ZeroNoise Logo zeronoise
Post
Build the system around the model
1 day ago
5 min read
137 docs
An organic recommendation list centered on the work between model capability and real outcomes: enterprise workflows, agent operations, chip/software co-design, founder education, and China-tech context.

A useful pattern runs through the strongest recommendations: the hard part is not access to a capable model but the design of the workflow, organization, or hardware that makes it useful. Aaron Levie’s endorsement is the clearest anchor: he says the gap between AI models and enterprise workflows leaves room for applied-AI companies, and names context, change management, model routing, business-system integration, workflow UX, and evaluations as the work beyond model intelligence.

Standout recommendation

Moats in the age of floods

  • Type / creator: Article by @nayakkayak. The article frames intelligence as a utility that needs systems to distribute it into real-world outcomes, rather than assuming the model layer captures all the value.
  • Link:Read the article.
  • Recommended by: Aaron Levie, who calls it a strong account of applied-AI strategy at scale.
  • Key takeaway: The article argues that real work remains difficult because it carries context, exceptions, incentives, approvals, accountability, legacy systems, and human coordination; adoption can move faster than the institutional change needed to turn capability into productivity.
  • Why it matters: Use it as a filter for applied-AI companies: what do they own in context, change management, integration, workflow UX, routing, and evaluation—not merely which model they access?

Infrastructure and inference

OpenAI Jalapeño

  • Type / creator: SemiAnalysis article.
  • Link:Read the article. Elad Gil shared it with the comment, “Impressive speed to tape out by OpenAI.”
  • Recommended by: Elad Gil.
  • Key takeaway: SemiAnalysis describes Jalapeño as a generalized LLM-inference chip developed with Broadcom from a blank slate; design began in mid-2024 and reached manufacturing tape-out in roughly 16 months. The article attributes its reported performance to tight hardware/software co-design rather than narrow specialization for OpenAI models.
  • Why it matters: The useful lesson is the bring-up loop, not a simplistic “OpenAI beat Nvidia” claim: SemiAnalysis says the team produced strong results within nine months of tape-out and three months on real silicon, but also notes that the performance numbers came from OpenAI, the full benchmark suite was not run, and production was only scheduled to ramp gradually during 2027.

DeepSeek’s published research on KV-cache compression

  • Type / creator: Research; DeepSeek.
  • Recommended by: Neil Movva, identified in the discussion as an ex-Nvidia GPU and kernel engineer who now runs Sail Research.
  • Key takeaway: Movva says the KV cache is still “quite uncompressed,” with storage potentially off by one or two orders of magnitude, and points to DeepSeek’s research as making roughly order-of-magnitude progress every year.
  • Why it matters: This is a concrete research direction for anyone studying inference cost: memory representation, not just raw compute, may be a major source of future efficiency gains.

Agents and company design

Coding with agents

  • Type / creator: A 50-minute podcast/video posted by @0xCodez, featuring a SpaceXAI engineer identified as ex-Cursor.
  • Link:Martin Casado’s recommendation.
  • Recommended by: Martin Casado calls it the “Single best discussion ever recorded on coding with agents.”
  • Key takeaway: The featured engineer says he runs 10–20 agents that automate 90% of his routine work, with a “Chief of Staff” agent that knows about and manages the other bots.
  • Why it matters: It offers an operating example of delegation and coordination to examine, rather than another abstract claim that agents will change software development.

Founder Mode and Lee Robinson’s product-engineer post

  • Type / creators: Paul Graham essay; Lee Robinson blog post, which Guillermo Rauch refers to as “the rise of the product engineer.”
  • Link:Rauch interview discussing both recommendations.
  • Recommended by: Guillermo Rauch. He says Founder Mode resonated because founders must go deep and stay in the weeds, then describes scaling that behavior through autonomous product-level owners.
  • Key takeaway: Rauch calls Robinson’s post “great” and “spot-on” for combining product management with P&L, customers, expansion, retention, and engineering judgment.
  • Why it matters: Together, the two recommendations form an organizational-design pair: preserve founder-level depth while giving capable people end-to-end ownership of products.

How Universities Should Prepare Founders

  • Type / creator: Essay by Paul Graham.
  • Link:Read the essay. Ryan Hoover shared it while saying he has never been a fan of whiteboarding a startup.
  • Recommended by: Ryan Hoover.
  • Key takeaway: Graham argues that universities should not create an entrepreneurship curriculum; they should teach subjects that make students good at building things, make starting a company feel viable, and encourage students to work on their own projects.
  • Why it matters: The project prescription is unusually concrete: projects deepen subject knowledge, reveal potential cofounders, make self-direction natural, and generate ideas that would look implausible on paper. Graham contrasts that with business-plan competitions, arguing that founders should build prototypes for users rather than practice stories for investors.

