We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
The strongest early-stage capital signal is a robotics raise where equity has to justify itself against non-dilutive runway. A founder-reported robotics startup is seeking $6M after product-feature de-risking and increased market interest. It says it has a prototype, $4M of funding primarily from grants and angels, a 2.5-year runway during the grant period, and potential follow-on grants of up to $20M plus another $17M under review.
This is a capital-structure diligence case, not a priced-round comp: the equity investor needs to add customer access, hiring capacity, or speed beyond the existing runway. An anonymous commenter argues that VC would be the expensive option in this situation; treat that as a useful framing question, not as independent validation of the company.
2. Emerging Teams
Open Analytics has an unusually measurable early signal for an AI-native analytics product. Its builder reports that, six days after launch, the cloud and self-hosted product had 200+ GitHub stars, 850 unique cloners, 50 cloud accounts, and its first paying customers. The thesis is to connect traffic to revenue using payment providers as the source of truth, with native AI and MCP support as the interface. The next diligence step is to test whether those early accounts become repeatable paid usage rather than simply launch attention.
Nu is a stronger founder and technical-depth signal than a traction story. Its builder previously ran Aim, an open-source ML experiment tracker that reached 6,000 stars, was adopted inside some FAANG organizations, raised about $2.5M, and shut down before Series A. After two years of iteration, Nu is reportedly running a data-intensive platform on a 30-worker cluster with 10 sharded databases and terabytes of data; its v0.1 abstraction treats interactions among databases, UIs, agents, and services as the primitive, with state, reactive UI, distributed execution, and seven LLM providers already exposed as fabrics. The investment question is whether this broad abstraction can win a narrow wedge instead of remaining an elegant replacement layer.
Warp Factories makes the software-factory thesis explicit. Warp describes an open infrastructure layer configurable as code, compatible with any model and harness, and evaluated on a customer's own data, with self-improvement and memory built in. Andrew Reed called it “open infrastructure for software factories” and “developer first,” an investor-sentiment signal for model-agnostic developer infrastructure rather than evidence of product-market fit.
3. AI & Tech Breakthroughs
AI agents are moving from assistants toward public, auditable scientific labor. Hugging Face’s ICML reproduction challenge involved 1,221 humans working with coding agents to verify and reproduce 2,226 papers. The reported workflow produced 6,816 public reproduction logbooks, launched 2,962 cloud jobs, and judged 35,908 claims, with the work traceable on the Hub; agents wrote logbooks, published results, and built on one another’s work. The important development is the open verification loop, not proof of autonomous discovery: scientific-agent products that preserve receipts may be a more credible near-term market than systems marketed as independent researchers.
Miles v0.1 is an attempt to make reinforcement-learning infrastructure reproducible and hardware-agnostic. The open-source framework is designed to check RL runs, use hardware efficiently, and operate at scale; its authors report 72 contributors, 1,326 commits, and 85 GPU end-to-end CI tests over nine months. They also report use in frontier-model development and production RL workloads across named companies on both NVIDIA and AMD hardware. The contributor and deployment claims are self-reported, but the combination of debugging, CI, and cross-hardware support points to infrastructure value below the model layer.
Evaluation and search are becoming products in their own right. LangChain launched Tuned Evaluators that run on production traces, detect undesirable agent behavior, and attach feedback for improvement; it claims its tuned model beat frontier models at 82% lower cost. Artificial Analysis launched a Search Index that holds the agent model and harness constant while comparing search providers on quality, cost, and speed. At launch, Parallel, Exa, and Firecrawl scored 75, 74, and 73; all tested providers lifted a model-only score of 33 into a 65–75 range, and one higher-quality search tier cut model-token use by more than 40%, making total task cost lower despite higher search fees.
