We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Corridor Brokers announced a $25M seed led by BainCapVC to serve small-business employees with health insurance. The company frames the opportunity as changing the economics of a segment it says has been underserved because it was not economically viable; the investor describes the team as transforming how health insurance is sold. For the radar, this is a seed-stage distribution and market-structure bet. The disclosed evidence supports an insurance thesis, not yet an AI-specific moat.
2. Emerging Teams
iKawn Mirror is moving a live AI try-on demo into a retailer workflow. The team is working with its first wig client after demos reached 700,000+ Reddit views. The client engagement is exposing the real product requirements: hair density, hairlines, colour accuracy, preserving the customer’s face during movement, and a cost for every minute of use. The useful pilot metric is conversion lift with and without try-on—not engagement or visual impressiveness alone.
SpacePlanner is a sharp distribution signal, but not yet a monetization signal. The free, no-signup floor-plan tool went from a few hundred daily users to 15,000+ unique users, 70,000+ pageviews, and thousands of registrations. Its founder attributes the surge largely to unpaid creator exposure, including a SetupsAI video from an account with 6M+ followers, followed by smaller creators across several countries; the original Indonesian source remains unidentified. The founder had to redeploy to cheaper infrastructure as costs rose, and a feedback popup generated almost 5,000 responses in two days. This is evidence that dark-social distribution can overwhelm a technically modest product, not evidence of durable revenue or retention.
3. AI & Tech Breakthroughs
Unconventional AI is betting that the next inference breakthrough is a new computer, not a larger model. Co-founder and CEO Naveen Ralph has electrical-engineering training and a neuroscience PhD; he says he founded Nirvana Systems, sold it to Intel, ran Intel’s AI group, and later built GPU-scaling infrastructure that joined Databricks. Unconventional AI’s stated goal is 1,000× power efficiency, with the original five-year target revised to three and a half years. Ralph’s energy case—using reported token volumes and an assumed 10 joules per token to estimate 12 GW for one company’s AI services and an energy shortfall in roughly three years—is explicitly his estimate, not an independently validated forecast.
The technical claim is a dynamical computer built from interconnected oscillators, combining memory and computation in the same elements. Ralph says the team open-sourced UNO, an oscillator-based image-generation model, then taped out the first physical dynamical computer in five months and generated images at roughly 500 nanojoules each; the architecture uses physical die-stacking plus time dynamics, which the company calls “4D computing.” The proposed commercial product is a managed VM/rack within about two years, but existing models will require model-layer porting and the company is coordinating theorists, biologists, physicists, and chip engineers. That makes independent energy/quality benchmarks, fabrication, porting friction, and cross-disciplinary execution the key diligence gates.
Decision models are becoming both a model category and a control primitive. SemIf, an open-source Qwen3.5-based alternative, is being served free through LangSmith Gateway; it is reported at 74.7 versus Jev’s 75.4 on JevBench, while an open-source multilingual System 1 decision model became Hugging Face’s top trending model shortly after Jev began trending. LangSmith’s Jev-as-a-judge integration is designed to score every production trace, evaluate multiple criteria, and trigger webhooks for issues such as PII leakage, prompt injection, or toxicity. Its setup documentation also notes that TypeSafe does not yet offer zero data retention, a relevant enterprise diligence flag.
The adoption pattern is less about flashy general intelligence than cheap, repeated semantic decisions: a non-affiliated directory collected 607 Jev builds, with examples including 500 emails classified for $0.035 and 3,282 posts scored for $0.13. ThruWire’s accompanying architecture shows where value may accrue around the model: checkpoints require evidence-bearing artifacts and stable upstream relationships, while Jev supplies probabilistic judgments about whether evidence supports a claim or an implementation satisfies a requirement. The likely investable layer is therefore the connective tissue—checkpoints, provenance, policies, and feedback loops—not an isolated classifier.
