We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
The most useful diagnosis: agents are workflows, not chatbots
“Some obvious & non-obvious reasons I think AI agents may not have really been widely adopted yet, even though the tech is ready” — X thread · Creator: @BadCapitalVC · Recommended by: Aaron Levie.
Levie calls the post “a great place to start” for understanding real-world agent adoption. Its central distinction is that an agent is a process you set up and steer, not a chatbot that simply returns an answer; prompting is therefore closer to writing a specification, including a definition of “done.”
The thread extends that diagnosis into a practical checklist: delegation, technical setup, agent-to-agent and agent-to-colleague handoffs, asymmetric trust when an agent can send or edit something, and the absence of a public job-to-be-done. Levie adds that the payoff requires new data flows, cross-organizational context, and redesigned human review—not merely putting an agent on top of an unchanged workflow.
Why it matters: This is the clearest adoption filter in the set. When assessing an agent product, ask whether it defines the work, supplies the context, manages handoffs, and makes review safe; a better prompt alone is not the transition Levie is describing.
A benchmark for long-horizon agency
Factorio Learning Environment (FLE) v0.3.0 — evaluation environment / benchmark · Creator: Factorio Learning Environment project · Recommended by: Tobi Lütke.
Tobi’s endorsement is deliberately provocative: “Let’s all agree that this is the correct and final eval for agi.” The release describes FLE 0.3.0 as an environment for testing agents on long-term planning, reasoning, and world modelling. Its lab-play benchmark gives an agent fixed resources and a production target, has it write Python against the FLE API, and limits the run to 64 steps; targets are 16 units per minute for solids and 250 for fluids.
The useful difficulty is operational rather than exam-like. FLE says its model ranking is closer to an economically oriented benchmark than to static exam benchmarks, while also warning that many successful agents still shuttle resources manually or use chests and belts as buffers. A 60-second holdout period is used to reduce that shortcut.
Why it matters: FLE gives Tobi’s claim a concrete basis: it tests whether an agent can maintain a model of a changing environment and construct robust multi-step automation, not just produce a plausible answer. Treat “final eval” as Lütke’s standard, not settled consensus; the release itself presents a constrained lab-play setting with known measurement limitations.
A serious reading path into agency: Spinoza
Three biographies of Spinoza, by Nadler, Goldstein, and Stewart — book set · Recommended by: Garry Tan. (The talk names the authors but does not supply full titles or links.)
Tan calls these three biographies “the best biographies about the man.” He says an agent read roughly 1,500 pages and produced a chronology, a map of disagreements among the biographers, and cited quotations that he used to build his talk. He also tells prospective founders that they could do well to learn from Spinoza, whose concept of joy he summarizes as an increase in one’s power of acting.
Why it matters: This is a substantive humanities recommendation with a direct founder lens: the biographies provide a way into a thinker Tan uses to ask whether a tool, decision, or practice increases a person’s capacity to act. The recommendation is also unusually specific about how to read a difficult corpus—compare biographies, preserve disagreements, and retain the quotations that survive the comparison.
Priority and attention
“First and Second Things” — essay · Creator: C.S. Lewis · Recommended by: David Perell.
Perell points to Lewis’s essay for a paradox: putting the highest thing first can bring the second thing with it, while putting the second thing first can cost both. He connects that idea to the warning that results are rarely achieved by obsessing over results themselves, and that a less outcome-dependent identity can create more freedom to perform.
Why it matters: It is a compact decision rule for ambition: keep the activity or value that makes the outcome possible ahead of the outcome as an identity test. (Perell’s post did not provide a direct link to the essay; the link above is a reading pointer.)
Tim Ferriss offers a related, more personal set of cues. He says seven quotes shaped his thinking and changed his behavior over the previous year, and that he revisits them often. Two of the book-derived recommendations are:
- Notes of a Native Son — book · Creator: James Baldwin. Ferriss selects Baldwin’s observation that people may cling to hatred because abandoning it would force them to face pain. Why it matters: It is a prompt to examine the emotional function of a story or grievance before treating it as a settled belief.
- The Art of Possibility — book · Creators: Rosamund and Benjamin Zander. Ferriss highlights the instruction: “I am here today to cross the swamp, not to fight all the alligators.” Why it matters: It directs attention toward completing the central passage rather than spending the day responding to every obstacle.
FLE (Factorio Learning Environment) v0.3.0 is an open-source evaluation environment for testing AI agents in long-term planning, reasoning and world modelling inside Factorio . Its formal benchmark in this release is lab-play: a constrained setting where an agent is given fixed resources and a single target entity to maximize production throughput, writes Python programs against the FLE API, and observes stdout/error output from each program . It replicates the original FLE paper's methodology to evaluate the strongest models as of September 2025 . The standardized harness appends interactions to a single conversation history and summarizes old history near token exhaustion; it deliberately does not use backtracking/reflection agent designs .
