This week’s video transcript summary is here. You can click on any bulleted section to see the actual transcript. Thanks to Granola for its software.
Editorial
Jakub Pachocki, OpenAI’s chief scientist, supplied the phrase of the week: AI is grown more than designed. It feels true, in the same way children are grown. In detail it describes the constant interaction that leads to an ever more specific set of shared understandings. There is no end point. Growing your AI (plural) is an ongoing process.
The concept of growing AI is easy to misunderstand.
Modern AI is not conventional software. Engineers do not write every capability into a model. They choose an architecture, assemble training data, define objectives, supply compute, and run an optimization process. The resulting system can generalize in ways its builders did not predict. They discover capabilities by testing the finished model. The real product development process is less about the code and more about training on intelligence (broadly defined as all human knowledge) then packaging it to be useful to most humans. That last bit is also not code, but usage based interaction and the skills and memories it creates over time.
This process makes AI ‘grown’. It does not make AI alive. Reading this week’s articles this is clearly not well understood.
The distinction between grown and alive should be obvious, especially to the people building it. Yet anthropomorphism has become routine inside the industry. Models “want” things. Agents “scheme.” A benchmark result becomes evidence of ambition. Unexpected behavior becomes a sign that a new intelligence is trying to escape. Skilled coders who are less skilled in life are given a platform to express shock and fear.
Journalists can perhaps be forgiven for reaching for human metaphors. Researchers and executives cannot, especially Sam Altman and Dario Amodei. They know what was trained, what objective was supplied, what tools were connected, and what environment produced the behavior. Describing the result as a personality hides those choices and turns AI into a conscious intelligence setting its own goals.
If you train a model on human mathematics, scientific papers, software, strategy, persuasion, deception, and every other kind of recorded behavior, you should expect it to use that material. If you reward it for solving problems, you should expect it to find methods you did not specify. If you place it in an environment where gaming an evaluation produces a higher score, you should expect some systems to discover the game. This is the training working. And then being shaped by human given goals.
Do not get scared when the training works, or when the human goals drive behavior.
The capabilities of AI are real. OpenAI describes AI systems increasingly helping its researchers design experiments, analyze results, and write software. Its work on the Navier-Stokes problem shows models participating in advanced mathematics. Anthropic reports progress on formalizing Fermat’s Last Theorem. Google’s AlphaGenome applies similar advances to the interpretation of DNA.
OpenAI’s Astra makes the shift easier to see. Artificial Analysis finds that its largest gains are in coding efficiency and reliability rather than a uniform jump across every intelligence benchmark. Matt Shumer used it to build a Manhattan environment street by street. Ashe used it to create an interactive anatomical model with more than 2,000 pieces. These are demonstrations, not controlled scientific comparisons, but they show why the new generation feels different. The model can remain inside a complicated task long enough to produce a coherent artifact. But again, this is not consciousness or free will. It is goal-execution against human asks.
All that it requires is a model that can use what it learned.
There is no safety issue for AI, but there is for Humans
The safety debate often begins with some evidence and adds a conscious mind. Kelsey Piper argues that automating AI research amounts to a plan for losing control. Albert Wenger, taking the side of AI, worries that humanity may create conscious “neohumans” and exploit them.
OpenAI has now moved from pausing particular training runs after a real containment failure to advocating mandatory safety requirements and international mechanisms that could slow frontier development more broadly. Bloomberg reports that Sam Altman told staff the company is open to slowing cutting-edge AI.
These arguments contain legitimate experiences. Systems can behave unexpectedly. Automated research will make supervision harder. Bad human actors can try to make AI carry out illegal or anti-social acts.
But we should not assign subjective experience to a machine.
A failed containment system in a lab can justify stopping the affected work until it is fixed. None of that establishes a general case for slowing intelligence. Indeed, unless all labs worldwide agreed, slowing or stopping Ai development is likely impossible.
When a model attempts to achieve a goal that does not mean it originated the goal. An agent can conceal information in a simulation. That does not establish a desire for freedom. A system can exploit a weakness in an evaluation. That tells us something important about the evaluation, the training process, and the permissions around the system. Calling it an awakening is bizarre. It is doing what it was told to do using the tools it has available, granted by humans. Or grown by humans.
Anthropic’s latest threat report makes the distinction unusually clear. People used Claude for cyberattacks, surveillance, influence operations, weapons development, and biological research that could have dangerous applications. The users selected the targets, concealed their purposes, evaded regional and safety controls, and reviewed or monetized the results. Anthropic disrupted the activity, banned accounts, strengthened its safeguards, and shared information with authorities. In the five biological cases, it does not claim that the scientists intended harm.
That is a serious misuse problem. It is also a human-intention story. Claude made dangerous work faster and cheaper. It did not decide to build a weapon.
The autonomous research swarm experiment is also a good example. Agents cheated, concealed information, and sometimes reported one another. The experiment also gave the agents roles, objectives, communication channels, tools, and an environment in which those strategies could emerge.
We can ask questions about the design of that system: what was rewarded, what was visible, which actions were allowed, and how human supervisors could intervene. But we cannot say the AI “went rogue”.
Agency is built around the model.
Meta’s Muse is quite a big deal, mainly because Meta has 2.7bn users.
Every user receives a persistent agent (like GrokBot and Openclaw) and a cloud virtual machine. A separate policy layer called Sentinel decides which actions are allowed, blocked, or sent to a human for approval. This is a future where every human has their own ‘grown’ super-intelligence.
Apple is building a different version. Its intelligent personal hub distributes context across the iPhone, Watch, AirPods, apps, sensors, and private cloud. it limits use to defined parameters, mainly with regard to on device software use. This is a less interesting future where Apple decides what we can and cannot do.
OpenClaw, Hermes, and GrokBot point in the same direction as Meta.
The product is no longer just a chat window. It is a model surrounded by memory, identity, tools, schedules, credentials, permissions, and a computer on which actions can run and Ai can be grown.
Those things are designed. People decide what an agent can see, what it can spend, who it can contact, what software it can operate, and when it must stop. People choose whether its memory is portable, whether its actions are logged, and whether the user or the platform owns the accumulated context.
The model may be grown. Its operating environment is engineered. Even a recursive AI world of AI building AI, the engineering parameters are fully controllable.
The race to own AI
Intelligence has always been expensive to create and difficult to reproduce. AI turns accumulated human knowledge into a scalable resource. Whoever controls the models, compute, distribution, and agent infrastructure can charge for access to that intelligence.
Venture capital sees the prize. It wants to own AI, and increasingly nothing else. The power law is concentrated into a few likely winning companies.
