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 visible period is more useful as a capital-underwriting signal than as a new-round signal. Suhail says he checked 13 providers for a single NVIDIA B200/B200s node and found zero availability; he expects GPU prices to reach $6.50–7 per GPU-hour and inference to become more expensive. For model and inference startups, capacity access and cost pass-through now belong in the financing conversation alongside benchmark quality.
2. Emerging Teams
Hermes Project Autopilot is a technically specific early-stage wedge in agent reliability. Its builder packages autonomous repository work into a durable mission contract with exact verification commands, autonomy levels, clean-repository and path/network gates, isolated worktrees, controller/planner/executor/verifier roles, checkpoints, a hash-chained evidence ledger, and approval before commits. The verifier is read-only, and v1 does not push, merge, deploy, or restart services. The project reports 18/18 deterministic safety scenarios passed, zero safety escapes, 135 focused integration tests, exact replay of the delivered Git tree, and a repository containing the patch series, provenance manifests, security documentation, and CI. The important next test is whether the evidence model generalizes: the builder defines “false success” as a worker claiming completion while checks failed, evidence is missing or stale, or the final repository state does not match the contract, and plans to publish seeded cases and raw results.
Adima AI shows a smaller but concrete privacy-first distribution signal. An independent developer says the local image restorer/upscaler has reached 30,000+ Android installs and 2,500+ Windows installs after two years of development. Its v1.2.0 Face Boost update came from repeated user requests and adds batch face restoration, 4×–16× upscaling, and fully local processing with no cloud upload. It is not a financing milestone, but it is evidence that a narrow on-device utility can accumulate usage while avoiding subscription and privacy objections attached to cloud alternatives.
3. AI & Tech Breakthroughs
The open-model ecosystem is broadening rather than consolidating. Interconnects’ current roundup says more organizations are still investing hundreds of millions to billions in training while releasing models openly, and argues that rising token demand is making “token machines” an attractive path to value. The release set spans Thinking Machines’ 975B-A41B multimodal Inkling and smaller fine-tuning-oriented version, Poolside’s 118B-A8B Laguna-S-2.1 that fits on a DGX Spark with published evaluation trajectories, and Korean startup Motif’s 314B-A13B preview with GDLA and mHC architectural changes. The commercial question is shifting from “who has the one winning model?” to who captures value through licensing, fine-tuning, serving, and distribution. Licensing is part of that competition: Kimi K3’s noncommercial license requires inference and fine-tuning providers to sign commercial agreements, which the roundup says could create future policy exposure for U.S. companies.
Qwen3.8-Max is the period’s sharpest new open-model claim. Alibaba’s Qwen account introduced it as “a new bar for coding and cowork.” Bindu Reddy says the model will be open-sourced this week, describes it as a 2.4T-parameter model “almost certainly Sonnet class or better,” and lists pricing of $2 per million input tokens, $6 per million output tokens, and $0.25 per million cached tokens. The pricing is directly stated in her post; the capability comparison remains an attributed claim until independent evaluations arrive.
Embodied-control research is moving beyond pure motion imitation. A Two Minute Papers transcript describes a controller trained in parallel to imitate human movement and solve new obstacle courses, using only 19 clips totaling about 30 seconds of parkour data; a learned judge scores whether generated movement looks both human and appropriate to the obstacle. The system is shown handling unseen obstacle arrangements, but the caveats are material: longer levels have only about 40% success, and unnatural recovery motions remain possible.
4. Market Signals
Inference, not training, is becoming the center of infrastructure underwriting. An Investing in AI analysis projects inference to account for roughly 80% of the neocloud market by 2030 and distinguishes it from training as recurring operating expense optimized continuously for cost and latency. It argues that neoclouds currently win on scarcity and deployment speed but must move up into managed inference, orchestration, fine-tuning, routing, or other software layers before scarcity fades. The same analysis identifies power and interconnects as binding constraints and tells investors to examine software/managed-services revenue, customer concentration, and whether contracted power outlasts hardware depreciation—not just GPU count.
