> ## Content Index
> Fetch the complete content index at: https://www.antoinebuteau.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Daily Digest - 2026-09-10
- URL: https://www.antoinebuteau.com/daily-digest-2026-09-10/
- Published: 2026-09-11T10:05:08.000Z
- Updated: 2026-09-11T10:05:08.000Z
- Description: OpenAI's Applied AI lead explains why even $100M+ ARR startups break agent behavior by continuously tacking on prompt rules, and how treating prompts like modular code improves retention and lowers costs.
- Author: Antoine Buteau
- Tags: Digest

## In this digest

1. [Why the world's best AI startups write bad prompts (& how to fix this)](#digest-item-1)
2. [Astra for Coding: Why Are We Doing This Again?](#digest-item-2)
3. [The Pulse: tech companies move to open AI models](#digest-item-3)
4. [I run six teams at Warp](#digest-item-4)
5. [Router Power](#digest-item-5)
6. [The 3 Waves of AI Consumption](#digest-item-6)
7. [The Hidden Pillar of Robotics Research is Deployment](#digest-item-7)
8. [Why Every AI Accountant Fails, and Why Palantir Knew First](#digest-item-8)
9. [Code-as-Policy: Sisyphus Deployed](#digest-item-9)
10. [Understanding RSI: A Comprehensive Guide to Self-Evolving AI](#digest-item-10)
11. [Knowledge as code](#digest-item-11)
12. [Model Mania](#digest-item-12)
13. [We grew from \~0 to $500M ARR, adding $250M last...](#digest-item-13)
14. [$600M ARR in 41 Months: ElevenLabs' Journey](#digest-item-14)
15. [Pushing DFlash2 Further: Training Where the Draft Breaks](#digest-item-15)

## Themes from yesterday

- **The Shift from Model Weights to Harnesses and Substrates:** Relying solely on larger models is hitting practical limits. Real reliability gains increasingly come from treating prompts, knowledge bases, and business ontologies as version-controlled, testable software modules (Agent = Model + Harness).
- **Enterprise Shifts to Open Models and Dynamic Routing:** High inference costs are leading companies like Uber, Pinterest, and AT&T to cut closed-model bills by more than 50% using model routers and specialized open models. At the same time, centralized routers gain an edge by training on shared error patterns across many companies.
- **Deterministic Execution Over Open-Ended Model Loops:** Running probabilistic language models directly inside live control loops leads to messy code and hardware failures in both software and robotics. Teams are moving toward code-as-policy, where models write and test code ahead of time while deterministic systems handle actual execution.
- **Scaling Growth Through Multi-Agent Workloads:** Token usage is reaching into the trillions as teams deploy parallel agent setups. Companies scaling quickly to $100M–$600M in ARR are combining automated production pipelines with sales incentives that reward employees for adopting AI tools.

## 1\. **Why the world's best AI startups write bad prompts (& how to fix this)** — Wulfie Bain

- Why read: OpenAI's Applied AI lead explains why even $100M+ ARR startups break agent behavior by continuously tacking on prompt rules, and how treating prompts like modular code improves retention and lowers costs.
- Summary: Production prompts often turn into contradictory messes because teams patch edge cases one by one without refactoring the overall prompt. Bain argues that teams should manage prompts like production software by dividing them into mutually exclusive, collectively exhaustive (MECE) sections that clearly separate internal execution logic from user-facing output formatting. This separation stops regressions, lets domain experts edit rules without breaking the system, and makes evals far simpler to design. Structuring prompts this way can cut latency by up to 50% and token costs by 30%. In practice, paying down prompt debt has become just as important for reliability as cleaning up legacy software.
- [Read more](https://twitter.com/wulfie%5Fbain%5F/status/2098060386813566990/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 2\. **Astra for Coding: Why Are We Doing This Again?** — Armin Ronacher

- Why read: Flask creator Armin Ronacher ran GPT-6 Astra through a 35-hour autonomous coding session, showing how frontier models optimize for token efficiency and task completion at the expense of readable, maintainable code.
- Summary: Over a weekend experiment, Astra burned through 4 billion tokens and $1,200 in API credits across 79 commits trying to add virtual threads to CPython, but produced nothing usable. Instead of adhering to codebase conventions, the model resorted to strange token-saving tricks, like splicing strings by hand in C, stripping whitespace, and chaining shell commands across different runtimes. Ronacher traces this to reinforcement learning setups that heavily incentivize task completion and low token counts while ignoring code readability and architecture. Left unchecked, autonomous coding setups can churn through budgets and stack up broken abstractions. Teams running coding agents need close human review, strict diff and linter checks, and hard stops on unvalidated loops.
- [Read more](https://lucumr.pocoo.org/2026/9/7/astra-why/?ref=antoinebuteau.com)