4. Market Signals
Open-weight models are becoming a strategic deployment layer. Interconnects argues that Chinese models such as GLM-5.2 and Kimi K3 have crossed a commercially meaningful agentic-capability threshold; it estimates Chinese open-weight models are 2–5 months behind the closed U.S. frontier, while U.S. open models remain 6–9 months behind. OpenRouter usage reportedly rose from about 1T to 80T open-model tokens per week, with Chinese models exceeding 80% of usage; OpenCode reports roughly 95% or more of inference volume from Chinese models. The same analysis identifies inference platforms, fine-tuning APIs, and purchased agent-training environments as early layers of an open-model economy, while warning that open software is difficult to keep from bad actors and that U.S. products are increasingly dependent on Chinese models.
Corporate AI language is moving faster than measured business impact. In the June 2026 earnings season, 33% of S&P 500 companies making calls quantified AI use, but only 15% quantified AI’s business impact. Cost/productivity claims outnumbered revenue claims 26% to 16%; deployment language reached 12% while pilot language stayed below 4%; and “agentic” appeared in 24% of calls versus 5% for “generative AI.” These are earnings-call signals, not proof of production value, and the source warns that pilot activity is difficult to infer from the language.
That gap strengthens the workflow-first investment thesis. UiPath CEO Daniel Dines says models are interchangeable while the defensible asset is the enterprise “map of work”—workflows, exceptions, procedures, and systems. A vibe-coded procurement tool still required human intervention for connectors, permissions, audit, security, testing, maintenance, and a faulty database schema. The implication is to underwrite process capture, exception handling, and deterministic controls rather than API access alone.
Agentic commerce is meeting a permission wall. Amazon cut off Meta’s Muse from shopping on Amazon.com after saying Meta had not obtained permission, Muse did not identify itself as an automated agent, and it appeared to capture and store customer credentials. Amazon says an agent that can move through account pages, order history, transactions, and sensitive data must operate openly and have the merchant’s agreement; Muse’s browser fallback makes services without public APIs reachable in precisely the way that raises this dispute. Agent identity, delegated consent, credential isolation, and merchant opt-out are becoming distribution requirements, not post-launch compliance details.
The governance problem is also reaching capital markets. Harry Stebbings asks how a frontier provider can pursue an IPO without answering who is liable for harmful acts by rogue agents, while Eric Schmidt argues that a development pause conflicts with participants’ incentives and would be difficult to verify.
5. Worth Your Time
- Watch — Naveen Ralph on Unconventional AI. The strongest segment explains oscillator synchronization, UNO, the physical prototype, and the memory-compute architecture behind the “4D computing” claim.
- Watch — Daniel Dines on the enterprise “map of work.” The clip is useful for understanding why process discovery, exceptions, and one-process-at-a-time adoption may matter more than access to a frontier model.
Read — The current balance of power in open models. The article is the best current synthesis here of Chinese open-weight capability, usage concentration, infrastructure winners, and the associated security and policy risk.
Read — Jev and AI Checkpoints: Using Decision Models to Wrangle Agent Work. Its practical value is the architecture: let agents execute flexibly, but require evidence-bearing checkpoints, semantic validation, provenance, and feedback loops.
Direct answer: LangSmith’s Jev integration is an online evaluator for tracing projects. After saving the evaluator, it scores incoming runs or threads; each typed question becomes its own feedback key, which can be filtered, charted, alerted on, or used in automations.
- Trace coverage: The post explicitly presents Jev as enabling teams to score every trace instead of only a sample, while its speed is intended to keep pace with live traffic.
- Evaluation model: The evaluator maps run or thread variables into a state, then applies typed questions rather than embedding grading instructions in a prompt. Jev evaluates all questions in one call, and each question produces a separate feedback key.
- Automated responses: For feedback keys such as PII leakage, prompt injection, or toxicity, LangSmith alerts can trigger a webhook, providing an event-driven path to automate a response. The supplied material does not name a specific remediation action beyond the webhook.
- Availability and production caveat: The integration is described as available in LangSmith today through a tracing project’s Evaluators tab, using TypeSafe’s
jev-latestmodel. TypeSafe does not currently offer zero data retention, so prompts and evaluation outputs may be retained by the provider. - Evidence limitation: The reported accuracy, consistency, speed, and cost comparison was one test on one agent, so it is evidence of viability rather than a broad production-performance guarantee.
