We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
The financing signal this period is a rotation toward usage-linked revenue and security exposure. A SaaStr market review reports strong performance from CrowdStrike, Twilio, Snowflake, and Datadog, arguing that each either bills by consumption or sells into budgets expanded by AI; it also notes the downside of concentration when a large customer cuts usage. The same review contrasts that with declines in seat-priced HubSpot, monday.com, and Figma; Figma’s AI credits increased inference costs while its pricing remained per editor. Its conclusion is that the billing unit—not growth alone—did much of the sorting as CIOs reallocate budgets toward tokens. For early-stage deals, the diligence question is whether revenue and gross margin scale with customer usage or remain exposed to seat-budget optimization.
Seed robotics capital is now confronting an operating bottleneck. An unnamed seed-stage robotics startup reports $4.2 million in funding and a technically dense team of three Robotics PhDs, a BME PhD, a MechE PhD, two robotics MS holders, and a BME MS holder; its advisers include technical executives associated with billions in exits. The CEO is also acting as CTO, and the company says it needs a systems engineer or technology lead to set priorities and review acceptance criteria. A senior BME systems engineer reportedly rejected an offer of $130,000 plus 0.5% equity as too low. The investment issue is executional: technical pedigree is not yet translating into a repeatable decision and quality-control layer that can scale beyond the founder.
2. Emerging Teams
A vertical AI operating tool is starting from measurable margin leakage rather than a generic chatbot. A 40-year-old family commercial-cleaning business with roughly 40 accounts connected payroll hours, billing, expenses, wages, overhead, and pricing data, then found that some long-running accounts were barely profitable or loss-making. The founder reports that money retained after costs rose from roughly $30,000–$40,000 on $3 million of annual revenue to roughly $80,000–$90,000 without doubling the customer base. The proposed product would use each operator’s own financial history to flag weak accounts, recommend low/base/high-margin bids, and compare estimates with actual hours and costs. It is still a validation project, not a launched company. A few owners have already reached out; the decisive test is whether another operator can obtain useful results without the founder explaining the system personally.
A niche data/IQ platform has strong repeat usage, but its buyer is not the crowd. The founder reports 150 forecasting votes in seven days, with 130 repeat voters and 60 voting three times; separately, a B2B user said they would pay for three quarterly opportunities unavailable through LinkedIn Sales Navigator or Crunchbase, while a Big Four user has used the due-diligence feature twice monthly for three months. The more actionable signal is the specific paid outcome and recurring professional workflow: feedback warns that forecasting, opportunity sourcing, and due diligence may be three different products, and that the recurring due-diligence user is a better monetization test than the vote count.
3. AI & Tech Breakthroughs
Inference efficiency is moving into the runtime layer. A prototype Sliding Window Attention layer applies attention sinks plus a recent-token window to pretrained Hugging Face models without retraining. In a Qwen2.5-7B experiment, reported KV-cache memory stayed near 3.5 MB at both 16K and 32K context, versus about 923 MB and 1.84 GB for full attention; full attention ran out of memory at 64K while the bounded version remained usable. TPOT fell from roughly 38.4 ms to 30.5 ms at 16K in that setup. The trade-off is material: tasks needing information outside the active window can degrade, and the author is still separating inherent limitations from implementation or model effects.
Uncertainty estimation may become cheap enough for serving, but agreement can still be confidently wrong. A released benchmark reports that normalized exact-match entropy reached 0.889 AUROC on GSM8K across 7B–27B models, matching neural semantic entropy while running in under 2 ms on CPU; the neural approach reportedly takes 100-plus seconds on CPU and adds substantial VRAM overhead. On 120B models, however, the authors describe “Confident Mode Collapse”: identical incorrect answers across samples drove AUROC down to 0.091. These are benchmark results, not evidence of a solved reliability problem.
THREADS is a useful hybrid-system pattern: exact memory below a neural model. The open-source prototype accepts structured facts and relationships, tracks temporal changes, retractions, contradictions, and provenance, and reports exact results on tests including a 200,000-hop chain, a 128-hop query amid 1 million irrelevant events, 5,000/5,000 historical queries, and 40,000/40,000 ambiguity/contradiction cases. Its author explicitly says it is not a replacement for transformers, databases, or SMT solvers and does not understand arbitrary English; the proposed role is an exact memory/reasoning layer beneath a language model.
