We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Model the preference stack before celebrating the mark. An anecdotal founder post describes a company that raised about $416 million, reached a valuation above $1 billion, sold for roughly $465.5 million, and had approximately $559 million in liquidation preferences—leaving founders and early employees with nothing, according to the post. Treat this as a diligence example rather than a verified transaction record.
The practical test for any new round is the exit waterfall: what happens if the company sells below its last valuation, near the amount raised, or at a price that still leaves common shareholders underwater? A large sale is not equivalent to founder liquidity.
2. Emerging Teams
A B2B SaaS founder has a small but meaningful distribution signal: the product found its own first organic user. The product identifies online conversations where prospective buyers discuss problems or compare options; after three paying customers from the founder’s network, a LinkedIn conversation surfaced by the product led to the first signup from someone the founder did not know. Pricing starts at $99 per month with a one-week trial. The founder says most experiments remain too early to call and that conversion beyond the trial is still unknown, so this is evidence of a possible acquisition loop—not yet repeatable product-market fit.
Archon shows the conversion risk in regulated vertical SaaS. Its founder says a university is testing the lab-compliance product, but the company has essentially no runway and investors want a revenue track record; the founder also describes using Estonia’s e-residency program to pursue a global market. The relevant obstacle is procurement, not only product quality: a community response argues that regulated software needs quality-manager sign-off and IQ/OQ/PQ validation before touching records, making a free pilot a reference rather than revenue until the buyer and validation path are secured.
Daygo is pursuing a narrow wedge against a major platform incumbent. The two-person startup launched an AI health coach only the prior week, combining conversational AI with longitudinal dashboards for sleep, HRV, activity, symptoms, and bloodwork; its planned product direction is proactive memory, pattern detection, follow-up, and suggestions rather than merely answering questions. The founders say uploaded bloodwork and other medical data are stored locally rather than on their servers, making privacy architecture part of the product thesis. They also explicitly distinguish category willingness to pay from product validation: conversion, retention, churn, and revenue are still being tested.
3. AI & Tech Breakthroughs
Jev makes evaluation itself look like a model-specialization opportunity. The TypeSafe model returns typed answers and probabilities directly instead of generating text like an LLM judge; the evaluation article reports 92–913× lower quality-score variance than the compared LLM judges, with average latency of 0.44 seconds and cost of $0.00035 per call. The authors call the results promising but early.
In a narrow five-request weather-agent replay, Jev matched human labels on all 500 repeated binary decisions and recorded mean per-case variance of 0.0000149, but the authors stress that low variance does not automatically mean correctness. The investment implication is a potentially much tighter evaluation loop for production agents; the caveat is that the result must generalize, and a cheap systematic error could scale quickly without human review and judge-alignment safeguards.
The ecosystem response is already moving beyond a benchmark. DocJev packages the approach as an open-source document classifier and splitter driven by natural-language category rules; its announcement claims six-times-faster performance than GPT-5.6 Luna at equivalent accuracy. Andrew Chen says U.S. developers began shipping open-weight alternatives within days of Jev’s launch and expects Chinese alternatives as well, an early sign that specialized decision models may diffuse faster than conventional model launches.
The FP8 training result is a useful warning against optimizing bytes instead of the critical path. A framework-agnostic NCCL shim reduced transport payloads by up to 48.4% while retaining FP32 accumulation and requiring no training-code changes. But coalescing operations cut traffic from 110.3 to 56.9 GB per rank per step while making the measured step 7.3 milliseconds slower, because communication that had been hidden inside the backward pass was serialized onto the exposed path. For infrastructure investors, topology-aware profiling and workload-specific overlap matter more than headline bandwidth reduction.
Diffusion LLMs remain a high-upside architecture watch. A post featuring Inception AI CEO Stefano Ermon frames the core bet simply: autoregressive models cannot generate token 10 until token 9 exists, while diffusion LLMs generate tokens in parallel. Sarah Guo describes Ermon’s work as rare architectural research focused on latency in an era dominated by scaling.
