We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Airbound announced a $37M Series A led by Greenoaks to make “all movement” airborne. The company rejects the premise that flying is inherently more expensive; investor Leo Polovets called its manifesto unusually clear and described Airbound as a company to watch for the future of logistics and transportation. For a VC, this is a transport-tech thesis bet on modality and system-level economics; the underwriting question is whether the company can turn that thesis into lower-cost, broadly deployable movement.
2. Emerging Teams
Suhail’s still-unnamed venture is a high-compute bet on autonomous AI research. The founder’s build log says the venture has secured seed funding, is building an “autonomous AI scientist,” and has validated a basic RLVR post-training stack. The team grew from one to three, made its first hire, is looking for another specialist in post-training or low-level model optimization, and says it acquired 64 B300s. A current update says the software around the harness is becoming “the new browser” and that supporting customers requires testing multiple implementations to ensure their APIs work consistently. The signal is unusually concrete resource commitment, but the financing terms and company identity remain opaque; the diligence test is whether automated post-training produces repeatable gains rather than simply consuming scarce compute.
FretTrack is a clean vertical-SaaS signal: a guitar-repair operator turned an internal tool into a tenant-isolated product and got its first paid annual subscriber. The founder started from direct workflow knowledge, rebuilt the product around a real database, authentication, shop-level data separation, permissions, subscriptions, and migrations, and had a UK shop process an actual customer job. The founder explicitly distinguishes the payment from beta compliments or sign-ups, while noting that AI accelerated development and also created technical debt that had to be addressed with security and reliability checks. The investable lesson is domain-specific workflow ownership plus willingness to pay, not a generic AI wrapper.
Applied Compute released AC2 in private beta with a “model factory” thesis. It argues that fast model progress makes attachment to one set of weights less valuable and positions AC2 as infrastructure around models to train, run inference, and continuously improve them on a team’s target work. This is an infrastructure thesis rather than proof of traction, but it is a useful early signal that applied-AI teams may buy continuous model operations instead of a one-time model choice.
3. AI & Tech Breakthroughs
Perplexity’s Portable Computer makes local-first agents a full-stack operating mode rather than a smaller-model demo. It launched on NVIDIA DGX Spark with the orchestrator model, subagent model, and harness running on local hardware; the primary research describes the model, harness, conversation, and trajectory as local by default, with web search, connectors, and stronger-model escalation gated by the user. It also argues that the model and harness must be co-designed. The boundary is implemented, not merely promised: tools run in an OS-level sandbox, the harness disables itself if isolation is unavailable, and a deterministic orchestrator retains authority over approved tool calls.
The reported results support a hybrid rather than fully offline thesis. On a 53-task knowledge-work benchmark, base Qwen 3.8 27B in the Computer harness scored 82.6%, versus 77.6% for Pi and 74.0% for Hermes; post-trained PPLX 27B reached 85.4%. On harder coding tasks, user-approved escalation to Claude Opus 5 raised the score from 59.6% to 73.0% at an estimated $0.415 per rollout, versus 82.4% and $0.65 for Claude alone. The implication for investors is a product and infrastructure wedge combining local privacy and cost control with selective frontier access—not the immediate elimination of cloud models.
Ramp’s Inspect shows why enterprise coding agents are becoming internal platforms. Inspect combines remote sandboxes and unlimited concurrency with internal tools and context, then verifies backend changes through tests, telemetry, and feature flags and frontend work through screenshots and live previews. Ramp reports that 75% of merged PRs come from Inspect sessions, the system has passed one million sessions, its team numbers 5.5 people, more than 150 Ramp engineers have contributed, and more than 80% of Inspect itself was written in Inspect sessions; more than 200 additional agents run on the platform. This is internal-adoption evidence rather than a startup KPI, but it points to context, verification, and orchestration—not code generation alone—as the durable control-plane layer.
Evaluation tooling is moving from static benchmarks toward executable task factories. LangChain’s eval-engineering process separates human judgment about what a task should measure from agent-automated task construction, using versioned specs, real-agent trajectories, and multiple model tiers to catch design flaws and calibrate difficulty. The authors frame continuously refreshed environments built from production data as infrastructure for prompt tuning, harness tuning, post-training, and deciding where cheaper models are sufficient. In parallel, LlamaIndex’s ExtractBench tests 14 systems across 370 enterprise documents, 67 document types, and more than 4,800 pages, explicitly targeting the messy forms, nested tables, and long reports that clean-invoice demos avoid. The investable opportunity is evaluation tied to representative work and operating cost; the benchmark announcement itself is not evidence that any extractor has won.
