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 Atoms financing follow-up is notable for structure rather than disclosed terms. Travis Kalanick says he showed up to fundraise for several stealth companies, then merged separate entities with different investors and put them under one roof; Ben Horowitz says his conviction was fast because Kalanick “was still Travis.” The related a16z post describes Atoms as industrial AI spanning manufacturing, real estate, and logistics and calls the check Horowitz’s biggest ever. For investors, that is a founder/platform underwriting signal rather than a clean Seed/A pricing comp: the bet is on consolidating multiple physical-world opportunities behind one operating thesis.
2. Emerging Teams
Flue 2 is a founder-led bet that the agent runtime—not another model wrapper—is the product. Fred Schott, creator of Astro, whose company was acquired by Cloudflare in January, has released Flue 2 as the framework’s first stable version, built around React-style “Agent Hooks.” In Flue, an agent re-renders before every model call; its 16 built-in hooks can manage state, lifecycle events, tools, skills, and subagents. Schott’s central claim is that “there is no agent without a harness”: Flue is an opinionated layer on the open-source Pi harness, and the framework is explicitly open source and designed to run across hosts rather than optimize for one cloud. This is an early category signal, not yet a hosted-revenue story—the team says it is focused on the harness and is not currently planning a managed product.
Perseus has unusually strong founder–problem fit for air-gapped AI infrastructure. Its solo builder describes an MIT-licensed open-source system for environments where LLMs cannot connect to SaaS; the design combines local context, encrypted memory, an evidence layer, and MCP across agent clients. He brings 20-plus years of systems-architecture experience, including classified defense environments, is pursuing SBIR and government small-business opportunities, and says any monetization would come from air-gapped integrations and consulting rather than charging for the product. The investable question is therefore less “can this be a SaaS?” than whether the open-source project can become a trusted integration wedge into regulated deployments.
DISPELDA is a diligence caution, not an investable product yet. Its solo builder has a 200 Hz STM32/MPU6050 proof of concept comparing a conventional Madgwick filter with a Liquid Neural Network, and explicitly says it is not yet a sellable product. A technical response argues that attitude estimation is solved, GPS-denied position requires aiding sensors, and the current IMU can drift several meters within seconds; it also puts defense sales cycles at three to five years and says buyers want qualified modules or demonstrated capability. The immediate diligence bar is a narrow vertical, ground-truth testing, and reproducible ATE/RPE metrics—not a generic “drones, robots, or machines” platform.
At the smaller-product end, GainFrame’s builder reports $1,500 in MRR and roughly $4,000 in revenue for the month. It is a self-reported but concrete traction signal amid a much larger volume of AI-built prototypes.
3. AI & Tech Breakthroughs
BDH-CQ moves in-context learning toward recurrent latent computation. The paper’s abstract describes a 150M-parameter system whose recurrent memory is updated by demonstrations at inference time, then solves the query through iterative computation in a high-dimensional latent space without verbalizing intermediate reasoning. It reports 29.5% pass@2 on ARC-AGI-1 at a computed cost of $0.0007 per task, claiming a new cost–accuracy Pareto point. The result is a research lead rather than a settled benchmark conclusion: the cost is explicitly computed, and the frontier claim still needs independent reproduction.
Model routing is becoming a harness-level systems problem. Factory’s Router reports more than two months of production use, with aggregate cost 58% below frontier-pinned sessions, median-session savings of 76%, more than nine in ten sessions saving at least half, matching across eight production-quality measures, and median turn latency falling from 81 seconds to 49. Long sessions make the placement matter: the analysis models cache-blind gateway switching at 2.12x to 2.37x an all-frontier baseline on long turn ranges, versus 0.19x to 0.28x for cache-aware routing; its 423-turn missions show 37.8% savings. Jerry Liu’s accompanying formulation is the right architecture test: each task is solved by a co-optimized mixture of model and harness, while gateway-only routing loses the broader session context. The opportunity is infrastructure that owns task decomposition, cache state, evaluation, and model choice together.