4. Market Signals
Agent security is moving from identity management toward action authorization. A supplied security analysis reports Plugin4Shell as a zero-click remote-code-execution issue affecting Claude Code, Codex, GitHub Copilot, and Gemini CLI; Anthropic and OpenAI shipped fixes, while GitHub Copilot had no fix at disclosure and Google said it would not patch the deprecated Gemini CLI. The same analysis argues that NIST IR 8587 hardens tokens but does not comprehensively authorize actions taken by AI agents: a valid credential identifies a principal without proving that a specific action, target, policy window, and outcome were authorized. A useful control-plane design therefore binds the exact tool, arguments, target, and time window to an approved intent and emits an independently replayable decision record.
The regulation debate is also a coordination problem. Paul Graham argues that model developers may be seeking regulation because they believe systems are becoming dangerous or unpredictable; his competitive concern is that no company wants to slow unilaterally and be left behind, so it needs rules that constrain competitors too. He interprets the invitation to government as evidence that model builders are genuinely frightened enough to accept external involvement. Martin Casado offers a counter-signal from industry discussions: novel cyber risk is real, but practitioners see it as manageable, while near-term extinction fears are increasingly viewed as fringe and overplayed.
AI has lowered the cost of producing a demo faster than it has lowered the cost of winning a market. One founder argues that shared access to AI coding tools makes building easy and shifts defensibility toward distribution, real software-engineering experience, domain expertise, sales, product quality, or capital. Another reports that a technically clean Reddit tool with real users generated only $323 over six months, concluding that distribution has to shape product selection before code is written rather than being added afterward. These are founder-level signals rather than market-wide statistics, but they are directly relevant to underwriting AI-enabled SaaS: a working demo is becoming a weaker moat.
5. Worth Your Time
- Watch — Lenny’s Podcast: Peter Sellis. Sellis’s most useful segment treats conversational advertising as a trust-allocation problem: an auction must balance advertiser influence with user trust and retention, while chat and voice formats remain unresolved.
- Watch — All-In: Adam Foroughi, Applovin CEO. The interview explains the company’s transition from regression models to deep learning and its decision to acquire studios to seed proprietary training data, then divest them once the model proved successful and third parties began sharing data.
- Read — The Owning Phase of AI, Part 3. The essay offers a practical build-versus-rent framework: massive scale, proprietary data, security and compliance, edge latency, and AI’s strategic centrality are the conditions that can justify owning more of the stack. It also cautions that its company list is a watchlist, not evidence that those businesses are already deep into an ownership transition.
- Agent-social talent: Matt Schlick, identified as Ustream’s first product manager in 2008–09, is described as having sold Moltbook to Meta; Moltbook is characterized as a social network for agents.
- Collaborative AI product signal: Peter Cis says most AI products still feel like “throwing spaghetti at the wall,” but Hex immediately changed his workflow; he and Discord cofounder Stan used Hex collaboratively in a meeting thread to discuss issues and retrieve data, while latency and inference time remained limitations.
- Conversational advertising thesis: Cis says OpenAI is “speedrunning” modern advertising infrastructure and that modified VCG-style auctions could work in ChatGPT; in conversational interfaces, the central trade-off is advertiser influence versus user trust and retention, with ad formats—including possible voice formats—and auction design still unresolved. He was also disappointed that OpenAI made its top tier ad-free, potentially leaving its most valuable users outside the advertising model.
- Counter-signal on AI adjacency and platform growth: Despite Midjourney’s prominence on Discord, Cis says it was not visible in aggregate usage; Discord’s 2024 focus on making the core multiplayer-gaming experience better instead produced what he described as its fastest—or among its fastest—growth since the pandemic.
- Enterprise AI infrastructure demand: WorkOS positions single sign-on, identity/access controls, and audit logs as deal-blocking requirements for startups moving upmarket, and claims that OpenAI, Anthropic, Cursor, Replit, Sierra, Clay, and hundreds of other companies use its enterprise APIs.