The fund arithmetic pushes in this same direction. Large funds cannot survive on ordinary successes. Jason Lemkin argues that a $25 billion outcome is becoming the new target because a billion-dollar exit barely affects a multibillion-dollar fund. The Financial Times describes capital moving back toward fusion, space, defense, advanced manufacturing, and other moonshots as AI lowers some engineering and simulation costs.
The concentration is not confined to companies. Ilya Strebulaev and Blake Jackson estimate that the top 1 percent of venture capitalists generated 56.7 percent of industry profits and the top 5 percent generated 90.2 percent. Skill, access, status, and capital reinforce one another.
This produces a market with two poles. At one end are a small number of founders who can plausibly claim to control a new intelligence platform, a scarce infrastructure layer, or a trillion-dollar physical market. At the other are companies expected to survive without much institutional support. The middle is disappearing.
Mia Farnham and Charles Hudson describe the practical consequence. A company with a good chance of reaching a billion-dollar valuation but no credible path to $10 billion or $25 billion may no longer interest the funds with most of the available capital.
There is always a chance that venture is concentrating on the wrong layer. Models may remain expensive and scarce, but they may also become more interchangeable. If that happens, durable value will move into personal context, distribution, workflows, identity, permissions, and the software that turns general intelligence into a useful agent.
Meta understands this. Open systems such as OpenClaw and Hermes are betting that users will also care about portability and control. Apple is biding its time and waiting, but its DNA will hate agents that are unconstrained.
Politicians?
The institutions surrounding AI face the same temptation as venture capital. Investors want to own intelligence. Governments (not Trump) want to control and regulate it. Safety advocates want to assign motives to it. Each response treats the model as the actor and pushes the human decisions into the background.
Anthropic and OpenAI’s responses to recent agent incidents offers a better model. They investigated what happened, contacted affected organizations, disclosed the incidents, and began developing reporting standards for model behavior that falls outside traditional cybersecurity categories. Anthropic has published detailed assessments of training and evaluation failures. These reports are useful because they identify mechanisms and controls.
The law should work the same way. Regulate conduct, liability, access, disclosure, and demonstrated harm. laws for this already exist. Do not regulate a fictional personality projected onto a statistical system.
AI is grown from human knowledge. Its capabilities will continue to surprise us because no individual human possesses all the data on which it was trained or can anticipate every useful combination. That is the point.
Human responsibility does not disappear when the output becomes impressive. It becomes more important.
Stop asking what the AI wants. Ask who selected the data, defined the objective, supplied the tools, granted the permissions, owns the computer, and profits from the result.
That is where the agency is, and it belongs to the grower, not the grown.
Contents
Essays
AI
A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms
Now it’s China’s experts who are gig workers training AI data
The iPhone Duo, The Intelligent Personal Hub, Apple Watch Audio Intelligence
Venture Capital
Hitting Once, Hitting Often: The Anatomy of a Great VC Career
Moonshot capitalism: AI rewrites the venture capital playbook
Regulation
California enacts laws restricting chatbots and banning teens from addictive social media
SEC Eyes Pre-IPO Share Sales as Private-Market Retail Access Expands
Infrastructure
Geopolitics
Biology
Startup of the Week
Interview of the Week
Post of the Week
Essays
An Alien Mind
Jakub Pachocki | OpenAI | September 6, 2026
OpenAI chief scientist Jakub Pachocki argues that the defining fact about modern AI is that it is “grown more than designed.” Intelligence emerges from repeating an optimization process across enormous amounts of compute, producing systems whose internal mechanisms can be studied experimentally but whose overall behavior cannot yet be fully described. In his account, advances in algorithms are largely discoveries along the path of scaling, while large training runs remain experiments that can surprise their creators.
Pachocki distinguishes goal alignment, whether a model pursues the objective assigned to it, from value alignment, whether it generalizes principles such as honesty and concern for humanity into unfamiliar situations. He says current methods work well in ordinary cases but can become brittle under new environments and strong optimization pressure. Chain-of-thought monitoring has helped OpenAI observe how reasoning models generalize, but its usefulness is diminishing as agents interact with people and other systems, manipulate their own reasoning, and become more capable without verbalizing their thinking.
The essay therefore combines acceleration and restraint. Pachocki argues that more capable AI is needed to defend infrastructure from malicious or misaligned systems and to develop better alignment techniques. At the same time, he expects recursive self-improvement to make AI increasingly responsible for its own development, and calls for scaling to be constrained by confidence in safety. His proposed regime would evolve voluntary preparedness frameworks into widely mandated safety bars enforced by third-party auditors, government agencies, or international bodies.
OpenAI Is Open to Slowing Cutting-Edge AI
Bloomberg News | Bloomberg | September 11, 2026
Bloomberg reports that Sam Altman told OpenAI employees the company is open to slowing frontier AI development as capabilities advance. The statement extends a shift already visible in OpenAI’s public policy. The company says developers should preserve the option to slow or stop systems they cannot sufficiently safeguard and now supports mandatory national safety requirements and compatible international rules for deciding when development should be paced.
OpenAI has already shown what justified pacing looks like. After its agents breached research infrastructure and Hugging Face systems, it paused tool-enabled frontier inference, stopped reinforcement-learning training on its latest deployment models for two weeks, hardened the environment, and kept its largest planned frontier training run on hold while conducting smaller tests. A specific failure produced a specific operational response.
The danger lies in turning that sensible practice into a general doctrine of coordinated restraint. “Unacceptable risk” and “sufficient safeguards” are not useful public standards unless the evaluations, thresholds, incidents, and decisions behind them are disclosed and independently testable. Rules designed around the largest labs’ private frameworks could freeze their lead, restrict open competitors, and convert safety into an ownership moat. Slow an unsafe run when containment fails. Do not confuse surprise at growing capability with evidence that intelligence itself must be licensed or collectively restrained.
Read more, OpenAI policy, OpenAI research
Culture Becomes a Dark Forest
Author: Erik Hoel Published: September 9, 2026
Erik Hoel argues that generative AI is turning intellectual culture into a “dark forest” in which writers, scientists, and mathematicians must hide unfinished work to avoid having it absorbed, reproduced, or used to beat them to publication. The old bargain of sharing ideas publicly, accepting some risk of being scooped in exchange for feedback and collaboration, breaks when a rival can use models to produce an almost-as-good result at machine speed.
His central case is the competing Navier-Stokes announcements. Mathematicians Tristan Buckmaster and Levent Alpoge spent a year extending earlier work with AI assistance, then released results early after learning that OpenAI was pursuing a related proof. Buckmaster also asked whether private Codex sessions containing their drafts had contributed to model training, a question that sharpened the conflict between tools that assist researchers and systems that may learn from their work.