4. Market Signals
AI demand is bifurcating rather than diffusing evenly. a16z says the top 1% of AI spenders spend more than 600 times as much as the median company. Gamma provides a useful operating contrast: SaaStr reports $100M ARR, 50 million users, 600,000 paying subscribers, 50 employees, profitability, and no sales team for most of the journey. Its CEO’s warning is that self-serve growth can generate so much signal that a company stops making decisions; Gamma still has not meaningfully expanded its self-serve base, while the company describes the broader AI market as fragmented, with most users still taking a first step and APIs becoming a real user class. Underwrite the conversion from broad usage to expansion, not user counts alone.
The compute-centralization debate is becoming a competition and policy variable. Dario Amodei argues that scaling and compute access structurally concentrate AI power; he says Anthropic’s preferred rules would slow frontier labs while advantaging smaller competitors, including through more rigorous testing of frontier models. Amjad Masad counters that algorithmic and hardware efficiency could make advanced capability far less data-center-bound, and that scaling laws are empirical relationships rather than laws of physics. The diligence implication is to separate a company’s exposure to scarce chips and regulation from its ability to benefit if algorithmic efficiency changes the cost curve.
UK garden-leave rules are being framed as a startup-formation bottleneck. A UK AI founder thread contrasts California’s ability to start a company immediately after leaving Google with one-year restrictions for senior UK researchers and six months for junior researchers; it says some contracts are imposed at promotion and prevent researchers from starting companies, hiring, or competing during the leave. Whether or not policy changes, this is a concrete talent-mobility variable for European AI venture formation.
Agentic distribution will require optimizing for non-human buyers as well as human PLG. Matt Swulinski argues that PLG companies should understand how agents research products and select tools and APIs, while using the e-commerce acquisition model—Meta, Google, lifecycle—and treating distribution as the moat. That makes agent discoverability and channel execution part of product diligence, rather than a post-PMF marketing task.
5. Worth Your Time
- Watch — Matt Swulinski on building a $100M growth engine. The most useful segment connects agent selection of tools and APIs to the familiar PLG funnel, then makes the harder claim that distribution—not code—is the durable SaaS moat.
Read — Why model routing must be in the harness. Read it for the production cost/latency results and the cache-blindness failure mode in long-running sessions.
Read — React for Agents: Flue 2. It is the clearest current explanation of why composability, runtime state, and a built-in harness are becoming the agent-framework battleground.
Read — BDH-CQ: In-Context Learning with Recurrent Latent Reasoning. The abstract is worth reading as a low-cost reasoning hypothesis, with the important qualification that its cost and frontier claims remain self-reported.
Direct answer from the only supplied source (arXiv:2608.09888 abstract page): the abstract claims BDH-CQ combines in-context learning with recurrent latent reasoning; inputs presented at inference time continuously update the model's recurrent memory, and a query is solved through iterative computation in a high-dimensional latent space without verbalizing intermediate reasoning . It further reports that a 150M-parameter configuration reaches 29.5% pass@2 on the public ARC-AGI-1 evaluation set at a computed inference cost of $0.0007 per task, and claims this operating point breaks the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency . The evaluation also used controlled ARC-like interventions to study what is learned from demonstrations, how consistently an inferred transformation is applied, and which concepts remain difficult .
Verification limits: the bundle contains only the abstract and metadata; no full-text methods, cost-computation basis, frontier baseline, or experimental tables are supplied, so the architecture and all numbers above are self-reported and cannot be independently verified here . The cost is expressly a 'computed inference cost', not a measured or standardized cost . The only code/data pointer is the GitHub link in the comments: https://github.com/pathwaycom/arc-task-gen.