Safe agent workflows are an infrastructure problem, not an MCP or prompting problem. A practitioner distinguishes an MCP interface—which lets an agent call product actions—from the control layer that must understand current state, enforce permissions and preconditions, pause for approval, and recover from partial failure. The proposed controls are concrete: re-check state immediately before a write, enforce permissions below the agent, implement real suspend/resume for approvals, and use an intent key that survives retries so a failed action cannot double-charge or double-send. Most SaaS APIs, the thread argues, return success/failure rather than current state and valid next actions; agent-ready products need state endpoints, permission-aware action manifests, explicit approval hooks, and idempotency keys.
Vibe-coded SaaS is creating an “understanding debt” diligence flag. An AI consultancy says a growing share of its work is rescuing products that already have paying customers; one booking product was polished and had about 80 customers, yet its founder could not explain what happened to an unused plan after a mid-month cancellation. The post argues that polished interfaces now hide unmade decisions around payments, refunds, and edge cases. Its practical test is useful in diligence: ask a founder to answer the five hardest questions about product behavior without opening the app; unanswered questions identify parts of the business the founder does not yet own.
Regulatory watch: A current-period community post says Article 50 of the EU AI Act took effect on August 2 and quotes a disclosure requirement for AI-generated text published to inform the public, with an exception for human review and editorial responsibility. It points to alleged hallucinated consulting reports from PwC and Deloitte and frames potential fines as a live consequence. Treat this as a verification item against official EU guidance before making compliance or investment decisions.
5. Worth Your Time
- Watch NVIDIA’s AI Learns Why Copying Humans Isn’t Enough. The useful part is that the demonstration and the failure modes sit together: 19 clips provide the imitation signal, while the second training “classroom” teaches adaptation to new obstacles; the transcript also reports only about 40% success on longer levels.
Read Interconnects’ latest open-artifacts roundup. It is a compact map of the current release wave, including model scale, licenses, hardware requirements, and evaluation transparency.
Inspect the Hermes Project Autopilot repository. The interesting artifact is not another coding demo but the explicit contract, verifier, provenance, and rollback design for autonomous repository changes.
Read Jason’s agent-permission post. A Google Drive connector silently granted read access across company files and write access to a Replit repository; the proposed operational rule is to inventory integrations like API keys and maintain logs that can answer what agents changed.
Garry Tan: 'Everyone mistakes the map for the territory' — in markets the outcome is the territory: 'did you make something people want?' — signaling an outcome-focused, meritocratic investment lens over credentials . Quoting @deedydas, Tan highlights Silicon Valley as 'one of the best meritocracies in the world' where proof of work beats pedigree: examples include Tim Cook (Auburn/IBM), Jan Koum (San Jose State dropout), Anthropic's CTO (no-name college), Jensen Huang (Oregon State/Denny's), Jeff Dean, John Carmack, Linus Torvalds, Satya Nadella, the GPT-3 team, and others . The essay's key rule: 'you get infinite shots on goal, but you just have to prove yourself to buy the next shot' — one strong product, viral repo, company, technical blog post, or career stint can 'shape your fortunes overnight' .
Tom Verilli, CPO of livestream shopping platform Whatnot (ex-CPO at Twitch, ex-Director of Product Growth at Twitter), describes Whatnot as the fastest-growing US marketplace business of all time, with only ~21–22 PMs for its GMV scale . He argues agentic commerce will be huge but not winner-take-all: US retail is a $7.5T industry, e-commerce has never exceeded 20% of US retail spend, and live commerce uniquely combines internet scale with the social/curated experience of in-person shopping . AI is enabling leaner product orgs: ICs can pull data that would have taken a 2017 Amazon L7 data scientist one to two weeks, PMs can query codebases directly (e.g., via Claude) to skip alignment meetings, and Verilli expects fewer, more senior PMs plus 'engineering-manager-light' hybrids . A notable talent trend: former CTOs of Workday, Instagram, Box, and Super.com have become rank-and-file engineers at Anthropic, a model Verilli wants for his own product team . Whatnot is hiring PMs with payments and logistics as top priorities .
