Jacob Jackson

20% confidence prediction I've had for a while: labs will stop offering API access to frontier models altogether to force people to use their proprietary products. Builders that aren't competing will be fine (OSS, previous-gen models, etc), but it'd very tough to compete directly in the niches they care about.

49

6 Comments

And now for Trap #2 - The Capital Stack Mismatch There’s a specific moment in hardware companies where the product works, trials succeed, and orders start to ramp. The stars are aligning, and it's a great time to be at a startup. Then, suddenly and often unexpectedly, growth itself becomes the threat. Cash keeps running short. Payroll becomes a circled date on the calendar and an ongoing stress point. You start thinking about a bridge round just to "smooth things out”. These are signals that growth has potentially broken your capital stack. Early on, equity feels safe. No payments, no covenants, minimal pressure (except from very leaned-in board members). So teams rely on it too long without realizing that equity isn’t free, it’s just quiet in those early days. That is the trap: choosing the comfort of equity to fund the math of inventory. It feels safer in the short term, but you’re paying for 12+ weeks of inventory with a permanent piece of your company. Debt often scares founders, and for good reason. In the wrong application it can feel like a millstone, higher stakes, real payments and covenants you often tip-toe around. However, the strongest founders I’ve worked with are deliberate and right after their first real purchase order, when everyone else is celebrating they are planning how to fund their growth using a simple rule: >> Equity underwrites uncertainty (R&D, prototypes). >> Debt underwrites certainty (repeatable demand, inventory). When demand is repeatable, using equity to fund inventory makes growth heavier, not safer. Equity does nothing to fix your Cash Collection Cycle, leaving you out of pocket for 6+ months while you ship. In hardware, COGS are real and cash cycles are long. Debt isn't a sign of distress, it's often the only efficient bridge between booking an order and getting paid while not sinking your company with growth. Next week: We’ll end the year with a fundraising observation as teams ramp up for the Q1 cycle and continue our Traps series in January.

16

3 Comments

a friend just texted me: "Hearing there are 'senior devs' who are very skeptical of AI generated code (aka not using any AI tools) at a startup I’m interviewing with… interesting that they are still employed in a startup against this backdrop [see image]" my response: seems like there are just different games being played those who embrace AI as a medium for expressing ideas are playing the right game — they're not just treating it as a "better tool" if you've been in industry awhile, you may need an intellectual reboot to see it last night i went to a VentureBeat event; mostly enterprise folks… several people (from IBM, Walgreens, etc) asked about “change mgmt” and how to convince their teammates to adopt AI tools… someone raised the question of brown-field vs green-field opportunity w/r/t AI… is it worth it to upskill existing workers or better to just start fresh…? Great question! for people entrenched in a job fiefdom, e.g. where the power players and how to execute reliably, the perceived risks of novelty are infinite. they're not wrong, they're just not interested in changing the game they've mastered. aaand also, developers are the self-disclosed laziest people in the world. but they're incredibly productive because they'll happily spend a week automating some repetitive task so they never have think about it again. this is why generative AI and coding go so well together: the existing developer culture despises inefficiency, and is always looking for ways to cut corners in favor of a more elegant, efficient route. this is because developers are incredibly sensitive to the tiniest impediment to achieving persistent flow states. it's like noticing a typo in that green-on-black matrix screensaver and instead of ignoring it as a minor irritant, they'd prefer to rewrite the screensaver from scratch to fix it. but for non-developery roles, this gameplay is totally foreign. maintaining those impediments ensures job security. since they know how to navigate around known obstacles, their tacit knowledge is an insurance policy against obsolescence — for now. so it makes that sense that some senior devs, whose sunk costs could represent hundreds of thousands or millions of $$$ into existing, _stable_ systems might resist the quasi-non-deterministic output of AI coding tools. except (as they say) there is more than one way to skin a cat. here's a new game to invite resistants to play: have the skeptical senior devs review several blind PRs, all submitted by “new,” unfamiliar contributors (a mix of humans and AIs), and see if they prefer the human or AI-generated code. if the AI code really is as good as many of us believe, the experience might snap them out of their sunk-cost fallacies. because until you surprise yourself with what these tools can actually accomplish—in a blind taste test—you’ll never really see what winning this new game looks like, even if it's happening right in front of your eyes.