- Matt Swolinski, who scaled growth at Superhuman and was the only growth executor at Whisper on a $3–5M budget, recently joined Victor as growth lead; he argues PLG leaders for humans will be best positioned for agentic buying decisions and bets that in ~3 years companies will run like boards of directors with agents doing ~80% of execution (today ~80% manual) .
- Victor claims to be first to deliver on the 'AI employee' premise — positioned around knowledge work to differentiate from ChatGPT and other AI players, launched with the horizontal tagline 'AI employee for everyone' then verticalizes by ICP (agencies, e-com, SMBs first); it is token/credit-based, and Swolinski says 8-person teams spend $15–20k/month on it and it replaced all their hiring .
- Growth playbook: SaaS should follow the e-commerce model — every cent tied to purchase/add-to-cart — applied at Superhuman then WhisperFlow; paid is the fastest way to validate PLG (a ~$100k budget is reasonable after a $3–5M seed raise) . Core acquisition channels are Meta, Google, Lifecycle, enough to reach $1–10M ARR; at a $100k Meta budget he says you need 400–500 new creatives/month via UGC creator programs, since Meta's Andromeda update made creative the targeting .
- SaaS lacks e-commerce-grade attribution: no out-of-the-box Triple Whale/Elevar equivalent, stacks are homegrown (DB, Clickhouse, Hex), and ~90% of companies don't set up measurement before spending — an open startup gap . For usage-based AI SaaS, token/inference costs (Anthropic, Modal, etc.) are a primary cost line; model LTV/gross profit and track every cost item because Meta/Google aren't built for high-variance B2B usage products .
- Referral lessons: align referrals to usage limits/paywall moments (Superhuman's give-a-month/get-a-month; Whisper at word limits); Victor pays ~20% revenue share for company referrals and CPM credits for LinkedIn posts, and trial credits should count in fully loaded CAC .
- Affiliate is the most underappreciated growth channel: at Victor it drives 10–15% of monthly acquisition via 10–15% revenue share, with top affiliates earning $20–30k/month; he has yet to meet a SaaS growth leader who says X/Twitter ads work .
- Whisper growth: Google Ads was the main driver (non-branded search, PMAX, YouTube), with 30s–1min YouTube explainers generating hundreds of millions of impressions; Swolinski 5x'd the budget in a month to find ceilings/incrementality. He calls Whisper a Polish startup that raised a big Series A, and notes its original UGC viral program was built by Tobin, now at Polymarket .
- AEO is a new channel: YouTube reviews and external narrative get cited by ChatGPT; pumping 100–200 AI-slop pages/week is the wrong move; PR's value is now mostly AI citations. Full-AI video creative is 'slop' and algorithms denigrate it — use AI for variations only .
- Talent shift: the bar moved from deep channel specialists to AI-native systems thinkers; <1% of candidates clear it, and teams fail at becoming AI-native when people can't map their job's inputs/outputs. Swolinski's hot take: fire marketers who aren't systems thinkers; one strong person with agents replaces 3–5 person teams .
- Investor workflow signal: Harry Stebbings' fund uses an AI call grader that scores every IC call on ~5 variables, stack-ranks deals, and fully delegates prioritization to the AI ranking; he says it has never been wrong across 1,000+ companies in 12 weeks .
- A solo founder is building DISPELDA, an embedded decision-intelligence layer intended to keep drones/robots/machines operating when GPS or communications are jammed/lost; prototype is an STM32 Nucleo F446RE + MPU6050 running a 200 Hz malloc-free static-memory C loop that compares Madgwick filtering against a Liquid Neural Network. He is solo, without a degree, team, or funding, and calls it a POC, not yet an MVP .
- A commenter claims to have already built a similar EU AI Act-compatible system in 5 months, with a POC and 3 MVPs, a >€100B market by 2030, and recent benchmark wins over the incumbents he says dominate—ST's X-CUBE-AI and Google's TFLite Micro—and is seeking investors; claims to be solo, without degree/team/funding .