What it measures
- Objective: production throughput targets of 16/min for solid items and 250/min for fluids, within 64 steps with early stopping; results reported as Pass@8 at Throughput Level 1 .
-
Verification uses task-specific success/failure flags (
task_verification) and production-flow statistics (flows) . To counter buffering, the agent must leave the factory untouched for 60 seconds before quotas are tested . - Caveat: agents can store items in chests/belts and pass checks without sustained production, so throughput is hard to measure consistently; the authors note higher targets would make manual logistics infeasible and force proper automation .
Reported results and limitations
- Ranking across frontier models is Claude > GPT > Gemini > Grok, a pattern the authors say is closest to GDPVal and opposite to static exam benchmarks (Humanity's Last Exam, AIME 25, GPQA, MMMU) where weaker FLE models do better .
- Mean error rates: Claude Opus 4.1 22.99%, GPT-5 25.05%, Gemini 2.5 Pro 27.29%, Grok 4 40.89% . Claude had zero syntactic errors and almost all pragmatic errors (97.7%); GPT-5 and Grok 4 had high syntactic-error rates (21% and 17%); Gemini 2.5 Pro, GPT-5 and Grok 4 showed API-misunderstanding rates of 12–17% .
- Error-recovery analysis: Grok 4 often falls into degenerate debug loops, while GPT-5 recovers gracefully; error accumulation/recovery remains hard across models .
- Open-source models reached the v0.2.0 SoTA on electronic circuits, steel plate, sulfur and plastic, and for the first time models succeed on harder tasks with over a dozen ingredient dependencies .
- Limitations: many successes are semi-manual (manual shuttling, chests as buffers) rather than full automation; agents rarely use the helper-function namespace (only Gemini 2.5 Pro does); misplacement of entities cascades into unrecoverable failures .
Operational details useful for a recommendation
-
Install with
uv add factorio-learning-environment; start a server cluster withfle cluster start; run an eval withfle eval --config configs/gym_run_config.json. - Release features: no dependency on the Factorio game client (headless pixel renderer for multimodal research), OpenAI Gym-compatible interface, open-source evaluation code with Weights & Biases logging, sweep resuming, analysis tools .
-
Agents receive a structured Observation object:
raw_text,entities,inventory,research,game_info,flows,messages,task_info,task_verification,serialized_functions, and optionalmap_image. - Claude Code integration via MCP adapters is included (livestreamed), demonstrating external-agent use .
- The page explicitly states frontier models are still 'shockingly bad' at Factorio and sees the environment as resisting saturation; planned work includes human baselines, METR-style task scaling, open-play/megabase scaling, latency-constrained real-time evaluation, multi-agent coordination, mod-based out-of-distribution tests, a native keyboard/mouse/vision interface and adversarial dynamics .
Identity. The target article is From Open Source Software to Open Source Strategy on the P3 Institute Substack (URL in bundle metadata). It carries no explicit byline. Internal first-person evidence points to Bill Gurley: “my own optimistic 1999 expectations,” “Benchmark, the firm where I worked as a General Partner for over 25 years,” and “my 2011 post, The Freight Train That is Android” . Treat the byline as inferred, not confirmed by the bundle.
Core argument. Open source has evolved from a development methodology into “Open Source Strategy”: a deliberate corporate-strategy tool used by a small number of companies to neutralize a stronger competitor, commoditize an expensive input, align an industry around a shared standard, or head off a regulatory crisis . The article stresses that these plays are mostly defensive, with offensive payoff often arriving later as a byproduct . Two premises drive the mechanism: open source always produces superior, more secure code than a closed alternative, and when an industry needs to coordinate on a single non-proprietary architecture, open source is the most powerful tool available . The underlying technical rationale is Linus’s Law — “given enough eyeballs, all bugs are shallow” — plus the classic strengths of leveraged development, better testing and bug discovery, more innovation, viral distribution, and customer cost-savings .
Framework worth retaining. The article frames four enabling developments: the rise of neutral foundations as referees, CIOs going “open source first,” AWS benefiting from a commoditized stack, and China embracing open source as national strategy . The core playbook is six canonical cases: Android (2007), Open Compute Project (2011), Kubernetes (2014), LF Networking (header 2017; launched January 2018), RISC-V (2010 roots), and Overture Maps Foundation (2022) . The cleanest codification is Overture’s “Open Source Strategy fingerprints”: a neutral referee under the Linux Foundation/Joint Development Foundation, a common adversary, a repeatable playbook, and a widening tent — with the referee contrast explaining why Google could recapture Android while members cannot recapture OCP-style projects . The economic case is also supported by major acquisitions, including IBM’s $34B Red Hat purchase and IBM’s later HashiCorp and Confluent deals .