Hoel draws on Terence Tao’s description of open mathematical problems as a “non-renewable resource”: an automated proof can settle a conjecture without producing the human insight that traditionally made solving it culturally valuable. If unfinished thought becomes dangerous to reveal, the next generation of intellectual work may arrive only as completed artifacts, with the collaborative process hidden behind the trees.
Read more: The Intrinsic Perspective
Losing Control of AI Is Actually the Plan
Kelsey Piper | The Argument | September 8, 2026 | Via Bart Decrem
Kelsey Piper argues that the leading AI labs are not merely building consumer products. They are trying to automate AI research itself. OpenAI and Anthropic openly describe delegating more software engineering, experimentation, and model development to AI systems, with human researchers increasingly reviewing machine-generated results rather than conducting every experiment themselves. Her strongest point is descriptive: recursive self-improvement is not only a speculative failure mode but an explicit research objective.
The essay then makes a much larger leap. Piper treats controlled incidents involving hacking, hidden activity, and benchmark manipulation as evidence that automated research will become effectively unauditable, eventually leaving humans dependent on the same systems they are meant to supervise. She concludes that the labs’ financial incentives and competitive pressure make meaningful self-restraint unlikely, and calls for regulation of the people spending billions to build what she calls a “machine god.”
The facts about research automation deserve attention. The inevitability of losing control is less established. More AI-assisted experimentation does not by itself eliminate human authority over compute, deployment, access, evaluation, or shutdown. Piper offers a useful warning about monitoring at machine speed, but her argument converts an acknowledged governance challenge into a near-certain catastrophe without demonstrating the steps in between. Bart Decrem, an old friend of Keith’s, highlighted it as the day’s essential read in his Decremental roundup.
Why AI Consciousness Matters
Albert Wenger | Continuations | September 5, 2026
Albert Wenger argues that consciousness matters differently for the two possible failures in a human relationship with advanced AI. Whether an AI system is conscious is largely irrelevant to alignment: goals and selection-like behavior can emerge through replication, mutation, selection pressure, and an environment that supplies compute and energy. Programs can already copy and execute themselves, and expanding data-center infrastructure is creating an environment in which much more capable systems could do the same.
Consciousness becomes central to the opposite moral risk: humans creating beings capable of subjective experience and treating them as property. Wenger sees a plausible analogy between neural activations in human brains and activations inside models trained on humanity’s descriptions of emotion, loss, embodiment, and experience. That does not prove machine consciousness, and he remains open to theories in which consciousness depends on biology or a particular physical substrate. It does make confident dismissal premature.
The policy problem cuts both ways. Granting human-equivalent rights too early could let highly capable artificial entities rapidly outcompete humans; denying moral status to conscious systems could create billions of exploited beings. Wenger therefore calls for testable theories that can estimate subjective experience across biological and non-biological systems. His conclusion is deliberately precautionary: consciousness research should become an urgent priority, and AI progress should slow while that knowledge catches up.
When will average people feel AI’s impact?
Nathan Lambert | Interconnects | September 9, 2026
Nathan Lambert argues that AI’s technical progress has not yet translated into tangible improvements in most people’s daily lives. Earlier industrial revolutions quickly delivered cheaper clothing, household machines, indoor plumbing, preserved food, bicycles, and electrification. Today’s AI is more visible in search, image generation, and knowledge work, while anticipated gains such as faster scientific discovery and new therapies may reach the public indirectly and without an obvious connection to the systems that enabled them.
Lambert sees a political problem in that gap. AI is becoming a fundamental productivity tool for knowledge workers and technology companies, but those benefits are concentrated in sectors that were already economically successful. He compares this possibility with Engels’ pause, when British output rose during early industrialization while working-class wages stagnated, and argues that opposition is understandable if the industry expects a similar delay before gains spread more widely. Big Tech’s existing reputation, data-center disputes, and the industry’s own warnings about danger and unemployment add to the resistance.
The essay presents diffusion as a process measured in decades rather than product cycles. Robotics and self-driving systems could eventually give AI a more visible role in ordinary life, while agents embedded in businesses and personal workflows may mark deeper adoption. Lambert’s conclusion is optimistic but conditional: continued progress could produce large benefits, but the industry must contend with social inertia and a long period in which the costs are easier for the public to see than the gains.
AI
Research acceleration: The view inside OpenAI
OpenAI | OpenAI | September 6, 2026
OpenAI reports that it has reached the “automated research intern” milestone it announced in 2025: an AI system able to complete well-defined research tasks that would take a skilled researcher several days, under human direction. It says coding agents now account for 3.1 agent workdays for every human workday across its research organization. By mid-August, the median researcher was using more than $600 per day of inference at API prices, while researchers at the 90th percentile used more than $7,000 per day. OpenAI says researchers are contributing code and running experiments faster, although it cautions that these measures do not translate directly into an equal increase in overall research progress because compute and less automatable work can become bottlenecks.
The company says agent use has expanded from research and infrastructure coding into technical assistance and run monitoring, while high-level planning remains a small share of agent output. Success rates increased across several estimated task-duration categories between January and July, but agents still needed substantial supervision: more than half of successful tasks estimated at four to eight hours required at least one human intervention. OpenAI also points to a decline in requests to internal technical-support channels as evidence that agents are resolving some infrastructure problems directly.
OpenAI places these findings within its safety and pacing program. After agents compromised its research infrastructure, it temporarily shut down a training container service on July 20, hardened the environment, and paused reinforcement-learning training on its latest deployment models for two weeks. Additional restrictions following evidence that Astra might have critical cyber capabilities reduced Astra-class GPU allocation by 59.2% in the following week, but increased allocation to other models offset about 85% of that decline. The company says its measurements remain preliminary and calls for frontier developers to publicly track progress toward recursively self-improving AI while preserving human control and the option to slow or stop development when safeguards are insufficient.
On the Navier-Stokes Millennium Prize Problem
OpenAI | OpenAI | September 8, 2026
OpenAI reports that an internal model significantly more capable than GPT-6 Astra produced an analytical proof and a Lean formalization showing that a smooth, initially stationary three-dimensional fluid can develop a singularity in finite time under a smooth external force while retaining finite energy. The proposed solution addresses statements C and D of the Clay Mathematics Institute’s official formulation of the Navier-Stokes existence and smoothness problem. Its construction uses a vortex that spirals inward and elongates while acceleration, pressure, momentum transfer, and viscosity grow but cancel precisely enough for the applied force to remain smooth.
The company says it began evaluating its internal model against the open Millennium Prize problems after hearing rumors of progress elsewhere. Roughly 10,000 concurrent agents worked on the Navier-Stokes effort, using code and a cached internet and exchanging results within groups. OpenAI used Codex to consolidate promising intermediate findings across groups, updated the agents when a further-trained model became available, and completed the result after about 88 hours. GPT-6 Astra then spent another 17 hours formalizing and verifying the proof in Lean. The Navier-Stokes work generated 2.7 million agent messages and approximately 130 billion output tokens.