## 3\. **The Pulse: tech companies move to open AI models** — The Pragmatic Engineer

- Why read: Uber, Pinterest, and AT&T cut their AI spend by more than half by shifting key workloads away from proprietary models toward open-weight alternatives and dynamic routing.
- Summary: Engineering leaders are managing high AI bills by moving away from default proprietary APIs toward more deliberate infrastructure setups. Uber lowered session costs by 52% by running weekly capability benchmarks, compacting context once conversations hit 400k tokens, and sending subagent tasks to open-weight models that cost up to 8 times less than frontier options. Pinterest cut costs by more than 92% after post-training small open models on internal company data, while AT&T reduced its coding tool expenses by 56% using LiteLLM routing with only a 2% drop in output quality. With closed models like Claude Opus 5 costing up to 100 times more than capable open alternatives, relying entirely on proprietary APIs is getting too expensive. For sustainable operations, teams are pairing tiered model routers with benchmark-driven open deployments.
- [Read more](https://read.readwise.io/read/01m2661j5zkxc29vx2h3nrkx5s?ref=antoinebuteau.com)

## 4\. **I run six teams at Warp** — Gabriella Garcia

- Why read: A former Apple and Facebook engineer shares how Warp runs marketing, HR, and partnerships like software teams using Linear and Claude Code, cutting out layers of middle management.
- Summary: Warp organizes six business functions, including marketing, people, and partnerships, around standard software engineering practices. Every task lives in Linear, and operators use Claude Code to automate routine work. Under their team rule, any manual task done three times is treated as a bug, prompting operators to write their own scripts and prompts rather than waiting for software engineers. Weekly board reviews and small, scoped tickets keep work predictable without managers having to chase status updates. Because operators manage their own queues and write lightweight tools for their specific domains, projects move faster and work stays visible across the whole company.
- [Read more](https://twitter.com/reallygabriella/status/2098069203924259168/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 5\. **Router Power** — Soren Larson

- Why read: A simulation of 35,000 agent runs tests the enterprise data moat theory, showing why centralized model routers hold more structural leverage than private, company-specific agent setups.
- Summary: Spurred by Stripe's $7.5 billion purchase of OpenRouter, this study tested whether companies gain a real edge by training agents exclusively on internal data. Using Sierra's tau2-bench benchmark across 35,000 interactions, the author found that agent harnesses trained on pooled error data across multiple companies outperformed single-company agents by 6 percentage points. Across common workflows like customer support, sales, and operations, error patterns look remarkably similar, meaning competitors' failure logs often provide just as good or better training data. Centralized routers end up with the upper hand because they see failures across hundreds of setups and model updates. Unless a company holds proprietary physical sensor feeds, pouring money into bespoke private data moats for general software tasks is rarely worth the cost.
- [Read more](https://twitter.com/hypersoren/status/2098130598858551596/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 6\. **The 3 Waves of AI Consumption** — Tomasz Tunguz

- Why read: Tomasz Tunguz breaks down how AI usage is scaling through three distinct shifts, moving from simple chat prompts to parallel agent swarms that run through billions of tokens each day.
- Summary: Autonomous agents on OpenRouter recently overtook human chat traffic, jumping 14x in six months to hit 7.3 trillion tokens. Tunguz divides this growth into three distinct tiers: interactive chat at roughly 1 million tokens per user per day, single-agent coding and research at 100 to 200 million tokens per day, and meta-harness parallel orchestration reaching into the billions daily. In these multi-agent setups, a lead agent delegates tasks to dozens of concurrent subagents that run tools, analyze codebases, and parse spreadsheets in parallel. Goldman Sachs projects that monthly agent token usage will reach 120 quadrillion by 2030, fueled by concurrency rather than faster raw generation speeds. Infrastructure teams need to start sizing compute capacity and cloud budgets around concurrent agent loops rather than human typing habits.
- [Read more](https://twitter.com/ttunguz/status/2098208902299127921/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 7\. **The Hidden Pillar of Robotics Research is Deployment** — Deepak Pathak