21

14 Comments

Every AI model you've ever used ran on a chip. That chip took 2-3 years and hundreds of millions of dollars to design. Partcl is making that process 7,000x faster. Not a typo. They're a 3-person team out of YC, backed by top VC, building GPU-accelerated EDA tools that compress what used to take months into hours — using custom diffusion models and reinforcement learning trained on the hardest physical layout problems in chip design. In the last 18 months their first engineer: → Built a physics simulator that matched state-of-the-art legacy tools in 6 weeks (experts said it would take 10 years) → Took a process running in 5 minutes down to 20 seconds → Pushed their speedup from 700x to 7,000x They're hiring Founding Engineer- This is the role for someone who: — Has trained custom ML models from scratch (robotics, physics-informed, biophysics — not fine-tuning LLaMA) — Thinks in GPU memory layouts and kernel optimization, not just PyTorch APIs — Wants to be in the room when the VP of Engineering at NVIDIA gives feedback on something they personally built — Would rather rebuild something from scratch than maintain it Stack: PyTorch, Rust, Triton, CUDA, C++ Location: San Francisco, in-person The market is $15B. And they have paying customers. If you write CUDA kernels for fun, have trained physics-informed models, or built something from zero that no one thought was possible — I'd love to talk. DM me or drop a link to your GitHub.

5

RIGHT IN TIME: a ~46GB quantized local model qwen3-coder-next that can actually drive a coding agent harness (Claude Code / Codex-style workflows). That’s a big deal. Many local models technically fit under 64GB, but in practice fail at real agentic coding: planning, tool use, multi-step edits, long-horizon context. If this works, it suggests local models are finally crossing the agent reliability threshold. Not “local beats frontier models”, but local becoming genuinely useful for autonomous developer workflows. If this holds up, it’s a major step forward for local AI.

37

4 Comments

A lot of VCs use founder age as a proxy for how well they’ll leverage LLMs and the latest AI. In my experience, the stronger signal is how technical they are. Technical founders understand how LLMs work and can exploit them fully; non-technical founders see magic, not mechanisms.

49

11 Comments

Intuit just lost 42% of its market cap. Its defense? 40 years of proprietary data no generic AI can replicate. The SaaSpocalypse didn't punish companies with bad AI. It punished companies whose value lived in the interface, not the data underneath it. Atlassian down 35%. DocuSign cratered. Same logic plays out in enterprise CX every day. The AI vendors winning aren't running the best base model. They're trained on your product catalog, your resolution history, your edge cases. Generic AI answers generic questions. Domain-trained AI solves your actual problems. The model is the engine. The data is the moat.

46

8 Comments

200+ founders in a room last night in SF. Here’s the mistake I made as an early startup. They try to build a product for everyone. What actually works is the opposite: make 10 customers insanely happy. Not 1,000 mildly interested users. Not a wide ICP deck. Ten people who would be upset if your product disappeared. That focus creates clarity in product, messaging, and distribution. Another strong takeaway from the panel, learned from Bolbi: urgency is not optional. If your product does not create a now moment, users delay, pipelines stall, and growth never compounds. Focus + urgency > breadth + polish. That’s how real products scale. Thanks Alan Zong and the UpScaleX team for the invitation. It was a fun night.

70

13 Comments

Developing a new thinking: Memory > Pipes AI isn’t destroying software moats. It’s relocating them. The 3 Phases of Software Moats: 1) Vertical SaaS (Past) Moat = Data lock-in. Proprietary schemas. High switching costs. Backward-looking defensibility. 2) Embedded AI (Present) AI inside silos is context-starved. LLMs aren’t weak — system architecture is. Fragmented context = fragile outputs. 3) Horizontal, Agentic AI (Emerging) Tools like Claude Code and open agent frameworks point to a new layer: AI reasoning across systems. Interface collapses. AI becomes the orchestration layer. Apps become tools. Where the Real Moats Will Form: Not in integrations. Not in wrappers. Durable defensibility will come from: * Compounding Memory → persistent, longitudinal understanding of users and teams * Domain-Specific Orchestration → reliability in high-stakes workflows (finance, legal, healthcare) * Workflow Network Effects → shared institutional intelligence Trust Infrastructure → auditability + compliance + governance Hence: Memory > Pipes APIs are commodities. Context is compounding. The next $10B companies won’t just connect data — they’ll accumulate intelligence. Stateless systems are replaceable. Stateful, memory-rich AI systems are not. That’s where long-term enterprise value will accrue, I think :) Entrepreneurs Roundtable Accelerator Remarkable Ventures

62

8 Comments