A China-tech reading stack

Sinocism, China Talk, Hello China Tech, and Breakneck

  • Type / creators: Three China-focused blogs/newsletters—Sinocism, China Talk, and Hello China Tech—plus Dan Wang’s book Breakneck: China’s Quest to Engineer the Future.
  • Recommended by: Packy McCormick, who calls the three publications “great China-focused blogs for western audiences” and the book “great.”
  • Key takeaway: Packy’s rationale is contextual: readers tracking energy, AI, hardware, and global competition need a better understanding of China and should use sources that explain it from more than a cartoonish adversarial frame.
  • Why it matters: This is a practical starter stack for correcting a U.S.-only view of technology competition, rather than a recommendation to consume another general futurist essay.

A primary source with a founding consequence

Bitcoin white paper

  • Type / creator: Technical paper; Satoshi Nakamoto.
  • Recommended by: Brian Armstrong, whose endorsement comes as an origin story: after first reading the white paper while working at Airbnb, he spent nights and weekends building the prototype that eventually became Coinbase, then carried Airbnb’s practices into the new company.
  • Key takeaway: Armstrong presents the paper as a trigger for moving from curiosity to a working prototype.
  • Why it matters: It is a rare founder-level example of a primary technical document producing action, not just changing an opinion.
Build the system around the model
Research extraction

OpenAI Jalapeño is a clean-sheet inference ASIC developed with Broadcom for LLM inference, but SemiAnalysis emphasizes that it is generalized rather than limited to OpenAI models: it is intended to run varied models and workloads, including the InferenceX benchmark.

  • Why the tape-out speed stands out: design work began in mid-2024, and the article says OpenAI went from initial team hiring to manufacturing tape-out in about 16 months—an “extremely fast ASIC development cycle.” The November 2025 tape-out covered the CoWoS package design, not merely the top-die silicon; within nine months, and after only three months of bring-up on real silicon, OpenAI had already produced strong results despite starting from zero on the software stack.
  • Why that speed is strategically important: first-generation chips are generally not competitive, yet the article says Jalapeño was already industry-leading against the Nvidia, AMD, and Google chips SemiAnalysis had tested on multiple open-source models, attributing this to unusually tight hardware/software co-design. SemiAnalysis therefore treats the rapid tape-out-to-working-software path—not tape-out alone—as evidence that a frontier AI lab can compress both chip development and accelerator bring-up.
  • The AI-design signal: the article says the timeline supports claims that AI is accelerating chip design, while also crediting OpenAI’s spending, pragmatic design choices, and highly capable team; it separately reports OpenAI’s claim that AI assistance reduced SIMD area by 8% and matrix-engine area by 10%, with improved timing and power, though the exact PVT conditions were not clarified.
  • The comparison that makes the pace especially striking: Rubin’s CoWoS tape-out occurred about one month earlier, in October 2025, but SemiAnalysis says the early public results were only from CoreWeave engineering samples and interprets the difference as faster Jalapeño software bring-up rather than proof that Nvidia hardware is inferior. The article goes as far as saying the CUDA moat is “potentially dead” if OpenAI can bring up new models on its silicon this quickly.
  • Post-tape-out iteration reinforces the point: the team reportedly achieved more than 2× throughput improvements at some interactivity levels in under two weeks and enabled TP32, including a full rack-scale configuration, in eight days. SemiAnalysis links this rapid kernel and workload bring-up to Codex-assisted software development and says the approach is promising because it can turn broad architectural headroom into working kernels quickly.
  • Important qualification: the article’s performance numbers were supplied by OpenAI; SemiAnalysis verified InferenceX runs in person but did not run the full suite or see AgentX results. It also says Blackwell is an incomplete comparison because Jalapeño’s HBM4 positioning makes Rubin the more appropriate rival, while Jalapeño was still at the engineering-sample stage and production was scheduled to ramp gradually during 2027.
OpenAI Jalapeño: Better Than Nvidia Blackwell
Research extraction

Specific argument: Universities should prepare founders indirectly, by teaching powerful ideas and making students good at building things—not by creating an “entrepreneurship” curriculum. Graham says YC looks for people who can build and habitually do so, while startup success depends on product knowledge and execution.

