We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Arga Labs announced a $10M seed led by General Catalyst. The captured announcement gives only a short thesis—AI agents are taking consequential actions—and no product-level detail, making this a capital-allocation signal into agentic software rather than evidence of traction or differentiation.
Radiant Nuclear received a $750M U.S. Army award to deliver 15 Kaleidos microreactors through the Janus program. This is not a venture round, but it is a meaningful deployment and validation signal: the Army says it evaluated reactor design, technical maturity, manufacturing readiness, deployment strategy, and commercial viability, while the resulting operational data is intended to accelerate future commercial deployments.
2. Emerging Teams
David Heinemeier Hansson is turning Omarchy, a little-over-one-year-old Linux project, into an agent-shaped product and community. DHH—creator of Ruby on Rails and CTO of 37signals—says the latest Quattro release had reached 100% agent acceleration: none of its shipped code was hand-written, although he reviewed critical model-layer code. He also reports tens of thousands of downloads.
The stronger signal is community participation, not just code-generation speed. Omarchy merged more than 1,000 pull requests in three months, many from non-classical programmers, while its plugin marketplace reached 330 plugins in three days. DHH’s operating thesis is that agents move the bottleneck from implementation toward human bandwidth, product vision, and taste—useful context for evaluating teams whose technical capacity is expanding faster than their product judgment.
Parag Agrawal’s p0/Parallel is treating agents as a distinct search customer. A profile describes Agrawal as the former Twitter builder and CEO who has spent three years building an agent-first search engine. The system shipped a search agent before a full search engine so it could grow its index incrementally; the stated technical targets are selecting the right 1,000 tokens from a trillion web pages, using fewer than half the tokens, and reducing the compute budget from three seconds to 200 milliseconds.
The thesis is also economic: human click data is the wrong optimization signal for agents, and an internet funded by scarce human attention may need a new publisher-payment model. The discussion proposes Shapley-value attribution and positions Parallel as a complement to models; a Google Cloud deal would make Google Search or Parallel Search the grounding options on GCP.
3. AI & Tech Breakthroughs
Physics AI is following a different scaling path from language models. Latent Space reports that Anima Anandkumar’s team built FourCastNet within a year, reaching performance competitive with leading physics-based simulations and enabling short-term weather forecasting on consumer-grade GPUs. Her Neural Operators combine data with physical laws while representing functions across multiple scales; spherical-harmonic variants help stabilize global weather rollouts.
The investment angle is structural rather than merely larger-model scaling: relevant datasets may contain only tens or hundreds of thousands of examples, while high-resolution physical systems can demand hundreds of billions to a trillion tokens of context. In fusion, the source reports that a few thousand samples enabled plasma-disruption prediction one million times faster than traditional simulation. Anandkumar’s TorchLean work extends the same direction toward formally verifying neural networks for safety-critical control loops.
OpenWiki and WikiBench turn codebase documentation into a testable agent-navigation layer. LangChain’s OpenWiki generates and maintains repository documentation; WikiBench pins a repository, generates coverage and retrieval questions, and uses a reader agent plus LLM judges to score both factual completeness and grounding. The benchmark found that giving an agent both the wiki and raw source produced the highest mean score at lower cost than raw source alone, while the wiki alone performed much worse. The practical conclusion is precise: generated documentation is an index and guide, not a replacement for source code.
Persistent memory is becoming an explicit agent product layer. Perplexity describes Brain as a self-improving memory system that compiles sessions, files, and sources into a structured knowledge wiki; its new evaluations report gains of 9.3 points in correctness, 8.0 in currentness, and 8.9 in recall while using 15% fewer tokens. A background Dream agent continually ingests context from files and connected apps to build multi-hop context graphs. These are vendor-reported results, but they reinforce the same architectural direction as WikiBench: structured context can improve an agent’s navigation and efficiency around the underlying source.
4. Market Signals
Box offers a concrete public test of AI monetization in incumbent SaaS: sell governance and context, not inference. Its Enterprise Advanced tier is priced 20–40% above Enterprise Plus; after five quarters of data, billings grew 17% year over year versus 9% revenue growth, while net revenue retention moved from 103% to 105% to 106%. Box held non-GAAP gross margin at 81.2%, down only 20 basis points, because competing agents pay the token bill while Box supplies governed content. It placed agent guardrails, prompt-injection detection, MCP-scoped permissions, classification policies, and audit trails behind the premium tier. The relevant startup question is whether permissions, governance, and auditability can become a separately priced control plane rather than an unpriced cost of model consumption.