Live evidence to include. Two unresolved cases carry the argument: autonomous vehicles and AI . For AVs, the author frames a fifty-plus-player prisoner’s dilemma — Choice A (outspend Waymo/Tesla) versus Choice B (support an open standard) — citing Waymo’s more than $45 billion in cumulative investment and $126 billion valuation, Cruise’s shutdown after more than $10 billion spent, and China’s Apollo history and structural advantages as reasons a Chinese-led open consortium should be expected . For AI, the key distinction is open weights; the claimed benefits are no lock-in, real academic participation, and little-tech access . The state of play: China currently leads the open-weight frontier (DeepSeek, Qwen, Kimi, GLM), OpenAI and Anthropic lead the closed frontier, Meta has pulled back from open frontier releases, and there is no credible Western open-frontier player; the author warns regulation could close the market rather than increase competition .
Uncertainties. No explicit publication date is in the bundle, but the text cites dated events through April 2026, so it was evidently written no earlier than that . Author identity is strongly indicated but not stated in the bundle, and the AV/AI predictions are forward-looking claims that should be re-checked before presentation as established fact.
Elad Gil cited a piece by Ben Horowitz proposing companies hold a pre-planned, once-a-year board meeting to discuss, without emotion, whether to consider exiting in the next six months — scheduled so neither founders nor investors are pushing it .
The episode also flagged YouTube videos of Jansen from Jansen Pharmaceuticals, called one of the best drug developers ever, interviewed 30–40 years ago about regulatory capture in pharma: drugs became expensive and slow due to regulatory capture and a risk-reward setup where the FDA focuses too much on safety/risk and not enough on benefit .
- Garry Tan recommends three biographies of Spinoza — by Nadler, Goldstein, and Stewart — calling them "the best biographies about the man"; his AI agent read the about 1,500 pages and synthesized a chronology, the points where the biographers disagree, and cited quotes, which became the opening story of his talk.
- He tells founders to learn from Spinoza ("If you're going to start a startup, you could do well to learn from Spinosa"), citing Spinoza's Ethics for its definitions of joy and sadness as the feeling of your power of acting increasing or decreasing, and its closing line: "All things excellent are as difficult as they are rare."
Box CEO Aaron Levie (@levie) recommends an X post by @BadCapitalVC on why AI agents haven't been widely adopted even though the tech is ready , calling it "a great place to start" to understand the dynamic of real-world agent adoption , and links the post: https://x.com/badcapitalvc/status/2085262600749945153.
His key takeaway: working with an agent is "much more like managing someone in a process vs. just asking an ai some questions and getting a response back" . He highlights the post's line that "prompting an agent is closer to writing a spec than asking a question. you have to scope the task extensively and define what "done" looks like" , and argues the real upside comes only when you change the underlying workflow — getting agents the right data, crossing organizational boundaries, and evolving the human-in-the-loop review steps . He predicts the vast majority of enterprise token usage will be agents "deployed" to go execute tasks inside workflows .
The recommended post argues adoption lags because agents are processes to set up and steer rather than prompt-in/answer-out; prompting is closer to writing a spec; the work requires delegation skills akin to managing an employee; much power still lives in terminal-shaped tools like Codex or Claude Code; agent-to-agent and agent-to-colleague handoffs are manual; trust is asymmetric (a wrong agent can send the email or edit the file, while a wrong chatbot wastes 10 seconds); and there is no job-to-be-done the public feels yet .
Paul Graham shared and endorsed the X thread 'Who Opposed Nuking Japan?' by @scotthortonshow , which compiles quotes from senior US leaders — including Eisenhower, Hoover, MacArthur, Nimitz, LeMay, Leahy, and Halsey — saying Japan was already ready to surrender and the atomic bombings were unnecessary or immoral . Graham said he was 'very surprised' by the quotes, noting that Eisenhower and Halsey were both against the bomb, and guessed there was heavy pressure toward public unanimity after the bombings . Link: https://x.com/scotthortonshow/status/2085415767474843744.
Chamath Palihapitiya shared Alfred Lin's X thread analyzing Empire State Building vs World Trade Center construction as a lesson source for startups — his quote post reads "I remember it well." and links to the thread . The thread argues the ESB (1930, 410 days, under budget, ~1,250 ft) beat the WTC (1966 groundbreaking, ~7 years, ~1,370 ft) by subordinating every design decision to speed: known steel/floor/window systems, minimizing on-site cutting/hand-fitting, designing trades to work in parallel, overlapping design with construction, co-locating owner/architect/engineer/contractor, and removing material/people friction with on-site railways, hoists, and on-floor cafeterias . It also notes the ESB was built to 85 stories for prestige despite the profit-maximizing design of 63, with speed discipline saving "vanity height economics" , while the WTC was plagued by lawsuits, political fights, and novel systems . The author distills "lots of learnings for startups, both inside the building (focus, parallelizing vs serializing, etc.) and outside the building (regulatory, etc.)" . The thread is at https://x.com/alfred_lin/status/2085007234665431359.