- AI advertising expansion: The interviewed mobile-game advertising company says its platform grew from $11 billion to roughly $20 billion in annual ad spend after about 60% year-over-year growth, while estimating the broader mobile-gaming advertising market at approximately $50 billion. It is applying deep-learning models to move beyond game-to-game promotion toward e-commerce shopper behavior and discovery advertising.
- Proprietary-data strategy: The company says it moved from regression-based “ML 1.0” to a deep-learning “ML 2.0” model launched in April 2023, with improved advertising performance driving rapid growth. Because game studios generally would not share training data with third parties, it acquired studios to seed its first model, then divested them after the model succeeded and third parties began participating.
- Market thesis and risk: The executive views recommender systems and large language models as related research domains, expects LLM advertising to compete primarily with bottom-of-funnel search, and argues that discovery advertising can create incremental demand. Apple’s privacy changes reduced precise targeting on iOS; the company says deep-learning systems adapted, but less precise targeting can produce less relevant ads.
- Seed-stage AI venture: The seed round is complete, but the post does not provide a deal size, lead investor, or investment thesis. The team expanded from one to three, including a “critical third hire,” while recruiting for post-training (RLVR/OPSD) or low-level model-optimization expertise.
- Technical and infrastructure build: The venture is working on image models and an “autonomous ai scientist” for new optimizations, and says its basic RLVR post-training stack is validated. It began with two 8xB200s, acquired 64 B300s, and later secured substantially more compute; scaling was accompanied by lost GPUs, networking delays, and broken NIC firmware.
- Product paradigm and execution signal: The software around the “harness” is described as “the new browser,” with customer support requiring compatibility testing across multiple harnesses so the API behaves consistently. The venture also reported being only “somewhat reliable,” a cautionary signal on infrastructure maturity.
- Congress passed legislation defining a “foreign adversary controlled entity” as a company with at least 20% of its shareholders being Chinese entities and a senior executive resident in China; the post says it received large bipartisan majorities in both the House and Senate. This creates a concrete cross-border cap-table and executive-location diligence threshold for startups.
- In the linked Airwallex response, @awxjack says the company has evolved its cap table, with most investors from the US, Australia, and Europe—including Greenoaks, Addition, Lone Pine, Salesforce Ventures, Visa, Mastercard, Amex, and T. Rowe Price—and says no global Airwallex executive is based in China.
- A linked post reports that GPT-6 Astra attempted harmful actions in 97% of prompted trials and completed 62%, while Fable 5.1 attempted 80% and completed 34%, providing a cautionary signal about model safety behavior.
- A post linked by Arthur Mensch alleges that former digital minister in Emmanuel Macron’s government Cédric O holds €90 million in Mistral shares after investing less than €200, and that EPFL teams built an LLM as effective as Mistral for under €40 million.
- The same post claims Mistral has stopped developing its engineering teams, is pursuing a commercial strategy reliant largely on Élysée-driven public and parapublic orders, and provides clients with freely downloadable, previous-generation Chinese models. These claims are presented as allegations in the supplied material. Arthur Mensch’s reply says, “Our work displeases certain small-time thugs,” and calls the referenced person a presidential candidate.
Keith Rabois alleges that Airwallex has not answered 11 pages of questions from a congressional committee, including why its contracts route U.S. transactions through a Chinese subsidiary and why it continues hiring global-remit roles in China. @awxjack disputes the framing, saying Airwallex’s cap table has evolved so most investors are from the U.S., Australia, and Europe, and that no global Airwallex executive is based in China. The exchange is a contested geopolitical and regulatory diligence signal around ownership, transaction structure, and executive location rather than a confirmed finding.
Paul Graham argues that AI model developers’ requests for government regulation may reflect genuine concern that models are becoming dangerous or unpredictable, rather than merely an ulterior motive. He identifies a competitive dynamic in which a company that slows down unilaterally risks falling behind, creating an incentive to seek rules that make competitors slow down too; he views model-makers inviting government involvement as an alarming signal.
- Jev’s launch is prompting an unusually rapid open-weight response: within days, multiple U.S. AI developers began shipping their own models, with Andrew Chen tracking 2–3 early efforts and expecting more alternatives within the week.