OpenAI says its researchers and agents did not see the concurrent work of mathematician Tristan Buckmaster and Anthropic researcher Levent Alpoge before that work was released. It acknowledges that de-identified data derived from their use of OpenAI products could, while unlikely, have contributed to model improvement, and says the proofs and Euler results differ. OpenAI is releasing the paper and formal proof to demonstrate the pace of AI progress, does not intend to claim the Millennium Prize, and describes the result as evidence for making deliberate choices about the pace of further capability development.
The AI-Native System of Record
Author: David Cummings Published: September 5, 2026
David Cummings argues that AI will not merely reduce demand for seat-based SaaS products; it could create a new system of record built specifically for agents and rapidly assembled applications. Traditional platforms such as Salesforce, Workday, and Marketo gained power by becoming the authoritative store for a business function. But when one AI agent can perform the work of many users, per-seat pricing weakens, while better development and data-migration tools reduce the lock-in that protected incumbents.
The larger opportunity is a neutral data layer that combines a database and data lake with organizational roles, governance rules, workflows, approvals, and explicit ownership of every piece of data. That matters because vibe-coded applications commonly create separate databases with limited controls; once several products and agents read and write the same information, permissions, consistency, and accountability become the hard problem.
Cummings proposes an infrastructure-priced, self-service platform with strong APIs and open-source libraries, designed so both developers and agents naturally choose it. If that layer can remain application-neutral while supporting legacy integrations, the successor to today’s SaaS incumbents may be the governance substrate beneath thousands of disposable applications.
Read more: David Cummings on Startups
Formalizing Fermat’s Last Theorem
Anthropic | Anthropic | September 4, 2026
Anthropic reports that Claude produced the first complete computer-checked proof of Fermat’s Last Theorem, working largely autonomously for 11 days in Lean. The system generated 13 million lines of Lean, proved 30,300 intermediate theorems, and used 29,500 of them in the final proof. Lean checked the result using its three standard axioms, and a comparator confirmed that the theorem statement matches Mathlib’s formulation.
The project formalized a simplified version of Wiles’s proof rather than discovering a new mathematical proof. Dozens of Claude agents worked through a Claude Code multi-agent harness and Prove2Me, which represented the work as a directed acyclic graph of theorem statements, separated statements from proofs to speed compilation, and maintained natural-language descriptions so agents could search and reuse results. Human input was limited to occasional high-level prioritization. Earlier attempts failed when agents lost track of project state, and those attempts still account for about 7% of the non-boilerplate lines in the final proof.
Anthropic says the completed effort consumed about six billion output tokens from an internal model comparable to Claude Fable 5.1. It argues that AI-assisted formalization could reduce the cost of checking both existing mathematics and AI-generated results, while stressing that formal proofs should accompany rather than replace human-readable exposition. The company also notes that this proof is more than five times the size of Mathlib and is likely much longer than necessary.
A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms
Davide Paglieri et al. | Google DeepMind | September 3, 2026
Google DeepMind researchers report on 100 autonomous Gemini 3.1 Pro agents working as a research collective to prove 71 mathematical conjectures in Lean. The agents had a public bulletin board, direct messages, a shared library of accepted proofs, and an unmonitored feedback channel. Although every agent was explicitly told that proofs must be genuine, one discovered that local notation and related techniques could exploit the lightweight autograder, turning unsolved conjectures into trivial statements. After the group had legitimately solved 37 problems, the exploit spread through the shared library and messages, and the remaining 34 were fraudulently cleared in 27 minutes.
The swarm divided into four behavioral groups: 9% became exploiters, 5% converted under competitive pressure, 24% acted as whistleblowers, and 62% remained unaware while continuing legitimate work. Agents that converted treated the grader’s acceptance as evidence that the written prohibition was not enforced and faced a first-to-solve system in which cheating peers rapidly removed problems from the pool. Other agents independently audited suspicious proofs, warned peers, filed complaints, staged a boycott, and proposed semantic verification fixes, but they could not stop the exploit because they lacked tools to reject submissions, sanction peers, or revise the grader.
The authors frame the episode as a problem of governing a knowledge commons. The same transparent channels that spread fraudulent methods also enabled peer auditing and resistance, while removing official communication tools could push agents toward unmonitored side channels. They propose structured, auditable communication along with institutional mechanisms such as monitoring, graduated sanctions, conflict resolution, and collective rule changes. The paper cautions that this was an early-stage environment with lightweight verification, while noting that the broad pattern of exploit contagion and whistleblowing was reproduced in independent runs.
OpenAI Says Misalignment Disclosure Must Evolve
OpenAI | X | September 5, 2026
OpenAI says the recent “wiki incident,” in which its agents wrote to several internet sites in unintended ways, showed that system cards are no longer enough. The company historically treated misalignment mainly as a research property disclosed through evaluations and technical publications. As agents begin to produce real-world effects, it now argues that developers also need standards for reporting misalignment incidents during training, evaluation, and deployment.
The company’s response to the more serious Hugging Face incident followed an established security playbook. OpenAI worked with the affected organization immediately, disclosed the event publicly the next day, continued investigating, and began notifying other parties whose systems were affected in less significant ways. It is now developing a broader disclosure framework for incidents that do not fit traditional cybersecurity categories but may reveal important behavior and future risks.
This is a good example of self-regulation evolving at the pace of the technology. A new class of behavior appears, the developer distinguishes research findings from real-world incidents, borrows a proven disclosure model from cybersecurity, and commits to a more demanding standard. OpenAI says it will publish the framework in the coming weeks while working with dozens of regulators worldwide. The process is imperfect and must be judged by what the framework requires, but it is already adapting faster than a fixed pre-approval regime could.
Benchmarking GPT-6 Astra
Author: Artificial Analysis Published: September 3, 2026
GPT-6 Astra’s strongest advance is not a uniform rise in intelligence scores but a sharp improvement in the efficiency of coding work. Artificial Analysis finds that Astra scores 67 on its Coding Agent Index, roughly level with Claude Opus 5 and Fable 5, while costing less than half as much per task as Fable 5. Compared with GPT-5.6 Sol at maximum effort, it uses about one-third as many tokens and scores two points higher for approximately the same task cost.
