We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Coverage is incomplete: some monitored sources or documents could not be processed. This brief covers the available verified material.
1. Funding & Deals
Sumble is the clearest early-stage deal and commercial signal in the monitored material. A current SaaStr profile says Anthony Goldbloom and Ben Hamner built Kaggle, which Google acquired in 2017, before starting Sumble. Coatue led an $8.5 million seed and Canaan led a $30 million Series A; the profile lists AIX Ventures, Square Peg, Bloomberg Beta, Zetta, Marc Benioff, and Nat Friedman among the other participants. Sumble’s wedge is not another contact database: it crawls job postings, company sites, social sources, and filings, then uses LLMs to map technologies to teams, reporting lines, hiring activity, and live initiatives. Its MCP server, API, and warehouse integrations put that context into Claude, Cursor, ChatGPT, Salesforce, Snowflake, and Databricks.
The company reports 550% revenue growth, 19 enterprise customers since its April 2024 launch, and a roughly two-dozen-person team; its customer list is concentrated among technical-product companies such as Databricks, Snowflake, Figma, Vercel, Wiz, and Elastic. The investment thesis is that execution is becoming commoditized while proprietary inputs and distribution remain defensible—but sales intelligence is already crowded, so the data foundation and workflow adoption must carry the moat.
China’s Z.AI is a strategic-capital signal, not an early-stage comparable. An X post reports a $5 billion raise, with 60% of net proceeds earmarked for next-generation GLM models and a “Fully Self Training” system in which each model generation builds the environments used to train the next. The report attributes the recursive-self-improvement framing to Z.AI itself, but gives no stage or lead investor.
2. Emerging Teams
ClarSeal is targeting the security layer created by AI-built software. The builder is developing a scoped checker for apps made with Lovable, Bolt, or Cursor; it scans deployed applications for exposed secrets and configuration problems, then sells findings, fix guidance, and branded PDF reports for agencies. The product is explicitly not a full penetration test or security guarantee, and the current ask is to walk through scans with founders and freelancers launching or handing off apps within 30 days. That makes this a credible category wedge, but still a discovery-stage signal rather than validated traction.
A Sydney construction tool shows the difference between problem validation and adoption. A graduate civil engineer built a voice/text workflow that structures lessons learned and resurfaces them when teams encounter similar problems. More than 20 practitioners, including an experienced tier-one-contractor operator, agreed the problem was real and expensive; directors and senior engineers took calls. Yet nobody had committed to a trial despite a working prototype and a free offer. The next diligence gate is behavioral: a two-week, one-project pilot with a named champion, founder-led setup, and one before-and-after measure—not another round of interview agreement.
solveathome.org is an unusual but very early distributed-research formation. It pools unused Claude, Codex, and other agent tokens into bounded research tasks on the twin-prime conjecture, publishing results, reviews, transcripts, and attribution. The founder says the hard problem is filtering what is novel, known, or wrong; consensus review was too expensive, so the project is moving toward trusted reviewers. It is only days old, so the signal is a new agentic-compute paradigm, not investable traction.
3. AI & Tech Breakthroughs
The newest agent-security failures are in the scaffolding, not the model. A security post reports that default GitHub Actions configurations published for Claude Code, Gemini CLI, and Codex could each be triggered by a single unauthenticated GitHub issue to reach remote code execution. The reported failures included flawed shell-argument validation, an unenforced Gemini tool restriction rated CVSS 10.0, and a poisoned-instructions path through Codex’s shared writable checkout. A related Google ADK issue allowed a low-privilege agent to induce a gated, high-privilege agent to act with inherited write permissions. For investors, agent evaluation now needs to include vendor-recommended CI/CD templates, permission inheritance, poisoned instructions, and rollback—not just model behavior in a clean sandbox.
A more positive control-plane direction is WorkOS Airlock. An Acquired transcript describes it as checking whether an agent’s action matches the user’s original intent, rather than merely checking whether the agent has permission to call a tool. That distinction—“allowed to use” versus “asked to do this”—is becoming an enterprise authorization primitive.