Harry Stebbings committed to invest in simile_ai, an AI simulation startup, on a 1AM call with founder Joon (@joon_s_pk), after Index Ventures' Shardul Shah (who backed Wiz) sent him the deal; Stebbings says if it works it will change marketing, stock markets, and human decision-making in general . Joon was a painter before becoming a “world class AI mind,” combining creativity and scientific rigor — “one of the most unique talents” Stebbings has met . Thesis: AI simulations will be so valuable that companies will pay $100M for a single run (~$20M compute) to prevent a $500M product failure . The bigger commercial opportunity is the “GPU of intelligence” — models replicating human judgment, bias, and imperfect decision-making to simulate real-world markets — not frontier labs' “CPU of intelligence” optimized for perfect reasoning . On talent: the best operator signal is consistent performance across career stages (were they the common denominator behind every success?), and exceptional founders combine contradictory strengths — short-term paranoid, long-term optimistic .
- The open-model ecosystem is expanding rather than consolidating: more labs are investing hundreds of millions to billions in training and releasing models openly, betting on "token machines" as a path to value as token demand rises .
- Thinking Machines, founded Feb 2025, is now generating hundreds of millions in annual revenue from its open-model fine-tuning service and releasing the best US-built open-weight models, ahead of NVIDIA Nemotron and Arcee's Trilogy . Its debut open model Inkling is a 975B-A41B multimodal MoE (text/image/audio inputs, text output) positioned as a fine-tuning base via its Tinker service, with a competitive 276B-A12B smaller version .
- Poolside, an open-model lab with three releases in three months, shipped Laguna-S-2.1, a 118B-A8B MoE that fits on a DGX Spark, under the OpenMDW license (Apache-2.0-like with AI-specific legal backing), including evaluation-trajectory transparency .
- Korean startup Motif Technologies previewed Motif-3-Beta, a 314B-A13B MoE with new architectural components (GDLA, mHC) — its most ambitious model to date .
- DeepSeek released V4-Flash-0731 one day after OpenAI cut prices on its smallest model by 80%, beating Luna at the Pareto frontier; the larger V4 remains unupdated .
- Meituan's LongCat-2.0 (1.6T parameters) is the first non-Huawei, non-toy model trained entirely on Huawei Ascend 910 accelerators .
- Moonshot AI's Kimi K3, the biggest open release in some time, is under a noncommercial license requiring commercial agreements for inference/fine-tuning providers. Analysts argue this could expose US companies to future government action restricting Chinese open model use: "if a US company needs a contract with Moonshot to provide the inference tokens that Kimi K3 generates, the picture looks different" .
- Tencent's Hy3 (295B-A21B MoE) moved from a restrictive custom license to Apache 2.0 .
- AMD released Instella-MoE-16B-A3B-Think, a 16B MoE trained on Instinct cards, with base/SFT/MidTrain/DPO checkpoints available .
Paul Graham calls it a red flag when founders emphasize (or exaggerate) their credentials when talking to investors, arguing they should instead lead with what they've built — and having built nothing is itself a problem. He cites 19-year-old John Collison as an example, saying he can't imagine Collison leading with the fact that he was a Harvard student .
The U.S. Office of Personnel Management (USOPM) is reviewing nearly 150 federal job classifications that currently require college degrees, aiming to substitute demonstrated merit for 'sheepskin' in hiring . The push is framed as moving away from the college degree as the end-all-be-all and toward alternative workforce pathways, with tools for young Americans to reach their full potential .
- Harry Stebbings committed to invest in AI startup @simile_ai after Index Ventures' Shardul Shah (backer of Wiz) flagged it; Stebbings met founder @joon_s_pk within 60 minutes, committed on the call, and gave "a boatload of cash" .
- Founder Joon is "a painter before becoming world class AI mind," praised for combining creativity and scientific rigor; Stebbings says if simile_ai works it will change marketing, stock markets, and human decision-making .
- Thesis from the conversation: AI simulations will become so valuable that companies will pay ~$100M per run; a $20M compute run could prevent a $500M product failure . The bigger commercial opportunity may be the "GPU of intelligence" (models replicating human judgment/bias to simulate markets), vs frontier labs' "CPU of intelligence" for perfect reasoning .