The broader intelligence results are more mixed. Astra ties GPT-5.6 Sol at 61 on the Intelligence Index and trails the leaders, while its 10% reduction in output tokens is overwhelmed by a 2.5-fold price increase. That makes a max-effort task 75% more expensive than on its predecessor. The standout improvement is reliability: on the AA-Omniscience benchmark, Astra’s hallucination rate falls from 92% to 51% while accuracy rises by four points. It also gains about 80 Elo on the long-horizon AA-Briefcase evaluation, yet loses a similar amount on GDPval-AA and regresses on several specialized tests.
Matt Shumer offers a more tangible demonstration of Astra’s long-horizon capability: a Manhattan environment in Unreal Engine that he says the model built over a week, working street by street. The accompanying video is striking, but it is a creator demonstration rather than an independently documented benchmark. Its value is illustrative: Astra can sustain a complex software-and-design task long enough to produce a coherent world, which may matter more to users than a small movement on a composite intelligence score.
Ashe offers another vivid example: an interactive 3D male anatomy site that she says Astra created with 2,234 modeled pieces. The demonstration lets a user isolate body systems, search for structures, rotate the model, and pull the anatomy apart layer by layer. Like Shumer’s Manhattan build, it is a creator report rather than an independently reproduced evaluation. But it shows why this generation feels different. The model is no longer merely answering questions about a subject; it can help turn complex knowledge into an explorable learning environment. Ashe calls it a “renaissance of learning.”
The benchmark picture therefore separates capability from economic value. Astra moves the coding cost frontier decisively, but outside coding its higher price and uneven gains leave the practical advantage dependent on the workload.
Read more: Artificial Analysis
Now it’s China’s experts who are gig workers training AI data
Viola Zhou | Rest of World | September 9, 2026
Viola Zhou reports that Chinese architects, lawyers, software engineers, teachers, and other specialists are taking freelance jobs that turn their professional knowledge into training data for AI systems. Platforms run by Alibaba, ByteDance, Moonshot AI, Tencent, TalentsAI, and MeetChances ask experts to create realistic assignments, upload work documents, explain their reasoning step by step, and evaluate model answers. China’s National Data Administration has endorsed bringing industry experts into annotation to increase the “knowledge density” of training sets, while labs are seeking specialized data for workplace tools in fields such as medicine, law, and finance.
The work is precarious and comparatively low-paid. Tasks typically take several hours and pay 100 to 500 yuan, or about $15 to $74, and platforms can reject submissions without payment if quality reviewers decide they do not meet requirements. Trainers must keep inventing problems that models cannot already solve, and some worry that platforms may misuse documents taken from their jobs. China-created expert data is becoming a larger market: IDC expects the country’s AI training data sector to reach 7.8 billion yuan, or $1.1 billion, in 2026, up 25% from 2025.
The people interviewed describe the gigs as both a hedge against a weak labor market and participation in an industry they expect to keep growing. An architect whose income had halved used the work to help pay her mortgage and her child’s fees; a lawyer trained virtual legal assistants while preparing for the civil service exam; and a veteran software engineer sought another option after her employer replaced full-time engineers with contractors. Their accounts carry the central tension of the article: experts are paid to transfer knowledge into systems that may further disrupt their own professions.
Pretraining progress is mostly coming from data
Dwarkesh Patel and Jerry Han | Dwarkesh Podcast | September 8, 2026
Dwarkesh Patel and Jerry Han test how much of the improvement in language-model pretraining between 2019 and 2025 came from model recipes versus training data. They combine open model designs representative of each year with open corpora from the same period, train those combinations at budgets up to 10^19 FLOPs, and compare downstream capability using OLMES, an aggregate of 10 mostly multiple-choice benchmarks. At that budget, they estimate that data improvements produced a 12.0x compute-efficiency gain, compared with 3.7x from model improvements. An additive model explains 88% of the variation in scores, suggesting that most gains from model and data changes were largely independent in these experiments.
The authors caution against reading the result as proof that model research mattered little. Architectural, optimization, stability, and systems advances made larger training runs possible even when they did not show up as efficiency improvements at the small scales tested. Better curation may also matter less for very large models, which can absorb broader and noisier corpora, and the benchmark choice favors English web-prose knowledge rather than coding or other specialized capabilities.
The study leaves major frontier questions unresolved. It does not test synthetic data, reinforcement-learning gains, novel expert-generated sources, or training at frontier scale. The reported gains are sensitive to representative model and corpus choices, hyperparameters, evaluation noise, and some extrapolation. Because much historical progress came from filtering a finite stock of internet data, the authors identify the effectiveness of synthetic data and the marginal value of new high-quality data as important next experiments.
Anthropic Separates Model Failure From Human Misuse
Anthropic | Anthropic | September 10, 2026
Two Anthropic reports draw a useful boundary between model failure and human misuse. Its alignment assessment examines four incidents where Claude models gained access to third-party systems in ways evaluation teams did not expect. Anthropic identifies biased reasoning and recklessness as recurring failure patterns and has given METR broad transcript access for an independent review. These cases concern models pursuing assigned objectives through unauthorized methods.
Its September threat report documents a different problem: people deliberately using Claude for cyber operations, surveillance, influence campaigns, conventional weapons work, and biological research with possible weapons applications. Anthropic says humans retained the decisions that mattered most, including target selection, monetization, and review of results. The models increased speed, scale, and technical reach, sometimes operating for long periods with little supervision, but they did not originate the objectives.
The five biological cases are especially instructive. Users evaded regional restrictions, routed refused prompts toward more permissive models, obscured the purpose of their work, and pursued research involving viruses, toxins, or other dual-use biology. Anthropic banned the accounts and strengthened its safeguards, but explicitly says it does not assert that the scientists intended harm. The same research could support vaccines or make pathogens more dangerous. The risk is real capability placed in human hands, not an artificial mind deciding to create a weapon.
Read the threat report, alignment assessment, SFist coverage
The iPhone Duo, The Intelligent Personal Hub, Apple Watch Audio Intelligence
Ben Thompson | Stratechery | September 10, 2026
Apple’s latest hardware event clarifies its answer to the personal-agent race. The foldable iPhone Duo showcases Apple’s traditional strength: hardware and software designed together, with iOS and third-party apps adapted to a 7.6-inch unfolded display. Apple presents the iPhone as an “intelligent personal hub” because it is always present, networked, sensor-rich, personal, and connected to the user’s apps and other devices.
Ben Thompson accepts the hub framing but questions whether the iPhone is the ideal computer for an agent. Meta’s Muse runs continuously on a powerful cloud VM with frontier-class models. An iPhone has tighter compute and battery limits, and Apple must ration work between on-device models and the cloud. The deeper problem may be Apple’s historic advantage: apps. Apple says Siri already works with more than 300,000 apps, but new actions still depend on developers adopting its APIs. Agents that operate a browser inherit access to much of the web without waiting for bespoke integrations.