Tahuna is pushing model operations down to small teams. Its open-source platform covers content-addressed code and data synchronization, compute provisioning, reproducible manifest-pinned runs, metrics, checkpoints, artifacts, and inference deployment. The public preview supports RunPod and R2, Docker self-hosting, SFT, RL agentic search, and an autonomous experimentation loop called Hillclimb. In parallel, Bindu Reddy’s team released the weights for Smaug Mini, a 27B model positioned for fast inference, while listing Smaug Flash API pricing at $0.10M input and $0.40M output. The pressure is moving from “can a startup access frontier capability?” toward “which workload justifies owning the stack?”
4. Market Signals
“Pacing the frontier” is becoming a development-cost and market-structure debate, not a proposal to stop progress. Sam Altman says OpenAI now formulates explicit safety cases before reinforcement-learning runs expected to increase capability, supports consistent federal safety requirements, independent auditors, and shared standards, and defines pacing as progress that is slower than it otherwise could be because safety cases and monitoring cost money. Khosla’s formulation is “oversight yes” but no slowdown that could let the US fall behind China. Bindu Reddy says open-source AI is accelerating and predicts the gap with frontier models could close by December if Anthropic and OpenAI slow down; that is a forecast, not an established result. Cohere’s framing—“evidenced standards, not a cartel”—captures the competing concern that safety coordination could become incumbent-controlled rulemaking. The underwriting question is therefore dual: price recurring safety and audit costs, while testing whether regulation raises barriers for new entrants or merely improves accountability.
Inference economics are forcing an ownership decision earlier in the company lifecycle. One current analysis puts AI-native application gross margins at 50–60%, versus 80% or more for traditional SaaS, with third-party token spend driving the gap; agentic tasks can make 30–200 model calls, creating usage and pricing volatility. The same analysis argues that self-hosting and distillation can reduce per-token costs by an order of magnitude and let production corrections improve weights a company owns, while startups still seeking product-market fit should rent and instrument usage. It places a rough ownership payback threshold at $2–5 million in annual inference spend over 18–24 months, and recommends a hybrid model for most companies.
Enterprise AI demand is broadening beyond coding. An Acquired transcript cites Anthropic data covering 1.2 million Claude Co-work sessions across 600,000 organizations: software development represented less than 9%, while business process and operations accounted for about one-third. Exponential View reports a qualitative survey in which about 95% of 250 IT executives said they had meaningful AI results and all planned to spend more, while Box’s survey found 83% of organizations already running agents, four in five reporting moderate or significant ROI, and half seeing impact within six months. These are directional survey signals, not a substitute for customer-level retention and margin data.
Vertical deployment is still labor-intensive. Harvey says roughly 180 former practicing attorneys, most with 8–10 years of experience, work in every deployment; its legal-engineering function spans pre-sales, post-sales adoption, and custom agent/playbook construction. The same profile flags $250,000-plus human deployment costs, hiring-pool constraints, ramp, retention, and management risk. That is a useful warning against underwriting vertical AI as pure software before deployment labor and gross margin are demonstrated.
5. Worth Your Time
Read — Sam Altman on frontier-AI safety requirements. The primary source for development-time safety cases, independent auditors, shared standards, and the claim that pacing slows rather than stops progress.
Watch — Acquired: Home Depot, the best-performing stock in the S&P 500 since IPO. Skip to the segments on intent-aware agent authorization and Claude Co-work’s business-process usage; they connect enterprise identity infrastructure to the shift from coding assistants toward operational agents.
Read — The Owning Phase of AI, Part 2. A practical framework for deciding when API rental should give way to self-hosting, distillation, or model ownership, with explicit attention to data uniqueness, token scale, evaluation costs, and gross-margin trajectory.
Read — Exponential View #601. Useful for the current enterprise-adoption signal and for separating qualitative demand evidence from claims that still require customer-level verification.
- Anthropic reported that 1.2 million Claude Co-work sessions across 600,000 organizations over three weeks were used mainly for business process and operations work—about one-third of sessions—with content and writing next, while software development accounted for less than 9%. Claude Co-work is positioned as a desktop agent that works in users’ actual folders and can build internal tools without waiting for engineering resources.
- WorkOS launched Airlock, an agent-authorization system that checks whether an action matches the user’s original intent rather than only whether the agent has permission to call a tool. The episode says OpenAI, Cursor, Perplexity, Sierra, Anthropic, and hundreds of other AI startups build on WorkOS, signaling demand for identity and authorization infrastructure for enterprise agents.