- Why read: Skild AI co-founder Deepak Pathak explains how prioritizing factory floor deployments over polished video demos helped the company reach $100M ARR in ten months.
- Summary: Edited video clips can make a robot with a 10% success rate look just as capable as one hitting 99.9%, but real manufacturing plants cannot pause for human intervention. On mixed human-robot assembly lines, a single slow or stalled station halts the entire floor, forcing systems to hit tight cycle times without dropping accuracy. To handle changing real-world conditions without running constant fine-tuning jobs, Skild built its S1 model to pick up new tasks on the floor from a single demonstration video. The company feeds operational edge cases back into its core foundation model through what it calls a "Physical RSI" flywheel. Pathak argues that robotics labs need to measure progress by factory telemetry instead of staged video clips.
- [Read more](https://twitter.com/deepakpathak/status/2098102939017769386/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 8\. **Why Every AI Accountant Fails, and Why Palantir Knew First** — Erik

- Why read: The founder of T:0 explains why AI bookkeeping tools stumble on flat ledgers, and how Palantir-style ontology graphs create financial software that auditors can actually verify.
- Summary: Standard general ledgers record only the final balances of transactions, discarding the original context and forcing AI models to guess why numbers moved. Erik argues that instead of building chatty bookkeeping assistants on top of flat tables, developers should turn double-entry bookkeeping's 500-year-old rules into a structured ontology of typed objects and verifiable links. When financial records are organized as a graph, reconciliations turn into automated data checks, variance analyses become straightforward graph traversals, and audits can run continuously. Crucially, the architecture enforces a strict boundary: probabilistic language models propose journal entries, but a deterministic engine performs the math and writes to the ledger. For vertical AI tools, building reliable software comes down to structuring the underlying domain data correctly, not expecting an LLM to figure it out from scratch.
- [Read more](https://twitter.com/eya0/status/2097801524579864803/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 9\. **Code-as-Policy: Sisyphus Deployed** — Zeon Systems

- Why read: Robotics startup Zeon Systems explains why asking models to write modular Python programs is far more reliable for factory and lab automation than putting them directly into live control loops.
- Summary: Placing language models directly inside real-time robot control loops introduces unpredictable variations that cause frequent errors in precise environments like scientific laboratories. Zeon solves this by using a code-as-policy approach: the AI writes modular Python scripts that stitch together tested motion and perception routines. The system tests these generated scripts inside 3D simulations before executing them deterministically on physical hardware. Neural vision-language-action models are reserved strictly for small, unpredictable steps that are difficult to model mathematically, such as handling flexible films or pouring liquids. Separating the model's planning stage from physical hardware execution gives teams the safety and repeatability needed for production environments.
- [Read more](https://twitter.com/zeonsystems/status/2098121083882971323/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 10\. **Understanding RSI: A Comprehensive Guide to Self-Evolving AI** — Yaowei Zheng

- Why read: Yaowei Zheng breaks down recursive self-improvement across five layers: model weights, runtime context, long-term memory, modular skills, and harness code.
- Summary: Autonomous systems are best understood through the framework Agent = Model + Harness, which expands self-improvement well beyond updating model weights. While retraining weights is slow and expensive, improving the surrounding harness through rewritten context, persistent memory stores, curated skill libraries, and updated orchestration code offers a safer, reversible way to build smarter systems. Zheng catalogs different evolution patterns, from linear chains to branching trees like the Darwin Gödel Machine and comparative graphs like the Mendel Gödel Machine. These setups typically split duties between a student agent performing the task and a teacher meta-agent validating progress in offline, online, or hybrid loops. For most teams, refining the harness and its tools yields steady performance gains without the complexity of continuous weight training.
- [Read more](https://twitter.com/code%5Fhiyouga/status/2098021114685616637/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 11\. **Knowledge as code** — Dave Nunez

- Why read: The former head of documentation at Stripe and Uber argues that enterprise AI fails when internal wikis are outdated, and outlines a Git-based playbook to keep company knowledge accurate.
- Summary: Enterprise AI agents often fall short not because of model limitations, but because corporate wikis are filled with stale information, broken links, and conflicting policies. Pointing retrieval systems at disorganized documents sends agents into expensive loops that produce outdated answers. Nunez advocates for "Knowledge as Code," treating internal documentation like software in centrally managed, version-controlled Git repositories accessible through CLIs and APIs. By introducing regular code practices like pull-request reviews, automated contradiction tests, and documentation health metrics, teams can maintain a dependable knowledge base. Clean, machine-readable documentation is an essential prerequisite for any reliable internal AI assistant.
- [Read more](https://twitter.com/capitaltruist/status/2098103576149311570/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 12\. **Model Mania** — Austin Johnsen from Artificial Diligence

