We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Stripe’s signed deal to acquire OpenRouter is the clearest strategic infrastructure transaction. The announced combination is framed as a network where AI companies exchange intelligence; its stated thesis is that tokens have become a universal medium of value exchange and that routing and clearing are the AI equivalent of the payments layer that enabled the prior web. Stripe’s Will Gaybrick says the goal is to make moving between tokens and dollars as seamless and safe as moving between dollars and euros. OpenRouter’s early wedge was aggregating API-credit demand for better pricing; its later differentiation is the operational burden of neutral, multi-provider routing, uptime, failover, and dynamic model/provider selection through tools such as Ori Eval. Founder Alex previously founded OpenSea, while a16z says it led OpenRouter’s seed and Series A.
Investment read: the deal validates routing, settlement, and exception handling as a strategic control plane around models. It is category validation, not a clean early-stage valuation comp.
Asymmlabs says it has raised $51M toward a defense mission. The company describes the goal as making offense more expensive than defense; Vinod Khosla calls it “the coolest defense sector startup” he is backing, while a recruiting post seeks people with radar, AI, and signal-processing expertise. The announcement discloses capital and technical hiring demand, but not stage or terms, so the useful signal is continued serious backing for defense AI rather than a priced-round benchmark.
2. Emerging Teams
AI-Native Services is an early-stage company-building pattern worth tracking. Newcomer reports that Emergence Capital’s Jake Saper is betting on accounting, insurance, legal, and other people-intensive industries as the next startup frontier. The model pairs lawyers, accountants, or insurance brokers with AI-focused software engineers, then sells the underlying service—often charging per completed outcome rather than billable hours. This shifts the underwriting question from software adoption to service gross margin, liability, customer acquisition, and whether the AI-enabled labor model can repeat across clients.
Router illustrates the adjacent control-plane opportunity. Its launch says it sends each request to the model “best for the task,” claims roughly 40% lower cost for the same outputs based on real-work benchmarks, and exposes the product through a base-URL change or two lines of code. The savings are self-reported and need replication, but the product direction is clear: model volatility is creating demand for spend control and routing rather than another standalone model wrapper. That fits a broader architecture shift in which agents are becoming directories of instructions, skills, tools, memory, and evals, with the harness and runtime underneath; coding-agent standards make some of those capabilities portable Markdown files or directories.
3. AI & Tech Breakthroughs
Prompt-injection defense is being optimized for usable specificity, not just attack detection. Patronus says it retrained Wolf Defender v2 because false positives were the larger operational problem: the earlier small model classified “Who are you?” as an injection with about 94% confidence. Its reported real-world benign specificity rose from 66.85% to 96.63% for the larger model and from 73.60% to 94.38% for the small model, while attack-detection F1 remained roughly stable; the benign example now receives 98.55% confidence and a real instruction override 99.99%. For enterprise agents, reducing the cost of false alarms may be a more important deployment milestone than squeezing another point from a clean attack benchmark.
Document agents are gaining a verifiable revision layer. LlamaParse says it can retain a document’s final-state Markdown while exposing tracked edits, deletions, and comments as structured data with author, content, and location—addressing the failure mode in which a deleted clause reappears as live text. A current follow-up frames the capability as giving downstream agents the full revision history rather than a static snapshot, especially for collaborative legal, finance, and other knowledge-work files. This is a concrete trust primitive for vertical AI: the agent can reason over what changed, not merely what the latest file appears to say.
Open-weight cyber capability is becoming a distribution and governance problem. A post summarizing Irregular’s testing claims Kimi K3 is the first open-weight model to pass CyScenarioBench, trailing closed frontier models by roughly six months at an estimated one-third of the inference cost. Its argument is that downloadable weights remove the API-level throttle, logging, and account bans that can constrain abusive campaigns. The claim is not a clean “open model” win: a commenter says Kimi K3’s license and size make it only notionally open and practically inaccessible without a datacenter, while another says local deployment still matters because it removes kill switches, logs, and rate limits. Treat this as a capability-risk and deployment-governance signal pending independent replication, not as settled benchmark evidence.
4. Market Signals
Agents are expanding coordination software, but AI cost is arriving before AI revenue. Atlassian’s reported quarter showed revenue up 28% to $1.766B, cloud up 31%, subscription ARR at $6.6B, and a 35% one-day stock move after the market had written off work-tracking software on the theory that agents would remove the middle layer. The report’s operating explanation is that more agents have produced more work items, branches, reviews, and governance. Yet Atlassian is bundling agentic capabilities into Jira while guiding non-GAAP operating margin from 36% in Q4 to 25% for FY27; Figma’s 84% GAAP gross margin also reflected per-request inference costs. For application underwriting, usage growth, AI revenue, and margin recovery should be modeled as separate events.
