We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Mundo AI raised a $20M Series A led by GreatPoint Ventures, with participation from Y Combinator, Next Frontier Ventures, and E12 Ventures; a previously unannounced $4M seed brings total capital to $24M.
Mundo is positioning the company as a data-and-evaluation layer for “perceptual intelligence”: AI that understands speech, video, gestures, environments, and human interaction. It says its datasets and evaluations are already used by leading AI labs for speech-to-speech interaction, fine-grained video understanding, and emerging modalities, and argues that progress depends on a continuous feedback loop between better data and better evaluations. The investment question is therefore less “which model wins?” than whether Mundo can make multimodal data and evaluation demand repeatable and defensible.
2. Emerging Teams
OSOA is an early paid test of an autonomous sales closer. The product finds and verifies prospects, writes outreach in the owner’s voice from the owner’s accounts, negotiates within preset limits, and closes without human involvement after setup; the founder says four businesses paid in the first week at $59–$179 per month, from 20 founding spots. The failure report is more informative than the demo: the system needed hard guardrails to avoid closing unwanted deals, while prospect verification produced false positives and nearly sent outreach to dead leads before a rebuild. The founder is raising a $500K SAFE, and the product began as an internal sales system for the founder’s first company. For diligence, policy enforcement and prospect verification matter more than message quality.
MarketOwl is a distribution signal, but not yet product-market-fit evidence. A first-time European founder who has been building solo since 2023 says an earlier version generated five to six demo calls a week but retained high churn; after shifting to Reddit and iterating toward an AI marketer, the founder reports testing with 12 companies at an average 15% positive-reply rate and getting zero-follower Threads accounts to posts with more than 1,000 comments. The proposed moat is a playbook layer combining social posts, DMs, ads, and SEO, built from three years of campaign experience. That makes durable customer impact—and avoiding channel or account bans—the key validation test, not raw automation volume.
A newly out-of-stealth earthquake-prediction startup is a useful diligence caution. Its team combines a serial technology founder who runs an SMB-focused IT company with a doctor of seismology; the founders say their methodology, developed using Greek seismological data, predicts California earthquakes up to 48 hours ahead. They later define the target as within 125 km, one magnitude, and 24 hours, reporting 80% accuracy and 83% recall. The claim is not yet underwriting-grade: a commenter who said they read the paper characterized it as off-the-shelf ML trained on historical Greek data, while another questioned the accuracy definition and requested false-positive and false-negative rates.
3. AI & Tech Breakthroughs
SPADE turns synthetic-environment design into a learnable post-training loop. Its Environment Designer writes executable, long-horizon training environments while a Reasoning Agent solves them; on Qwen3-30B, the reported game-suite average was 58.3, or 8.1 points above base and 5.3 above the strongest fixed-environment baseline, while tool-use environment design improved every tested backbone. Code and checkpoints are available. The investable implication is a potential reduction in the cost of creating diverse post-training tasks, although the framework’s gains remain bounded by the capabilities of the model generating the environments.
Hawkeye packages hardware-specific kernel optimization into a test-driven coding-agent workflow. Researchers from Harvard, Stanford, Together AI, and Caltech describe an open-source framework whose unit tests pair a human-authored solution kernel with profiling metrics and a usage guide. They report matching or exceeding torch.compile on tested workloads across NVIDIA Ampere, Hopper, Blackwell, and AMD MI350, and an 18.9× geometric-mean speedup over expert-authored Triton kernels for emerging attention variants. If reproducible, this moves scarce accelerator expertise into reusable agent infrastructure rather than leaving every new chip or kernel pattern to specialist manual optimization.
Protege makes healthcare evaluation an outcome-measurement problem, not simply a model-quality problem. The article argues that medicine lacks a shared absolute ground truth, that clinical datasets are guarded, and that benchmarks often measure test design or physician preference rather than clinical merit; objectively evaluating a decision requires following the patient forward over time. The context gap is concrete: the median patient record contains about 8,500 tokens, while five of six public healthcare-AI benchmarks provide less than median context and several provide fewer than 200 tokens per case. Protege says it uses real-world partner data to build tests tied to health gains for clinicians and patients.
The broader calibration remains uneven. Import AI’s summary of a METR analysis reports major acceleration in vulnerability discovery, minor and hard-to-measure acceleration in mathematics, and no measurable acceleration across seven AI algorithmic-progress areas. That makes domain-specific evaluation a prerequisite for distinguishing a real capability phase change from a compelling demo.
4. Market Signals
Agent demand is rising while open-weight usage and model costs are repricing the stack. Exponential View reports that open-weight tokens’ share doubled in the last year and is approaching a 1:1 ratio with closed-weight tokens, even as the number of closed-weight tokens grew sevenfold; it also says agents now use 14× their February token volume while human token usage grew 2.8×. A separate essay argues that AI-token costs could fall by more than 1,000× in under a decade, with cost declining 2–5× per year as utility improves, shifting value toward applications that become viable at lower prices.
The infrastructure corollary is a power and supply-chain contest. The essay characterizes Nvidia’s CUDA and accumulated tooling as owned, packaging capacity as rented, and power generation as absent; it says Nvidia’s four largest customers have contracted roughly 9.8 GW of nuclear power while also shipping their own silicon.
Systems of record are being pushed toward agent-native interfaces, with evaluation emerging as the routing moat. Jerry Liu says agents should use existing software such as Slack rather than rebuild it or rely on the vendor’s agent, because software and systems of record need to become agent-native. Garry Tan’s current formulation preserves deterministic APIs, ACLs, SQL, and data structures but says vendors must add the AI harness and full customer solution or risk being subsumed. In the adjacent model-routing layer, Brendan Foody argues that 80% of building a good router is building a good evaluation, with routing logic the easy part.
Launch attention remains weak evidence of software durability. An audit of 2,291 Product Hunt launches reports that 24.4% were hard-dead and 28.9% no longer existed as independent products; AI products died at essentially the same rate as non-AI products, 24.5% versus 24.4%. Most deaths occurred in the first year or two, and the author warns that website availability is only a proxy and may understate the true failure rate.
5. Worth Your Time
Read — Import AI 470. The most useful single synthesis in this period: uneven real-world acceleration alongside SPADE’s synthetic environments and Hawkeye’s hardware-aware kernel agents.
Read — The Oracle Problem. A strong case for longitudinal, outcome-based health-AI evaluation and for testing against full patient-record context rather than compressed vignettes.
Inspect — AQuA’s model-development loop. Its bounded configuration diffs and sealed evaluator are a useful pattern for making agent-led model search auditable, with explicit caveats around unequal compute, undisclosed features, and label construction.
Read — AI fact-checker citation audit. The author found 12 dead or nonexistent URLs among 215 citations and traces the problem to letting the prose model author provenance; the proposed remedy is retrieval-owned citations plus mechanical URL and text-support checks.