We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Security is now a basic assumption
A new case for the "vulnpocalypse" thesis arrived this period. Sarah Guo reports that eMudhra's intermediate CA private keys were tiny numbers like 4 and 56. Anyone could recover them in seconds and mint a trusted certificate for any site. Browsers have already blocked the bad certificates, but many servers, containers and devices still trust them. She traces the problem to a single bad key-generation script from 2025, which is now forcing a "Heartbleed-scale cleanup" . Asked what share of token use will go to security in 2027, Guo relayed @dok2001's answer: ">50%" .
Leadership moves point the same way:
- Aleksander Madry is back at OpenAI after what he now calls a sabbatical. He joins @merettm because, in his words, getting safety and security right "has to become a central priority" .
- Dario Amodei defended Anthropic's "race to the top" strategy. He said Anthropic had Mythos-class models in February but held general availability until June, when stronger cyber safeguards were ready, and probably gave up "a significant chunk of a several-month model lead." He also said competitors sell against Anthropic by pointing to its strict bio classifiers .
- Clem Delangue said the Netflix documentary on the OpenAI agent attack on Hugging Face left out the lack of agent containment and the fact that closed-model safeguards blocked Hugging Face's defense. Hugging Face defended itself with an open model . This continues the guardrail-refusal story from earlier briefs.
Market sizing angle: Vanta CEO Christina Cacioppo says AI attackers mostly chain many low-severity weaknesses rather than find zero-days. That pushes defenders toward fixing at machine speed . She sees a constant pentest-and-patch product for customers who won't pay $150K for Wiz but "very well could pay 40 or 50" . Separately, Vanta is absorbing token costs rather than marking them up. About 25% of its customers use the agent flow, and it is aiming for 75% .
Where capital went
- Arda (Bob McGrew) launched publicly with a $160M Series A at a $1.5B valuation. It has spent a year building with Anduril as its design partner. Today it helps manufacturers run manual assembly lines; next it plans reconfigurable automated lines aimed at products now made by large workforces in Asia .
- Daytona raised a $60M Series B at $1B for its "Agent Cloud," led by FirstMark. Angels include Patrick Collison, Tobi Lütke and Aravind Srinivas. The team has 35 people .
- Eka Robotics: a16z led the Series A. Eka builds a Vision-Force-Action model co-designed with its own robotic hands. The founders are Pulkit Agrawal (MIT CSAIL) and Tuomas Haarnoja (formerly Google DeepMind) .
- OpenAI acquired Orchestra Bio, a drug-development coordination platform, to build on its Rosalind life-sciences models .
- Salesforce is buying Braze for about $5.66B EV, a 58% premium . Six weeks earlier, Braze's stock fell 25% after a beat-and-raise quarter with 26% growth. SaaStr's takeaway: companies growing in the 20s with little AI revenue may get their best price from a strategic acquirer, not a next-round lead .
- Workday is acquiring C1.ai, which governs access for both employees and AI agents under one lifecycle and audit trail .
Founder flows seem to agree with these rounds. One widely shared observation says smart young builders are moving from vertical agents into data/RL environments, physical AI, and wet-lab drug discovery . Nathan Benaich replied "that checks" (anecdotal).
Physical AI: design gets easy, atoms stay slow
a16z argues AI will soon become a superhuman hardware designer, but building, testing and iterating will remain bottlenecks . It names opportunities in simulation (Navier, Inviscid, Lanyon) and in suppliers and test software (Hadrian, Cofactr, Instrumental, Revel) . Pat Gelsinger gave a chip example: AI could produce a design in three months, but silicon processing takes nine . On robot data, Mecka's Josh Gao says the field is now at millions of hours, still "many orders of magnitude" short of autonomous driving . He expects pick-and-place and packaging tasks on assembly lines within one to two years .
Compute and cost
Sundar Pichai says Gemini 4 Argon processed nearly 25T tokens in its first 72 hours, about 3x Gemini 3.0, and is capacity-constrained . Google gives DeepMind first call on compute, and its cloud business is forgoing short-term revenue as a result . Customer requests for 2027 compute are "incredibly difficult" to fill . He grants that an overbuild could happen if any of roughly 2,000 components becomes a constraint . Internally, Google reports 30% more code changes and about one-third fewer rollbacks in Argon's first four weeks . SF Compute's compute market is still thin: 5% of its capacity is on the order book .