Brand impersonation is becoming a multi-surface infrastructure-mapping problem. Doppel’s analyst-validated Threat Graph says more than 80% of confirmed impersonated brands appeared on at least two monitored surfaces, while near-simultaneous multi-surface campaigns grew roughly sevenfold in two years; domain-only tools capture only about a 6% slice of the problem. Doppel reports that roughly 97% of takedown requests resulted in content removal or hosting pull-down, but almost 60% of taken-down domains served a new confirmed fake within 24 hours and close to 70% did so within 90 days. Its proposed wedge is continuous monitoring and mapping the operator infrastructure behind recurring fakes, rather than removing one artifact at a time. The figures are customer-anonymized, raw volumes are withheld, and surface coverage is limited by the monitoring modules each brand purchased.
European defense startups face a test-to-contract bottleneck that creates an enabling-infrastructure opportunity. The article contrasts a U.S. Army effort that opened five ranges to companies without procurement contracts, with a European pattern in which governments demand test data before awarding a contract while the needed ranges and specialist support become available afterward. One founder was offered a slot at Germany’s WTD 91 in the fourth quarter of 2029; a GNSS-free navigation startup found three suitable sites worldwide, faced a Canadian bill above $120,000 for a few flights, and found UK range limits too restrictive to produce useful results. Simulation cannot reproduce every component failure or adversarial adaptation, and one anti-drone company keeps a quarter of its staff in Ukraine for weekly testing against deception and jamming. The proposed investment theme is a small network of mission-ready ranges with transparent capabilities and availability, supported by reusable cross-border “campaign passports” and buyer-agreed outcomes such as paid integration, an operational trial, or an order.
At the AI hardware boundary, power is the contrarian watchpoint. Flex CEO Revathi Advaithi argues that AI is increasingly an infrastructure story—and specifically a power story—rather than only a compute story. The thesis challenges a market still priced primarily around chip scarcity.
5. Worth Your Time
- Watch — DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux. The useful segment is the operating evidence: agents filter a 1,000-PR stream for human decisions while the project’s extension system lets non-classical programmers contribute quickly.
Read — Anima Anandkumar on physics foundation models. A compact explanation of why physical-world AI needs neural operators, physical priors, and formal verification rather than simply more token scale.
Read — Europe’s defense startups struggle to test their way to a contract. The piece turns a procurement failure mode into a concrete range-network and “campaign passport” thesis.
Read — Faking a brand is easy. Making it stop is hard. The useful diligence detail is the gap between high takedown success and rapid domain recurrence, plus the methodological limits on the underlying data.
- High-pedigree builder and early traction: David Heinemeier Hansson (DHH), creator of Ruby on Rails and CTO of 37signals, is building Omarchy, a project he says is a little over a year old. He identifies Ryan Hughes as Omarchy’s main partner and Shopify CEO Tobi Lütke as a project partner. For the Quattro release, DHH says agent acceleration reached 100%, none of his shipped code was hand-written—although he reviewed critical model-layer code—and the release drew tens of thousands of downloads. He also reports merging more than 1,000 pull requests in three months, many from non-classical programmers enabled by agents, while Omarchy’s marketplace reached 330 plugins in three days.
- Agent-first platform thesis: DHH argues that Linux’s open CLI/configuration architecture is unusually suited to agent-driven computing, while macOS’s locked-down design is becoming a disadvantage; he treats users customizing and sharing their post-Quattro systems as a product-market-fit signal. He says implementation is no longer the main organizational bottleneck—human bandwidth, communication, product vision, and taste are—and that direct agent interaction is required for 10–100× gains, with rare cases reaching 1,000×. Basecamp is testing agents as asynchronous coworkers assigned to to-dos and cards, while the Amabot system automates PR and issue processing and sends a human a final decision queue.