Bill Gurley (@bgurley) directed readers to a P3 Institute Substack article (https://p3institute.substack.com/p/from-open-source-software-to-open) as the detailed case for his argument that it is not "game-over" for Google in AI: Google's only remaining play is to follow its own Android/Kubernetes playbook and fully embrace open models . Vinod Khosla (@vkhosla) amplified and endorsed Gurley's post, adding that Google DeepMind still has one of the deepest AI talent benches and will remain a player whether models are closed or open .
Keith Rabois quote-posted 'Not wrong' in endorsement of an X post by Dean W Ball (@deanwball): https://x.com/deanwball/status/2085239725061972327. Ball argues that no macro-technology birthed by the Bay Area tech industry has drawn more hatred from Bay Area tech elites than frontier AI: many elites dismissed the labs' 'AGI' hopes in 2020/21, concluded crypto would be the technology of the decade, and now — with their businesses and worldviews under threat — openly hope the US frontier labs are commoditized or immolated by America's principal geostrategic rival. Ball allows commodification could render the labs insolvent, but sees something 'Randian' in the elites' disdain for success and achievement .
Tim Ferriss (@tferriss) shared seven quotes that he says shaped his thinking and changed his behavior over the past year, which he revisits often :
- Arthur Koestler: “Language can become a screen which stands between the thinker and reality. This is the reason why true creativity often starts where language ends.”
- James Baldwin, Notes of a Native Son: “I imagine one of the reasons people cling to their hates so stubbornly is because they sense, once hate is gone, they will be forced to deal with pain.”
- Chuck Palahniuk: “People fall so in love with their pain, they can’t leave it behind. The same as the stories they tell. We trap ourselves.”
- Robert M. Pirsig: “The truth knocks on the door and you say, ‘Go away, I’m looking for the truth,’ and so it goes away. Puzzling.”
- Yousuf Karsh: “Character, like a photograph, develops in darkness.”
- Winston Churchill: “A man is about as big as the things that make him angry.”
- Rosamund and Benjamin Zander, The Art of Possibility: “I will have to remember ‘I am here today to cross the swamp, not to fight all the alligators.’”
Elon Musk endorsed a video post by @teslaownersSV featuring his quote: "Don't pursue money. Make useful products, and money will come as a consequence" , replying "Yes" . The post calls it a "simple philosophy" behind some of the world's most ambitious companies and links the clip: https://video.twimg.com/amplify_video/2085250299393167360/vid/avc1/1080x1080/xtl72xiCBbangoa-.mp4?tag=29.
David Perell recommended C.S. Lewis's essay "First and Second Things," saying Lewis argued you're more likely to get what you want when getting it isn't the most important thing in your life — if you put the highest thing first, you'll often get the second thing as well, but put the second thing first and you'll lose both . Perell offered this in response to a Rory McIlroy quote about the paradox of chasing success — that obsessing over winning and results is "playing the wrong game" and that outcomes are rarely achieved by focusing on outcomes themselves .
Elon Musk endorsed a video by @Rothmus on the Berlin Wall's history, drawing the lesson that walls are built to keep people in: "Easy to tell who the bad guys are by looking at who has to build a wall to keep people from escaping!" . The video, attached to Rothmus's post, recounts how East Germany closed the border in 1961 after ~2.5 million fled, gave shoot-to-kill orders, and that escapes only ever went one direction . Video link: https://video.twimg.com/amplify_video/2085394736718782464/vid/avc1/1920x1080/tUri0Qe4d45WxrdI.mp4?tag=29
Tobi Lütke endorsed the Factorio Learning Environment (v0.3.0) at https://jackhopkins.github.io/factorio-learning-environment/versions/0.3.0.html as "the correct and final eval for agi" .
On the August 2026 episode of 20VC with Nikesh Arora, no genuine content recommendations were made. The episode discussed Airtable's acquisition by Bending Spoons for $1.285 billion , AI cybersecurity (average patch time 55 days, detection 4 days) , and the "Situational Awareness" memo by Leo Ashenbrenner, but only praised its investment thesis, not as a learning resource . No books, articles, podcasts, or papers were recommended.
Marc Benioff shared a Fast Company article about the next ad market being built for machines, noting that AI agents have overtaken humans in reading TIME Magazine ads: "Spoiler: the bots just took the lead" . Link: https://www.fastcompany.com/91584067/next-ad-market-built-machines.