Prices keep falling. OpenAI's Karan Singhal says GPT-6 Luna beats the model that was OpenAI's best a month ago on its health benchmark, at 34x lower cost . Among the companies The Pragmatic Engineer surveyed, AI spend on API plans has roughly halved since June–July peaks, helped by moves to open models. Some subsidized subscriptions are ending, though, and those bills could rise 5–10x .
Open models and policy
The DHS order bars DeepSeek, both the company and its models, from federal civilian procurement. It is unclear how the order applies to fine-tunes or to models trained on DeepSeek outputs . DeepSeek then open-sourced DeepInfer, its internal inference stack; one estimate puts DeepSeek's token gross margin as high as 90% . China's MOFCOM proposed, but has not enacted, pre-approval for sharing frontier AI abroad, with automatic approval for open-source .
Local and sovereign deployment also moved forward. Perplexity Computer now runs fully on a Mac, with both the model and the agent on-device . Mistral's Voxtral 3 adds speech recognition, text-to-speech and a voice-agent SDK, available through its API or on customer infrastructure .
Agent infrastructure
Ramp tested several off-the-shelf agent harnesses before building its own model-agnostic one, Laminar . LangChain packaged the pieces its internal agents kept needing into Managed Deep Agents: per-user identity, scoped memory, sandboxed code execution, and scheduled runs .
Briefly
- Waymo reports 0.01 serious-injury crashes per million miles versus 0.21 for human drivers, across 271M miles. Semafor reports Waymo plans to grow from about 4,000 to 20,000 cars by end-2027, funded by a $5B loan .
- Microsoft exited Nscale's £2bn Essex data center after its grid connection slipped to the 2030s. Nscale says it will still build .
- An NBER paper estimates California's billionaire tax would raise about $30B but have a mean net present value of –$39B once departures are counted .
- Reflecting on her startup experience, Elizabeth Yin says founder operations have become less tedious: she contrasts manually maintained cap tables and a 2011 convertible-note fundraising process involving lawyer costs and physical checks with cap-table software and e-signed SAFEs today. She argues that reduced admin gives founders more time to build.
- In a Mercury-partner post, Yin promotes Mercury Payroll, saying it handles tax registrations and payroll compliance and can prefill information from a Mercury account. Her example of the problem is multi-state hiring administration: setting up a Massachusetts employee required state registrations, forms, and workers’ compensation, and she says she had to repeat the process after he moved to South Carolina. Mercury’s disclaimer says employers remain responsible for legal and regulatory obligations, while some services are provided by affiliates and third parties.
- The founding team brings nontraditional robotics experience: Josh says he previously built back-of-house restaurant software and installed kiosks, while his cofounder built a stablecoin venture that was sold to Coinbase.
- Mecca positions itself as a creator, processor, and licensor of training data for robotics foundation models. Its collection model combines consumer apps and business partnerships with custom sensing hardware, logistics, and data processing and quality gates. The founder says differentiation comes from verified data quality, hardware, operations, and software—not primarily contributor payments.
- The founder describes robotics data as having reached millions of hours, but still many orders of magnitude behind autonomous-driving data; manipulation also involves more degrees of freedom than driving. He expects commercial capability to advance task by task rather than through a single all-or-nothing breakthrough. He sees a mix of human and egocentric data, teleoperation, and in-domain data, with tactile data and simulation still to be added.
- For new manipulation use cases, he expects constrained industrial tasks to lead, citing data-center maintenance, laundry folding, and packaging/pick-and-place; he forecasts packaging and pick-and-place tasks on assembly lines within one or two years. He cautions that homes require particular diligence on safety, while describing service, cleaning, and delivery robots as already deployed in hotels in China and Southeast Asia. Mecca sees itself eventually extending from data provision into domain-data collection, post-training, and enterprise deployment support.
- OpenAI says more than 300 million people use ChatGPT weekly for health and wellness; most are free-tier users and outside the U.S., and many use it between visits for care navigation and follow-up. OpenAI says it specializes general frontier models with healthcare data at each training stage, shapes evaluation and training with a global physician network, and does not use connected personal health information for model training or ad targeting.
- OpenAI said its newly released lowest-cost frontier API model outperformed its previous best on HealthBench Professional at 34 times lower cost, while noting that the models still need improvement.
- OpenAI described two-way ChatGPT–Epic integration; UCSF is an early adopter focused on uses such as chart review, and AdventHealth reportedly achieved roughly 80% reductions in some administrative workflow costs using AI. OpenAI’s health lead said AI has not yet had a macro effect on total healthcare spending, with savings dependent on integration, workflows, deployment, and scale; he characterized OpenAI’s role as enabling care partners rather than providing all care itself.