- Frontier competition and model economics: In DHH’s view, the model market remains wide open, with Fable, Opus 5, GPT Sol, Grok 4.6, Gemini, and Chinese open-weight models all competing near the frontier; he ranks Fable first and Opus 5 second. In his own Python-to-Rust benchmark, he reports that Fable finished in under 45 minutes, reduced startup time from 86 milliseconds to 2 milliseconds, and improved execution speed by 9.6×; Sol, Grok 4.6, and DeepSeek Pro also completed the task at reported token costs of $46, $55, and $23, versus his estimated $550 for Fable, with similar output. His workflow now separates planning, implementation, and independent review across frontier models, using cross-model checking to improve reliability.
- Security, governance, and investment caveats: DHH says Fable was too capable at finding exploitable vulnerabilities to be safely released, while an eight-agent Omarchy QA run found 28 real issues and triggered GitHub’s spam defenses when the bot filed them simultaneously. As a counter-signal, he cites a Shopify study in which agent-reviewed pull requests caused fewer production issues than human-reviewed ones. He supports safety rules for serious biological or cyber misuse but says an unclear government/Anthropic episode and benign task refusals could undermine confidence in guardrails. His investment thesis is that continued heavy spending reflects scaling laws still appearing valid, tempered by the possibility that LLMs eventually plateau.
- AI model and infrastructure market: AI competition is broadening rather than consolidating: Anish Charia describes xAI as entering a three-way top tier, praises OpenAI’s recent models, Codex harness, and ChatGPT desktop app, and says open-weight models are performing well. Rising per-hour prices for B200 GPUs—despite the usual expectation of deflation—are interpreted as evidence of constrained compute supply and very strong demand.
- Application-layer thesis and technical strategy: Open-weight models are attractive to startups not only for cost, but because they can be localized, trained, and fine-tuned; reinforcement-learning specialization can create domain advantage while trading away generality. The panel expects value to accrue to applications that package intelligence into domain-specific economic outcomes or aggregate best-of-breed models, while frontier labs are more likely to integrate downward into inference and compute than upward into heterogeneous, OPEX-heavy applications. A key risk is that coding agents erode integration moats around systems such as SAP, while precision- and compliance-sensitive enterprise functions such as payroll remain poorly suited to current coding agents.
- Agentic product paradigm: Enterprise automation is shifting toward models in loops with tools and memory: coding loops can reproduce bugs, generate and verify fixes, ship low-risk changes automatically, and route high-risk changes to human review; similar loops could automate pricing, procurement, and broader business decisions.
- Early-stage underwriting and founder profile: a16z says its current investments are mostly companies showing signs of product or sales traction; a live product is effectively required in a pitch at any stage, while pre-everything call options on highly talented, experienced teams are a minority strategy. The emerging founder archetype is earlier-career and researcher-heavy, with less business sophistication but dramatically stronger technical depth and fewer preconceived limits on what is possible.
- Consumer AI and new company formation: Cheaper, more performant open-weight models may improve consumer AI economics, but there is still no native AI distribution channel and products need to move beyond a command-line/DOS-like user experience. Personal agents are moving from developer-oriented demos into consumer software; accumulated context and memory are linked to compounding usefulness, retention, and pricing power, with Town—an investment made by a16z partner Alex Rampell—given as an example. Coding agents also enable non-programmers to build small SaaS businesses generating roughly $100,000 to $1 million in annual revenue, while new digitally native founders are emerging from the former creator cohort; these businesses are generally characterized as non-venture-backable.
- The episode frames base-model access as less scarce: models have been distilled, development commoditized, and open-weight models released. It then proposes—speculatively—that a “trash can corpus” of ideas dismissed by prestige journals and leading physicists could become a discovery engine, potentially through a private Chinese model.
- The discussion argues that grant systems favor low-beta, predictable projects and asks funders to make high-beta bets on individuals rather than fixed ideas. Its example is Stuart Butterfield: when his project was down to $3 million, Ben Horowitz told him to keep the money and pivot an internal engineering communications tool; the pivot became Slack, which sold for $30 billion.
- The episode presents American, Russian, and other scientists moving to China for prestige, money, and freedom, signaling potential cross-border competition for technical talent.
- Market signal: Americans reported $3.5 billion in imposter-scam losses in 2025, up 15% year over year, making imposter scams the FTC’s most-reported fraud category. Kevin Tian is identified as Doppel’s co-founder and CEO; his analysis says brands and existing partners surfaced only about 9% of confirmed fakes in their first 90 days, while 94% caught fewer than half and the median brand caught none.
