versus 0.1% for the US), and is offsetting its shrinking workforce by installing 54% of all industrial robots worldwide as part of a push toward software-defined production. Harrison frames the decade as an asymmetric race: the US trying to reindustrialize while China races to automate before its labor pool craters, noting that in a conflict, factories and shipyards count for much more than profit margins. For founders and investors in defense, hardware, robotics, and supply-chain infrastructure, the essay argues for building domestic industrial capacity rather than banking on Chinese stagnation.
- Read more