4. Market Signals
The security-response window is collapsing at the same time that agent governance is lagging. a16z reports that roughly 87% of software bugs hackers exploit are attacked on or before the day the bug becomes public, up from 23% in 2020. Separately, a practitioner post argues that enterprise security teams are shipping their own threat-detection and triage agents faster than governance teams can track; it claims shadow agents are already common, with tool access, data connections, and runtime behavior rarely monitored centrally, while internally built agents may sit outside the 80-plus regulatory and security frameworks enterprises use. The investable control-plane opportunity is therefore not just better model detection, but faster vulnerability intelligence, agent observability, and enforceable permissions.
AI infrastructure is creating real labor demand while increasing land and permitting risk. An a16z post relaying Goldman attributes more than 300,000 construction jobs to the AI buildout since 2022, including roughly 75,000 in the past year; electrician and HVAC trades are growing about 2% annually, roughly twice the rate of construction overall. At the same time, a current post says rural land prices are rising as some owners sell into data-center development while others resist it. At a July protest in Lubbock, Texas, Agriculture Commissioner Sid Miller was quoted saying developers were taking prime farmland and sometimes offering up to 10 times its value. Land acquisition, grid access, permitting, and local support now belong in the deployment case alongside power and chips.
Coding capability and coding economics are sending different signals. Martin Casado says recent models show a meaningful step in computer use but no comparable improvement in coding for his work, and suspects further model advances for high-skilled development tasks may no longer make economic sense even as cybersecurity improves. A separate builder reports that the most capable coding models sit behind expensive tiers, consume more tokens, and can cost more than $100 to improve a single function, potentially giving venture-backed and corporate teams an advantage over budget-constrained developers. The cost report is anecdotal, but it supports a narrower underwriting question: where does agentic coding create enough accepted output to justify frontier-model spend?
5. Worth Your Time
- Watch — Why companies are becoming a series of loops | Anish Acharya. The useful segment treats agents as loops across coding, growth, sales, support, and legal, but says loops plateau at local maxima and still require human out-of-distribution judgment. The same conversation offers a practical model mix: cheaper open-weight models for bounded tasks and expensive frontier models for high-upside work such as research and engineering.
- Read — The “Owning” Phase of AI, Part 1. The essay is a useful screening framework for vertical-model companies: proprietary non-scrapable data, an API-cost tipping point, low tolerance for mistakes, workflow ownership, and privacy mandates. It cautions that a custom-model announcement is not enough; the investable combination is captive data, captive distribution, and sufficient scale.
- A16Z consumer-AI thesis: Anish Acharya is an A16Z general partner and product founder whose companies Social Deck and Snowball were acquired by Google and Credit Karma, respectively; he later led product and the broader consumer product/credit-card business at Credit Karma. He says A16Z’s bar has shifted from rejecting overly ambitious ideas to avoiding ideas that are too small, and identifies three major consumer-AI opportunities: coding agents/general-purpose problem solving, personal agents, and entertainment, creative, and companionship products. A16Z’s cited portfolio example Wabi is a mini-app platform where users can create, consume, and share apps.
- Emerging AI architecture and product paradigm: Acharya expects companies to organize work as cascading agent loops across coding, growth, sales, support, and other functions, but says loops reach local maxima and still require human intuition for out-of-distribution decisions. He predicts a split between cheaper open-weight, narrow-task models for bounded-upside functions and expensive frontier models for high-upside or harder-to-evaluate work such as research, engineering, sales, and support; he expects multiple specialized model families rather than a single winner.
- Competitive signals for early-stage investors: Acharya argues that AI-company moats are often discovered through shipping rather than designed upfront; Cursor’s product wedge generated reasoning traces that enabled proprietary model training, while momentum, craft, and growing engagement can justify investment even before a clear durability story exists. Distribution is increasingly driven by organic word of mouth because existing networks resist new network formation, but startups may benefit from pursuing product directions incumbents avoid; he also reports unusually high willingness to pay, including roughly $200/month consumer plans and million-dollar enterprise ACVs. The main caveat is that AI adoption may reshuffle near-term winners without radically changing industries whose economics are not intelligence-bound.
- World Labs’ Atlas is presented by co-founders Justin Johnson, Fei-Fei Li, and Ben Mildenhall, alongside a16z’s Martin Casado, as a world model for spatial intelligence built on new-view prediction; it unifies pixel generation and 3D reconstruction in one model.
- Atlas natively handles text, images, videos, and camera poses, using 3D as a native modality and viewpoint estimation to unify reconstruction with generation.
- The post claims the approach reduces digital 3D-capture requirements by 50–100x: a single room that previously required 100–300 photos can be captured from three. The discussion also frames robotics as data-bottlenecked rather than chip-bottlenecked and positions new-view prediction as potentially AI-complete.
Sam Altman amplified Jakub’s “An Alien Mind” post. The post discusses the state of AI, expresses concern about the next few years, and argues that choices are needed to keep the future in humanity’s hands.