4. Market Signals
Enterprise AI is reallocating software budgets before it creates a new vendor category. SaaStr’s account of a 141-CIO Redpoint survey says 45% of respondents fund AI from existing software budgets, 54% are running vendor-consolidation programs, only 3% expect AI to create more vendors, and 58% say AI feature additions are the leading driver of software-spend increases. The same survey says 46% expect usage- or outcome-based pricing to become more common and 29% expect seat-based pricing to decline.
That repricing creates concrete product requirements: consumption vendors need buyer-set hard caps, rollover blocks, threshold alerts, and an agent-readable budget API; resolution pricing instead bills for a countable completed result rather than an API attempt. Allie Miller’s reported examples of F500 and digital-native businesses imposing monthly token limits—from roughly $75 to $3,000 in cited cases—come from online threads, so they are directional, but they reinforce that AI spend governance is becoming part of procurement rather than an afterthought.
Agent security is shifting outside the model. A post summarizing UK NCSC guidance says containment should match autonomy, deployments should choose among different human-oversight modes, sandboxing should be layered, and activity should be logged with attribution; it also says model-level safety training can be bypassed once an agent has tools, credentials, and a goal. A practitioner describes the corresponding runtime pattern: middleware tracks token spend and revokes scoped API keys or tool access mid-run, with process termination as a last resort. That makes identity lifecycle, sandboxing, observability, and kill-switch infrastructure plausible investment wedges, while the underlying security post is still a secondary summary and the implementation account is anecdotal.
The capital layer is concentrating even as local products try to decentralize inference. Steven Sinofsky and Martin Casado describe computing as “capital-bound,” arguing that a team of about 20 can put $1B to productive use; Casado says he underestimated how long scaling laws would hold and warns that a $100B training run could concentrate resources in ways whose consequences are difficult to control. At the hardware boundary, Sarah Guo’s formulation is that AI can compress chip-design cycles but cannot fix supply, while Bindu Reddy argues that OpenAI, Google, and Amazon’s proprietary chips—and an expected Anthropic chip—are pushing Nvidia to build a broader AI-startup ecosystem. The resulting barbell is clear: open-weight and local-first systems can reduce inference dependence, but power, silicon, and capital remain concentrated bottlenecks.
5. Worth Your Time
Read — Perplexity’s Portable Computer research. The primary write-up has the sandbox boundary, on-device document results, the 53-task knowledge-work benchmark, and the user-gated advisor cost tradeoff needed to separate local-first architecture from “fully offline” hype.
Read — Why Ramp built Inspect. The useful detail is not the coding-agent label but the remote execution, internal integrations, verification loop, and adoption metrics behind an internal agent platform.
Read — How we Build Agent Environments & Tasks. This is a practical blueprint for separating human-reviewed task specs from automated environment construction and keeping evals aligned with production behavior.
Read — The 3 New Pricing Models in B2B. It connects budget reallocation to consumption controls, outcome pricing, and resolution pricing, with concrete implications for new entrants versus installed-base vendors.
Watch/listen — the lab-economics discussion. Treat it as scenario thinking rather than market data, but it is a useful frame on compute concentration, inference-to-training shifts, capex, and the possibility of centralization around a few labs.
Direct answer: The source describes Portable Computer as a local-first, user-gated system: the model, harness, conversation, and trajectory run on the user’s machine by default, while web search, connectors, and stronger-model escalation are optional off-device actions. It reports strong local results on research, multimodal document understanding, and a 53-task knowledge-work benchmark, with advisor escalation improving difficult coding performance at a stated API cost.
- Local architecture and privacy: Local inference and private-document processing are on-device by default; external services are invoked only when necessary and with user permission. The source says sensitive data does not leave the device without permission and that local inference carries no per-token inference fee.
- Harness design: The harness is co-designed with the local model to compensate for weaker long-horizon and context-handling performance. It uses a minimal prompt and core tool set, loads task-specific skills on demand, supports context compaction, and converts commonly used MCP connectors into compact command-line tools.
