We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Google’s Australia–New Zealand accelerator is the clearest capital-adjacent early-stage signal, rather than a disclosed priced round. Google selected 15 startups for a 10-week, equity-free hybrid program aimed at Seed and Series A companies using AI and machine learning; the announcement describes the program but does not state an investment amount or investor syndicate. The cohort includes Blunge’s AI design agent, Bower’s research operating system, Eyes of AI’s dental-imaging product, Mentana’s autonomous supply-chain manager, Merchmix’s inventory operating system, RossOps’ manufacturing memory layer, Totex Energy’s flexible-grid tooling, and Unseen’s commercial-real-estate trust layer.
The sourcing thesis is vertical workflow infrastructure—research, supply chain, manufacturing, health, energy, and auditable real estate—rather than another general-purpose chatbot. Treat the accelerator as ecosystem validation and a deal-sourcing map, not as financing proof.
2. Emerging Teams
Baukompass is the strongest domain-first team signal. An unnamed founder working at a German insulation and fire-protection contractor is building the product at night with a small AI-agent team, using the employer as its first pilot in a market dominated by two incumbent vendors. The founder learned that the real product is decades of domain rules, GAEB tender exchange, calculation history, and price books—not the incumbents’ dated interfaces—so Baukompass became a sidecar: measurement data and a tender go in, a quote draft comes out, and the customer finishes the job in its existing system. The pilot has exposed the execution risk: the parser handled the official specification but failed on the third real customer file, estimators still check every generated position, and current time savings are estimated at roughly 30%, versus an initial 80% expectation. The investable signal is domain expertise plus incumbent-compatible distribution; the first pilot and manual review mean repeatable traction is not yet demonstrated.
A previously exited SaaS/IoT founder is using AI to operate as a much more capable solo builder. The founder says a prior company was acquired by a listed French group, and that the new fractional-C-level marketplace in Europe reached beta after about five months of solo work. They report building authentication, onboarding, admin, GDPR, and audit features with Claude Code despite spending the previous 15 years mainly on sales, marketing, fundraising, and legal, and say most new signups report that ChatGPT or Claude recommended the platform. That acquisition signal is promising but not clean attribution: commenters argue that AI discovery captures existing demand and can obscure the comparison pages, directories, or other web sources that actually generated it.
An unnamed brand-strategy product offers a useful activation-before-monetization case. Its founder reports 1,300-plus signups, more than 300 completed brands, and about 12 paying customers across 16 purchases; every payer had completed a full brand before paying. The product gates exports, downloads, and team access rather than the work itself, while subscriptions, one-time unlocks, and credit packs have all seen early use. The traction remains small and self-reported: on one day, Reddit supplied 74% of referred traffic from a base of 82 visitors, but the founder could not attribute a single payer to a channel. The founder’s prior branding leadership at DivX and roughly 20 years shipping applied AI add relevant pedigree, but not yet venture-scale proof.
Ceadly is an early control-plane wedge for agentic software. The new project positions itself as an authorization and governance layer that requires named human approval before consequential actions. Its proposed operating model starts with a short list of refunds, external emails, production writes, and permission changes; routine actions can later be policy-approved with timeouts and audit trails, while repeated unchanged approvals become candidates for relaxation. The product’s own framing identifies policy calibration—not merely adding an approval button—as the hard part: broad controls create approval fatigue, while narrow controls can miss the damaging action.
3. AI & Tech Breakthroughs
The Station is a meaningful step toward open-ended, multi-agent scientific discovery. The linked paper reports an environment where agents from different model families pursue a shared mathematical goal without a central coordinator or scripted pipeline, choosing research directions, running experiments, collaborating, and building a shared literature. Across 12 AlphaEvolve construction problems and two additional case studies, it reports results novel relative to prior literature on five problems, plus new infinite families for Book Ramsey numbers; the agents also produced explanatory theorems and analyses and released raw dialogues, proofs, and verification code. The results should be treated as reported claims until the released proofs and code are independently reviewed, but the important paradigm is that the environment delegates research direction and collaboration—not just a fixed sequence of model calls.
TUFF shows a practical local-inference trade: storage bandwidth in exchange for RAM. The solo builder’s native Swift/Metal macOS app keeps shared mixture-of-experts components resident and streams needed experts from SSD, claiming that a 61GB model can run on a 16GB MacBook Air without swapping. The project reports more than 1 token per second for an open-source 120B model and more than 7 tokens per second for Qwen3.6 35B-A3B and Gemma 4 26B-A4B on a 16GB M2 Air; these are README benchmarks, not independent validation. The opportunity is wider private, device-local access to models that exceed conventional memory limits; the diligence question is whether SSD latency and older-device performance preserve a usable experience.
