We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Nvidia is reportedly combining capital, talent, and open-weight model development in one Poolside transaction. A post citing the WSJ says Nvidia plans to spend $6 billion on a powerful open-weight model, license Poolside’s technology, bring more than 100 Poolside employees into Nemotron, and invest another $1 billion in Poolside at a $12 billion pre-money valuation. The stated target is to challenge DeepSeek and Kimi while competing with OpenAI and Anthropic.
This is strategic corporate financing rather than a clean seed or Series A comparable. The diligence question is whether the combination of Poolside technology, transferred talent, and Nvidia’s model-building program produces a durable open-weight advantage; confirm the reported terms before underwriting.
Stripe’s reported OpenRouter purchase makes routing an exit thesis—but not a settled one. SaaStr’s 20VC recap says Stripe paid around $7 billion for OpenRouter four months after its $1.3 billion round, describing it as a leading LLM-routing layer. The same analysis says enterprises may prefer a few models and in-house routing, and gives only a one-in-three chance of a standalone routing business versus absorption into Stripe infrastructure. For early-stage investors, routing needs defensible control of model flow or a distribution advantage; a model menu alone is unlikely to be enough.
2. Emerging Teams
A repeat founder is testing persistent roles as the company operating system. A developer and founder with more than 10 years of experience who previously ran a SaaS business to roughly $5 million ARR says a new company is seeing “real traction.” He gives persistent agents CEO and CMO roles, feeds them company context, lets them disagree, makes the final decision himself, and then has agents handle execution, research, and testing. He says the value is memory that carries decisions and observed results forward, and forces agents to distinguish BUILT, VERIFIED, NOT VERIFIED, BROKEN, and UNKNOWN. Treat this as an operating-model signal rather than an underwriting datapoint.
Intimassy shows how basic distribution fixes can unlock early monetization. Its 11-year software-engineer founder says Reddit feedback led to a proper domain, tripling organic search visits, and an iOS launch that produced three paying users from 30 organic downloads on day one. The post headline reports 1,500 daily users and about $100 per day. The signal is discoverability and platform coverage—not another model feature; retention, paid acquisition, and durability remain unshown.
MUON is an early bet on coordination as an agent primitive. A research engineer at a YC startup built an open-source desktop app, MCP integration, and CLI that connect Claude Code, Cursor, Codex, and OpenCode into a shared memory and coordination graph. The author says it is still early, has broken features, and is being dogfooded on itself; its Polyform Noncommercial license permits personal or day-job use but excludes a company-wide shared brain without an enterprise arrangement. The category signal is stronger than the current traction evidence.
3. AI & Tech Breakthroughs
ShardFlow attacks inter-region inference latency with speculative decoding. The framework splits HuggingFace transformers across multiple GPU machines; in a two-T4 setup across GCP regions with roughly 86 ms round-trip latency, K=8 drafting commits 4.07 tokens per round trip instead of one. On Qwen2.5-7B, the builder reports 4.92 TPS for the non-speculative baseline versus 28.10 peak and 20.31 average TPS with a neural drafter and CUDA Graphs; graph capture reduced draft latency from 112 ms to 25 ms. These are builder-reported results, but reproducibility would make distributed inference over public WAN a meaningful deployment option.
Agentic document work is converging on retrieval first, vision second. Jerry Liu describes a two-pass pattern: a cheap open-source parser scans tens to thousands of files, then a just-in-time VLM screenshots and dissects only the relevant pages. That avoids running expensive VLM OCR over entire file dumps, while exposing current weaknesses in grounding, parser versatility, and cost.
Ship Safe packages the agent attack surface into a local developer tool. The MIT-licensed scanner checks for prompt injection and agent hijacking, dangerous MCP configurations and permissions, secrets, CI/CD and supply-chain risks, auth/API vulnerabilities, and RAG or memory poisoning. It runs locally with npx ship-safe; its agent workflow proposes a fix, shows the diff, asks for approval, applies it, and verifies the result. The project is still seeking security-tooling feedback, so coverage and false-positive economics are unproven.
4. Market Signals
Codex adoption is spreading beyond tech into functions with domain-specific workflows. a16z reports that its fastest-growing Codex adopter categories since February are legal at 108x, sales and recruiting at 41x each, marketing at 26x, and healthcare at 24x. This is a vendor-reported chart, so use it as a directional vertical-discovery signal rather than market-share data.
Inference price competition is an adoption tailwind and an application-margin trap. Suhail says Chinese model subsidization at the inference layer is helping new agentic coding products compete with Anthropic and OpenAI, and calls the trend healthy. A separate SaaS discussion argues that every AI action adds variable cost: raising prices causes churn, usage caps frustrate customers, and routing to cheaper models only partly closes the gap—quietly turning conventional SaaS into usage-based pricing. Underwrite token cost, pricing architecture, and gross margin together.
Agents are putting systems of record and link economics under pressure. Garry Tan predicts that systems of record will need to become AI harnesses or face replacement by agents. Separately, a current post says UK publishers asked the CMA to keep ChatGPT and Perplexity off Google’s default-search choice screen because chat answers generate no click-through or referral traffic; the CMA has yet to decide whether chatbots count as search engines. The practical risk is that incumbents retain data while losing the interaction and distribution layer.
Governance demand is real, but the headline statistic is not clean enough to underwrite. A post claims that 78% of organizations had not taken meaningful AI-compliance steps despite deploying agents on sensitive data. A commenter says that figure conflates at least three different measurements, while agreeing that PII leakage and prompt injection receive less attention than visible accuracy failures. The practical control discussed is explicit human confirmation before sensitive execution rather than trusting a model’s silent “processed successfully” claim.
5. Worth Your Time
Read — Jerry Liu’s two-pass RAG thread. The clearest current architecture note on cheap broad parsing followed by just-in-time VLM inspection, including the grounding and tool-quality gaps.
Inspect — ShardFlow. The repository and implementation notes behind the WAN-inference benchmark, including the Rust relay, KV-cache handling, and model slicing.
Read — The Speed of Thought: What If Intelligence Has a Universal Limit?. A contrarian capital-allocation frame arguing that frontier gains are roughly logarithmic in compute and that the next value may accrue more to applied AI, tooling, integration, and cost reduction than to recursive self-improvement.
Read — 20VC x SaaStr’s AI deal analysis. Useful valuation framing: the article contrasts systems-of-record software at roughly 5.3x revenue and slower or non-SOR software near 2.7x with OpenRouter near 70x trailing revenue.