- Execution and boundary controls: The orchestrator is deterministic harness code rather than an LLM; it assembles context, enforces policy, and executes approved calls. Tools run in an always-on OS-level sandbox restricting processes, filesystem paths, and network access; if sandboxing is unavailable, the harness disables itself rather than running unsandboxed.
- Common benchmark setup: The harness comparison uses the same Qwen 3.8 27B model with medium reasoning on an NVIDIA DGX Spark, isolating harness effects before model post-training. For mixed private/public work, local files remain the authoritative source while web search adds context, and web search can be disabled for fully offline work.
- BrowseComp methodology and result: On 1,266 BrowseComp tasks, Computer uses Perplexity Search as Code while Pi and Hermes use Brave, their recommended provider. Computer scores 66.7% accuracy versus 50.2% for Pi and 43.9% for Hermes.
- BrowseComp efficiency and caveat: Reported mean wall time and token use are 402.1 seconds and 852k tokens per Computer task, versus 1,020.9 seconds and 1.01 million tokens for Hermes and 826.0 seconds and 2.82 million tokens for Pi. The chart notes that incomplete outcomes score zero and that time/token averages exclude rollouts without recorded measurements; score whiskers are 95% Wilson confidence intervals.
- Document-work methodology and privacy claim: ParseBench-100 is a 100-task subset with 20 tasks each for charts, layout, tables, text content, and formatting. The harness passes document pages and images directly to the multimodal model and combines visual evidence with extracted text; the source says on-device processing keeps sensitive documents and extracted content private.
- Document-understanding results: Computer scores 65.1% mean on ParseBench-100, versus 34.6% for Hermes and 13.9% for Pi, and averages 60.6 seconds and 20.1k tokens per task, versus 108.3 seconds and 32.1k tokens for Hermes and 410.5 seconds and 829.1k tokens for Pi. It leads all five categories, with the largest advantage on charts, while layout remains difficult for all three. The reported category scores are Computer: 76.5% charts, 16.2% layout, 72.7% tables, 87.9% text content, and 72.4% formatting; Hermes: 29.3%, 2.9%, 44.1%, 61.5%, and 35.2%; Pi: 2.5%, 0.1%, 11.0%, 29.7%, and 26.1%, respectively.
- Hybrid cloud escalation mechanics: A local model can call an advisor for planning, ambiguity resolution, repeated-failure recovery, or final-result verification. The user chooses whether escalation is enabled and whether calls require manual or automatic approval. Before a call, the harness selects relevant context, applies a PII classifier, shows the user what would leave the device, and sends only approved context; the advisor returns text guidance and has no direct access to local files, tools, or conversations.
- Escalation benchmark and scope limitation: Terminal Bench 2.1 contains 89 coding-agent tasks. The comparison uses Qwen 3.8 27B locally, Claude Opus 5 alone in the Computer harness as the frontier baseline, and Qwen 3.8 27B paired with Claude Opus 5 as the hybrid system. Pi and Hermes were not tested with escalation because they lack an equivalent advisor tool, so adding one would no longer represent their off-the-shelf harnesses.
- Escalation score/cost tradeoff: Advisor escalation raises Computer from 59.6% to 73.0%, a 13.5-point gain, at an estimated $0.415 API cost per rollout. Claude Opus 5 alone scores 82.4% at $0.65 per rollout. The source characterizes escalation as recovering roughly three-fifths of the frontier gap at about two-thirds of the frontier cost; it also states that fully local inference has virtually no runtime API cost because it occurs on the user’s hardware. These are reported API costs per rollout, not a complete hardware or ownership-cost analysis.
- Knowledge-work benchmark methodology and scores: The Local Knowledge Work Bench is a held-out set of 53 representative day-to-day tasks spanning seven categories; the benchmark reports three trials per task and 95% confidence intervals. With base Qwen 3.8 27B, Computer scores 82.6%, versus 77.6% for Pi and 74.0% for Hermes; PPLX 27B, post-trained for the Computer harness, raises Computer to 85.4%.
- Knowledge-work efficiency and coverage: On that benchmark, base Computer uses 520k tokens per task versus 681k for Pi and 634k for Hermes; Pi is fastest at 176 seconds, versus 218 seconds for Computer and 292 seconds for Hermes. PPLX 27B uses 678k tokens and has an estimated 250-second wall time. The 53 tasks cover deep research; data, finance, and procurement; documents, presentations, and design; engineering, IT, and incidents; contracts, evidence, and compliance; dashboards, software, and visualization; and people, projects, and meetings.