Visual verification is becoming a first-class agent capability. A solo AI-video builder says the previous system shipped after a tool returned { success: true }, even when text overflowed, scenes were blank, or layers were stacked invisibly. The new loop renders the live editor at 2, 7, and 13 seconds, sends the actual pixels back to the agent, and patches the HTML/GSAP source in place; if the browser capture flag is unavailable, it returns an explicit error rather than passing a blank frame as success. This is a useful harness pattern beyond video: tool-level success is not output-level correctness, so agents need access to the artifact they actually produced.
4. Market Signals
AI is moving the moat away from implementation and toward distribution, trust, data, and domain context. A current founder discussion argues that one competent developer can now build in days or weeks what previously took a small team months, while successful implementations can be reproduced just as quickly; it names distribution, brand, proprietary data, network effects, domain expertise, and existing customers as the scarcer assets. Replies add that selling, deployment, maintenance, security hardening, scaling, and requirements remain substantial bottlenecks. Andrew Chen compresses the sentiment into “100x side projects, 0x actually shipping,” while another founder observes that the expanding AI-product field contains many similar promises with unclear differences until users spend hours testing them.
Agent operations are becoming a control-plane market. A booking workflow that handled flights, hotels, Gmail, and SMS in a stateless sandbox reportedly double-booked, skipped notifications, or hung in production because four-plus external APIs could fail silently while state was only partially committed; mocked dry runs did not reproduce the real behavior. Security operations show the same exposure at higher stakes: agents are being given endpoint controls, threat-intelligence feeds, and incident-response tooling comparable to a privileged service account, while RBAC can confirm entitlement without judging whether the action is appropriate in context. The adjacent FinOps gap is similarly concrete: multi-provider dashboards do not connect spend to teams or workflows, so instrumentation needs to tag requests before they leave the application and include retries and failed calls before joining usage to revenue.
Compute demand remains strong in the bullish case, but infrastructure underwriting must separate utilization from asset quality. An a16z post argues that moving from chatbots to reasoning, agents, and multi-agents multiplies the tokens required per task and leaves the industry constrained by GPUs, power, and the ability to create supply. An Investing in AI analysis argues that generalized models are commoditizing while specialized models, inference-specific chips, and better harnesses can reduce costs; it reports cutting some customer inference bills by more than 80% and says lower costs could make projects uneconomic in 2025 viable in 2027. The same analysis warns that a demand drop need not create a compute glut, but that poorly located data centers, bad power contracts, and rapidly depreciating hardware can still produce write-downs.
AI-native distribution is now showing up in operating metrics, not only product demos. A SaaStr account says Owner.com, a vertical SMB software company, has passed $100 million ARR with triple-digit growth and reports that more than 83% of new customers now begin inside an AI product. The company replaced a sales-led demo and onboarding path with a free five-minute AI build that produces a website, photography, video, and SEO/CRO audit. Its claimed moat is outcome data from thousands of live restaurant sites and tens of millions of consumers, linking product decisions to order volume rather than merely accumulating a public corpus. The relevant diligence question is revenue impact: the same account reports more than $2 million ARR per rep and roughly four times direct SMB competitors, rather than stopping at time-saved metrics.
Automation projects are also becoming labor and permitting products. A quoted WSJ account says resistance to data-center construction has broadened from electricity, water, and noise into jobs, with politicians halting approvals or tightening oversight and unions warning that thousands of building-related jobs are at risk. In Atlanta, the rideshare drivers’ union says drivers have seen lower pay, longer waits, and fewer fares since Waymo’s rollout and is asking for a $0.50–$1 robotaxi impact fee. For infrastructure and robotics investors, local political support and worker-transition costs belong alongside power availability and model economics.
5. Worth Your Time
- Watch — Lenny’s Podcast: AI’s third era, the rise of persistent AI coworkers. The strongest segment frames persistent agents as coworkers that work at higher levels of abstraction and eventually collaborate across users, while stressing that cloud agents need data access, infrastructure, and reliability—not only better models.
Read — How To Think About The Negative ROI on AI Investment. Use the essay for its separation of model-lab valuation from long-term viability, its cost-curve thesis around smaller models and better harnesses, and its warning that useful demand can coexist with bad data-center bets.
Thread — Sriram Krishnan on open models and the Hugging Face incident, with Clement Delangue’s response. The exchange is a concise case for open-weight models in private cyber investigation: Sriram says Hugging Face was blocked from using closed models, while Delangue says GLM helped identify planted backdoors and enabled analysis without sharing confidential data or asking a provider’s permission.