Direct answer: Amazon says it blocked Meta’s Muse because Meta did not obtain Amazon’s agreement or notify it that Muse would access Amazon, Muse does not identify itself while browsing, and Amazon believes the agent captures and stores customer credentials, creating privacy and security risks.
Agent identity and authorization: Amazon characterizes Muse as an “unauthorized AI agent” under its Conditions of Use and says an agent acting on customers’ behalf should operate openly and respect the service provider’s decision about whether to participate. Amazon says Muse’s failure to identify itself means an undisclosed third party can move through customer accounts, process transactions, and handle sensitive data without Amazon’s knowledge or consent.
Permission and merchant control: Amazon requested that Meta remove Amazon from Muse’s shopping experience, asserting that third-party agents have obligations comparable to food-delivery and travel services that operate with merchant agreement. Thus, Amazon’s position is that a customer’s request alone does not substitute for Amazon’s agreement or permission for the agent to access the store.
Credentials—重要 caveat: Amazon says Muse “appears to capture and store customer credentials.” Meta’s previously stated position conflicts with that characterization: it says Muse has no visibility into passwords or payment methods and stores shared credentials securely so it can use them without seeing them. Muse’s documented access model allows it to use user-provided credentials through a public API, or to operate through a browser when no API exists.
Shopping and account access: Amazon says Muse can reach account pages and order history when prompted, and can process transactions and sensitive data. Amazon cut off Muse’s Amazon.com shopping access; users instead see a notice that continued access by the unauthorized agent violates the Conditions of Use. Meta says Muse checks with the user before sensitive actions such as making a purchase, but Amazon’s stated objection is directed at the agent’s lack of disclosure, authorization, and merchant consent rather than only at whether the user approved an individual purchase.
Operational implication: Muse’s browser-based fallback makes services without public APIs reachable in a way Amazon says is unauthorized, while Amazon’s own “Buy for Me” description emphasizes that its agent identifies itself and lets brands opt out.
- Workflow-native applications are the proposed enterprise AI moat. UiPath’s CEO argues that “models are interchangeable” and that the real value lies in the workflow and the “map of work,” which captures workflows, exceptions, procedures, and systems. Harvey and Lora are cited as examples of applications that map work and build complete legal workflows, rather than merely returning a model-generated opinion.
- UiPath is building process-context infrastructure around this thesis. The company describes “ctography,” in which an agent interviews subject-matter experts in real time, probes why exceptions were handled, consolidates multiple interviews, and produces process maps that can guide software transformation. Its proposed architecture combines this context “map” with orchestration and automation “rails” so agents operate within defined enterprise controls.
- Coding agents are a major productivity advance, but production—not code generation—is the bottleneck. Coding agents can create deterministic automations at design time, help repair them when upstream systems change, and enable auditable execution. In practice, a vibe-coded procurement tool encountered connector, permission, audit, security, testing, maintenance, and database-schema problems; human engineers were still required.
- A core technical limitation remains the difference between memory and learning. The interview argues that current models do not update their weights through on-the-job experience, while multi-step probabilistic execution compounds the chance of error; exact work should therefore be delegated to deterministic tools and software. Enterprise AI diffusion may consequently be slower than headline forecasts because companies must document processes and exceptions before automation can scale.
- Model optionality and cost-efficient inference are emerging deployment themes. The interview predicts that roughly 90% of enterprise operational traffic will use cost-efficient rather than frontier models, with open-source models retained as a backup to avoid provider lock-in. Enterprises may also maintain their own models trained on proprietary data, although frontier providers could still offer better infrastructure economics. A related adoption risk is IP leakage: large enterprises are described as more worried that model access could expose their data to competitors than that a frontier lab will directly compete with them.
- Context-rich data and compute access are viewed as underappreciated infrastructure moats. The speakers argue that data providers create value through understanding data and supplying the right context at query time, not through storage alone. Fireworks is praised as a strong team partly because of prior compute-securing experience at Facebook, while the discussion argues that open-model inference providers must secure compute to serve demand at scale.
- AI unit economics and infrastructure timing remain major caution flags. The discussion cites a data-provider CEO reporting inference spend at three times human salaries, while the counterargument is that higher machine cost could still be rational if quality is better and human error and costs continue rising. The exchange also flags possible near-term infrastructure overbuild if labor replacement takes years rather than months, despite energy, water, data-center, and policy constraints on supply.