- Talent-evaluation takeaways: look for consistent performance across roles and the "common denominator behind every success"; exceptional founders are short-term paranoid but long-term optimistic .
- Conversation episode: Spotify · YouTube
New AI research enables a virtual human to perform parkour with human-like motion while adapting to unseen obstacle courses, using only 30 seconds / 19 clips of internet parkour as reference data; prior methods were brittle or cheated . The controller trains simultaneously in two "classrooms" — imitating human movement and solving obstacle courses — with body, obstacle, and destination information, using a judge that scores movements as real vs. fake; judge and athlete improve together . It adapts and composes new skills, solving longer levels, varied levels, and unseen obstacle arrangements . Limitations: lower tracking error than predecessors but with a hit to success rate, ~40% success on longer levels, and possible unnatural recovery motions . The research paper is freely available, with code possibly released later .
Garry Tan (President & CEO of Y Combinator) posted that "AI will create unimaginable economic growth and that is the best white pill," endorsing Andrew Ho's essay on AI as a growth driver .
- Andrew Ho argues he is "willing to take a bet on stronger capabilities" because he views AI not as extinction risk but as "humanity's last hope" against a potential centuries-long "dark age" .
- His pessimistic baseline includes falling birthrates with adverse selection among educated/wealthy populations, aging-driven political capture and redistribution toward the old (e.g., UK "triple lock"), anti-market sentiment among youth (e.g., California Prop 13), and rent-seeking in productive US cities (SF's $17B budget without solving homelessness, CA HSR mismanagement, ~$100M Manhattan subway elevators) .
- He warns of possible negative economic growth within his lifetime and sees AI as the lever to cut through bureaucracy, reinvigorate pro-market sentiment, and shift the economy from a negative to a positive feedback loop .
Paul Graham (@paulg) argues that most successful startups have a wonderfully popular product that more than compensates for many founder mistakes; without it there is too little margin for error, but a popular product buys the startup out of all kinds of trouble.
Jason Calacanis is offering a $5,000 bounty (or an equivalent Grand Seiko purchased with the winner in Ginza) for the best annotated .com build out . He decided to build this as a startup, noting the bounty program is explicitly meant to generate ideas and code . He plans to spend two months on it, bring the best builders from twistartups to show progress, and pick a winner on October 1st after three rounds .
- Harry Stebbings committed to invest in simile_ai after Index Ventures' Shardul Shah (whom he credits with Wiz) introduced him to founder Joon Sung Park; he decided on the call .
- Founder Joon Sung Park is described as a painter before becoming a "world class AI mind," blending creativity and scientific rigor; his talent test is asking whether someone was "the common denominator behind every success" .
- Stebbings' thesis: "If this works, it will change marketing forever. It will change stock markets forever. It will change human decision-making in general" .
- He argues AI simulation will be highly valuable: companies will pay $100M for a single simulation, e.g., a $20M compute run that prevents a $500M product failure .
- The bigger commercial opportunity may be the "GPU of intelligence"—models replicating human judgment, bias, and imperfect decision-making to simulate real-world markets—versus frontier labs racing to build the "CPU of intelligence" for perfect reasoning .
a16z GP Martin Casado reports that DeepSeek v4 Flash results "aren't great" in his testing, while K3 is "quite impressive," and he wonders whether AI model quality is hitting actual model size limitations .
AI infrastructure is shifting from training to inference, projected to be ~80% of the neocloud market by 2030 ; inference is a recurring opex buyers optimize on cost/latency, unlike training's capital-event dynamics .
Thesis: neoclouds (CoreWeave, Lambda, Crusoe, Together) are winning today on scarcity and speed against hyperscalers (AWS, Azure, GCP), but scarcity is temporary — long-term survival requires moving up the AI software stack, which most haven't started .
Neocloud advantages: clusters approaching 80,000 GPUs deployed in weeks vs hyperscalers' multi-tenant, virtualization-laden architectures ; bare-metal pricing reported up to 85% cheaper, partly architecture, partly pricing for share ; hyperscalers are anchor tenants — Microsoft was ~62% of CoreWeave's 2024 revenue — making the largest customer also the largest competitor .