- Overall caveat: The source says Computer matches or exceeds Hermes and Pi across the reported benchmarks, but local Qwen 3.8 27B still trails the frontier model on Terminal Bench 2.1; escalation narrows rather than fully closes that gap.
- China’s industrial and investment trajectory is shifting toward vertical integration: investors who once copied American VCs are now running state-backed deeptech playbooks; battery-cell companies are expanding into grid, datacenter, and asset-owning businesses; and former contract manufacturers are becoming brands.
- Shenzhen offers an unusually dense hardware ecosystem: a local contact described 100,000 people making something new each day, unmatched supplier and supply-chain density, and strong university manufacturing programs. The same guidance flags structural risks for foreign operators, including lack of access to government subsidies, IP exposure, and supplier overcharging unless investors understand the full production process.
- The author frames China’s Greater Bay Area and infrastructure buildout as a long-term competitive force, citing more than $1 trillion in planned infrastructure spending over the next decade.
- Persistent memory across multiple AI harnesses is a potential tooling theme: Klaus Byskov reports that integrating Garry Tan’s GBrain across Claude Code CLI, Grok Build, and Codex was his biggest productivity boost, reducing the need to argue with models or repeatedly remind them of context; he says users without permanent AI memory are missing out. Garry Tan frames the concept as “cross-harness memory maxxing.”
Sam Altman stated, “we made a chip and it is fast,” signaling a possible chip or AI-infrastructure development. The post provides no company attribution, architecture, benchmark, funding details, or development stage, limiting its value for evaluating an investable opportunity.
AI is shifting startup economics toward capital- and compute-backed challengers. Martin Casado argues that AI removes the distribution and demand advantages historically held by incumbents, while access to sufficient capital lets startups reach competitive footing with Microsoft and Meta; he points to token/GPU demand and the rapid growth of Cursor, Anthropic, and OpenAI as examples. The accompanying a16z discussion claims that teams as small as 20 people can put $1 billion to work productively, while warning that the capabilities of a model built with $20 billion are difficult to predict.
- A highlighted Grok use case is a cross-application personal AI agent that analyzes email, calendar, and Slack, then suggests actions intended to improve the user’s happiness; its example recommendation was to prioritize time with friends and family.
- The idea received an enthusiastic endorsement from the post author, who wrote “Man I love this” and “Shamelessly stealing,” signaling interest in this agentic product paradigm.
- AI is described as changing startup economics: as few as 20 people can put $1 billion to work productively, AI addresses the distribution constraint that historically kept startups small, and challengers now have a level playing field with Microsoft and Meta.
- Startup formation is becoming faster and less infrastructure-intensive: founders can launch without building a data center, egress infrastructure, or telecom connectivity and can be operational within the first hours. Sinofsky also argues that large-company culture remains structurally resistant to crushing startups because incumbents often focus on one another instead.
- Capital-scale AI carries major forecasting uncertainty: the discussion argues that nobody can predict the capabilities of a model built with $20 billion.
Steven Sinofsky and Martin Casado argue that AI is shifting startups from an engineering- and distribution-constrained environment toward a capital-driven one: teams as small as 20 can deploy $1B productively, AI removes the distribution bottleneck that historically kept startups small, and challengers can compete with Microsoft and Meta. The discussion also emphasizes that the capabilities of a model built with $20B remain difficult to predict.
- Suhail’s new venture has secured seed funding and acquired its domain/name . Its technical direction is an “autonomous ai scientist” for new optimizations , with a basic RLVR post-training stack already validated .
- The team has expanded from one to three , made its first hire, and is seeking another specialist in post-training, including RLVR/OPSD, or low-level model optimization . The project began with two 8xB200s , later acquired 64 B300s , and secured much greater quantities of compute ; GPU loss and networking issues have delayed scaling .
- The software around the harness is described as “the new browser,” and the team says it needs to test multiple implementations to ensure the API works like testing IE and Firefox .
- a16z frames AI as enabling a “fat startup” model: Martin Casado argues there is now a discipline for small teams to deploy very large sums productively, with engineering no longer serving as the same natural limiter; he also presents increased private-market capital as potentially market-expanding rather than zero-sum.