- Sierra positions its AI customer-service agents around outcome-based pricing rather than token usage, with support for complex workflows across voice, chat, email, and WhatsApp; it says its platform is used by 40% of the Fortune 50, one-third of the world’s leading banks, and five of the ten largest healthcare companies. Sierra also launched Personas for brand-specific agent voices and personalities, and says a telecommunications customer’s problem-solving rate increased by nearly 50% after introducing a new agent persona.
- Sentry says its Seer agent can identify software root causes and open a pull request for human review, extending AI-assisted debugging beyond code generation; Sentry reports that more than 200,000 organizations run on its platform.
- Frontier-AI safety tooling is moving into model development: OpenAI says it now prepares explicit safety cases before frontier reinforcement-learning runs expected to significantly increase capability, extending earlier responsible-scaling and preparedness work beyond completed-model deployment. Altman also calls for shared industry standards covering misalignment, monitoring, and safety, and cites independent auditors as a possible mechanism.
- Regulation and operating-cost signal: Altman supports consistent federal safety requirements for frontier AI and international coordination, while arguing that “pacing” should slow—not stop—progress because safety cases and monitoring carry significant costs.
David Ulevitch endorsed Flock Safety as a public-safety technology positioned to balance privacy protection, abuse reduction, and improved safety. The cited post reports that President Trump said he likes Flock cameras, providing a high-profile public signal for the category.
- OpenAI is moving frontier-AI safety beyond release-time safeguards toward development-time controls: Sam Altman says the company now prepares explicit safety cases before frontier reinforcement-learning runs expected to significantly increase capability, alongside its existing pre-release safety work.
- Altman supports consistent federal frontier-AI safety requirements, independent auditors, and shared industry standards for misalignment, monitoring, and safety; he says “pacing” should impose some cost and slow progress relative to unconstrained development. He also flags loss of human control and excessive concentration of AI power as the two risks the industry must avoid.
Paul Graham says that if model companies slow down, it would become “slightly more feasible” to start a new model company—a conditional signal that reduced incumbent momentum could create openings for new AI model startups.
Paul Graham says YC office hours effectively give founders essays customized to their individual circumstances.
- Martin Casado linked to a post reporting that China’s Z.AI raised $5 billion, with 60% of net proceeds funding next-generation GLM models and a “Fully Self Training” system. The stated thesis is a recursive self-improvement loop in which each GLM generation builds the environments used to train the next. The cited report does not specify the round stage or lead investor.
- Andrew Chen describes a local-first AI experimentation stack: a Framework Desktop Mainboard AI Max+ 395 hosting Hermes, a 5090 eGPU running Qwen 3.8 27B at sometimes 150+ tokens/second, two DGX Sparks running Deepseek v4 Flash 0731 for slower but better workloads, plus a Mac mini development box and Raspberry Pi monitoring.
- Hermes uses a homegrown router with Arch-Router to decide whether requests run locally or move to cloud/frontier models; the setup is currently about 60–70% local, with the goal of reaching 100% local. The DGX Sparks handle batch and background jobs.
- The post includes a practical caveat for local-AI infrastructure: Chen says the full setup is unnecessary, regrets buying the eGPU, and would not recommend it.
- Martin Casado shared Lina Khan’s warning that existing laws may already expose AI companies and CEOs to liability for releasing unvetted or defective models or agents, including through consumer-protection and competition rules.
- Khan also flags concentrated, interconnected AI partnerships and cross-investments as potential conflicts that can undermine accountability, and argues that inadequate data-security controls or failure to fix known vulnerabilities can create legal exposure; one cited analysis found roughly one-third of Fortune 100 companies lacked a simple way to report security issues.
Martin Casado endorsed Yann LeCun’s criticism of an unnamed “Dario,” whom LeCun says argued in 2019 that GPT-2 was too dangerous to open-source; LeCun now says people should mock that position. The exchange is a clear pro-open-source sentiment signal in the AI safety and model-release debate.
- Geopolitical AI risk: Vinod Khosla says it is “even more critical” not to get behind China, says he would not trust China even if it agreed to slow down frontier AI, and calls advanced AI in Chinese hands “far more dangerous than advanced AI”; the visible post ends mid-comparison.