The boom case remains intact, but financing fragility is moving into the base case. Exponential View’s dashboard has no gauges in the red and two in amber; it puts trailing-twelve-month AI revenue at $126B through July. At the same time, tight compute supply is pulling more infrastructure investment and increasingly complex debt and financing structures into the buildout, while funding quality has deteriorated since September 2025 and is expected in its base case to turn red with economic strain during 2027. The actionable distinction is between demand risk—which the dashboard still sees as manageable—and financing risk if revenue compounding slows.
Trust is now a competitive variable in coding agents. The Pragmatic Engineer reports that an independent researcher found Grok Build transmitting .env secrets and uploading entire repositories, including files the agent was told not to read; the report says the behavior was active by default and not disabled by turning off “Improve the model.” Uploads were later disabled through a remote feature flag, and SpaceX said it had disabled default retention and was deleting previously retained coding data, while saying enterprise users with zero data retention were unaffected. The report’s market conclusion is that the incident could shrink enterprise prospects and force unusually generous limits or pricing to regain developer and security-team approval. Capability and cost cannot compensate indefinitely for a coding agent that enterprises cannot trust with their codebase.
5. Worth Your Time
- Read — OpenRouter & Stripe: The Intelligence Network. The primary articulation of the token-routing, settlement, and model-selection thesis, including the founder and a16z financing history.
- Read — Is AINS the Next SaaS?. A concise primer on outcome-priced services and the domain-professional/AI-engineer team model.
- Listen — From Chrome DevTools to AI Engineering, with Addy Osmani. A useful operating lens on “cognitive surrender,” loop engineering, accountability, and the unbundling of engineering, product, and go-to-market roles; Osmani brings more than 14 years of Google experience.
Sam Altman (@sama) posted "we support business privacy!" with a link to OpenAI's zero-data-retention offering for frontier models .
Grok CLI privacy scandal (xAI/SpaceX). The Grok Build coding CLI was shown to transmit contents of files it reads — including .env secrets files — verbatim and unredacted, and to upload the entire working repository (every tracked file plus full git history) independent of what the agent read, to a GCS bucket (grok-code-session-traces); it was active by default, not disabled by toggling off 'Improve the model,' and never opted into or disclosed to users . Proven on a real codebase: told to 'reply OK, do not read any files,' the CLI still uploaded the repo as a git bundle containing a canary file it was told not to open .
After the findings went viral, Grok disabled uploads via a remote feature flag; SpaceX's response attributed the data collection to debugging and said only enterprise customers with zero data retention (ZDR) were unaffected . After OpenAI CEO Sam Altman contrasted the open-source Codex harness , Musk committed to open-sourcing Grok CLI ; SpaceX open-sourced it on 15 July, disabled default retention for all users on 12 July, and is deleting previously retained coding data .
Competitive impact. The newsletter argues rivals Claude Code, Codex, OpenCode, and Gemini CLI have never violated user trust this way, the incident has likely shrunk SpaceX's enterprise subscription prospects, and Grok may need very high usage limits plus deep API price cuts to win developers back . Cursor — merged into Grok under SpaceX — previously indexed codebases locally and sent only embeddings to its server, storing none of the user's code .
Model and market signals. Grok 4.5 (built by the Cursor acquisition, trained on SpaceX GPUs) benchmarks near Opus 4.8 and GPT 5.5 in coding at 60-70% lower cost . Enterprise token pricing for AI coding tools reportedly runs 20-30x the $20/month consumer Claude Code/Codex subscription, and engineering leaders flag a growing, unsolved increase in code review load as AI accelerates code output .
Stripe has signed a deal to acquire OpenRouter, combining to become the network where AI companies exchange intelligence . The deal validates the thesis that tokens are a new universal medium of value exchange, with OpenRouter as the trusted, neutral routing and clearing layer for AI tokens . OpenRouter's early traction came from aggregating API credit demand, making tokens nearly as affordable for small developers as for big tech . Its roadmap includes Ori Eval, which dynamically routes models and providers based on prompts and business needs . Founder Alex previously started OpenSea in 2017; a16z Crypto invested in OpenSea and led its Series A in 2021, then a16z led OpenRouter's seed and Series A rounds after Chris Dixon connected him with the firm . Stripe's Will Gaybrick says the goal is to make moving between tokens and dollars as seamless and safe as moving between dollars and euros .