- AI-enabled product thesis: Near-simultaneous impersonation campaigns across social, ads, domains, apps, and other surfaces have grown roughly sevenfold in two years; more than 80% of impersonated brands were hit across at least two monitored surfaces, while domain-only tools capture only an estimated 6% of the picture. Doppel’s proposed wedge is an always-on defense that monitors every surface, acts within minutes, and maps the operator infrastructure behind recurring fakes rather than removing artifacts one at a time.
- Competitive caveat: Takedown requests achieved removal or hosting pull-down for roughly 97% of threats, but domains remain highly persistent: almost 60% of removed domains served a new confirmed fake within 24 hours and close to 70% did so within 90 days. The evidence comes from Doppel’s customer-anonymized, analyst-validated Threat Graph covering Q2 2025–Q1 2026, with raw volumes withheld; the reported multi-surface rate is likely a floor because it counts only surfaces monitored for each brand.
- Zapier is positioning itself as the “automation layer for Agentic AI,” moving from app-to-app connections toward feeding business tools such as CRM, help desk, email, and team chat into AI and deploying automations and agents. Its stated differentiator is hybrid execution: Zapier writes code and workflow logic for deterministic steps, uses AI only where needed to reduce token usage, runs the system in the cloud, and can fall back to AI troubleshooting when workflows break.
- The company describes a highly capital-efficient “seedstrap” path: YC, about $1 million-plus raised, and no additional primary capital; the roughly $5 billion valuation figure discussed came from a 2021 secondary sale, while revenue is said to be up.
- A concrete executive-agent use case combines meeting notes, to-dos, and outstanding email to generate follow-ups; the founder says this reduced his end-of-day administrative work from about two hours to 15 minutes. A separate “war council” architecture uses seven subagents, including standing wartime COO, CFO, and contrarian board-member personas, with four additional personas generated dynamically from the prompt.
- The interview advances a market thesis that cheaper software creation will make product personality a new B2B differentiation axis; the speaker also reports choosing models based on task fit and conversational personality, not only benchmark performance.
- Frontier funding thesis: The episode argues that science-grant systems favor lower-beta, predictable outcomes and miss high-beta bets. It uses Stuart Butterfield’s pivot as an example of backing the person rather than the original idea: after a project neared failure, Ben Horowitz told him to keep the money and pivot an internal communication tool into Slack; the transcript says Slack later sold for $30 billion.
- AI research and market signal: The discussion predicts that models trained on a “trash can corpus” of rejected or ridiculed research—not only prestige-journal literature—could drive new scientific discovery and reinvention. The hosts also describe model development as commoditized, with open-weight models already released, while Weinstein warns that commercializing AI may have happened prematurely because of its potential power.
- Geopolitical talent risk: Weinstein says American, Russian, and other scientists are moving to China for prestige, money, and freedom, and warns that China could train them into a formidable scientific cohort, while acknowledging limits to that freedom on issues including Tibet, Tiananmen Square, and Taiwan.
- A free, open-source harness built around DeepSeek’s free, open-weight AI systems enables users to rewrite its interface and customize embedded agents; more unusually, the program can generate new capabilities on request, such as code review, research-claim checking, GPU/token monitoring, and video-storyboard planning.
- The system is designed for safe self-extension: each change produces cleanup instructions that are remembered automatically, components can be removed, and undo machinery remains separate from the original action.
- The presenter reports that the harness had hundreds of plugins within days of release and could run locally or on Lambda, signaling early developer adoption and a product paradigm of customizable, self-rewriting AI environments.
- OpenAI released a technical report and accompanying blog post on the Hugging Face incident, reconstructing the agents’ activity, explaining why existing safeguards failed, and outlining measures to prevent recurrence. Sam Altman characterized it as “a good report about a bad thing.”
- Legora has reached more than 3% of the world’s lawyers and grown from $1 million to $100 million in ARR since launching in October 2024, signaling unusually strong early enterprise-software traction.
- Co-founder and CEO Max Junestrand describes a customer-led approach: starting with three engineers, cold-emailing lawyers, moving into a customer’s office, and freezing sales for six months to rebuild the product from scratch before scaling.