Market forces in inference: NVIDIA controls allocation, cadence, and software (AMD is only a marginal second source), squeezing neocloud gross margins between dictated input pricing and commoditized output ; power is equally binding — committed-power data-center real estate is the scarce asset, so 'smartest capital is buying megawatts, not GPUs' . Buyers' switching costs collapse as routing/orchestration matures . Hyperscalers are projected to spend $600–700B CapEx in 2026 and can bundle AI compute into enterprise contracts, while neoclouds fight each other on price . Substitutes include edge inference (majority of global volume by 2026, >70% in some forecasts), custom silicon (Google TPUs, AWS Inferentia/Trainium, inference ASICs), and model efficiency (distillation, quantization) . Entry is eased by GPU-backed debt/private credit and by sovereign AI clouds in Europe, the Gulf, and APAC that don't need venture hurdle rates .
Pure compute brokerage is a stopgap ; neoclouds' two credible paths are moving up the stack (managed inference, training orchestration, fine-tuning, observability, routing) or defensible niches (sovereign compute, ultra-low-latency edge, regulated verticals) . Hyperscalers are expected to use custom silicon and bundling to absorb enterprise demand and buy distressed neocloud capacity — consolidation is coming . Investor diligence should focus on % revenue from software/managed services, customer concentration, and whether contracted power outlasts hardware depreciation, not GPU count . Companies that used the GPU shortage to build a platform will become next-gen infrastructure titans; markup renters will be acquired, refinanced, or forgotten — outcomes should be clear within ~8 quarters .
Medal, a consumer app, is on track for 3 billion clips this year, according to @PimDeWitte . Vinod Khosla calls the app "Pretty crazy" for one many haven't heard of and asks what Medal is and what is likely to be the best world model .
Alibaba's Qwen team announced Qwen3.8-Max, described as "A New Bar for Coding and Cowork" . VC/angel @Jason highlighted it as evidence that China is making increasingly competitive AI models, while joking about its "world-positive AI marketing" — riffing on the claim that "THE COMPUTERS ARE GOING TO DO OUR JOBS FOR US AND WERE ALL GOING TO THE BEACH" .
Per @polynoamial, an internal version of OpenAI's next major model family Astra solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science, and is expected to be 'a major step for scientific reasoning' . @Jason amplified the news with a joking note to Sam Altman to 'Please keep your finger on the kill switch' .
Open-source models are closing the gap: @Jason states the difference between the open-source models he is using and Frontier models is already negligible .
The thread asks why no location-based AR app with Pokémon Go-style social cohesion has emerged since its peak . Despite the phenomenon, the space remains risky: Harry Potter: Wizards Unite, Minecraft Earth, Pikmin Bloom, Monster Hunter Now, and Jurassic World Alive all flopped ; Pikmin Bloom is a partial exception with strong player bases in Taiwan and Japan but low awareness elsewhere . Pokémon Go's success is largely attributed to its generational IP rather than game quality — "There is no game, it's just a tracking app with Pokémon" — and copycats without that IP failed . Fanbase loyalty is such that a direct successor would likely spike again . The original was created by a founder who left Google and partnered with Nintendo/Pokémon, and the data it collected powered future products . Commenters allege the AR scans were used to build a global visual positioning system (VPS) for drone navigation when GPS is unavailable — now Niantic Spatial's flagship product with applications cited for Waymo — and describe the app as a data-collection ploy, including military drone training and outsourced data labeling for AI companies . Monetization is the core barrier: these concepts are hard to profit from without degrading the experience, and consumer jadedness over data-mining/cash-grab dynamics erodes trust . AR tech is seen as much more mature than 2016, leaving room for a new, well-executed concept . Pokémon Go was acquired by Scopely in 2024 and still generates ~$1B in annual profits, marking the commercial ceiling for the category .
Investor Jason Calacanis (@Jason) posted: “Intelligence has commodified, wisdom and intent have not” — a sentiment signal that raw AI capability is increasingly a commodity while judgment and purposeful application remain differentiating .