- The GPS-denied / alternative-PNT market has been hot since 2022 (jamming, spoofing), but it is defense-heavy: 3-5 year sales cycles, clearances, export and sovereignty constraints, and buyers want qualified modules or demonstrated capability, not a lone developer's software layer . Commenters question the novelty: autonomous drones switching to inertial navigation when GPS is lost/jammed have existed for a long time , though defense companies with large R&D budgets are interested in doing it well .
- Technical caveats are strong: IMU attitude estimation has been solved for 20 years; the real GPS-denied problem is position, and a $2 MPU6050 drifts meters in seconds; a Liquid Neural Network will at best tie Madgwick at 100x CPU cost with no certification path; without aiding sensors (VIO, LiDAR, terrain) an 'embedded decision layer' is window dressing . A hardware-founder commenter says the stack is solid but not the moat, calling the problem systems-level sensor fusion that big defense players are pouring money into . Existing systems already cover GPS-denied modes: PX4/ArduPilot EKF2/EKF3, VINS-Fusion/OpenVINS/ORB-SLAM3, ModalAI VOXL2, Skydio, Shield AI, Emesent/Exyn .
- Vertical/business guidance: 'drones, robots or other machines' is not a market; pick one vertical (industrial inspection, agriculture, research), with indoor industrial (mines, sewers, warehouses, tank inspection) a better first market than defense . Credibility requires ground-truth benchmarks (ATE/RPE), and a realistic solo path is subcontracting/integration with a drone OEM . In France, non-dilutive options include BPI French Tech/i-Lab, DGA RAPID, ASTRID, AID, and NATO's DIANA .
- Seed round closed for a new AI venture; founder is "excited to be back in the game" after previously working on image models and reviewing fundamentals .
- Compute ramp: started with two 8×B200 systems, later acquired 64 B300s, then lost all GPUs ("GPU poor") with replacements delayed by networking issues; subsequently locked down "much greater quantities of compute" and studied the "frontier of the datacenter industry" .
- Technical focus: "autonomous ai scientist" running optimizations; validated a "basic RLVR post-training stack"; got a key research piece working and needs to scale it .
- Team building: first hire made; hiring #2 for post-training (RLVR/OPSD/etc.) or low-level model optimization; team at 3 .
- Sentiment: Silicon Valley "very much alive" with people generous to help/share .
Travis Kalanick is publicly re-emerging with Atoms, an industrial AI company that frames manufacturing, real estate, and logistics as the "CPU, storage, and network of the physical world" . Kalanick says he never actually left — he spent eight years and thousands of employees working across multiple industries without posting on LinkedIn . The a16z investment is described as "the biggest check Ben Horowitz has ever written" . During fundraising, Kalanick had several stealth companies at once and had to merge separate entities with different investors; a16z wanted everything under one roof . He compared this consolidation to Elon Musk managing multiple companies and "starting to put the pieces back together" . Ben Horowitz said his conviction was fast because "he was still Travis" .
- AI cost opacity and cost deflation: Amazon spent ~$1.8M on a Claude project that ran five months unnoticed; a senior employee said "It's difficult to figure out how much anything [AI-related] costs" . The author's OpenClaw agent blew through ~$500/day before an audit found older, higher-tier models like Opus 4.5 running unnecessarily; after switching to cheaper models it cost ~$6/day vs ~$50–60/day before, raising the question of whether ~$494/day has "disappeared from genAI revenue" . A price war between Anthropic and OpenAI in response to Chinese advances drove costs down .
- Enterprise AI spend and model pricing: Ramp reports July median AI spend per employee of $7,400 for the top 1% of businesses, $650 for the top 10%, and $11.95 for the median firm; Opus 5 is driving a large chunk of Anthropic's revenue growth . SpaceXAI is picking up a pricing fight with Grok 4.6, which undercuts top rivals by more than 60% .