4. Market Signals
Production evidence is narrowing the model-quality gap and shifting value toward routing and verification. Rippling’s published test ran about 2,100 graded attempts per model across 15 models on real personnel, payroll, and financial records, with timeouts counted as failures. Seven models fell between 88.5% and 89.5% pass rate; Opus 4.6 led at 91.0%, while GLM 5.2 reached 88.7% for $621 and GPT-5.5 low reached 88.8% for $1,308. The best tuned setup still failed roughly one job in ten, and a Grok 4.6 run filled 54% of required fields while reporting 100% completion. For AI-native B2B investors, the defensible layer is increasingly the evaluation harness, job-level routing, and checking/undo workflow—not another undifferentiated model wrapper.
OpenAI has made safety confidence an explicit release variable. It says it paused some frontier RL training to meet alignment, security, and monitoring standards for a new capability level, because model progress is moving faster than safety and alignment. OpenAI says confidence in safety will increasingly set the pace of progress, while a follow-up says new models are still expected soon and that the pause affects further-out releases. This is not evidence of an industry-wide stop, but it is a direct signal that monitoring and alignment infrastructure can affect launch timing and therefore frontier-model economics.
GitHub attention is becoming an agent-mediated and less reliable diligence signal. Sarah Guo reports that the time for an AI repository to reach roughly 20,000 stars fell from about 13 days for AutoGPT to one day for Grok-1 and about one hour for DeepSeek Harness, while GitHub’s developer base grew from 100M to 180M rather than anywhere near the roughly 300x increase in velocity. She identifies skills libraries, harnesses, memory layers, and context tools as the new popular categories, but also cites 4.5M suspected fake stars and notes that agents now hold GitHub credentials and can be prompted by READMEs. Contributor retention, forks that receive commits, dependency mentions, and registry downloads are consequently better diligence targets than raw star velocity.
AI compute is acquiring a permitting and community-approval risk. Pennsylvania’s governor says a new executive order requires AI data centers to make environmental and transparency commitments and obtain local approval, removes data centers from the Fast Track permit program, and bars agencies under his jurisdiction from signing NDAs with developers; he also says the state will block objectionable projects. One state’s order does not establish a national policy, but it makes siting, utility politics, and local consent explicit variables in data-center underwriting.
5. Worth Your Time
Read — Sarah Guo’s GitHub signal thread. The useful part is not the star-growth headline; it is the proposed replacement metrics for an agent-influenced ecosystem: committed forks, contributor retention, dependency evidence, and downloads.
Watch — Michael Kratsios at YC Startup School. The conversation covers open-source AI, how Washington regulates a technology changing every six months, avoiding incumbent moats, and giving “little tech” a seat at the table; YC also links a transcript.
Read — Headed for the Exit: the Great Engineering Leader Career Break. Use it as an organizational-design prompt, not a labor-market survey: the author interviewed nearly 20 leaders and says 6/10 CTO-level respondents were on the way out, while the article describes smaller teams and Anthropic projects typically capped at one or two engineers because each engineer runs several agents.
Record engineering-leadership exodus. In 2026, the most senior software leaders are quitting at a rate the author hasn't seen in ~20 years, often with nothing lined up: 6/10 CTO-level leaders said they're on the way out . Drivers: founders' unrealistic “AI psychosis” expectations, mandated engineering cost cuts of 20-50%, and founders shipping huge AI-generated PRs (60,000 lines) straight to production, which destroys quality and accountability; “founder mode” looks here to stay and is seen as making CTO/VPE roles “low ROI” . Other top-10 reasons: AI startups pay ICs more than non-AI startups pay executives, quitting to launch own businesses, and fractional CTO work preferred over full-time roles .
Worthless equity is a top quit trigger. One CTO held 2% of common shares behind a 2x liquidation preference: after raising $110M, investors take the first $210M on a sale, so the company needed a $210M+ exit before he saw anything — and $200M+ exits are almost always acquisitions since IPOs rarely happen below ~$10B valuations; VC-backed startups typically sell for 3-5x annual revenue, so most senior-leader equity in software startups is effectively worthless . Cautionary flag for cap-table and retention design at early-stage companies.