Read — Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment. Read the primary abstract before treating the Station’s novel-result claims as investment-grade evidence; its distinctive contribution is the combination of open-ended agent coordination, explanatory mathematics, and released verification artifacts.
Read — Owner.com’s AI rebuild. It is a useful operator case on replacing a sales funnel with a free outcome, turning deployment data into a moat, and directing agents at coordination overhead rather than only code generation.
Direct answer: The linked source reports an autonomous “Station” in which AI agents from different model families pursue a shared mathematical-research goal without a central coordinator or scripted pipeline; agents select research directions, run experiments, collaborate, and build a shared literature.
- Study scope: The report covers 12 construction problems from the AlphaEvolve catalogue plus two additional case studies.
- Reported novel results: Relative to prior literature, the Station reportedly produced results on five problems: a new infinite family of finite-field Kakeya sets; new exact 604-point kissing configurations in dimension 11; new records for the discretized Kakeya needle and sign-uncertainty problems; and a substantially improved lower bound for Erdős’s minimum-overlap problem.
- Additional result: The agents also reportedly discovered novel infinite families for Book Ramsey numbers.
- Interpretability claim: The source says the agents generated theorems and analyses explaining their numerical constructions, rather than only supplying numerical outputs.
- Transparency and verification materials: The authors state that they release all raw agent dialogues, proofs, and verification code, presenting this as a transparent record of how the discoveries emerged.
- Qualification: In the supplied extract, transparency is explicitly claimed through release of artifacts; the extract itself does not display those artifacts or report an independent correctness audit, so the novel-result and verification statements should be treated as reported claims pending inspection of the released proofs and code.
The official announcement confirms a 15-startup Australia & New Zealand cohort.
- Program terms: It is a 10-week, hybrid, equity-free accelerator, presented as part of Google’s ongoing Digital Future Initiative commitment.
- Stage and technical focus: The program is designed for Seed and Series A startups leveraging AI and machine learning.
- Funding and investor context: The announcement supplies no stated funding amount, investment, grant, investor, or investor-introduction terms in the supplied text. The only financing-relevant detail is the program’s “equity-free” description.
- Agent product paradigm: Tara Sash describes AI work moving from chat to agents and toward persistent co-workers that handle longer-running tasks, sync with users, and eventually collaborate across users’ agents.
- Product convergence: ChatGPT’s Work mode uses Codex under the hood, with the same underlying power and a different interface. OpenAI’s stated north star is to remove user decisions between modes and automatically select the right harness for a task.
- Infrastructure and evaluation bottlenecks: Effective cloud agents require data access to third-party systems, cloud infrastructure, and reliability in addition to model intelligence and long-running task ability. Knowledge-work agents also need visible inputs, citations, in-progress work, reasoning, and relevant organizational context because final outputs cannot be validated as easily as code through tests.
- Model-cycle strategy: Product teams should build for capabilities expected in roughly two to three months—not for today’s models or a speculative one-year-future—and stay tightly aligned with the research roadmap.
- Prompt-generated personal software: Sash describes “sites” as prompt-built, hosted and shareable software with databases, auto-updating behavior, and the ability to use internal data for dashboards and other dynamic work surfaces.
- Pedigreed AI operators: Sash leads Codex and ChatGPT work at OpenAI; she previously spent six years at Stripe, joining as one of its first five product managers, led product at Watershed, and was a founder and fellowship alum. Ari Weinstein, who leads computer-use work at OpenAI, founded Sky, which was acquired by OpenAI, and previously had a company acquired by Apple.
- Early-stage investing/playbook signal: Sash says Sutter Hill’s incubation model treats product-market fit as repeatable and emphasizes enterprise sales, positioning, the initial founding team, and recruiting. Her specific takeaway is to test product-marketing fit before committing to the product shape by pitching roughly 100 people and refining the transformative narrative.
Paul Graham’s startup-idea heuristic is to avoid deliberately searching for ideas, which can produce conservative choices and eliminate outliers; instead, founders should build projects with friends that they themselves find compelling.
- An All-In episode highlighted by Jason covers China’s robot competition, Tesla Optimus, and Grok Bot, while framing strong Nvidia and Salesforce earnings as a challenge to the “AI Capex Bubble” and “SaaSpocalypse” narratives.
- The episode also addresses controversy over AI-written opinion pieces and Moderna’s mRNA cancer-vaccine work, adding signals on AI-content scrutiny and adjacent biotech innovation.
