Joel Oxman

Your Team's Attention Is the Actual Bottleneck, Not Your Data Every private lending org runs into the same ceiling eventually: reps can only make so many calls in a day, marketing has a fixed budget, and leadership only has so much bandwidth to review the pipeline. Given that ceiling, "how do we find everyone who could possibly borrow" is the wrong question to be optimizing for. You already know approximately who could borrow. That list has existed for years. The question that actually determines revenue is: where does the next unit of attention go? That reframe changes what "good data" even means. A system that hands your team 5,000 more qualified names hasn't solved anything if your reps can only work 200 of them this week. What actually helps is a system that tells you which 200 to work first, and, just as importantly, which names to leave alone until the timing changes. Attention is the scarce resource in this business. Everything else is just infrastructure around it. Full breakdown – link below 👇

2

1 Comment

The "New MarTech" isn't about tracking impressions. It’s about enforcing Market Honesty at the settlement layer. Most federal and enterprise clearing operations accept 120 bps of friction as the standard "cost of doing business." They allow legacy aggregators to leak value through slippage and inefficient routing. At Orkid Labs (18NQ1), we view that leakage as a failure of attribution. We treat every disbursed dollar as a high-fidelity event stream. We just pulled our latest live production data, and the quantitative reality of our deterministic clearinghouse is absolute. Here is how we are internalizing value for the federal balance sheet: 1.78 bps Median Execution: Our live "Clip Size 2" operations clear at 1.78 bps. Compared to the 120 bps market median, this is a 67x efficiency advantage. We do not simply optimize; we offer a Structural Privilege. 60% Toxic Flow Extinguished: Our system operates on strict Rust/Prolog logic gates. By enforcing a 2.0 Z-score threshold, we reject 60% of attempted transactions. We do not pass out-of-band friction to the ledger; we correct it before settlement. Deterministic Attribution: We have adapted a high-velocity event-streaming architecture to federal value clearing. We collect a fee solely for providing this service: enforcing structural integrity and returning the rebate directly to the client. (Our data proves that while standard market volatility expands exponentially with scale, our logic gates maintain a flat, deterministic risk profile.) We are currently opening a 14-day execution window for federal supply chain nodes and healthcare disbursement programs to integrate this architecture. If your agency or prime contract is bleeding basis points to legacy middlemen, the next non-negotiable step is to audit your clearing stack. Send me a direct message to initiate a structural rebate assessment and stop funding market inefficiency. #GovTech #FinTech #GovCon #MarTech #SupplyChain

3

This master class from Anish Agarwal is a great opportunity to learn how to remediate code at scale in production

18

1 Comment

In the span of a few weeks, Salesforce priced both ends of this market. Two checks, opposite directions, and the gap between them is the entire thesis. Today Salesforce signed to acquire Fin (formerly Intercom), for $3.6 billion. Fin's last priced venture round valued it at $1.28 billion. That was 2018. So Salesforce is paying nearly three times an eight-year-old mark, in cash, for a company that spent those eight years turning a support inbox into a proprietary AI agent that now resolves roughly 76 percent of customer queries end to end across more than 30,000 businesses. Now hold that against the other Salesforce deal a few weeks back. Contentful, valued at $3 billion in 2021, bought for roughly $1.1 billion. A third of its peak. Same acquirer. One asset re-rated up nearly 3x from its last private mark. The other re-rated down to a third of its. The market did not lift both. It sorted them. The rule behind the sort is simple. Fin built something a buyer that size could not replicate fast enough to matter: a model post-trained on real support resolution, the proprietary data underneath it, and the team that made it work. Contentful was good 2021-vintage SaaS that never grew into its last round. One is scarce and compounding. The other is solid and available. Scarcity took the premium. Availability took the discount. This is the MaintainX lesson from last month, now proven by a single buyer writing two checks in the same quarter. A reopening market does not raise all valuations. It re-rates the data-driven, growth assets a strategic cannot live without, and it marks everything else to where it actually trades. For those of us deploying into private companies, the instruction is the same. Underwrite to scarcity, not to the index. The premium goes to what cannot be rebuilt or bought twice.