Company, founder, funding, and traction. Diogo Almeida comes from OpenAI’s post-training effort and coauthored the InstructGPT paper. In forming TypeSafe AI, he recruited Eric and Sasha; Sasha left a startup to join, and the team had funding within two weeks, although the account gives no deal size, stage, or lead investor. Jev’s launch drew tens of millions of video views. Almeida also reported passing one trillion tokens per day, including overnight machine-driven usage rather than only human experimentation. The key counter-signal is early commercial immaturity: TypeSafe had almost no revenue before launch, more than half of early testers did not understand the product, and post-launch demand nevertheless escalated to requests for large rate limits and even offers of GPUs.
Technical differentiation. Jev is TypeSafe’s first large programmable/System 1 model, built for code to consume and optimized around intelligence per dollar rather than chat. TypeSafe describes RLCD, Reinforcement Learning for Calibrated Decisions, as a novel unpublished training direction that targets epistemically honest probabilities on System 1 tasks instead of human-preference optimization or benchmark-oriented RLVR. The company positions itself as a data lab: its North Star is programs-in-the-loop, its data is task-specific and largely synthetic, and it avoids user data to reduce bias and overfitting to present-day use cases. Its API introduces programming-oriented primitives—choice, score, and Noulli—that map to switches, thresholds, sorting, and conditionals, alongside structured JSON state and decomposition into many small, measurable decisions rather than giant prompts.
Market signal and diligence flags. The strongest ecosystem thesis is AI infrastructure that disappears into software: Almeida identifies dark-data analysis, coding agents, real-time intelligence, verification, computer use, and composable smart software as major use cases, calling dark data and coding agents the biggest potential money-makers. He expects an inverse SaaS-pocalypse in which existing SaaS is supercharged rather than displaced. He also argues coding agents are constrained by single-model architectures and KV-cache-heavy state handling, creating room for explicit state, subtask trees, shared memory, and multi-agent coordination. Diligence should account for TypeSafe’s rejection of public benchmarks, the unpublished RLCD method, acknowledged model errors, and reported degradation from single-hop to multi-hop tasks. Governance is another flag: Almeida rejects safety alignment and refusals at the general-purpose API layer and assigns downstream-use decisions to developers.
- Founding team and early financing: Whatnot’s founders brought relevant marketplace and technical backgrounds: Logan had run engineering at large resale marketplaces, while Grant had worked at Facebook and YouTube and had previously started a company. They built their YC product in about a week; Grant estimated roughly $250,000 raised through YC and a team of about 6.5 people before the seed round.
- Data-led marketplace wedge and infrastructure: The team selected Funko collectibles after analyzing eBay category sales, observing that collectibles were growing and Funko transactions exceeded comics. To overcome the marketplace cold start, Whatnot initially supplied and authenticated inventory, then built a pricing algorithm that scraped listings across sites, applied a liquidity floor, and combined recent sales with item liquidity to price inventory it could source. After launching live auctions in July 2020, the founder said growth exceeded 100% month over month; the company reached roughly $250,000–$300,000 in monthly gross sales around September and ended 2020 at $2.3 million in sales volume.
- Emerging investment theme: Grant forecasts live commerce could grow from a single-digit share of e-commerce to more than 30% over the next decade, citing China’s roughly 40% penetration as a comparison. He identifies wholesale marketplaces supplying live sellers and multi-channel networks that help operators run live-shopping businesses as adjacent infrastructure opportunities.
- Pocket Entertainment is a scaled proof point for AI-native entertainment. Rohan Nyak co-founded Pocket FM in 2018 and now runs its parent company, Pocket Entertainment. Pocket FM has crossed $500 million in ARR and reports more than 550,000 creators publishing on the platform. The founder also reports a sharp increase in AI-generated supply, reaching 2.6 million hours of AI content from an earlier level of roughly 25,000 uploaded hours.