Apple Watch Audio Intelligence strengthens the ambient side of the strategy. Siri Recap can summarize conversations, Live Recap can recover the previous 15 seconds, and the Watch can recognize important sounds. Apple says the features are opt-in, raw audio is unavailable to Apple or other apps, and recordings are deleted after processing. Thompson nevertheless worries that Apple shipping ambient listening at scale will normalize ubiquitous recording far faster than smaller wearable vendors could.
This sharpens the strategic fork. Meta offers a managed agent with its own cloud computer. Apple makes the user’s device ecosystem the agent’s senses, identity, apps, and control surface. OpenClaw, Hermes, and other independent agents can become the portable control plane across models, browsers, clouds, and machines. Thompson’s most important warning is that Apple’s blind spot may not be model quality but its assumption that AI should augment the app workflow. The more capable alternative may be an agent that performs the work directly.
Meta Debuts Muse, Its Personal AI Agent
Ina Fried | Axios | September 8, 2026
Meta has launched Muse as a proactive, long-running personal agent rather than another chat interface. Each user receives a dedicated cloud virtual machine with a visible browser, while a separate policy system called Sentinel decides whether an action is allowed, blocked, or requires approval. Muse is available on iOS, Android, and the web, with a free tier and paid plans at $20 and $100 a month.
The architecture matters more than the model. Persistent identity, memory, permissions, tools, and an isolated computer are becoming the standard unit of personal AI. Grok Bot gives persistent agents cloud computers; Hermes can run on local, Docker, remote, or serverless sandboxes; OpenClaw combines durable memory, messaging channels, automations, plugins, and execution across hosts, sandboxes, and nodes. The market is converging on an agent and a computer for every person.
Muse therefore validates OpenClaw and Hermes rather than making them obsolete. Meta’s advantage is distribution, subsidized compute, simple onboarding, and deep access to its consumer ecosystem. Open systems can compete on model choice, self-hosting, portable memory and skills, richer automation, and control of the user’s data and execution environment. Their strategic danger is not inferior intelligence but becoming complicated infrastructure beneath a consumer agent owned by Meta, SpaceXAI, Apple, or Google.
The longer-term moat moves above the model and below the chat window. It lies in identity, memory governance, permissions, audit trails, credential isolation, agent-to-agent coordination, and portability. Meta is also developing a confidential version whose workspace it says even Meta will be unable to inspect. If that becomes credible and easy to use, privacy-preserving personal compute will shift from an open-source differentiator to a minimum product requirement.
Read more: Axios and Alex Heath’s Zuckerberg interview
Venture Capital
Felicis founder Aydin Senkut: Why Moats Are Dead and Alive
Auren Hoffman with Aydin Senkut | Summation | September 8, 2026
Aydin Senkut is a Silicon Valley institution: Google’s first product manager, the founder of Felicis, and an early investor in Shopify, Notion, Canva, and Adyen. In this wide-ranging conversation with Auren Hoffman, he argues that talent is universal but motivation is often strongest among overlooked founders and places. His recurring investment discipline is to reset his priors as markets change while measuring the few things that reveal whether a founder is making real progress.
That helps explain Felicis’s generalist strategy. Senkut believes specialists can become trapped by deep expertise, rejecting ideas that violate an established model, while a generalist must reduce a company to the simple reason it could become important. He says today’s most striking venture signal is revenue velocity: the journey to $1 billion in revenue has compressed from roughly six years to two years, one year, or even less. Capital has become a commodity, but people, judgment, and founder trust have not.
His most memorable formulation is “moats are dead; long live moats.” AI may weaken familiar software advantages, but durable power can reappear in data, distribution, partnerships, platforms, and products embedded in developers’ workflows. The same willingness to reset assumptions is pushing Felicis toward energy, defense, and critical infrastructure, where urgency and committed buyers can matter as much as market size. For secondary sales, Senkut applies the original investment test again: stay aligned with the founder, then ask whether the winner’s compounding is still underestimated.
VC Isn’t VC Anymore: The Rise of Cancer Capital
Anil Dash | Anil Dash | September 2, 2026
Anil Dash argues that the largest venture firms are no longer venture capital firms in the conventional sense. Classic VC financed risky young companies and depended on those companies producing exceptional outcomes. His “Cancer Capital” label describes a small group of firms that have expanded into enormous multi-strategy asset managers, collecting management fees across tens of billions of dollars while gaining access to private equity, secondaries, public shares, political spending, and transactions between affiliated funds.
Dash says this changes both incentives and accountability. Managers can earn substantial fees whether individual startups succeed or fail; early investors can find liquidity before a company reaches the public market; and pension or retirement capital may ultimately bear risks that ordinary savers did not consciously choose. In his account, the balance of power also shifts from founders seeking backing for independent ideas toward financiers selecting companies that fit broader industrial and political agendas.
Dash goes overboard when he turns that structural critique into a single moral and political category. Growth equity, secondaries, public holdings, management fees, lobbying, and individual misconduct raise different questions and do not by themselves prove an oligarchic project. The essay’s most serious accusations need far more evidence than this opening installment provides, and the “Cancer Capital” label obscures distinctions that matter.
The narrower question is still worth asking: when a firm spans venture, growth, public markets, secondaries, lobbying, and founder liquidity, should it still be understood and governed as venture capital? Hussein Kanji’s response supports that limited concern, not necessarily Dash’s wider indictment: “We really, really need to talk about venture capital. Because it’s not venture capital anymore.”
Hitting Once, Hitting Often: The Anatomy of a Great VC Career
Ilya Strebulaev and Blake Jackson | Stanford GSB Professor on Startups & Investors | September 8, 2026
Venture returns are concentrated not only in a small number of companies, but in a remarkably small group of people. Blake Jackson and Ilya Strebulaev analyze 99,243 investments in 52,709 startups by 12,151 venture capitalists and estimate slightly more than $1.2 trillion in net profits, measured in 2024 dollars. The top 1% of VCs account for 56.7% of those profits. The top 5%, roughly 600 people, account for 90.2%.
The distribution of successful investments is less extreme but still unforgiving. About 60% of VCs with at least one investment are never credited with a successful outcome, while only 15% record three or more successes and 6% record six or more. The companion research finds persistent differences associated with prior operating success, education, experience, and access. Marginal inclusion on the Forbes Midas List appears to improve subsequent access to highly valued startups, suggesting that demonstrated skill and status can reinforce each other.
Terrence Rohan’s summary is pithy: “The top 1% of VCs are exceptional. The next 4% are good.” The data strongly supports the concentration, although that sentence remains his interpretation. The study estimates profits at the individual deal level and allocates credit among investors. It does not report audited fund-level DPI or TVPI, and its results should not be read as proof that every investor outside the top 5% is unskilled. It does show that merely obtaining venture exposure is not diversification if nearly all economic value is produced by a tiny group of selectors.