37

6 Comments

🚀 What if AI agents could run the entire R&D loop—from deciding what’s worth trying to proving that an improvement is real? That’s what we’re building with SFR-AutoR&D at Salesforce AI Research. The idea is simple: give agents a goal, let them form hypotheses, invent methods, implement them, run controlled experiments, and keep only the changes that beat the baseline. SFR-AutoR&D brings together three pieces: 🔧 Auto-Engineer searches for methodological and systems improvements in heavily optimized code. pandas: up to 1.84× faster, with 136K differential checks and 0 mismatches USearch: up to +17.5% QPS hnswlib: +4.1% QPS with byte-identical results Our RL stack: peak memory 128.5 → 78.6 GiB and generation up to 3.14× faster with identical output tokens 🧠 TrainForge goes beyond choosing from predefined training recipes—it proposes and implements new training methods from scratch, evaluates them, and promotes only those that improve frozen held-out benchmarks. Its best discovered methods delivered +11 to +14 percentage-point gains across math, code, and research-agent tasks. ⚙️ AutoInfra provides the training backbone: a Tinker-compatible API that scales from lightweight LoRA experiments to large distributed training. It supports full-parameter training at 1M-token context for models up to 1T parameters, and LoRA training of a 1T model on just 16 H200s. The longer-term goal is to build agents that can discover improvements nobody explicitly asked for, across algorithms, training methods, systems, and the open-source infrastructure AI teams depend on—and attach the evidence needed to trust those improvements. 🌐 SFR-AutoR&D: https://lnkd.in/g4-awAYU 🔧 Auto-Engineer: https://lnkd.in/gikA637r 🧠 TrainForge: https://lnkd.in/g_8AFM3U ⚙️ AutoInfra: https://lnkd.in/gwx-FtPs Great work by awesome Shrey Pandit, Yiran Zhao, and Xuan Phi Nguyen at Salesforce AI Research. #AIResearch #AgenticAI #MachineLearning #ReinforcementLearning #AIEngineering #AutoResearch

69

San Diego-based Manifold (AI Detection and Response platform) has raised $8 million in seed funding; lead backer Costanoa Ventures, along with several institutional investors including Cherry Ventures, Rain Capital, Modern Technical Fund, and angel investors.

4

1 Comment

Yesterday Clay hit $100M ARR and I shared a post on the GTM bets that got us here. Today, I'm sitting down with our Head of Finance to dive deep on the business - metrics, our view on the AI market, and the path to $1B. Karan often gets this question from candidates: "Look, I have to dedicate my livelihood to Clay. All the upside of my equity is baked into one asset. How should I frame the real risk/reward?" It's a great question! So I wanted to use this video to go beyond the tactics and share how we think about the business holistically. The fundamentals, market dynamics and where we're really going. Some of what we cover: - The AI market and why GTM is actually the least crowded - Cohort data that shocked even us (enterprises never churn!) - How we've built an entire economy around Clay - The three things that create our customer flywheel - Why we're not burning cash while growing at an insane rate - The vectors to $1B: new geos, personas, use cases, channels, and verticals At the end, we get candid about the real risks and why World of Warcraft helps explain them. Watch here 👇

460

44 Comments

A big thank you to CNN and Nayeli Jaramillo-Plata for highlighting Samsara’s latest analysis on winter crash risk during the Thanksgiving travel surge. With airlines signaling capacity constraints and weather-driven delays, millions of Americans are expected to drive instead of fly—AAA projects roughly 73 million travelers will be on the road. Our analysis of anonymized operational data across billions of miles shows a clear pattern: Winter crash risk is highest when heavy commuter and delivery traffic combines with fast-changing pavement conditions—such as the Sunday after Thanksgiving. What to know if you're hitting the road this holiday: • Crash rates increase 13% during Thanksgiving week (Mon–Fri) • If you can, avoid traveling the Sunday after Thanksgiving, when crash rates increase 65%—the single most dangerous travel day of the period • With winter weather arriving earlier than usual, the risk is even higher this year Grateful for the coverage and the opportunity to help travelers and fleets better understand these shifting winter hazards. Story here, link to analysis in the comments: https://lnkd.in/gseTPq2f

56

1 Comment

On September 1, John Ternus became Apple's CEO and inherited more than an iPhone roadmap. Apple's most invisible operating system has repair approvals, invoices, keys, vendors, and 3,700+ active drivers. That figure comes from a 2026 Apple Fleet Operations Support Specialist posting, where Apple said the team authorized and supported those drivers. The same post described approximately 2,000 fleet assets across more than 40 countries and 123 business units, with over $20 million in allocated spend. The scale caught my eye. Then I read the actual work: maintenance schedules, recalls, towing, telematics, and keeping records current. Those tasks touch an entire room: fleet managers, local site teams, drivers, maintenance vendors, procurement, finance, compliance, and system owners. Each person owns part of the outcome. One missed handoff can leave a vehicle sitting even when every team did its piece. Apple has spent decades making complicated technology feel simple. Its fleet posting showed how many operators sit behind an equally simple outcome: a ready vehicle. At Proaction, we start by mapping who owns each step and which decisions need a person. Then we connect the work across the tools a fleet already uses. People will judge Ternus by the products Apple ships. Apple's fleet team has a less visible standard. The vehicle is ready when a driver reaches for the keys, and the work behind it is traceable.

19

1 Comment