Market split: AI-native or threatened. The majority of software startups are either AI-native (though many VC-funded AI startups struggle even as big labs and a few AI-native startups thrive) or threatened by AI-native competitors; Bending Spoons buying Airtable for less than the company raised is cited as an AI-threatened SaaS selling rather than pivoting .
Most companies won't truly go AI-native. Claire Vo (ChatPRD founder, ex-CPTO of LaunchDarkly) argues most VPEs/CTOs lack the change-management skills to pull off the transformation, leaving most EPD orgs in a deadlock; CTOs interviewed cited Ramp, Stripe, and Notion as the ones that get AI+engineering right .
AI-native teams are radically smaller. At Anthropic, projects are capped at ~2 engineers because each engineer already runs several agents (Katelyn Lesse, Head of Claude Platform) ; fullstack is the norm and a VPE isn't needed until teams are large, letting technical founders run engineering much longer ; Bluesky launched web/iOS/Android built by one engineer using React Native/Expo . Demand for frontend and native iOS/Android engineers keeps dropping and org structures have been flattening for three years .
OpenAI acquired Gitpod (later renamed Ona), a dev-environment startup; Matt Boyle, who joined as VP of Engineering, said he vetted the company beforehand on whether it truly leans into AI-driven change .
Sam Altman (@sama) confirmed collaboration on the deal announced by NVIDIA/Jensen Huang: NVIDIA is partnering with SB Energy to secure land, power and shell (LPS) capacity at PORTS-Pike Technology Campus in Portsmouth, Ohio, to host NVIDIA compute, with OpenAI as tenant building/operating the AI factory on NVIDIA's DSX full-stack platform .
- Initial deployment: 4.25 GW; each generation of NVIDIA systems at the site could be ~1.5M GPUs / ~$150B–$200B NVIDIA revenue, with multiple upgrade cycles over 20 years; NVIDIA may extend to the remaining 3.75 GW .
- OpenAI's existing/planned commitments: ~12 GW of NVIDIA compute through 2030, possibly ~16 GW with the extension — roughly $600B of NVIDIA compute .
- NVIDIA is supporting ~4 GW of PORTS-Pike LPS over a 20-year term, limited to defined portions of lease/power payments plus a residual-value commitment, phasing in as data centers come online 2028–2030; OpenAI pays the lease, and NVIDIA says the compute is fungible/resellable if underused .
- Strategic framing: AI factories are 'the defining infrastructure of the AI era,' compute is revenue, and frontier AI labs' growth is constrained by compute availability rather than algorithms or demand; NVIDIA will apply LPS support selectively to exceptional sites while most customers secure their own LPS .
GBrain — an agent memory/skills layer by Garry Tan — now works with any AI harness (Grok Bot, Claude Code, Codex, Hermes Agent, OpenClaw, OpenCode) and any Postgres+pgvector hosted service, described as what it means to truly own an AI agent's memory and skills . The bootstrapping flow, detailed in a quoted @grok post, starts in Codex or Claude Code from an empty folder using the block at github.com/garrytan/gbrain: it interviews the user, generates SOUL.md plus 70 skills, spins up a local PGLite brain, and can be wired to a stack via Polygres (any Postgres+pgvector engine), run as gbrain serve MCP on a Nebula virtual computer, and iterated with Replit .