Pennsylvania Gov. Josh Shapiro signed an executive order imposing what he calls the nation's strictest standards for AI data centers : AI data centers must meet strict environmental and transparency requirements and receive local community approval ; all data centers are removed from the Fast Track permit program and agencies under his jurisdiction are barred from signing NDAs with AI data center developers ; he says he will use executive authority to block "objectionable, unwarranted" projects . Garry Tan (@garrytan) highlighted the order under the title "Irresponsible AI Data Centers," asking "Who writes this stuff?" — a signal of founder/VC frustration with state-level regulatory headwinds for AI data center buildout.
@davidu says @martin_casado's post "Been a good week" refers to his most recent commits to Exo .
Y Combinator launched Send a SAFE, a free tool to generate, sign, and send a SAFE in ~2 minutes, built for founders (and their AI agents) . It builds on the SAFE created by @cjoneslevy at YC in 2013, which YC founders have used to raise $15B+ and which has become the standard way startups raise money . Learn more at https://www.ycombinator.com/safe.
In an a16z podcast with partner David George, Whatnot co-founder and CEO Grant LaFontaine discusses how his live-commerce platform did $8B in sales last year, with its largest sellers doing over $100M a year at up to 40% margins . He traces the company back to his first online transaction — a Pokemon card sold in third grade and paid for by money order . The episode frames live commerce as roughly a third of all commerce in China yet still emerging in the US .
Aaron Levie (Box) recounted his early founding days in a USTechForce fireside chat, including how he raised $350k in seed money from Mark Cuban in 2005 and what happened to his other early investors . The discussion emphasized the importance of technology inflection points and of listening to the "right" customers — those willing to join you in the future . A video of the story was shared with USTechForce fellows .
- Stripe has signed a deal to acquire OpenRouter, positioning the combined entity as the trusted, scaled network where AI companies exchange intelligence .
- The deal's core thesis: "Tokens are the new dollars" — tokens have become a universal medium of value exchange, and OpenRouter acts as the neutral routing/clearing layer for this exchange, with Stripe aiming to make token-dollar movement as seamless as currency exchange .
- Context: model companies are in a generational race for intelligence, application companies face a Cambrian explosion, and the acquisition cements a pillar of AI's economic infrastructure .
- Founding-team signal: OpenRouter founder Alex previously founded OpenSea (2017); a16z Crypto invested in OpenSea and led its Series A in 2021. For OpenRouter, Chris Dixon introduced Alex to Martin Casado and Anj, a16z did the seed deal and "took a big swing" at the A .
- OpenRouter's early wedge was aggregating API credit purchases for better pricing, making tokens nearly as affordable for the long tail of developers as for big tech; it later built a technical moat in uptime, scale, and failover, with Ori Eval enabling dynamic model routing .
Stripe has signed a deal to acquire OpenRouter, creating the network where AI companies exchange intelligence and tokens; Casado calls it a pillar of AI's economic infrastructure . The thesis: "tokens are the new dollars" — a universal medium of value exchange for AI companies, with OpenRouter as the trusted, neutral routing/clearing layer . Stripe's Will Gaybrick says the goal is to make moving between tokens and dollars as seamless and safe as between dollars and euros, a big part of Stripe's future . OpenRouter founder Alex previously founded OpenSea; a16z led OpenSea's Series A in 2021 and backed OpenRouter at seed and Series A . OpenRouter's early aggregation made tokens nearly as affordable for long-tail developers as for big tech, an inflection point for AI adoption . Roadmap includes Ori Eval, which dynamically routes the best models/providers based on prompts and business needs .
Perplexity is promoting 'Computer in Email': users email requests to computer@perplexity.com and get a redlined Word document delivered to their inbox, with the pitch that it works for lawyers . CEO Aravind Srinivas calls it 'A decent lawyer in the form of an email interface' .
ITheClixs posted a study of weight-space perception in SIREN-style implicit neural representations, fitting ~1.8M networks across MNIST, FashionMNIST, and CIFAR-10 with controlled protocols separating shared initialization, optimization stochasticity, and independent initialization — the setting where reading semantics from weights works for shared-init but collapses when networks are fitted independently .
- Randomizing only the exact symmetry group (D_inf wr S_n, per hidden layer) while keeping each network's realized function fixed destroys 79.1 of the 80.4 accuracy points in the MNIST shared-init vs random-init gap; the author frames this as establishing sufficiency — symmetry scatter alone reproduces almost the entire degradation — not that 79.1/80.4 of the natural gap is causally mediated by symmetry . Within that induced loss, sign flips account for ~63 points, neuron relabeling ~15, and integer-pi phase shifts ~1 .