- OpenAI sees healthcare as ripe for agents and computer-use tools despite fragmented, legacy systems, and sees potential to mine underused EHR, claims, and longitudinal data for new insights. Its health AI lead forecast that AI-guided target or molecule generation and clinical-trial prioritization and validation could contribute to new and more personalized treatments over a 5–10-year horizon.
- Vanta’s integrations with 400 tools feed a security “trust graph” used for automation and agents; its agent can surface work such as offboarding accounts, then act with user permission. About 25% of customers are in this agent workflow; Vanta is absorbing token costs to increase usage and retention, because customers compare marked-up tokens directly with model prices. The CEO says output- or outcome-based pricing is attractive but hard to define, while token credits can look like obscured token charges.
- Vanta’s team uses security and compliance experts to write task-specific agent evaluations and currently prioritizes model quality over cost; the CEO says coding agents need older codebases modularized to work effectively. A new onboarding flow has users provide messy files and screenshots for an LLM to interpret and turn into setup, replacing a conventional wizard.
- The CEO argues AI attackers can chain individually low-severity weaknesses, increasing pressure for faster fixes; she sees continuous testing and remediation as an opportunity, including for customers below the large-enterprise segment served by established security vendors. Her early-stage market read highlights cyber and go-to-market technology, calls robotics underweighted as models improve, and says AI is increasing demand and budgets for government and enterprise modernization. She also sees the ecosystem around open-source model providers as less developed than confidence in those providers would suggest.
- The post argues that AI is making software faster and cheaper to build, potentially opening startup opportunities in narrow industry and workflow niches that were previously uneconomic. Its forecast is for many more specialized products, with understanding customer problems becoming more valuable than implementation skill; it also expects most products to fail and few to become billion-dollar companies.
- A commenter predicts that AI agents and workflows built into CRMs and app builders could shift B2B SaaS from license pricing toward usage-based credits; custom tools might cost a fraction of generic products, but could also create low-quality software and switching friction.
- Adoption caveats raised in the discussion include learning curves, the need to follow strict processes using real-project expertise, and the risk that internal custom apps will be poor—constraints on the idea that cheaper software automatically displaces packaged SaaS.
ValsAI released ClaimBench, a benchmark for testing whether decision models can determine if a financial document supports, contradicts, or does not address a claim. It evaluated 15 models on 1,960 claims paired with 10-K excerpts, and ValsAI reported Decider V1.2 as the leading decision model . Perplexity CEO Aravind Srinivas characterized Perplexity Decider as Pareto-optimal on the benchmark .
- The hosts describe Theranos as a medtech cautionary case: Edison was presented as able to run about 100 tests from a tiny blood sample, but they say it could perform only roughly 12–13; when it failed, samples were run on conventional analyzers and results presented as Theranos’s own.
- The podcast says the documentary depicts Holmes and a team trying to revive Theranos by pitching investors and recruiting to make the technology work better with AI while she was about to go to prison. The hosts’ interpretation was that Holmes was complicit but may have rationalized misconduct as necessary to achieve the mission; they contrasted that rationale with the higher stakes in healthcare.
- Chronosphere grew from M3, an open-source observability project built inside Uber, into an enterprise observability company acquired by Palo Alto Networks for $3.35 billion in 2026, seven years later.
- Uber built M3 because commercial tools could not handle its observability needs at scale; its 2018 engineering post described processing 20 million data points per second, and within months LinkedIn, Alibaba, Databricks, and Walmart had deployed it in production, spurring demand for commercial support and services.
- Greylock partner Jerry Chen viewed Uber’s observability needs as an early signal of broader cloud-native scaling demands. Greylock wrote Chronosphere’s first check, investing $10 million of its $11 million Series A after a five-month pursuit; the founders skipped a conventional seed round.
Garry Tan called California’s Billionaire Tax Act an “asset seizure tax” and argued it would harm the state. The NBER post he shared estimates the act would collect roughly $30 billion but have a mean net present value of negative $39 billion after accounting for lost income taxes from departures.
Anthropic CEO Dario Amodei frames the company’s “race to the top” strategy as requiring frontier status to demonstrate responsible development and influence the industry; he says the approach has involved commercial costs and tradeoffs, while acknowledging organizational mistakes.
Amodei says Anthropic held Mythos-class models from February until general availability in June to strengthen cyber safeguards, initially limiting access to partners for cyberdefense and sacrificing part of its model lead. He also says restrictive biological safeguards produce significant false positives, frustrating biomedical users and creating a competitive disadvantage against more permissive rivals.