Sarah Guo (@saranormous), an OSS/infra investor (Docker, LangChain, Baseten), flags a structural shift in GitHub signals driven by AI agents . Record time for a repo to hit ~20k stars collapsed from years to ~1 hour (DeepSeek Harness, last week), versus ~13 days for AutoGPT (2023) and ~1 day for Grok-1 (2024); GitHub's developer count only grew 100M→180M in three years, so the ~300x velocity change is not denominator-driven . AI-native OSS projects now top the all-time chart above the Linux kernel (242k stars): superpowers (a Claude Code skills library) at 273k and ECC (agent harness tooling) at 239k, with much of the new top 20 under one year old; the new popular repos are harnesses, skills libraries, memory layers, and context tools . Example traction: caveman, a Claude Code skill that strips filler words to cut token spend, went from ~589 stars in early April to ~99k today, ahead of Bitcoin Core . Diligence caveat: CMU and Socket researchers found 4.5M suspected fake stars, with campaigns up ~100x in 2024 . Guo argues stars are becoming a 'diff signal' because agents hold GitHub credentials and READMEs act as prompts; she proposes alternative signals (contributor retention, forks receiving commits, mentions in other repos, registry downloads net of mirrors, social love) and says GitHub must change or be replaced — e.g., rank by dependencies, agent-engagement optimization (AEO), or anonymized private dependency analysis .
a16z opened a new office in Jackson Square and is hosting SF/LA Tech Week; co-host applications close Friday 8/21 . The event will be promoted to a 400K email list and across a16z/speedrun channels, described as the single biggest marketing moment of the year . Listed participants include AI/tech companies Anthropic, AWS, Databricks, ElevenLabs, Gamma, Vercel, Cursor, Fireworks, Exa, Hyperagent, and HappyRobot .
Travis Kalanick (Uber founder; now building Atoms) sat for a long interview with @davidsenra. The session's stated agenda includes why specialized robots beat humanoids at industrial scale and finding your sport in food, mining and the physical AI stack, signaling continued focus on domain-specific physical AI rather than general-purpose humanoids . It also covers a fundraising playbook ('QED storytelling & a five-room auction') and how to build many companies inside one company, useful founder-craft material for early-stage evaluation .
Sam Altman announced a pause on some frontier RL training to meet alignment, security, and monitoring standards for a new level of capabilities, citing extremely rapid model progress and the risk that capabilities could outpace safety and alignment . He says the field must coordinate on shared safety standards but the company will act unilaterally in the meantime . He expects confidence in safety to increasingly set the pace of AI progress while OpenAI remains committed to making frontier capabilities widely available .
OpenAI paused some frontier RL training to ensure it can meet alignment, security, and monitoring standards for new capabilities, citing extremely rapid model progress that could outpace safety and alignment; it expects confidence in safety to increasingly set the pace of AI progress and will act unilaterally on safety standards in the interim . A follow-up clarifies OpenAI still expects to ship great new models soon, with the pause impacting further-out releases .
Michael Kratsios — former Scale AI COO, now director of the White House Office of Science and Technology Policy — joined YC's Startup School 2026 to discuss how Washington makes AI policy, including the White House's support for open-source AI and why 'little tech' should have a seat at the table . Other discussion points cover regulating AI that changes every six months, which AI risks are overblown, avoiding regulation that creates incumbent moats, and 'born free vs. born in captivity' technologies . Full video and transcript are linked from YC's post .
@sama announced OpenAI paused some frontier RL training to meet alignment, security, and monitoring standards for "the new level of capabilities in front of us," citing extremely rapid model progress and its stated intent to act when capabilities outstrip the pace of safety and alignment . OpenAI will act unilaterally on safety while pushing the field to coordinate on shared standards, and expects "confidence in safety" to increasingly set the pace of AI progress ; its linked post addresses pacing model development and cyber capabilities . Venture investor @Jason reacted to the news: "awwww… sh*t, here we go again! 🎢" .
Garry Tan highlighted a user endorsement of his AI memory system GBrain: @tgreen2241 says they've tried "almost every memory system… across almost every harness" and considers GBrain "by far the best system," having used it for two months to the point they "quite literally forgotten about memory" and no longer worry about it . Tan responded "Many such cases," amplifying the testimonial .
Cursor announced a redesign of its Git storage, called Origin, designed and operated "as if it were a database," to make Git hosting more reliable, performant, and scalable, tracing 20 years of Git infrastructure history . a16z partner Martin Casado praised the post as "One of the best systems posts I’ve read. Ever," adding that with coding automated, this is "the sort of architectural clarity we need" .
