We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Vals is turning model evaluation into a financed infrastructure layer. Vals announced a $40M Series A at a $400M valuation led by a16z, with existing investors 8VC, Pear VC, and Bloomberg Beta, plus HRT Ventures and Next Ladder. The round accompanied the general availability of Vals Smith, new risk benchmarks including the RSI Index with CoreWeave and ReverseEngBench with Columbia, Tufts, UC Berkeley, and UCLA, and broader coverage in Vals Index 2.0. Vals says revenue is up 8x versus all of 2025, customers doubled, and the team tripled in six months; a16z names Rayan Krishnan and Langston Nashold as the founders it is backing.
The thesis is unusually clear: domain experts turn real workflows into benchmarks and automated graders, replacing leaderboard theater with evidence that a model can do useful work. The diligence question is whether Vals becomes a recurring decision layer for model buyers rather than a catalog of benchmark products.
Uplift is an early hard-tech bet on manufacturing economics in housing. The company launched with a $7M seed led by a16z. Its co-founders say they previously built at SpaceX and Tesla and are applying a Starlink/Cybercab-style design-for-manufacturing playbook to homes: deploy the first units over the next three months, then scale toward production rates comparable to the largest car companies. The announcement is a strong pedigree and thesis signal; the stated deployment and production milestones are the execution proof to underwrite.
2. Emerging Teams
HomeTurf is an early mortgage-marketplace experiment with real, if still modest, supply-side traction. The founders say the product lets borrowers anonymously post a mortgage need while NMLS-verified lenders bid against one another, with contact information withheld until a winner is chosen. It is live in five states with about 30 lenders actively bidding; borrowers are free, lenders are temporarily free, and the planned monetization is a flat lender subscription rather than per-lead fees or paid placement. One co-founder spent years at Wells Fargo and another has long mortgage-startup experience.
The acquisition loop is currently paid ads, Facebook groups, word of mouth, the co-founder’s industry network, and cold calls; the founder describes uptake as early but gaining momentum. The main diligence risk is regulatory and operational: the team says state-by-state compliance requires legal work and creates substantial monetization hurdles.
Suhail’s unnamed inference startup is a watchlist signal, not yet an investable story. The public update is hiring model-optimization and inference-performance engineers, says compute is already locked down, and calls for an in-person San Francisco team. It gives no product or traction detail, but the infrastructure-first hiring and secured compute suggest an attempt to move quickly from zero to a production-backed company.
3. AI & Tech Breakthroughs
The model race is moving into execution economics. A Gemini 3.7 Flash launch post says the workhorse model targets coding and agentic workflows at half the introductory price of 3.6 Flash, with reported gains in software engineering from 37.0% to 65.3% and enterprise automation from 13.4% to 30.4% after moving from 3.5 to 3.7 in three months. OpenAI is previewing GPT-5.6 Sol’s Ultrafast mode at up to 14x the speed for select API customers, while Perplexity is moving Sonar into an Agent API with grounded search, multi-step research, code execution, built-in tools, and multi-model access; it reports more than doubling the best Sonar score on BrowseComp and WideSearch.
The technical implication is not simply “better models”: latency, token cost, tool orchestration, and access to multiple models are becoming the product surface. That favors infrastructure and applications that can route work intelligently, rather than products whose only differentiation is access to one model.
World-model evaluation is exposing where attractive metrics stop being useful. The open-source worldproof project compares predicted rollouts with ground truth and physical invariants rather than scoring task success. On real SO-101 robot-arm footage, a copy-last-frame baseline reached 0.983 SSIM and 53.9 dB PSNR over a six-step horizon, with flat error across the horizon—meaning the setup could not distinguish a good model from a do-nothing predictor. On DROID footage, the only discriminative region was roughly steps 8–24; both the short and long ends tied.
For robotics and physical-AI diligence, benchmark design is itself an investable layer: ask whether the horizon, frame rate, task speed, and metric can actually rank systems. The project is Apache-2.0 and runs its evaluation path on a laptop without a GPU, which lowers the cost of independently checking model claims.
AI-for-science may need records of decisions, not just records of results. Nathan Benaich argues that papers show the “happy path” while omitting failed experiments—the history models need to learn scientific taste. A useful record must distinguish an underpowered assay, degraded reagent, and wrong hypothesis; it must also address the counterfactual problem that only the chosen experiment produces an outcome. He proposes that funders reserve part of a grant for testing branches that would otherwise be discarded, so labs do not repeatedly rediscover the same dead ends.
4. Market Signals
Model access is becoming a switchable input rather than a durable relationship. Suhail says he stopped using Anthropic models and has no loyalty to whichever model he uses next; his criteria are “fast, cheap, reliable.” Martin Casado separately says AI code generation feels saturated and that his remaining work is architectural and semantic—scale, performance, and trade-offs. Paul Graham says startup interest in tuning open-weight models has returned after a period when it was considered a waste of time. These are directional operator and investor observations, not a market-wide measurement, but they favor companies that own distribution, data, evaluation, and cost control rather than thin model wrappers.
Enterprise AI GTM should follow buyer exposure and proof portability, not prestige by default. a16z’s Lighthouse/Landgrab framework puts regulated, high-consequence markets where proof travels in “lighthouse” territory: a few credible customers make the category safe to buy. Established-budget markets with recoverable mistakes and fragmented buyers are “landgrab” territory, where math and coverage matter more than a marquee logo. The current repost makes the operating implication explicit: founders can sell in Ohio, Chicago, or St. Louis instead of competing for the same San Francisco logos.
For diligence, ask which motion the company has earned and whether it can execute it. POCs should have fixed milestones and a defined conversion point; otherwise an enterprise “trial” can become an open-ended science project that consumes the founding team.
Dual-valuation rounds are becoming a cap-table diligence issue. Newcomer reports that Starcloud’s $170M Series A was split into a first Benchmark-led tranche priced at a $250M valuation and a second tranche that closed at more than four times that price. The article says roughly 25% of recent deals seen by MVP Ventures used dual valuations, a view echoed by six other early-stage investors who described the mechanism as moving from rare to pervasive. Investors should request tranche-by-tranche pricing and model how the structure affects employee reference prices and future dilution.
Background agents are becoming an operational product category. LangChain’s Harrison Chase says agents running in the background will move work beyond direct prompting, with cron schedules as an early implementation; a companion post says scheduled self-prompting agents can be built in a few lines. Sequoia’s current framing groups the system into an open agent layer, a compounding evaluation loop, and a governed runtime. The commercial opportunity is shifting toward state, scheduling, observability, and control—not just another chat surface.
5. Worth Your Time
- Watch The GTM Advice Behind Billion-Dollar AI Companies. The useful framework maps buyer exposure against whether proof travels, then separates lighthouse from landgrab; it also gives a practical warning against POCs that never convert.
- Watch When to Build Your Own Agent Harness | Harrison Chase, LangChain. Chase’s compact framing—an agent is a harness, a model, and context, with model switching as protection against lock-in—is a useful architecture checklist for agent investments.
- Read AI needs science’s search history. It is the clearest current argument for collecting failed and unchosen scientific paths rather than training only on polished results.