Paul Graham highlights an unnamed startup growing around 8× annually, although its revenue graph understated the strength of that performance; no company, stage, funding, or team details are provided.
Founder advantage heuristic: founders who have experienced a company’s weakness may retain a stronger user-delight orientation, while hired CEOs may be more likely to take the company’s power for granted.
A quoted AI outlook shared by Nathan Benaich forecasts that current AI progress could continue into recursive self-improvement; if development stays on its current path, systems in the next few years could deliver capability jumps of equal or greater magnitude and increasingly contribute to their own development. This is a forward-looking capability thesis rather than a startup funding or product announcement.
- The AI buildout is creating substantial demand for infrastructure trades: Goldman attributes more than 300,000 construction jobs to it since 2022, including approximately 75,000 in the past year; electrician and HVAC-related trades are growing about 2% annually, roughly twice the pace of construction overall.
- Data-center construction is accelerating, with spending rising by more than $25 billion in six months—about equal to the increase over the prior two years—and construction job openings recovering from roughly 200,000 to above 300,000.
- Khosla’s early-stage hiring heuristic is to hire for the business’s key risks rather than simply filling open roles; he argues that industry experts may know the previous version of the world and that the first 10 hires shape the next 500.
- Founders should remain obstinate on vision but flexible on tactics, avoid attractive revenue detours that do not advance the core mission, and deliberately spend to test assumptions they may not realize they hold.
- Burn should be managed for optionality: delay braking while experimenting, but retain the ability to cut marketing spend or contractors quickly when conditions change.
The post flags perceived saturation in coding-model progress: while recent models are described as “amazing” and computer-use capabilities show a meaningful step forward, the author reports no comparable improvement in coding for their own work, with the caveat “Maybe skill issue :)”.
The software vulnerability-response window is collapsing: a16z reports that ~87% of bugs hackers exploit are attacked on or before the day they become public, up from 23% in 2020. This points to rising demand for faster vulnerability intelligence, detection, and patching infrastructure.
David Sacks alleged that social-media influencers were offered money to promote AI “doomer” messages, describing the efforts as well organized and well funded; this is an investor-sentiment signal about contested AI narratives, not evidence independently verifying the campaigns.
Martin Casado argues that the economics may be reaching a point where further model advances for high-skilled developer tasks no longer make sense; he sees clearer capability gains in cybersecurity than in raw software engineering.
Bindu Reddy signals an expected near-term AI release sequence: Fable 5.2, followed by Gemini 4.0; she describes “Bel from OpenAI” as hopeful rather than confirmed. The post provides no funding, team, technical, or precise timing details beyond “shortly.”
Harry Stebbings reports switching from Google to ChatGPT 24 months ago and from ChatGPT to Instinct this weekend; he says each transition sharply increased his search volume and expanded the tasks he asks, providing an anecdotal signal of changing AI-mediated search behavior.
Bindu Reddy framed Meta as a cautionary example of AI strategy, alleging that it spent $5B on Anthropic, saw its stock decline over the prior year, conducted layoffs, faced employee rebellion, and “bench-maxxed Muse Spark 1.3.”
Bindu Reddy flags a potential agent-mediated workplace paradigm, describing “zombie employees” who rely on AI to record meetings, handle email and Slack responses, and produce decks, documents, or PRs; she argues that agents may increasingly communicate directly while human oversight becomes largely nominal.
Cliff Weitzman argues that human-computer interaction is moving toward voice: Google and ChatGPT demonstrated the power of a text box and button, while future users may converse continuously with computers, phones, and wearables and use screens less. This points to voice-first interfaces as an emerging product paradigm.
- Replit CEO Amjad Masad said cybersecurity may be the most important technology area to get right, linking to an external post as supporting context. The linked post claims its author audited approximately 120 web and mobile apps over two months and found no high- or critical-severity vulnerability in an app built with Replit. This is an anecdotal signal relevant to security in AI-assisted software development, not a formal benchmark.
- AI coding-agent switching costs may come from accumulated user context—skills, routines, system instructions, project setup, conversation history, and persistent memory—rather than model access alone; externally maintained project documentation can transfer some context but loses conversational nuance. The proposed product opportunity is a “self-improving context graph” that efficiently supplies the right context to the agent.
- This moat is fragile when model quality improves: the author ported skills and conversational context to Codex after trying GPT-6 Astra and planned to use Codex as the daily driver, suggesting early adopters may face near-zero switching costs when a new model is sufficiently better.
An anecdotal signal for a possible new game-creation workflow: the post says that “everyone's building some video game with astra” and that the author “made gta 6 at home,” alongside a video attachment.