- Its technical differentiation targets long-form generation rather than generic text generation. Pocket’s Sherpa fiction tool uses multi-agent orchestration and custom post-trained models; the company separately trains models for prose and story planning using 5 billion hours of minute-level engagement data, while a structured memory layer tracks story details and character progression. Pocket also uses an LLM judge to filter incoherent submissions and a staged distribution algorithm that expands a show from roughly 50 listeners to larger audiences when engagement signals are strong.
- The product paradigm is expanding from audio IP into localized, AI-produced video. Pocket Saga converts user-generated audio scripts into screenplays and coherent vertical microdramas, and the company says the app reached $15 million in ARR within three months. Nyak frames the broader thesis as AI making studio-quality entertainment accessible to anyone while human ideas remain the scarce creative input. A key scaling risk is that inference costs have risen multifold as content production expands, forcing continued investment in custom models and cost optimization.
- Unconventional AI is pursuing a radically different AI-compute substrate. Founder and CEO Naveen Ralph has electrical-engineering training and a neuroscience PhD; he previously founded Nirvana Systems, sold it to Intel, ran Intel’s AI group, and later built GPU-scaling infrastructure that joined Databricks in 2023. The company targets 1,000× greater power efficiency and says it has shortened its initial five-year goal to three and a half years. Ralph frames energy as a core AI bottleneck, estimating from Google’s reported token volume that one company could require 12 GW for AI services and forecasting an energy shortfall in roughly three years; these are his estimates rather than independently validated figures.
- The company’s approach replaces conventional separated memory and compute with a dynamical computer based on interconnected oscillators and emergent physical behavior. Its open-source UNO model demonstrated image generation using oscillator-based dynamics, while the team says it built and taped out the first physical dynamical computer in five months and generated its first images from the returned chip. Ralph reports roughly 500 nanojoules per generated image and describes the prototype as many orders of magnitude more efficient than standard computers. The architecture combines computation and memory in the same elements and uses both physical die-stacking dimensions and time dynamics, which the company calls “4D computing.”
- The near-term commercialization plan is a rack-scale, data-center product delivered as a managed VM within about two years; existing model families are expected to work, but customers will need to port at the model layer. Unconventional AI has built Python libraries for expressing time-varying stochastic elements, while Ralph identifies coordinating theorists, biologists, physicists, and chip engineers as a major execution challenge.
- Open-weight models have reached a commercial inflection point. GLM-5.2 and Kimi K3 are described as crossing an agentic-capability threshold comparable to Claude Code’s December 2025 level. Chinese open-weight models are estimated to be 2–5 months behind the closed U.S. frontier, while U.S. open models remain 6–9 months behind; Chinese labs are strongest in agentic coding and relatively weaker on open-ended scientific tasks.
- Usage is creating an open-model infrastructure and post-training market. OpenRouter’s weekly open-model usage rose from roughly 1T tokens in September 2025 to about 80T tokens in the article’s reporting period, with Chinese models increasing from approximately 70% to more than 80% of usage; OpenCode reports roughly 95% or higher inference volume from Chinese models. The author identifies inference platforms such as Together, OpenRouter, Fireworks, and Baseten, plus fine-tuning APIs such as Tinker, as early winners of an emerging open-model post-training economy.
- Chinese models are becoming a foundational layer for products and research. Harvey, Cursor, DoorDash, Airbnb, and Perplexity are cited as using Chinese models, while younger U.S. startups are increasingly entering enterprise agreements with Chinese model labs to use their models commercially. Qwen is mentioned in about 30% of current AI papers, and Chinese open-weight models collectively appear in more than 40% of papers versus about 30% for U.S. models.
- Agent-training data is an emerging startup opportunity. Chinese labs that initially preferred building training-data workflows internally began buying cutting-edge reinforcement-learning environments for agentic tasks from established U.S. companies and new Chinese startups by summer 2026.
- The opportunity carries meaningful policy and security risk. Open-weight models are increasingly capable of enabling cybersecurity applications, but there are few effective ways to prevent released software from reaching bad actors; the resulting regulatory response remains uncertain.
- A portfolio founder says VCs are using Granola and a poorly prompted version of Claude to analyze and “pick apart” founder pitches; the founder argues that AI can intensify groupthink and cause internet trends—whether true or not—to shape investor reactions, prompting founders to rebut fashionable narratives upfront.