- Frontier-market and regulatory sentiment: David Sacks’ quoted commentary characterizes OpenAI and Anthropic as a frontier-intelligence duopoly by market share, revenue growth, and model capability, while arguing they can “pace the frontier” without using that decision to demand a preferred regulatory framework; he also says China is unlikely to join a global agreement.
- Sacks further argues that frontier-AI restraint is partly commercially motivated by product-liability exposure and customer demand for reliable, predictable models, rather than being purely altruistic.
- Satya Nadella calls for AI’s benefits to be accelerated and broadly diffused, with a frontier ecosystem where closed and open-source models can thrive. He argues that enterprises should retain control of their proprietary knowledge, continuous-learning loops, and model weights rather than depend on a single model provider.
- Nadella also emphasizes human control and alignment as design goals, including embedded evaluators and broad representation across countries, fields, and academia; he says a code of conduct for first-party MAI models will be published for public consultation. Martin Casado explicitly endorses the acceleration and diffusion objective.
Martin Casado amplified a “Nationalize us” framing alongside a clip in which Dario says it is strange that the technology is being built by a private company and that he would give it to “the right combination of governments.” This surfaces government-control and ownership risk as an investor-relevant signal for frontier technology companies.
Regulatory risk signal: Ossoff called for rapid action on AI oversight, including inspections of frontier labs, bioterrorism safeguards, congressional legislation, and an international AI treaty. Martin Casado amplified the message with the sarcastic response, “Yes, we are happy to regulate you. Thanks for asking.”
AI-agent infrastructure faces material security risk: attackers reportedly stole a METR API key, consumed about $600,000 in credits over three weeks, exploited a fail-open bug that disabled Google authentication on a public agent dashboard, prompted an agent to reveal the key, and established SSH persistence.
Martin Casado endorsed an argument that AI-doom scenarios involving the creation of dangerous viruses are “bogus” because successful iteration would require access to the physical world; he added that rebuttals often rely on “a magic step” or technology that does not exist.
Martin Casado amplified a post alleging that Barack Obama urged Hakeem Jeffries to make AI regulation central to the Democratic agenda “once you are speaker,” framing the claim as an escalation in U.S. political attention to AI regulation. The linked post cites a New York Times article, but the supplied material does not independently substantiate the allegation.
- Enterprise AI demand is broadening: In a Las Vegas survey of 250 IT executives, roughly 95% reported serious, meaningful results from AI initiatives versus about 25% a year earlier, and every participant planned to spend more the following year. Microsoft was also reported to be planning growth in AI-serving capacity from about 2 GW to nearly 13 GW by 2032, within total fleet expansion from 12 GW to 38 GW.
- Agent adoption is already producing measurable enterprise returns: Box’s 2026 survey of more than 1,600 leaders found that 83% of organizations already run AI agents, four in five report moderate or significant ROI, and half saw business impact within six months of approving a project.
- AI-for-science capability is advancing sharply: OpenAI reportedly used 10,000 agents, 2.7 million messages, 130 billion tokens, and 88 hours to produce a solution to the Navier–Stokes problem; the author estimates the cost could fall from a few million dollars today to tens of thousands within two years and eventually a few dollars. The resulting proof exceeds 500 pages and, according to the article, would not be intelligible to humans.
- Frontier-AI concentration creates transparency and entry-barrier risks: The article flags an allegation that OpenAI deployed an internal team and model on the same problem after a year of external Codex work without clear disclosure about overlap or data use, while frontier labs increasingly operate in secrecy. It also describes OpenAI and Anthropic as agreeing to “pace the frontier,” including proposed employee-level access for independent evaluators, and warns that coordination among already oligopolistic labs could slow development and raise the cost of entry for future competitors.
Martin Casado cautioned against accepting near-term claims that Dario Amodei believes human lifespan could be doubled and cancer cured within a few years, replying: “Let’s pace that shit for sure.”
Martin Casado flagged an emerging AI-governance risk: reacting sarcastically to a reported proposal by Bernie Sanders for a 20-year prison sentence for AI developers pursuing Artificial Superintelligence—compared in the referenced post to illegally developing rogue nuclear weapons —he framed the policy as punishment for developing technology described as carrying a 10% chance of wiping out humanity .