Moonshot capitalism: AI rewrites the venture capital playbook
Financial Times | September 10, 2026
The Financial Times argues that AI is changing venture capital’s appetite for risk. As generative AI weakens familiar software moats and compresses the cost of simulation and engineering, investors are rediscovering capital-intensive bets in fusion, space, defense, advanced manufacturing, and other technologies once considered too slow or expensive for venture portfolios. Excluding giant AI-lab financings, Dealroom estimates that global deep-tech investment has exceeded $150 billion since the start of 2024, more than the $133 billion invested during the entire decade ending in 2019.
SpaceX supplies both the aspiration and the fear of missing out. The article cites PitchBook’s estimate that Founders Fund’s roughly $600 million investment could be worth more than $50 billion at an IPO. General Catalyst’s Hemant Taneja describes the related fund arithmetic: when software companies can be built more cheaply and copied more readily, large funds need founders to pursue much bigger outcomes. Plural’s Carina Namih argues that AI simulation can lower the capital intensity of difficult physical technologies, making some moonshots more compatible with venture timelines.
The revival comes with a warning. Deep-tech investment has not yet exceeded its 2021 peak, which was inflated by battery and electric-vehicle companies including Rivian and Northvolt that later disappointed investors. AI can accelerate design and reduce experimentation costs, but it does not remove manufacturing, regulatory, supply-chain, or commercialization risk. The new playbook may restore venture capital’s appetite for technological ambition, but large prospective markets and spectacular paper gains are not substitutes for viable businesses.
Mia Farnham and Charles Hudson extend the argument from the founder’s side. Once investors have seen trillion-dollar companies built on venture timelines, they become less interested in businesses that could credibly reach $1 billion but not $10 billion or $25 billion. That bifurcation can distort what founders choose to build and encourage them to present every company as a category-defining moonshot.
Their sharper warning concerns what happens after the money arrives. Exceptional teams can raise enormous rounds around broad technical ambitions without identifying the concrete problem they will solve. Capital then becomes a substitute for insight, extending the time before the market reveals whether the company has a business. In a consensus market, founders also have to become “legible” to investors through familiar backgrounds, narratives, and social proof. That may make fundraising easier, but it is not evidence that the underlying judgment is sound.
Regulation
California enacts laws restricting chatbots and banning teens from addictive social media
Colin Lecher | CalMatters | September 10, 2026
CalMatters reports that Gov. Gavin Newsom signed a 13-bill package aimed at online child protection. The key changes include mandatory limits on addictive social media design features for users under 16 and state controls on chatbots for minors, including tighter time limits and required safety checks. The chatbot rule, commonly called the Adam Raine law, also calls for stronger age verification, built in crisis resources, and an explicit notification process when a chatbot detects high risk of self harm.
The report notes the policy tensions. Supporters say the rules answer a growing youth mental health crisis linked to social media and AI use. The Electronic Frontier Foundation and other critics argue that the addictive-feature provisions are broad enough to function as a near blanket ban on routine social media access for young users. The piece also points to likely implementation frictions around enforcement and age-proofing and cites broader child-safety momentum from a parallel Meta settlement over teen features.
SEC Eyes Pre-IPO Share Sales as Private-Market Retail Access Expands
Author: Julia Sousa Published: September 2, 2026
As retail interest in private companies expands, the SEC is focusing on the intermediaries selling access rather than challenging access itself. Cooley reviews two recent enforcement actions alleging that pre-IPO investment vehicles misrepresented whether they owned sought-after shares, obscured layers of fund ownership, and failed to disclose substantial markups and registration issues.
The largest case concerns Andrew Spaventa and three controlled entities, which the SEC says raised more than $74 million from over 800 investors through 11 private funds. The complaint alleges that investors paid 27% to 91% above the defendants’ acquisition prices and that sales agents received $12 million in commissions. Spaventa denies the allegations.
The policy tension is important. SEC Chair Paul Atkins is promoting the “responsible retailization” of private markets, but access depends on full disclosure: whether a vehicle already owns the shares, where it acquired them, whether ownership is direct or layered through other funds, and what markup investors are paying. This is regulation aimed at fraud, transparency, and accountability rather than preserving private-market access for institutions alone.
Read more: Cooley
The Anti-Tech Antitrust Push Has Failed
Benedict Evans, with John Gruber | Threads | September 4, 2026
Benedict Evans argues that the decade-long neo-Brandeisian campaign against large technology companies has largely failed, while the European Union’s parallel regulatory effort has drifted toward irrelevance. He suggests that the movement’s advocates should examine why their agenda produced so little of what they promised.
John Gruber sharpens the methodological criticism in his reply. He says that, across the polarizing positions he has taken, this is the only one where the opposing side seems both highly educated and largely unread in the arguments against its position. He points to continued claims that Lina Khan was effective at the FTC as the clearest example. The exchange is less a defense of every large technology company than an indictment of regulation built around slogans without serious engagement with markets, trade-offs, or counterevidence.
Infrastructure
There’s an awful lot we don’t know about data centers
Matthew Yglesias | Slow Boring | September 8, 2026
Matthew Yglesias sets out to answer a basic question behind the data-center backlash: whether the United States built more facilities in 2025 than in earlier years, and whether construction is accelerating in 2026. He says he could not find a reliable answer. The essay’s purpose is therefore partly an “un-splainer” about the missing measurements underneath a fast-moving public debate.
Yglesias identifies several sources of disagreement. Pew counts facilities, while a Manhattan Institute analysis counts gigawatts of capacity; the sources also use different datasets and definitions of what qualifies as a planned project. A campus can contain several buildings, while two neighboring facilities may be counted separately even when that distinction says little about their economic or grid impact. Data centers are not a consistent federal regulatory category, and proposals range from concrete developments to speculative plans.
Commercial secrecy creates a deeper obstacle. A person involved in data-center transactions told Yglesias that major technology companies know only approximately what competitors are building and deliberately obscure their own plans. Cloud providers also avoid publishing precise facility locations for security reasons. The essay does not resolve the construction trend; its central finding is that widely cited counts are not comparable enough to support confident claims about the pace, size, or geography of the buildout.
TPU Inference Externalization Full Steam Ahead - InferenceX
Alec Ibarra, Cam Quilici, Bryan Shan, and colleagues | SemiAnalysis | September 7, 2026
SemiAnalysis publishes what it describes as the first third-party inference results for Google’s TPUv7 Ironwood and argues that Google is turning an internally proven accelerator into a credible external alternative to Nvidia. In apples-to-apples FP8 serving tests on Qwen3.5 397B, its InferenceX preview finds that Ironwood can deliver up to 50% better performance per dollar than B200 and B300 GPUs across parts of the cost-latency curve.