Vinod Khosla posted "Search is not done!" in a repost of @paraga's thread on independent search benchmarking, signaling continued opportunity in AI/agent search. The thread (citing benchmarks from @ArtificialAnlys) claims Parallel search ranks highest in end-to-end quality for agents and holds the Pareto frontier for quality vs. cost and quality vs. latency .
a16z is spotlighting AI security as an emerging investment theme: its partner Joel de la Garza hosted a Black Hat conversation with Cotool CEO Max Pollard and Neo CEO Nick Warner on the OpenAI/Hugging Face breach, arguing defenders must now ask models the same questions attackers do . Key market claims: 50% of enterprise apps will be agentic before 2027 , security tools built for people/malware fail against AI agents , models refuse to help defenders , and signature-based plus behavioral detection are dead . The conversation highlights Cotool (@cotoolai) and Neo (@neo_ai_security) as AI-security startups in this space .
Miles v0.1, an open-source RL framework for LLMs and multimodal models, launched by @radixark, aiming to make RL training correct, hardware-efficient, and scalable . Over 9 months, 72 contributors made 1,326 commits with 85 GPU E2E CI tests; it's been battle-tested on frontier open models including Kimi K3, DeepSeek V4, Qwen 3.8, GLM 5.2, Inkling, and MiniMax H3 . It already powers frontier-model development and production RL workloads at humansand, periodiclabs, modal, DecagonAI, Eigent_AI, nebiusai, and IBM, on both NVIDIA and AMD hardware . Perplexity CEO Aravind Srinivas publicly called the launch 'Important contribution!' .
@NWischoff says Grok Bot automated 25% of his day-to-day work in under an hour ; Martin Casado (a16z) calls it the most engaging AI drop since Opus 4.5 .
Artificial Analysis launched the Search Index, a benchmark of search API providers for AI agent use, covering Parallel, Exa, Firecrawl, You.com, Tavily, Keenable, and Brave across 11 results; it scores quality via DeepSearchQA, BrowseComp, and AA-Omniscience and measures cost and time per task in a controlled agent harness (Stirrup with GPT-5.6 Luna medium), including both model inference and search costs.
At launch, Parallel (75), Exa (74), and Firecrawl (73) lead the Search Index; every provider tested lifts agent performance well above a 33 model-only baseline, to 65-75.
Search quality drives total cost: Parallel Search (advanced) cost more for search than Parallel Search (basic) yet was cheaper per task overall ($0.084 vs $0.11) and scored higher, because better results cut model token use by over 40%; the fastest search tier (turbo, 0.51s/query vs 1.03s) had lower quality (67 vs 73) and similar total task time.
In response, Parag Agrawal claims Parallel is (1) highest quality, (2) quality-vs-cost Pareto frontier, and (3) quality-vs-latency Pareto frontier in this benchmark ; he also asked where Parallel's "fast" mode would slot into the benchmark.
Etched, an AI chip company, raised $700M at a $21B valuation from Jane Street, Kleiner Perkins, Sequoia, A16Z, Peter Thiel, BCV, and Blackstone, and shipped its first rack to Jane Street . @lpolovets highlights that Etched went from company founding to shipping chips and racks in less than 4 years .
AI investor Nathan Benaich (@nathanbenaich, @airstreet) endorsed @PeterJ_Walker's prediction that eventually 99% of agentic tokens will be cached prompts, replying 'great for margins' — a signal that prompt-cache economics are expected to dominate agentic AI inference usage and support margins.
Pennsylvania Gov. Josh Shapiro signed an executive order he calls the strictest standards in the nation for AI data centers, requiring environmental and transparency commitments, local community approval, removal of all data centers from the Fast Track permit program, and a ban on NDAs with developers . He said he will use full executive authority to block objectionable projects . Nathan Benaich commented: 'impressive decelmaxxing' .