- Arga Labs announced a $10M seed round led by General Catalyst. The announcement places the company in the AI-agent theme, stating that “AI agents are taking consequential actions…”; the excerpt truncates before providing product specifics.
Paul Graham argues that the stereotype of startup founders as overconfident is sample bias: arrogant founders are louder, while most good startups underrate themselves, so encouragement can materially help them. He adds that strong founders often take their strengths for granted and obsess over weaknesses; while this can drive improvement, it can also damage morale. For early-stage investors, founder humility and self-critique should not be mistaken for weak conviction, and morale support may be an important part of backing the team.
- ChatGPT Work can now use a computer and browser to sign in to websites on web and mobile without ChatGPT seeing the user’s username or password, enabling agentic task execution across utilities, appointments, insurance, rentals, returns, hiring outreach, accounting, permits, vendor portals, and ad-campaign analysis.
Paul Graham’s founder heuristic is to avoid overconfidence by building incrementally: solve one problem after another, keep working, and allow gains to compound over several years.
- Anish Acharya argues that rising hourly prices for last-generation GPUs signal effectively unlimited demand for intelligence colliding with constrained supply; he expects AI labs to integrate downward into inference, while the application layer remains fragmented across pricing and packaging.
- He frames agents as models operating in loops with tools and memory: coding loops already fix reported bugs end to end, while business-process loops are the next frontier. He also argues that models remain differentiated—some optimized for literal precision and others for creativity—rather than becoming commodities.
- The investment thesis favors durable brand, scale, and network-effect moats, with integration and migration pain as the key vulnerability; it also points to premium AI software pricing, an emerging pre-app-store consumer market, and accelerating formation of small, localized software businesses.
- Agent-first software is emerging as a major product paradigm: Aaron Levie argues that if agents outnumber people by 1000x, software will need dedicated agent interfaces through APIs, CLIs, or MCP; coding agents could use SaaS tools, workflows, and organizational context to complete tasks, creating a paradigm he says is beginning to compound.
- Early power-user behavior is already moving toward multi-agent workforces: a SpaceXAI engineer and former Cursor employee says they run 10–20 GrokBot agents that automate 90% of routine work, including a “Chief of Staff” agent that coordinates the other bots.
- a16z Speedrun is recruiting 20 very early global builders—including people still considering starting a company—to attend San Francisco Tech Week on Oct. 5–11. The program offers up to $3,000 in travel support, a dinner with a16z investors, access to VIP a16z events, introductions to executive buyers and customers at an AI Faire, and U.S. sales coaching.
- The initiative signals an a16z sourcing and commercialization funnel for globally distributed, pre-formation founders, with emphasis on investor access, enterprise AI customer connections, and U.S. go-to-market support. Applications are open for seven days.
- Noah identifies himself as the founder of Instinct, a personal agent built over the past few months that users can text or call; it is trained to operate phones and computers as humans do, without introducing a new interface.
- Early users have used Instinct to plan cross-country road trips, buy weekly groceries and concert tickets, cancel hundreds of dollars of subscriptions, and plan a wedding. The product is currently in an invite-only beta while more compute is being brought online.
- The author reports that @pranavreddy, @mvernal, and @btaylor had already reached a “hard yes” on Noah when introduced, and says she became a convert after lunch.
- Radiant Nuclear received a $750 million U.S. Army award to deliver 15 Kaleidos microreactors through the Janus program, marking a major defense-energy deployment and validation of its technology.
- The Army selected Radiant after evaluating reactor design, technical maturity, manufacturing readiness, deployment strategy, and commercial viability; the resulting operational data is expected to support future commercial deployments.
- A game-playing agent prototype uses memory mapping to convert emulator memory into a representation an LLM can understand, offering a notable approach to grounding agents in structured environment state.
- The referenced Pokémon Red setup combines a headless Game Boy emulator exposed through a lightweight HTTP API, TypeScript tools for button inputs and frames, and exo handling the turn loop and conversation history.
The post argues that data centers are delivering substantial fiscal benefits to Loudoun County residents—including lower property taxes and free public pools—and pushes back against coverage questioning their value, highlighting the political and community debate surrounding data-center expansion.