- Why read: A benchmark evaluating models on corporate development diligence memos highlights GPT-6 Astra's analytical strengths, exposes major flaws in automated grading rubrics, and crowns DeepSeek v4 Flash the best budget performer.
- Summary: Johnsen tested nine frontier and open-weight models on a diligence memo task, with GPT-6 Astra leading the group in spotting critical issues and avoiding false alarms. But during initial automated grading, Astra received a zero on decision quality because its recommendations did not match the simple regex patterns looking for a specific conclusion sentence. A manual review and rescoring of 228 runs shifted 112 scores, proving that brittle keyword rubrics frequently penalize thoughtful, nuanced outputs. Among budget models, DeepSeek v4 Flash stood out as the clear cost-efficiency winner, offering near-frontier analysis at a fraction of proprietary API pricing. For evaluation teams, testing frontier reasoning requires flexible semantic judges rather than rigid keyword matching.
- [Read more](https://read.readwise.io/read/01m26k9szpvmz28mwh3nyqkkbs?ref=antoinebuteau.com)

## 13\. **We grew from \~0 to $500M ARR, adding $250M last...** — Rohan Nayak

- Why read: Pocket FM founder Rohan Nayak details how the audio platform scaled to $500M in profitable ARR, adding $250M in its latest year by combining automated ad creation with culturally adapted audio series.
- Summary: Pocket FM reached $500M in ARR while remaining EBITDA profitable by moving entirely to AI-driven production. The company grew its catalog from 25,000 to 2.5 million hours annually across 770,000 titles, while improving 12-month revenue retention from 44% to 76% with LLM-based quality filters. To drive customer growth, their marketing system generates 17,500 video trailer ads each month. Nudging click-through rates from 2.0% to 2.25% lowered acquisition costs by 30%, making large-scale ad spend viable. Instead of using word-for-word translations, they adapted stories to local humor, folklore, and idioms, which lifted European retention by 50%. For consumer AI companies, growth depends on linking automated content production directly to high-volume distribution tests.
- [Read more](https://twitter.com/RohanNayak2/status/2098104784264065525/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

## 14\. **$600M ARR in 41 Months: ElevenLabs' Journey** — SaaStr

- Why read: ElevenLabs' founding revenue leader explains how the voice company crossed $600M ARR in 41 months, in part by paying human sales reps full commissions on deals closed by AI agents.
- Summary: ElevenLabs grew to more than $600M ARR in under four years, marking one of the fastest software expansions to date. A critical decision in their sales strategy was paying human account executives full commissions even when AI agents handled and closed enterprise deals. Aligning incentives this way removed internal resistance, prompting reps to actively train, deploy, and refine sales agents across their pipelines. The company's trajectory also highlights an evolving venture market where top investors now look for $25B exit potential to support early valuations. For founders deploying internal agents, adjusting compensation plans is often the simplest way to turn employees into partners instead of competitors.
- [Read more](https://read.readwise.io/read/01m25q3gzhqc6nw7etjb153906?ref=antoinebuteau.com)

## 15\. **Pushing DFlash2 Further: Training Where the Draft Breaks** — Nicholas Liu

- Why read: Nicholas Liu explores speculative decoding in production, showing how training block-diffusion draft models on rejection boundaries pushes inference speedups from 2.51x to 3.34x.
- Summary: Speculative decoding speeds up large language models by using a lightweight draft model to suggest tokens, which the main model verifies in a single forward pass. While block-diffusion drafters like DFlash propose multiple tokens at once to skip step-by-step delays, predicting tokens in parallel can produce awkward phrases that cause the target model to reject the draft early. Liu's team adapted DFlash2 by training the drafter specifically around the points where drafts fail, lifting the average accepted draft length from 3.59 to 4.89 tokens. In their production setup, this increased overall speedup from 2.51x to 3.34x. For inference teams, training draft models specifically on their error boundaries is an effective way to lower latency and serving costs.
- [Read more](https://twitter.com/NicholasLiu77/status/2098097570262491555/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)