- The discussion claims that roughly 20 people can put $1 billion to work productively, that AI removes the distribution constraint that historically kept startups small, and that challengers can compete on a more level playing field with Microsoft and Meta; it also highlights uncertainty about the capabilities of a model built with $20 billion.
Aravind Srinivas described an early demo of “Portable Computer” on DGX Spark, shown to Jensen; Jensen gifted the team a DGX Station. He says the local machine can serve frontier models such as GLM 5.3, positioning “unmetered frontier intelligence” on users’ own hardware as an upcoming product paradigm.
- The discussion frames AI as a fundamentally new, human-level abstraction layer: users delegate reasoning itself rather than merely specifying tasks that compute, networks, and storage execute.
- The speakers argue that AI is changing startup economics by solving distribution constraints, enabling teams of around 20 people to deploy $1 billion productively, and giving challengers a more level playing field with Microsoft and Meta; they also caution that the capabilities produced by a $20 billion model investment are difficult to predict.
A product-market and retention heuristic emphasizes hourly active users (HAUs), not just daily active users (DAUs): products with high-retention HAUs are claimed to have trillion-dollar potential.
- Capital intensity is becoming a defining AI market shift: Steven Sinofsky and Martin Casado characterize computing as moving from engineering-bound to capital-bound, arguing that teams of about 20 can put $1B to productive use, AI reduces the distribution bottleneck that kept startups small, and challengers can compete on a more level playing field with Microsoft and Meta.
- Scaling-driven upside comes with concentration risk: Casado says he underestimated how long scaling laws would hold and now questions the implications of a potential $100B training run; he says concentrated resources could be directed toward beneficial goals or weapons and calls the concentration potentially dangerous.
- AI-driven workforce transition is creating demand for career-navigation products. The newsletter cites World Economic Forum estimates that by 2030, 39% of today’s skills will change or become obsolete, with more than 90 million jobs displaced and 170 million new jobs created; it argues that AI is also redrawing role boundaries and day-to-day responsibilities.
- Vocation is a career-navigation platform built by Cliff Maxwell and Bobby Moesta. Maxwell’s background spans CPU building, education research, product management, venture capital, and serving as Chief of Staff to Clayton Christensen; Moesta co-created the Jobs to Be Done framework. The product is informed by more than 1,000 interviews with professionals changing jobs and hundreds of coaching engagements, and frames the individual as the customer buying their next career move.
- A SpaceXAI engineer identified in the quoted post as a former Cursor employee says they run 10–20 GrokBot agents that automate 90% of routine work, including a “Chief of Staff” agent that manages the other bots; the post presents a 50-minute podcast on building this kind of 24/7 agent team.
- a16z’s Martin Casado gave the discussion a strong endorsement, calling it the “single best discussion ever recorded on coding with agents.”
An anecdotal agentic-commerce use case: a bot scans wine-auction sites, bids automatically when a deal meets a bargain threshold, and texts the user for pricier purchases. This points to a consumer product paradigm combining autonomous transactions with human approval for higher-value actions.
- Martin Casado and Steven Sinofsky argue that AI is shifting computing from engineering-bound back to capital-bound: a 20-person startup could deploy $1 billion productively, changing the relationship between capital, innovation, competition, and defensibility.
- The discussion frames AI as removing distribution constraints and giving challengers a more level playing field against Microsoft and Meta, while noting that the capabilities of a model built with $20 billion remain unpredictable.
- Airbound announced a $37M Series A led by Greenoaks, with a thesis of making “all movement” airborne and rejecting the assumption that flying is fundamentally more expensive than other forms of movement.
- Leo Polovets highlighted Airbound as a company to watch, praising its manifesto as unusually clear and inspiring for the future of logistics and transportation.
- The post surfaces a product thesis for voice-first, proactive, goal-oriented AI agents that use smarter routing, work without an always-open laptop, present well-designed escalations instead of relying purely on delegation or chat, and reliably complete real-world tasks such as medical and personal-service appointments.
- It highlights infrastructure needs around local AI workload limits, on-demand provisioning of hundreds of VMs, smarter orchestration, eliminating manual context compaction, and more reliable MCP integrations.
- Emerging agent-product opportunities include concise and natural communication, visual rather than code-like interfaces, harsh self-evaluation, multiplayer AI, and “agent fusion” that combines one person’s Claude with someone else’s.