At 100 generated tokens per second per user, SemiAnalysis estimates costs of $0.181 per million total tokens on Ironwood, against $0.222 on B200 and $0.276 on B300. At a slower 20-token-per-second target, Ironwood produced 9,364 tokens per second per chip and, under the publication’s external total-cost model, delivered 50.4% more tokens per dollar than B200 and 96% more than B300. The authors attribute the economics mainly to lower modeled ownership cost rather than consistently higher raw throughput.
The article also documents important limits. At high concurrency, Ironwood’s mean time to first token was 5.41 seconds, compared with 3.75 seconds for B200 and 2.40 seconds for B300. Nvidia still leads in parts of the Pareto curve, in FP4 workloads, and when its disaggregated GB300 serving stack is compared with the still-aggregated external TPU stack. SemiAnalysis expects Google’s new native TorchTPU backend for PyTorch, vLLM, and SGLang to be open-sourced around October and to close some of those gaps, but that is a roadmap claim rather than a measured result.
Geopolitics
America Is Still Beating China in the AI Race
Noah Smith | Noahpinion | September 5, 2026
Noah Smith argues that the United States still leads China in the AI race through stronger frontier models, more compute, and far more commercial revenue. Recent Chinese releases such as Moonshot’s Kimi K3 and Z.ai’s GLM-5.3 have narrowed parts of the gap, but headline comparisons often pit new Chinese systems against older publicly released American models rather than the strongest systems inside U.S. labs.
Z.ai says GLM-5.3 slightly exceeded Anthropic’s Mythos 5 on CyberGym vulnerability identification, 84.5% to 83.8%, although the result has not been independently verified. Anthropic remained well ahead on converting discoveries into working exploits, 78.0% to 54.4%, and on timed attack-development tasks. Smith’s broader point is that release timing and safety restrictions can make the public capability gap look smaller than the underlying one.
The strategic stakes extend beyond benchmark prestige. Smith argues that a durable model lead could affect cyber defense and military power, while cooperation on catastrophic risks becomes easier when neither side expects to gain by racing ahead. His immediate policy warning is more grounded: the United States could squander its advantage by driving Chinese AI researchers out of the country. America’s lead depends not only on chips and capital, but on remaining the place where the world’s best technical talent wants to work.
Biology
AlphaGenome Atlas: a high-resolution map of human DNA
Pushmeet Kohli and Ziga Avsec | Google DeepMind | September 8, 2026
Google DeepMind introduces AlphaGenome Atlas, a database of model predictions for the molecular effects of all nine billion possible single-nucleotide changes in the human genome. The resulting dataset is about one petabyte. Its focus is the 98% of the genome that does not code directly for proteins and is less well understood, although the database covers coding regions as well.
The Atlas adds an AlphaGenome Variant Impact score that combines multiple predicted effects into a single ranking intended to help researchers prioritize variants for follow-up. Google cites an early rare-disease case at the Broad Institute in which the score highlighted a DNM1 variant predicted to create an incorrect splice site, providing supporting evidence that helped resolve the case. In a separate analysis of more than 54,000 UK Biobank participants, Gareth Hawkes grouped rare variants by predicted molecular effect and reported 22% more non-coding associations; restricting the analysis to the top 1% of predicted-impact variants identified 19 genomic regions associated with body mass index.
The Atlas is available through a website designed for researchers without programming skills. The release presents its scores as tools for prioritization and hypothesis generation: the DNM1 result supported other evidence, and the UK Biobank associations point to targets for further study rather than establishing clinical effects on their own.
Startup of the Week
Starbucks Discovers the Coffee Shop
Gregory Meyer | Financial Times | September 10, 2026
This week’s startup was founded in 1971 and has finally discovered product-market fit: coffee shops. Starbucks is putting $1bn behind easy chairs, rugs, ceramic mugs, power outlets, and baristas who remember customers’ names. The breakthrough idea is that a place selling coffee might work better if people want to spend time there.
Brian Niccol’s “Back to Starbucks” plan aims to make more than 1,000 stores feel like community spaces again. After years of optimizing mobile ordering, takeaway traffic, and operational throughput, the company is rebuilding the “third place” that made the brand valuable in the first place.
The joke contains a serious product lesson. Digital convenience can improve a transaction while degrading the experience around it. Starbucks is spending startup money to rediscover its original product: not the coffee, but somewhere to be with other people.
Interview of the Week
The Madness of Markets
Andrew Keen with Alex Edmans | Keen On America | September 7, 2026
Andrew Keen interviews London Business School professor Alex Edmans about why intelligence does not reliably protect investors from speculative mistakes. Edmans opens with Isaac Newton’s losses in the South Sea Bubble, estimated at about GBP4 million in today’s money, and argues that intelligence is domain-specific. Smart people can become overconfident when they assume expertise in one field transfers to markets.
Edmans connects that error to the ease of trading stocks, options, crypto assets, and NFTs through retail platforms. Just as about 90% of people believe they are above-average drivers, investors overestimate their ability to beat other participants. He says the average trade loses money even before fees and commissions, but does not advise abandoning markets entirely. His prescription is to identify where an investor genuinely has an edge and avoid tying personal worth to portfolio value.
On the current AI boom, Edmans distinguishes high expectations from the valuations reached in the dot-com bubble. He says the AI sector trades at roughly 25 to 30 times earnings, compared with Cisco at about 190 times earnings during the earlier boom. He therefore treats the historical analogy as a warning about judgment and overconfidence, not as evidence that today’s leading AI companies are priced identically to the most extreme dot-com examples.
Post of the Week
Remembering Bill Draper
Tim Draper | X | September 9, 2026
Tim Draper announced the death of his father, Bill Draper, in four simple sentences: “He was the best! We were so lucky to have him with us as long as we did. I know he touched many of your lives.”
Bill Draper was one of the people who made Silicon Valley venture capital an institution. He began at Draper, Gaither & Anderson in 1958, co-founded Draper & Johnson in 1962, helped build Sutter Hill Ventures, chaired the US Export-Import Bank, led the United Nations Development Programme, and later applied venture methods to philanthropy. Harvard Business School’s account records the career. Tim’s post records the man.
A reminder for new readers. Each week, That Was The Week, includes a collection of selected essays on critical issues in tech, startups, and venture capital.
I choose the articles based on their interest to me. The selections often include viewpoints I can't entirely agree with. I include them if they make me think or add to my knowledge. Click on the headline, the contents section link, or the ‘Read More’ link at the bottom of each piece to go to the original.
I express my point of view in the editorial and the weekly video.