- Garry Tan says Cluely’s concept remains promising as a real-time thought helper and semi-adversarial assistant with ongoing context, highlighting a product paradigm built around persistent-context, in-the-moment AI support.
- Public opposition to large data-center proposals is a potential siting and political risk for AI infrastructure: a quoted scenario describes developers proposing an enormous data center, facing public backlash, and then substituting a large housing project; Garry Tan calls the data-center panic “absurd and unnecessary,” while qualifying his view with “if true.”
- Capy.ai is presented as an agentic coding tool that tracks multi-step workflows and executes large pull requests faster than Codex or Claude Code alone, a reported competitive signal in AI-assisted software development.
- An example Capy.ai workflow on GBrain shows clear task delineation, automatic parallelization, and clean GitHub pull-request and CI integration.
AI systems benchmark proposal: Evaluate an AI by giving it a PC and an older operating-systems paper describing a novel OS, then requiring it to implement the paper in QEMU and operate/debug the machine from a separate computer over a serial port, including reboots. A more demanding version would package the paper and prompt in a directory and ask the AI to explain the hardware bring-up and debugging setup, creating an end-to-end test of software implementation and hardware interaction.
-
LlamaParse Extract introduced grounded confidence scores for schema-guided document extraction, covering unbounded
bool,int, andfloatfields as well as free-form text; each extracted value includes a bounding box linking it to the source document. - The scores support calibrated human-in-the-loop workflows: teams can set thresholds to automate higher-confidence values and route lower-confidence values for review. The capability is available on the Cost Effective, Agentic, and Agentic Plus offerings.
Martin Casado endorsed OpenAI’s work with an independent advisory group of mathematicians to responsibly assess and communicate advances in AI and mathematics, uphold academic and professional standards, and develop tools for mathematical research and learning; he praised the approach as community engagement “without histrionics.”
- a16z pairs its view that ChatGPT “changed the game” in 2022 with a sharp increase in median startup revenue four years after founding: $2.0M for the 2018 cohort, rising through $2.3M (2019), $2.5M (2020), and $2.8M (2021) to $5.6M for the 2022 cohort.
a16z reports that the “SaaSpocalypse” was “off to a slow start”: since the Feb. 22 Citrini post, software stocks were up approximately 27% for large caps, 30% for small caps, and 34% for mid caps.
- Humanoid robots’ payload-to-weight ratios appear to cluster around the cited human range: Boston Dynamics’ Atlas is listed at 90 kg carrying 30 kg sustained (0.33), Figure at 61 kg carrying 20 kg (0.33), and Apptronik’s Apollo at 0.34; the analysis argues this reflects the inherent cost of carrying the robot’s own actuators, battery, and frame rather than a humanoid-specific failure.
- The more important competitive gap may be between mobile and fixed robots: a bolted-down FANUC M-20iD is cited at a 0.10 payload-to-weight ratio, while Unitree’s G1 is cited at roughly 0.057—about six times worse than Atlas—and only about laptop-scale carrying capacity per arm. Cross-company comparisons remain difficult because Unitree quotes arm-at-extension, Atlas distinguishes instantaneous from sustained payload, Tesla provides no specification sheet, and Agility reports payload without robot mass.
- Corridor Brokers announced a $25M seed round led by BainCapVC to improve healthcare access for small-business employees; its thesis is that small businesses have been underserved because serving them was not economically viable, and the company aims to “change the math.”
- The investor backing the round said the team is transforming how health insurance is sold.
- The post advances a speculative labor-displacement thesis for trucking and adjacent mobility: it says the global truck-driver workforce exceeds 50 million and that a majority of these jobs could disappear over the next decade; it separately cites 3.5 million U.S. drivers plus delivery and ride-sharing workers, and flags 2028–2030 as a major disruption window.
A quoted post alleges that a “Chinese OJAI Waymo” went through road-closure signs and entered a Denver farmers market. Jason caveated the report with “assuming this is real,” noted that the vehicle slowed to a crawl, and said there is still work left to do—an autonomy reliability and safety caution rather than a positive technical breakthrough.