- Infrastructure and biotech signals: Each additional data center within 25 miles raises a neighboring town's bond spread by ~10 bps, and a town with the average number of nearby data centers issues ~$34M more debt; after state data-center incentives, school transfers fall ~$673 per student . Microsoft teams are exploring regenerative, biomimicry-based data center designs . Anthropic is hiring a chip design team . Other signals: protein logic gates that can trigger cancer cells' self-destruction , an AI that designed 16 synthetic viruses better at killing E. coli , and a hidden prompt injection in a court filing asking AI to side with the plaintiff .
Paul Graham spent 80% of office hours reassuring founders who were doing well; he says if they weren't doing well he'd be telling them what to worry about . He argues investors judge startups partly by their confidence, so telling startups when they're not making mistakes matters — founders with reason to be confident should be .
Martin Casado recommends a Latent Space podcast with Alex Krentsel on RSI (recursive self-improvement) for agents, calling Krentsel "one of the clearer thinkers on the topic" and "doing some amazing work on it" . The episode is available on YouTube .
Garry Tan (President & CEO of Y Combinator) posted a "whitepill timeline cleanser" citing record year-on-year crop production as evidence that progress and abundance come from markets and technology working together over millennia, linking to Hannah Ritchie's USDA-based global food tracking .
LlamaIndex — CEO/cofounder Jerry Liu — released ExtractBench, billed as the most comprehensive benchmark for information extraction from complex enterprise documents, built by its applied research team . It evaluates 14 systems (frontier VLMs, coding agents, specialized extraction APIs) on 370 enterprise documents: 4,869 pages across 67 document types and finance, energy, gov, and auto domains . The harness is 100% deterministic with no LLM judges, scoring value accuracy, long-record completeness, spatial grounding, and per-page cost .
Headline finding: commercial VLMs collapse below 35% recall on documents past 50 pages due to silent list truncation — they keep high precision but drop most table rows ; the post also frames viable per-page token cost (not paying upward of $1/page in tokens) and exact spatial citations for auditability as production requirements for scaling to millions of docs .
In tandem, LlamaIndex launched Agentic Plus, a new LlamaParse extraction tier debuting at #1 on the ExtractBench leaderboard with 95.6% value accuracy at less than a third the cost of the closest peer . It targets massive field extraction (10k–100k+ fields) from long documents (100–500 pages), tested on the FTX creditor matrix (75k fields, 114 pages) .
Benchmark artifacts (blog, GitHub, HuggingFace) are public, with LlamaIndex planning to evolve both the benchmark and the extraction harness . Together the posts position specialized extraction tooling as necessary where frontier models fail on long-document production workloads .
a16z spotlights Dario Amodei's recent remark that he was concerned AI employees were beginning to prioritize money over mission, adding "He may be on to something..." . It links this to its Charts of the Week edition, whose URL slug reads "head-in-the-neoclouds," pointing to neocloud infrastructure as a focus area .
- Jason disclosed a $750k personal bet on Bittensor, noting it has been ±25% since entry; he estimates a 10% chance of 100–500x and is "ok going to zero" .
- Jason predicts Bittensor-native models will consistently rank top 20 vs contemporaries over 12 months at below-average cost; within 24 months he sees a non-zero chance of consistent top-ten placement and an accelerating amount of network compute .
- He urges adding extra compute to Bittensor for "AI sovereignty" and frames buying bittensor:native as a 10–100x lottery ticket .
- Supporting signal: Teutonic-I, a 10B pretrained model trained on Subnet 3 (first in a release sequence), reportedly outperforms every other decentralized pretrained model, including 40–70B models 7x its size .
- Fred Schott, creator of Astro (his company acquired by Cloudflare in January), released Flue 2, its first stable release, built on React-style "Agent Hooks" . Agents are JS functions that re-render before every model call; hooks let agents manage state, listen to lifecycle events, and attach resources/capabilities at runtime, with 16 built-in hooks (useSkill(), useTool(), useSubagent()) . Schott positions Flue as "the React for agents" after realizing composability, not routing, is the core need .