  • The two institutional changes he recommends are to make students feel that starting a startup is a viable option and to encourage them to work on their own projects. He suggests showing students relatable founders to make entrepreneurship seem both desirable and accessible.
  • Student projects are central because they deepen subject knowledge, help potential cofounders discover whether they work well together, make the self-directed nature of startups feel natural, and generate promising startup ideas that students might otherwise reject as implausible.
  • He explicitly rejects teaching startup formation through ordinary university classes or business-plan competitions: starting a startup must be learned by doing, and investor-pitch exercises misleadingly prioritize fundraising stories over building prototypes that users want.
  • The ideal program should therefore look relatively quiet and cost little: give students more unscheduled time, let projects remain unofficial rather than over-managed, and build an organic startup culture instead of adding deans, buildings, or “innovation” infrastructure.
How Universities Should Prepare Founders
Brian Armstrong
Profile
  • Bitcoin white paper (paper): Brian Armstrong says reading it while working at Airbnb led him to spend nights and weekends building a prototype that would become Coinbase; he also carried Airbnb’s startup practices into Coinbase.
  • Unnamed recent documentary about Satoshi Nakamoto’s identity (documentary/video): Armstrong calls it “good,” says it argues that Hal Finney and Len Sassman were likely Satoshi, and says he tends to accept that theory while acknowledging that nobody can know for sure.
The Man Behind Coinbase: Interview with Brian Armstrong | KMP Ep.52
Guillermo Rauch
Profile
  • Paul Graham — “Founder Mode” (essay). Guillermo Rauch says the essay resonated strongly because it argues that founders must go deep in what matters and stay in the weeds; he applies the principle to scaling by giving product-level leaders autonomous ownership.
  • Lee Robinson — post on “the rise of the product engineer” (article/blog post). Rauch calls Robinson’s post “great” and “spot-on,” valuing its model of combining product management, P&L, customer, expansion, retention, and engineering thinking in one role.
Inside Vercel's AI-era playbook: "Recursive Founder Mode" | Guillermo Rauch (CEO)
Invest Like The Best
  • DeepSeek’s published research on KV-cache compression(research; creator: DeepSeek; recommended by the interview guest) — The guest called the work “really interesting” and said it is making good progress. He highlighted its reported order-of-magnitude advances roughly every year as evidence that KV-cache storage has substantial room for improvement, estimating current storage may be off by an order of magnitude or two.
Ex-NVIDIA Engineer: Why AI Is About to Get 1000x Cheaper
Not Boring by Packy McCormick

Packy McCormick recommends Sinocism, China Talk, and Hello China Tech as “great China-focused blogs for western audiences” seeking to understand China better. He also endorses Dan Wang’s book Breakneck: China’s Quest to Engineer the Future as a “great book.”

An American Engineer in China
Elad Gil
  • “OpenAI Jalapeno: Better Than Nvidia” — SemiAnalysis newsletter article shared by Elad Gil. He highlighted OpenAI’s “impressive speed to tape out” and linked to the article. Read the article
Impressive speed to tape out by OpenAI [https://newsletter.semianalysis.com/p/openai-jalapeno-better-than-nvidia](https://newsletter.semi…
Aaron Levie
  • Article — “Moats in the age of floods” by @nayakkayak: Aaron Levie recommends it as a strong account of applied-AI strategy at scale, highlighting its argument that the gap between increasingly capable models and enterprise workflows creates substantial opportunity for applied-AI companies. Levie says the opportunity requires domain context, change management, routing across models, integration with critical business systems, workflow-specific UX, and strong evaluations—not model intelligence alone. Read the article.
Good post on what the applied AI strategy looks like at scale. It’s clear that there’s a wide gap between the AI models and the underlyin… Moats in the age of floods [https://x.com/i/article/2089796903403745280](https://x.com/i/article/2089796903403745280)
martin_casado

Martin Casado recommended an @0xCodez-posted discussion on coding with agents as the “Single best discussion ever recorded on coding with agents,” linking to https://x.com/0xcodez/status/2091980766372639135. The linked 50-minute podcast/video features a SpaceXAI engineer describing a team of 10–20 agents that automates routine work, including a “Chief of Staff” agent that coordinates the others.

Single best discussion ever recorded on coding with agents. [https://x.com/0xcodez/status/2091980766372639135](https://x.com/0xcodez/stat… SpaceXAI engineer (ex-Cursor): "right now I'm running 10-20 GrokBot agents that automate 90% of my routine i have a Chief of Staff agent.…
Ryan Hoover

Ryan Hoover shared Paul Graham’s essay How Universities Should Prepare Founders (https://paulgraham.com/prepare.html), highlighting that he has “never been a fan of whiteboarding a startup.”

How Universities Should Prepare Founders: [https://paulgraham.com/prepare.html](https://paulgraham.com/prepare.html) I've never been a fan of whiteboarding a startup From [@paulg](https://x.com/paulg)'s latest essay ![](https://pbs.twimg.com/media/HQk09P…