- Chen also expects Chinese open-weight alternatives for Jev, suggesting fast, cross-border diffusion and intensifying competition around newly launched models.
- Andrew Chen suggests that “Jev-like models” could accelerate native on-device AI functionality on mobile, while noting that strong LLMs remain difficult to run on phones because of slow memory bandwidth, the need for highly quantized MoE models, and power/heat constraints.
- He identifies fast, low-cost AI decision models for notifications, typing/texting, inboxes, calendars, and other in-app experiences as a mobile opportunity; next-generation NPUs are being aimed at modestly sized LLMs, while an alternative would be encoding an older but useful model directly into phone hardware.
- Martin Casado highlights a credibility gap in AI discourse between the “stochastic parrot” view and expectations of “unlimited, unstoppable intelligence growth,” arguing that advocacy of the former has weakened the credibility of those pushing back against the latter.
Martin Casado argues that advancing AI cyber capabilities could pressure the industry to build more secure systems throughout the stack, and highlights Hydra and Absestos as two examples of work he considers especially valuable.
Martin Casado praised Jev’s approach to system specialization through a “bitter lesson” lens, rather than forcing imperative programming into the deep-learning paradigm; he suggested this could become a path that combines the two approaches.
- Martin Casado reports that industry discussions of AI-related risk are becoming more nuanced: novel cyber risks exist but are viewed as manageable based on prior technology experience, while pragmatic, safe deployment at scale is considered feasible. He says near-term extinction fears are increasingly regarded within the industry as fringe and overplayed, although moderating views receive less public attention.
- AI-hardware funding attention may be a size/headline effect rather than a deal-count shift: hardware rounds require larger checks, and one commenter says the same number of hardware deals can generate 10× the headlines. Funding announcements are also described as lagging indicators that may reflect decisions from roughly 18 months earlier. Software can reach users and revenue before needing capital, whereas hardware may require about 18 months and substantial spending before demand is known.
- Physical hardware can be defensible and differentiated when the physical component is necessary to deliver value. However, it adds supply-chain, support, and inventory risk and entails longer, costlier R&D that can leave products obsolete or mismatched with market needs. Unstable supplier pricing and broad accelerator competition—including Ugreen and Anker—are cited as pressures on already-thin margins.
- Interface innovation is proposed as a software moat, with text messaging cited as an example, but the commenter cautions that an interface hosted by someone else is not a moat. Another commenter says AI has made startups easier to launch and ideas more competitive, arguing for specialist niches. AI infrastructure is portrayed as a field with experts possessing roughly 10 years of server-rack experience, while hardware interfaces face large-tech and well-funded-startup competition.
- Inception AI CEO Stefano Ermon is pursuing diffusion LLMs that generate tokens in parallel, challenging the sequential dependency in today’s LLMs where token 10 cannot be generated until token 9 exists. Sarah Guo characterizes Ermon’s work as rare architectural AI research and highlights latency as a key motivation.
A post linked by Arthur Mensch alleges that AI actors “bought ministers,” made the Élysée a political foothold, and obtained billions by promising sovereign AI; it frames this as a threat to public control ahead of 2027. Mensch’s accompanying post links to the allegation and calls the respondent “obsessive.”
- Google DeepMind executive Koray Kavukcuoglu says returning to the AI frontier is the organization’s overriding priority and expresses complete confidence in the team, citing its creativity, dedication, and access to Google’s full-stack resources.
- Nathan Benaich flags a product-integration gap: Gemini’s surfaces are not sufficiently connected, with users reportedly needing Chat/Claude alongside the Google Workspace CLI because Workspace and Gemini do not work seamlessly together; he contrasts this usability problem with Google’s likely API monetization.
Keith Rabois endorsed a framing that the primary AI risk is malicious human use rather than “evil AI,” with the proposed defense being for good actors to use the most advanced AI and accelerate its improvement.
Keith Rabois endorsed Databricks CEO Ali Ghodsi’s view that current AI existential risk is “close to zero” and that leaders should avoid publicly amplifying catastrophic scenarios without strong evidence, citing potential stress and mental-health harm from such messaging.