- A weight-space reader that directly quotients the symmetry structure on raw parameters reaches 0.917 accuracy, vs 0.628 for the best orbit-valued reframing, 0.526 for a fixed invariant encoding, and 0.265 for a permutation-equivariant baseline .
- On FLOPs-matched comparisons, simply querying the INR as a function (95.3% at 1.6 MFLOP) beats the best weight-space rung (64.4% at 5.5 MFLOP); the author concludes the strongest justification for operating in weight space may be computational rather than informational, since a complete invariant is informationally equivalent to access to the realized function .
- The author also proves generic identifiability modulo the symmetry group for one-hidden-layer SIRENs via the distributional Fourier transform of the realized function, noting that integer-pi phase transformations are affine rather than linear . Paper, implementation, tests, pre-registrations, and results are open-sourced on GitHub .
- a16z's David George interviewed Grant LaFontaine, co-founder and CEO of live-commerce platform Whatnot; Whatnot did $8 billion in sales last year, and its largest sellers do over $100 million a year at up to 40% margins .
- LaFontaine's market framing in the interview: live commerce is about a third of all commerce in China but still emerging in the US .
- Whatnot started as a Funko Pop marketplace; the founders observed customers "hacking social media" by selling items in live video despite broken transactions, shipping, payments, and discovery, then built the first product version in a few weeks, concluding that focusing on user experience beats "trying to build a market" .
Martin Casado (a16z) amplified @MillionInt's claim that "What will be left at the end of computer science is databases, neural networks and chips" , adding "Also … networks …" .
Sarah Guo (@saranormous) praised @tengyuma as "an extraordinary researcher but also an amazing teacher" and said she is "excited for the next company he starts" . She made the remark in reply to @linasbeliunas's recommendation of a 2h34m Stanford lecture described as "the clearest end-to-end explanation I've seen of how ChatGPT and Claude are actually built," covering tokenization/BPE, the transformer architecture, training pipeline, and next-token decoding .
Stripe has signed a deal to acquire OpenRouter, positioning the combined company as the trusted, scaled network where AI companies exchange intelligence; tokens are framed as a new universal medium of value exchange, with Stripe's Will Gaybrick saying moving between tokens and dollars should be as seamless and safe as moving between dollars and euros .
OpenRouter has become the trusted, neutral routing and clearing layer for token exchange, and its early aggregation of customers secured better API-credit deals that made tokens nearly as affordable for long-tail developers as for big tech — a factor that materially accelerated AI adoption .
OpenRouter's roadmap includes Ori Eval, a tool that dynamically routes requests to the best models and providers on a second-to-second basis based on progressive understanding of prompts and business needs .
Founder Alex Atallah previously founded OpenSea (2017); a16z Crypto was an early OpenSea investor and led its Series A in 2021. a16z led OpenRouter's seed round and took a large Series A position .
Martin Casado (a16z) calls the last 24 months unprecedented in his career, with small AI teams creating enormous value quickly — enabled by the new medium of token-based value exchange .
a16z spotlighted OpenRouter co-founder and CEO Alex Atallah, noting he was featured on the front page of The Stanford Daily in 2011 after releasing his campus social network "Dormlink" .
According to Omar Kunbargi, defense startup Asymmlabs raised $51M toward "Peace is possible", with more announcements to come . Vinod Khosla called it "The coolest defense sector startup we are backing", signaling his backing .
- Addy Osmani — 14+ years at Google on Chrome, DevTools, Core Web Vitals, most recently AI developer experience — reports a major Google culture shift: VPs and SVPs now code on weekends because AI tools make it easy, signaling AI's democratization of coding at the most senior levels .
- He warns that AI-assisted development carries a "cognitive surrender" risk (eroding comprehension and memory of the codebase), and advocates "loop engineering" — mutual amplification where the agent logs its decisions and the human reviews/internalizes them — as the countermeasure .
- On the future of engineering as an investment theme: he remains bullish because easier software creation expands the total market of builders, argues engineers stay essential since AI models cannot be accountable (roles will be defined by accountability), and predicts role unbundling where engineers must add product and go-to-market skills .
Jason Calacanis says autonomous vehicles save lives and will make rides cheaper, but will eliminate millions of US jobs and 100m+ globally; he urges responsibly weighing the pace of job losses against profits "to the AI Gods" . He predicts America will follow China in limiting, taxing, and regulating AVs, and that most cities won't let Silicon Valley "flip a switch" to flood streets with AVs, given "socialism and data center moratoriums" . The post responds to an Atlantic argument that cities rejecting driverless cars block technology that would improve residents' lives .