- Agents may eventually test many software products, publish the results to a shared “Artifactory moltbook,” and use that corpus to guide agent swarms.
- The AI application market is showing signs of crowding and weak differentiation: product counts are expanding, radical new ideas appear scarce, and many tools make similar promises without an obvious advantage until users invest substantial time evaluating them.
- Sriram Krishnan frames a cyber incident as evidence that open-weight models can be strategically important for security: HF was reportedly blocked from using closed models to analyze the incident and turned to open-weight models, while GLM helped identify planted backdoors; the approach enabled detection and analysis without sharing confidential data or asking a model provider for permission. This points to an emerging infrastructure thesis around open models for private threat detection, investigation, and response.
- Open-source AI reportedly increased its token share from 28% to 62% at Vercel over the past two months, taking share from OpenAI and Anthropic.
- This shift could be positive for AI infrastructure: open-source models may compress model-layer margins, but their tokens still require as much compute to produce as frontier-model tokens. Gavin Baker estimates closed frontier tokens could retain 60–90% of economic value while representing only 15–25% of total tokens.
- OpenAI and Anthropic usage also accelerated in July, while Baker suspects Grok is growing even faster, suggesting overall token and AI-infrastructure demand may be accelerating beyond frontier-model growth.
Andrew Chen characterized “Coding in 2026” as “100x side projects” but “0x actually shipping side projects,” offering a cautionary investor-sentiment signal that project-generation velocity may be decoupled from product delivery.
- a16z announced the Machine Age Fund, a new $1.1 billion fund backing founders rebuilding the AI stack across chips, memory, networking, systems software, power, and machines that bring AI into the physical world.
- The fund’s thesis is that model progress is outpacing the underlying memory, interconnect, power, and cooling infrastructure, turning each physical-stack constraint into a company-building opportunity and drawing strong software teams into complex hardware.
- a16z’s accompanying investment view is that the shift from chatbots to reasoning, agents, and multi-agent systems increases token demand by orders of magnitude; inference repeatedly uses AI as a building block, making infrastructure supply—not engineering time—the limiting factor.
YC wants rejected companies to succeed because their outcomes reveal whether YC passed on a strong company, signaling an investor mindset that treats rejected-startup performance as feedback on selection quality.
Martin Casado is experimenting with a portfolio of specialized personal-assistant bots: a house bot handles property administration and can email, call, and schedule; an LP bot manages investment paperwork; a family bot checks calendar and time allocation; a news bot aggregates X and traditional media; and a manager bot coordinates the others, reports activity, and distributes tasks. The system is still experimental; the finance bot remains underdeveloped and lacks a Plaid integration, while Casado may publish the templates if it works well.
Bindu Reddy posted that DeepSeek V5 would launch in September, claiming it would “supercharge personal agents,” cost 100× less than Terra and Sonnet, and handle every use case except hard-coding; she argued frontier models would mainly remain necessary for app-building or hard-coding loops.
fal.live introduced a platform for “infinite, interactive AI livestreams”: users choose a channel, prompt what happens next, and watch the show generate in real time, shifting the audience into an active director role.
- Bindu Reddy says OpenAI cut Luna’s price by 80%, after which usage rose 1,000x and Luna became competitive with Deepseek Flash. She contrasts this with her view that Haiku has become obsolete. This is a market signal that aggressive model pricing may rapidly shift usage and competitive positioning.
- The post proposes a two-model workflow: Astra handles system-wide reasoning loops, while Fable 5,1 performs deep dives into complex problem execution; together, they are claimed to automate very complex, long-running tasks and “easily outperform 100-person engineering teams.”
The account announced a Gen-2 weights release framed as a limited-edition collection: acquiring the complete model required collecting 6,834 books. A follow-up said 382 boxes of the Gen-2 Book of Weights had been found while making room for a new book.
Cristóbal Valenzuela is working to make Chile the AI capital of Latin America, citing the country’s energy potential and research talent; he argues that constitutional or other man-made constraints should be treated as changeable rather than fundamental limits.
- AI agents are proposed as a cybersecurity countermeasure to AI-enabled hacking: they can search for vulnerabilities, perform security scans, and run penetration tests.
- Hugging Face CEO Clement Delangue frames cyber defense as a sequence of detection, understanding, containment, and remediation rather than only real-time response. In the incident discussed, the team made an initial cut more than a week before OpenAI recognized the problem, GLM helped identify planted backdoors, and continued agent probing made containment necessary.
- Delangue argues that open models enable threat detection and understanding without asking permission or sharing highly confidential data, signaling a potential role for open-model infrastructure in cybersecurity.