- Flue's core thesis: an agent must have a harness — "There is no agent without a harness" — and Flue is built on Pi, an open-source minimal harness, as an opinionated take on it . Schott names Vercel's eve as the closest competitor (also harness-first), and contrasts older "OG agent frameworks" (Vercel AI SDK, Cloudflare Agents SDK, Mastra) that are adding harnesses as a feature .
- Flue is "an open source framework for every host," with host portability a defining principle, unlike Vercel's eve which is optimized for Vercel's platform .
- On meta-harnesses (e.g., Databricks' Omnigent, Exo), Schott argues one API across harnesses would muddle Flue's story — "the framework and the harness are very intertwined" .
- Competitive signal: LangChain launched Managed Deep Agents as a hosted agent platform; Schott says Flue is not planning a managed product yet, focusing on "building the best harness" .
- Harry Stebbings names growth leader Matt Swulinski top-3 among the 100 growth leaders he has interviewed (alongside Alex Schultz and Brian Hale); Swulinski scaled Wispr Flow to $100M+ ARR and built its UGC machine, and scaled Superhuman to $50M ARR .
- Stebbings argues the e-commerce playbook is the right model for modern SaaS as distribution becomes a critical moat in a crowded AI market — deploying UGC creators, constantly testing new creative, and diversifying channels . Paid acquisition is the fastest way to validate a PLG funnel, letting teams test positioning, messaging, and conversion within a single week .
- Swulinski's core guidance: scaling to $10M ARR requires only three channels — video intent on Meta, search intent on Google, and lifecycle retention via email and SMS — while running ten channels poorly ruins the acquisition engine .
- Stebbings predicts that within three years lean human teams will operate like boards of directors, spending 20% of time on strategy while autonomous AI agents handle 80% of operational execution; marketers stuck on manual tasks are becoming replaceable, and high-performing teams need systems thinkers who build self-improving AI workflows with 10x leverage .
a16z highlights a wide AI adoption gap: the top 1% of AI spenders spend more than 600x as much as the median company . The stat is featured in a16z's Charts of the Week .
Nathan Benaich calls total air defense 'a non-negotiable in today's world,' acknowledges it is 'super expensive and super complicated to build,' and urges 'more neoprimes' to work on it .
AI product Instinct — an AI that navigates websites and executes transactions on the user's behalf — drew strong endorsements: Sarah Guo said it is "amazing, even better than I thought" , and Elliot Comite called it "my most 'holy shit' AI product since Wisprflow" .
@nathanbenaich captures a sentiment about Europe's AI scene: Chinese companies ("the chinese uber eats") are shipping better models than European ones, and Europe's best hope prefers to work for those companies rather than build locally ("serve them rather than cook") .
Nathan Benaich reposted a satirical thread mocking AI founder/VC clichés, calling it a 'pretty accurate zeitgeist read' . The parody strings together tropes: leaving OpenAI for Anthropic (and reverse), YC acceptance/rejection, 'taste is the moat', pivoting focus from RL environments to data to 'personalizable AI for science and enterprise', and spending $10M on Claude tokens without building anything . Signal: skepticism toward hype-driven AI founder narratives and buzzword pivots.
Matt Swulinski, a top growth leader, scaled Wispr Flow to over $100M ARR and Superhuman to $50M ARR . His core thesis: SaaS should adopt the e-commerce playbook where every dollar ties to a purchase or conversion, and "distribution is the only moat" — using UGC creators, constant creative testing, and channel diversification . He advises that scaling to $10M ARR requires mastering only three channels (Meta video, Google search, email/SMS), producing 400–500 new creatives per month, using paid acquisition for rapid PLG validation, and running holdout tests to measure spend incrementality . He predicts that within three years, companies will operate like boards of directors, with AI agents handling 80% of operational execution and humans 20% strategy; marketers must become systems thinkers who build self-improving AI workflows .