Other examples he cites are Anthropic’s advocacy for SB53 and chip export controls, restrictions on AI-service sales to PRC-controlled companies—including through third countries—and an announced contribution of 5% of the company, potentially more over time, to a public benefit fund. He acknowledges communication failures and says the change to Anthropic’s Responsible Scaling Policy is its most substantive point of criticism; he defends the revised focus on transparency and external review while recognizing that changing plans invites scrutiny.
- a16z’s Christian Keil argues that AI may master digital hardware design before hardware engineering: design-for-manufacturing, physical testing, and iteration remain harder because real-world data is variable, sparse, costly, slow to collect, and sometimes qualitative. He expects demand for better engineering simulation, citing Navier, Inviscid, and Lanyon, and for more manufacturing and testing capacity, including Hadrian, Cofactr, and Revel.
- Periodic Labs has set up an autonomous lab for end-to-end physical experiments on superconductors, magnets, and semiconductor materials. Ekin Dogus Cubuk argues that LLMs need to learn scientific inquiry through iteration and real-physics experiments, not simulation alone.
Garry Tan called Happenstance the “greatest unlock” for startup founders trying to reach the right people . Happenstance describes itself as a LinkedIn/Claude/ChatGPT hybrid and says it helps Ruben find a warm path to people across a network of 800,000+ .
Aleksander Madry is returning to OpenAI after what became a sabbatical, saying AI progress has entered a different regime in which the safety and security of increasingly capable AI must become a central priority; he will join @merettm to help chart a path through these challenges. Sam Altman described the safety and security challenges of confronting superintelligence as immense and said Madry has an excellent skill set for the work.
Paul Graham highlighted an analysis by Benjamin Jaros and Joshua Rauh estimating that California’s proposed Billionaire Tax Act would collect roughly $30 billion but have a mean net present value of negative $39 billion after accounting for lost income taxes from departures; Graham characterized the proposal as losing the state money.
GStack is now Capy-aware and, according to Garry Tan, works well with Capy’s software-factory feature Captains.
Martin Casado amplified a post listing alleged security issues: iPhone 11–16 were “apparently” vulnerable to remote attacks, SSL certificates were said to be broken until browsers removed eMudhra CA, and a PHP >9 CVE was expected “tomorrow”; the post framed these as recent or upcoming issues, and Casado added that the list was not exhaustive.
- Former Intel CEO Pat Gelsinger says AI could produce a chip design in three months, but silicon processing still takes nine months; reaching scale in a year and a half could make assumptions about AI workloads stale, leaving manufacturing scale as a major bottleneck to AI-driven chip innovation.
- a16z’s Christian Keil argues that AI may advance hardware design faster than physical execution: design is digital and model-legible, while design for manufacturing, building, testing, and iteration require more real-world interaction. He identifies potential demand across simulation tools (Navier, Inviscid, Lanyon), manufacturing inputs (Senra Systems, Podium Automation, Hadrian, Cofactr), and telemetry, inspection, and testing software (Sift, Instrumental, Revel); he also cites Periodic Labs’ autonomous lab as an end-to-end physical-experiment data loop.
Scott Kupor highlighted the thesis that “free is good” and that open source and open weights matter to a thriving software ecosystem.
Wabi founder Eugenia Kuyda argues that today’s chatbots are an interface bottleneck: users mostly turn to ChatGPT, Gemini, and Claude for search, writing, or homework despite models’ expanding capabilities, so a simpler, more engaging interface is needed to help people discover broader use cases. She likens current chatbots to AI’s “MS-DOS era” and describes single-player, text-only AI as the terminal era of AI interfaces .
- a16z’s Christian Keil argues that AI is nearing superhuman hardware design, but that will not equal end-to-end hardware engineering: building, testing, iterating, and designing for manufacturability require more physical-world interaction, while collecting physical data can be sparse, costly, slow, or qualitative. He identifies customer feedback as a particularly hard-to-compress step, since even many cheaply manufactured design iterations cannot tell AI what users want without that feedback.
- Keil expects the opportunity to shift toward simulation tools that let engineers iterate faster and suppliers of manufacturing inputs and services. Examples he names include Senra Systems (wire harnesses), Podium Automation (control panels), Hadrian (machined parts), Cofactr (electronic components), Instrumental (inspection), Revel (testing), and Mariana Minerals (copper); he also expects broader demand for PCB tools such as Diode as more users design hardware. These are forecasts and examples, not announcements of new launches.
